xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c (revision cf6d949a409e09539477d32dbe7c954e4852e744)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include <linux/device.h>
25 #include <linux/export.h>
26 #include <linux/err.h>
27 #include <linux/fs.h>
28 #include <linux/file.h>
29 #include <linux/sched.h>
30 #include <linux/slab.h>
31 #include <linux/uaccess.h>
32 #include <linux/compat.h>
33 #include <uapi/linux/kfd_ioctl.h>
34 #include <linux/time.h>
35 #include <linux/mm.h>
36 #include <linux/mman.h>
37 #include <linux/ptrace.h>
38 #include <linux/dma-buf.h>
39 #include <linux/processor.h>
40 #include "kfd_priv.h"
41 #include "kfd_device_queue_manager.h"
42 #include "kfd_svm.h"
43 #include "amdgpu_amdkfd.h"
44 #include "kfd_smi_events.h"
45 #include "amdgpu_dma_buf.h"
46 #include "kfd_debug.h"
47 
48 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
49 static int kfd_open(struct inode *, struct file *);
50 static int kfd_release(struct inode *, struct file *);
51 static int kfd_mmap(struct file *, struct vm_area_struct *);
52 
53 static const char kfd_dev_name[] = "kfd";
54 
55 static const struct file_operations kfd_fops = {
56 	.owner = THIS_MODULE,
57 	.unlocked_ioctl = kfd_ioctl,
58 	.compat_ioctl = compat_ptr_ioctl,
59 	.open = kfd_open,
60 	.release = kfd_release,
61 	.mmap = kfd_mmap,
62 };
63 
64 static int kfd_char_dev_major = -1;
65 struct device *kfd_device;
66 static const struct class kfd_class = {
67 	.name = kfd_dev_name,
68 };
69 
70 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
71 {
72 	struct kfd_process_device *pdd;
73 
74 	mutex_lock(&p->mutex);
75 	pdd = kfd_process_device_data_by_id(p, gpu_id);
76 
77 	if (pdd)
78 		return pdd;
79 
80 	mutex_unlock(&p->mutex);
81 	return NULL;
82 }
83 
84 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
85 {
86 	mutex_unlock(&pdd->process->mutex);
87 }
88 
89 int kfd_chardev_init(void)
90 {
91 	int err = 0;
92 
93 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
94 	err = kfd_char_dev_major;
95 	if (err < 0)
96 		goto err_register_chrdev;
97 
98 	err = class_register(&kfd_class);
99 	if (err)
100 		goto err_class_create;
101 
102 	kfd_device = device_create(&kfd_class, NULL,
103 				   MKDEV(kfd_char_dev_major, 0),
104 				   NULL, kfd_dev_name);
105 	err = PTR_ERR(kfd_device);
106 	if (IS_ERR(kfd_device))
107 		goto err_device_create;
108 
109 	return 0;
110 
111 err_device_create:
112 	class_unregister(&kfd_class);
113 err_class_create:
114 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
115 err_register_chrdev:
116 	return err;
117 }
118 
119 void kfd_chardev_exit(void)
120 {
121 	device_destroy(&kfd_class, MKDEV(kfd_char_dev_major, 0));
122 	class_unregister(&kfd_class);
123 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
124 	kfd_device = NULL;
125 }
126 
127 
128 static int kfd_open(struct inode *inode, struct file *filep)
129 {
130 	struct kfd_process *process;
131 	bool is_32bit_user_mode;
132 
133 	if (iminor(inode) != 0)
134 		return -ENODEV;
135 
136 	is_32bit_user_mode = in_compat_syscall();
137 
138 	if (is_32bit_user_mode) {
139 		dev_warn(kfd_device,
140 			"Process %d (32-bit) failed to open /dev/kfd\n"
141 			"32-bit processes are not supported by amdkfd\n",
142 			current->pid);
143 		return -EPERM;
144 	}
145 
146 	process = kfd_create_process(current);
147 	if (IS_ERR(process))
148 		return PTR_ERR(process);
149 
150 	if (kfd_process_init_cwsr_apu(process, filep)) {
151 		kfd_unref_process(process);
152 		return -EFAULT;
153 	}
154 
155 	/* filep now owns the reference returned by kfd_create_process */
156 	filep->private_data = process;
157 
158 	dev_dbg(kfd_device, "process pid %d opened kfd node, compat mode (32 bit) - %d\n",
159 		process->lead_thread->pid, process->is_32bit_user_mode);
160 
161 	return 0;
162 }
163 
164 static int kfd_release(struct inode *inode, struct file *filep)
165 {
166 	struct kfd_process *process = filep->private_data;
167 
168 	if (process)
169 		kfd_unref_process(process);
170 
171 	return 0;
172 }
173 
174 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
175 					void *data)
176 {
177 	struct kfd_ioctl_get_version_args *args = data;
178 
179 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
180 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
181 
182 	return 0;
183 }
184 
185 static int set_queue_properties_from_user(struct queue_properties *q_properties,
186 				struct kfd_ioctl_create_queue_args *args)
187 {
188 	/*
189 	 * Repurpose queue percentage to accommodate new features:
190 	 * bit 0-7: queue percentage
191 	 * bit 8-15: pm4_target_xcc
192 	 */
193 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
194 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
195 		return -EINVAL;
196 	}
197 
198 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
199 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
200 		return -EINVAL;
201 	}
202 
203 	if ((args->ring_base_address) &&
204 		(!access_ok((const void __user *) args->ring_base_address,
205 			sizeof(uint64_t)))) {
206 		pr_err("Can't access ring base address\n");
207 		return -EFAULT;
208 	}
209 
210 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
211 		pr_err("Ring size must be a power of 2 or 0\n");
212 		return -EINVAL;
213 	}
214 
215 	if (args->ring_size < KFD_MIN_QUEUE_RING_SIZE) {
216 		args->ring_size = KFD_MIN_QUEUE_RING_SIZE;
217 		pr_debug("Size lower. clamped to KFD_MIN_QUEUE_RING_SIZE");
218 	}
219 
220 	if (!access_ok((const void __user *) args->read_pointer_address,
221 			sizeof(uint32_t))) {
222 		pr_err("Can't access read pointer\n");
223 		return -EFAULT;
224 	}
225 
226 	if (!access_ok((const void __user *) args->write_pointer_address,
227 			sizeof(uint32_t))) {
228 		pr_err("Can't access write pointer\n");
229 		return -EFAULT;
230 	}
231 
232 	if (args->eop_buffer_address &&
233 		!access_ok((const void __user *) args->eop_buffer_address,
234 			sizeof(uint32_t))) {
235 		pr_debug("Can't access eop buffer");
236 		return -EFAULT;
237 	}
238 
239 	if (args->ctx_save_restore_address &&
240 		!access_ok((const void __user *) args->ctx_save_restore_address,
241 			sizeof(uint32_t))) {
242 		pr_debug("Can't access ctx save restore buffer");
243 		return -EFAULT;
244 	}
245 
246 	q_properties->is_interop = false;
247 	q_properties->is_gws = false;
248 	q_properties->queue_percent = args->queue_percentage & 0xFF;
249 	/* bit 8-15 are repurposed to be PM4 target XCC */
250 	q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
251 	q_properties->priority = args->queue_priority;
252 	q_properties->queue_address = args->ring_base_address;
253 	q_properties->queue_size = args->ring_size;
254 	q_properties->read_ptr = (void __user *)args->read_pointer_address;
255 	q_properties->write_ptr = (void __user *)args->write_pointer_address;
256 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
257 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
258 	q_properties->ctx_save_restore_area_address =
259 			args->ctx_save_restore_address;
260 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
261 	q_properties->ctl_stack_size = args->ctl_stack_size;
262 	q_properties->sdma_engine_id = args->sdma_engine_id;
263 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
264 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
265 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
266 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
267 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
268 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
269 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
270 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_BY_ENG_ID)
271 		q_properties->type = KFD_QUEUE_TYPE_SDMA_BY_ENG_ID;
272 	else
273 		return -ENOTSUPP;
274 
275 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
276 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
277 	else
278 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
279 
280 	pr_debug("Queue Percentage: %d, %d\n",
281 			q_properties->queue_percent, args->queue_percentage);
282 
283 	pr_debug("Queue Priority: %d, %d\n",
284 			q_properties->priority, args->queue_priority);
285 
286 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
287 			q_properties->queue_address, args->ring_base_address);
288 
289 	pr_debug("Queue Size: 0x%llX, %u\n",
290 			q_properties->queue_size, args->ring_size);
291 
292 	pr_debug("Queue r/w Pointers: %px, %px\n",
293 			q_properties->read_ptr,
294 			q_properties->write_ptr);
295 
296 	pr_debug("Queue Format: %d\n", q_properties->format);
297 
298 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
299 
300 	pr_debug("Queue CTX save area: 0x%llX\n",
301 			q_properties->ctx_save_restore_area_address);
302 
303 	return 0;
304 }
305 
306 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
307 					void *data)
308 {
309 	struct kfd_ioctl_create_queue_args *args = data;
310 	struct kfd_node *dev;
311 	int err = 0;
312 	unsigned int queue_id;
313 	struct kfd_process_device *pdd;
314 	struct queue_properties q_properties;
315 	uint32_t doorbell_offset_in_process = 0;
316 
317 	memset(&q_properties, 0, sizeof(struct queue_properties));
318 
319 	pr_debug("Creating queue ioctl\n");
320 
321 	err = set_queue_properties_from_user(&q_properties, args);
322 	if (err)
323 		return err;
324 
325 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
326 
327 	mutex_lock(&p->mutex);
328 
329 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
330 	if (!pdd) {
331 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
332 		err = -EINVAL;
333 		goto err_pdd;
334 	}
335 	dev = pdd->dev;
336 
337 	pdd = kfd_bind_process_to_device(dev, p);
338 	if (IS_ERR(pdd)) {
339 		err = -ESRCH;
340 		goto err_bind_process;
341 	}
342 
343 	if (q_properties.type == KFD_QUEUE_TYPE_SDMA_BY_ENG_ID) {
344 		int max_sdma_eng_id = kfd_get_num_sdma_engines(dev) +
345 				      kfd_get_num_xgmi_sdma_engines(dev) - 1;
346 
347 		if (q_properties.sdma_engine_id > max_sdma_eng_id) {
348 			err = -EINVAL;
349 			pr_err("sdma_engine_id %i exceeds maximum id of %i\n",
350 			       q_properties.sdma_engine_id, max_sdma_eng_id);
351 			goto err_sdma_engine_id;
352 		}
353 	}
354 
355 	if (!pdd->qpd.proc_doorbells) {
356 		err = kfd_alloc_process_doorbells(dev->kfd, pdd);
357 		if (err) {
358 			pr_debug("failed to allocate process doorbells\n");
359 			goto err_bind_process;
360 		}
361 	}
362 
363 	err = kfd_queue_acquire_buffers(pdd, &q_properties);
364 	if (err) {
365 		pr_debug("failed to acquire user queue buffers\n");
366 		goto err_acquire_queue_buf;
367 	}
368 
369 	pr_debug("Creating queue for process pid %d on gpu 0x%x\n",
370 			p->lead_thread->pid,
371 			dev->id);
372 
373 	err = pqm_create_queue(&p->pqm, dev, &q_properties, &queue_id,
374 			NULL, NULL, NULL, &doorbell_offset_in_process);
375 	if (err != 0)
376 		goto err_create_queue;
377 
378 	args->queue_id = queue_id;
379 
380 
381 	/* Return gpu_id as doorbell offset for mmap usage */
382 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
383 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
384 	if (KFD_IS_SOC15(dev))
385 		/* On SOC15 ASICs, include the doorbell offset within the
386 		 * process doorbell frame, which is 2 pages.
387 		 */
388 		args->doorbell_offset |= doorbell_offset_in_process;
389 
390 	mutex_unlock(&p->mutex);
391 
392 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
393 
394 	pr_debug("Ring buffer address == 0x%016llX\n",
395 			args->ring_base_address);
396 
397 	pr_debug("Read ptr address    == 0x%016llX\n",
398 			args->read_pointer_address);
399 
400 	pr_debug("Write ptr address   == 0x%016llX\n",
401 			args->write_pointer_address);
402 
403 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
404 	return 0;
405 
406 err_create_queue:
407 	kfd_queue_unref_bo_vas(pdd, &q_properties);
408 	kfd_queue_release_buffers(pdd, &q_properties);
409 err_acquire_queue_buf:
410 err_sdma_engine_id:
411 err_bind_process:
412 err_pdd:
413 	mutex_unlock(&p->mutex);
414 	return err;
415 }
416 
417 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
418 					void *data)
419 {
420 	int retval;
421 	struct kfd_ioctl_destroy_queue_args *args = data;
422 
423 	pr_debug("Destroying queue id %d for process pid %d\n",
424 				args->queue_id,
425 				p->lead_thread->pid);
426 
427 	mutex_lock(&p->mutex);
428 
429 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
430 
431 	mutex_unlock(&p->mutex);
432 	return retval;
433 }
434 
435 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
436 					void *data)
437 {
438 	int retval;
439 	struct kfd_ioctl_update_queue_args *args = data;
440 	struct queue_properties properties;
441 
442 	/*
443 	 * Repurpose queue percentage to accommodate new features:
444 	 * bit 0-7: queue percentage
445 	 * bit 8-15: pm4_target_xcc
446 	 */
447 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
448 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
449 		return -EINVAL;
450 	}
451 
452 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
453 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
454 		return -EINVAL;
455 	}
456 
457 	if ((args->ring_base_address) &&
458 		(!access_ok((const void __user *) args->ring_base_address,
459 			sizeof(uint64_t)))) {
460 		pr_err("Can't access ring base address\n");
461 		return -EFAULT;
462 	}
463 
464 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
465 		pr_err("Ring size must be a power of 2 or 0\n");
466 		return -EINVAL;
467 	}
468 
469 	if (args->ring_size < KFD_MIN_QUEUE_RING_SIZE) {
470 		args->ring_size = KFD_MIN_QUEUE_RING_SIZE;
471 		pr_debug("Size lower. clamped to KFD_MIN_QUEUE_RING_SIZE");
472 	}
473 
474 	properties.queue_address = args->ring_base_address;
475 	properties.queue_size = args->ring_size;
476 	properties.queue_percent = args->queue_percentage & 0xFF;
477 	/* bit 8-15 are repurposed to be PM4 target XCC */
478 	properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
479 	properties.priority = args->queue_priority;
480 
481 	pr_debug("Updating queue id %d for process pid %d\n",
482 			args->queue_id, p->lead_thread->pid);
483 
484 	mutex_lock(&p->mutex);
485 
486 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
487 
488 	mutex_unlock(&p->mutex);
489 
490 	return retval;
491 }
492 
493 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
494 					void *data)
495 {
496 	int retval;
497 	const int max_num_cus = 1024;
498 	struct kfd_ioctl_set_cu_mask_args *args = data;
499 	struct mqd_update_info minfo = {0};
500 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
501 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
502 
503 	if ((args->num_cu_mask % 32) != 0) {
504 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
505 				args->num_cu_mask);
506 		return -EINVAL;
507 	}
508 
509 	minfo.cu_mask.count = args->num_cu_mask;
510 	if (minfo.cu_mask.count == 0) {
511 		pr_debug("CU mask cannot be 0");
512 		return -EINVAL;
513 	}
514 
515 	/* To prevent an unreasonably large CU mask size, set an arbitrary
516 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
517 	 * past max_num_cus bits and just use the first max_num_cus bits.
518 	 */
519 	if (minfo.cu_mask.count > max_num_cus) {
520 		pr_debug("CU mask cannot be greater than 1024 bits");
521 		minfo.cu_mask.count = max_num_cus;
522 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
523 	}
524 
525 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
526 	if (!minfo.cu_mask.ptr)
527 		return -ENOMEM;
528 
529 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
530 	if (retval) {
531 		pr_debug("Could not copy CU mask from userspace");
532 		retval = -EFAULT;
533 		goto out;
534 	}
535 
536 	mutex_lock(&p->mutex);
537 
538 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
539 
540 	mutex_unlock(&p->mutex);
541 
542 out:
543 	kfree(minfo.cu_mask.ptr);
544 	return retval;
545 }
546 
547 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
548 					  struct kfd_process *p, void *data)
549 {
550 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
551 	int r;
552 
553 	mutex_lock(&p->mutex);
554 
555 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
556 			       (void __user *)args->ctl_stack_address,
557 			       &args->ctl_stack_used_size,
558 			       &args->save_area_used_size);
559 
560 	mutex_unlock(&p->mutex);
561 
562 	return r;
563 }
564 
565 static int kfd_ioctl_set_memory_policy(struct file *filep,
566 					struct kfd_process *p, void *data)
567 {
568 	struct kfd_ioctl_set_memory_policy_args *args = data;
569 	int err = 0;
570 	struct kfd_process_device *pdd;
571 	enum cache_policy default_policy, alternate_policy;
572 
573 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
574 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
575 		return -EINVAL;
576 	}
577 
578 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
579 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
580 		return -EINVAL;
581 	}
582 
583 	mutex_lock(&p->mutex);
584 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
585 	if (!pdd) {
586 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
587 		err = -EINVAL;
588 		goto err_pdd;
589 	}
590 
591 	pdd = kfd_bind_process_to_device(pdd->dev, p);
592 	if (IS_ERR(pdd)) {
593 		err = -ESRCH;
594 		goto out;
595 	}
596 
597 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
598 			 ? cache_policy_coherent : cache_policy_noncoherent;
599 
600 	alternate_policy =
601 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
602 		   ? cache_policy_coherent : cache_policy_noncoherent;
603 
604 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
605 				&pdd->qpd,
606 				default_policy,
607 				alternate_policy,
608 				(void __user *)args->alternate_aperture_base,
609 				args->alternate_aperture_size,
610 				args->misc_process_flag))
611 		err = -EINVAL;
612 
613 out:
614 err_pdd:
615 	mutex_unlock(&p->mutex);
616 
617 	return err;
618 }
619 
620 static int kfd_ioctl_set_trap_handler(struct file *filep,
621 					struct kfd_process *p, void *data)
622 {
623 	struct kfd_ioctl_set_trap_handler_args *args = data;
624 	int err = 0;
625 	struct kfd_process_device *pdd;
626 
627 	mutex_lock(&p->mutex);
628 
629 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
630 	if (!pdd) {
631 		err = -EINVAL;
632 		goto err_pdd;
633 	}
634 
635 	pdd = kfd_bind_process_to_device(pdd->dev, p);
636 	if (IS_ERR(pdd)) {
637 		err = -ESRCH;
638 		goto out;
639 	}
640 
641 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
642 
643 out:
644 err_pdd:
645 	mutex_unlock(&p->mutex);
646 
647 	return err;
648 }
649 
650 static int kfd_ioctl_dbg_register(struct file *filep,
651 				struct kfd_process *p, void *data)
652 {
653 	return -EPERM;
654 }
655 
656 static int kfd_ioctl_dbg_unregister(struct file *filep,
657 				struct kfd_process *p, void *data)
658 {
659 	return -EPERM;
660 }
661 
662 static int kfd_ioctl_dbg_address_watch(struct file *filep,
663 					struct kfd_process *p, void *data)
664 {
665 	return -EPERM;
666 }
667 
668 /* Parse and generate fixed size data structure for wave control */
669 static int kfd_ioctl_dbg_wave_control(struct file *filep,
670 					struct kfd_process *p, void *data)
671 {
672 	return -EPERM;
673 }
674 
675 static int kfd_ioctl_get_clock_counters(struct file *filep,
676 				struct kfd_process *p, void *data)
677 {
678 	struct kfd_ioctl_get_clock_counters_args *args = data;
679 	struct kfd_process_device *pdd;
680 
681 	mutex_lock(&p->mutex);
682 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
683 	mutex_unlock(&p->mutex);
684 	if (pdd)
685 		/* Reading GPU clock counter from KGD */
686 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
687 	else
688 		/* Node without GPU resource */
689 		args->gpu_clock_counter = 0;
690 
691 	/* No access to rdtsc. Using raw monotonic time */
692 	args->cpu_clock_counter = ktime_get_raw_ns();
693 	args->system_clock_counter = ktime_get_boottime_ns();
694 
695 	/* Since the counter is in nano-seconds we use 1GHz frequency */
696 	args->system_clock_freq = 1000000000;
697 
698 	return 0;
699 }
700 
701 
702 static int kfd_ioctl_get_process_apertures(struct file *filp,
703 				struct kfd_process *p, void *data)
704 {
705 	struct kfd_ioctl_get_process_apertures_args *args = data;
706 	struct kfd_process_device_apertures *pAperture;
707 	int i;
708 
709 	dev_dbg(kfd_device, "get apertures for process pid %d", p->lead_thread->pid);
710 
711 	args->num_of_nodes = 0;
712 
713 	mutex_lock(&p->mutex);
714 	/* Run over all pdd of the process */
715 	for (i = 0; i < p->n_pdds; i++) {
716 		struct kfd_process_device *pdd = p->pdds[i];
717 
718 		pAperture =
719 			&args->process_apertures[args->num_of_nodes];
720 		pAperture->gpu_id = pdd->dev->id;
721 		pAperture->lds_base = pdd->lds_base;
722 		pAperture->lds_limit = pdd->lds_limit;
723 		pAperture->gpuvm_base = pdd->gpuvm_base;
724 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
725 		pAperture->scratch_base = pdd->scratch_base;
726 		pAperture->scratch_limit = pdd->scratch_limit;
727 
728 		dev_dbg(kfd_device,
729 			"node id %u\n", args->num_of_nodes);
730 		dev_dbg(kfd_device,
731 			"gpu id %u\n", pdd->dev->id);
732 		dev_dbg(kfd_device,
733 			"lds_base %llX\n", pdd->lds_base);
734 		dev_dbg(kfd_device,
735 			"lds_limit %llX\n", pdd->lds_limit);
736 		dev_dbg(kfd_device,
737 			"gpuvm_base %llX\n", pdd->gpuvm_base);
738 		dev_dbg(kfd_device,
739 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
740 		dev_dbg(kfd_device,
741 			"scratch_base %llX\n", pdd->scratch_base);
742 		dev_dbg(kfd_device,
743 			"scratch_limit %llX\n", pdd->scratch_limit);
744 
745 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
746 			break;
747 	}
748 	mutex_unlock(&p->mutex);
749 
750 	return 0;
751 }
752 
753 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
754 				struct kfd_process *p, void *data)
755 {
756 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
757 	struct kfd_process_device_apertures *pa;
758 	int ret;
759 	int i;
760 
761 	dev_dbg(kfd_device, "get apertures for process pid %d",
762 			p->lead_thread->pid);
763 
764 	if (args->num_of_nodes == 0) {
765 		/* Return number of nodes, so that user space can alloacate
766 		 * sufficient memory
767 		 */
768 		mutex_lock(&p->mutex);
769 		args->num_of_nodes = p->n_pdds;
770 		goto out_unlock;
771 	}
772 
773 	/* Fill in process-aperture information for all available
774 	 * nodes, but not more than args->num_of_nodes as that is
775 	 * the amount of memory allocated by user
776 	 */
777 	pa = kcalloc(args->num_of_nodes, sizeof(struct kfd_process_device_apertures),
778 		     GFP_KERNEL);
779 	if (!pa)
780 		return -ENOMEM;
781 
782 	mutex_lock(&p->mutex);
783 
784 	if (!p->n_pdds) {
785 		args->num_of_nodes = 0;
786 		kfree(pa);
787 		goto out_unlock;
788 	}
789 
790 	/* Run over all pdd of the process */
791 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
792 		struct kfd_process_device *pdd = p->pdds[i];
793 
794 		pa[i].gpu_id = pdd->dev->id;
795 		pa[i].lds_base = pdd->lds_base;
796 		pa[i].lds_limit = pdd->lds_limit;
797 		pa[i].gpuvm_base = pdd->gpuvm_base;
798 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
799 		pa[i].scratch_base = pdd->scratch_base;
800 		pa[i].scratch_limit = pdd->scratch_limit;
801 
802 		dev_dbg(kfd_device,
803 			"gpu id %u\n", pdd->dev->id);
804 		dev_dbg(kfd_device,
805 			"lds_base %llX\n", pdd->lds_base);
806 		dev_dbg(kfd_device,
807 			"lds_limit %llX\n", pdd->lds_limit);
808 		dev_dbg(kfd_device,
809 			"gpuvm_base %llX\n", pdd->gpuvm_base);
810 		dev_dbg(kfd_device,
811 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
812 		dev_dbg(kfd_device,
813 			"scratch_base %llX\n", pdd->scratch_base);
814 		dev_dbg(kfd_device,
815 			"scratch_limit %llX\n", pdd->scratch_limit);
816 	}
817 	mutex_unlock(&p->mutex);
818 
819 	args->num_of_nodes = i;
820 	ret = copy_to_user(
821 			(void __user *)args->kfd_process_device_apertures_ptr,
822 			pa,
823 			(i * sizeof(struct kfd_process_device_apertures)));
824 	kfree(pa);
825 	return ret ? -EFAULT : 0;
826 
827 out_unlock:
828 	mutex_unlock(&p->mutex);
829 	return 0;
830 }
831 
832 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
833 					void *data)
834 {
835 	struct kfd_ioctl_create_event_args *args = data;
836 	int err;
837 
838 	/* For dGPUs the event page is allocated in user mode. The
839 	 * handle is passed to KFD with the first call to this IOCTL
840 	 * through the event_page_offset field.
841 	 */
842 	if (args->event_page_offset) {
843 		mutex_lock(&p->mutex);
844 		err = kfd_kmap_event_page(p, args->event_page_offset);
845 		mutex_unlock(&p->mutex);
846 		if (err)
847 			return err;
848 	}
849 
850 	err = kfd_event_create(filp, p, args->event_type,
851 				args->auto_reset != 0, args->node_id,
852 				&args->event_id, &args->event_trigger_data,
853 				&args->event_page_offset,
854 				&args->event_slot_index);
855 
856 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
857 	return err;
858 }
859 
860 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
861 					void *data)
862 {
863 	struct kfd_ioctl_destroy_event_args *args = data;
864 
865 	return kfd_event_destroy(p, args->event_id);
866 }
867 
868 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
869 				void *data)
870 {
871 	struct kfd_ioctl_set_event_args *args = data;
872 
873 	return kfd_set_event(p, args->event_id);
874 }
875 
876 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
877 				void *data)
878 {
879 	struct kfd_ioctl_reset_event_args *args = data;
880 
881 	return kfd_reset_event(p, args->event_id);
882 }
883 
884 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
885 				void *data)
886 {
887 	struct kfd_ioctl_wait_events_args *args = data;
888 
889 	return kfd_wait_on_events(p, args->num_events,
890 			(void __user *)args->events_ptr,
891 			(args->wait_for_all != 0),
892 			&args->timeout, &args->wait_result);
893 }
894 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
895 					struct kfd_process *p, void *data)
896 {
897 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
898 	struct kfd_process_device *pdd;
899 	struct kfd_node *dev;
900 	long err;
901 
902 	mutex_lock(&p->mutex);
903 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
904 	if (!pdd) {
905 		err = -EINVAL;
906 		goto err_pdd;
907 	}
908 	dev = pdd->dev;
909 
910 	pdd = kfd_bind_process_to_device(dev, p);
911 	if (IS_ERR(pdd)) {
912 		err = PTR_ERR(pdd);
913 		goto bind_process_to_device_fail;
914 	}
915 
916 	pdd->qpd.sh_hidden_private_base = args->va_addr;
917 
918 	mutex_unlock(&p->mutex);
919 
920 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
921 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
922 		dev->kfd2kgd->set_scratch_backing_va(
923 			dev->adev, args->va_addr, pdd->qpd.vmid);
924 
925 	return 0;
926 
927 bind_process_to_device_fail:
928 err_pdd:
929 	mutex_unlock(&p->mutex);
930 	return err;
931 }
932 
933 static int kfd_ioctl_get_tile_config(struct file *filep,
934 		struct kfd_process *p, void *data)
935 {
936 	struct kfd_ioctl_get_tile_config_args *args = data;
937 	struct kfd_process_device *pdd;
938 	struct tile_config config;
939 	int err = 0;
940 
941 	mutex_lock(&p->mutex);
942 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
943 	mutex_unlock(&p->mutex);
944 	if (!pdd)
945 		return -EINVAL;
946 
947 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
948 
949 	args->gb_addr_config = config.gb_addr_config;
950 	args->num_banks = config.num_banks;
951 	args->num_ranks = config.num_ranks;
952 
953 	if (args->num_tile_configs > config.num_tile_configs)
954 		args->num_tile_configs = config.num_tile_configs;
955 	err = copy_to_user((void __user *)args->tile_config_ptr,
956 			config.tile_config_ptr,
957 			args->num_tile_configs * sizeof(uint32_t));
958 	if (err) {
959 		args->num_tile_configs = 0;
960 		return -EFAULT;
961 	}
962 
963 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
964 		args->num_macro_tile_configs =
965 				config.num_macro_tile_configs;
966 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
967 			config.macro_tile_config_ptr,
968 			args->num_macro_tile_configs * sizeof(uint32_t));
969 	if (err) {
970 		args->num_macro_tile_configs = 0;
971 		return -EFAULT;
972 	}
973 
974 	return 0;
975 }
976 
977 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
978 				void *data)
979 {
980 	struct kfd_ioctl_acquire_vm_args *args = data;
981 	struct kfd_process_device *pdd;
982 	struct file *drm_file;
983 	int ret;
984 
985 	drm_file = fget(args->drm_fd);
986 	if (!drm_file)
987 		return -EINVAL;
988 
989 	mutex_lock(&p->mutex);
990 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
991 	if (!pdd) {
992 		ret = -EINVAL;
993 		goto err_pdd;
994 	}
995 
996 	if (pdd->drm_file) {
997 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
998 		goto err_drm_file;
999 	}
1000 
1001 	ret = kfd_process_device_init_vm(pdd, drm_file);
1002 	if (ret)
1003 		goto err_unlock;
1004 
1005 	/* On success, the PDD keeps the drm_file reference */
1006 	mutex_unlock(&p->mutex);
1007 
1008 	return 0;
1009 
1010 err_unlock:
1011 err_pdd:
1012 err_drm_file:
1013 	mutex_unlock(&p->mutex);
1014 	fput(drm_file);
1015 	return ret;
1016 }
1017 
1018 bool kfd_dev_is_large_bar(struct kfd_node *dev)
1019 {
1020 	if (dev->kfd->adev->debug_largebar) {
1021 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1022 		return true;
1023 	}
1024 
1025 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1026 	    dev->local_mem_info.local_mem_size_public > 0)
1027 		return true;
1028 
1029 	if (dev->local_mem_info.local_mem_size_public == 0 &&
1030 	    dev->kfd->adev->gmc.is_app_apu) {
1031 		pr_debug("APP APU, Consider like a large bar system\n");
1032 		return true;
1033 	}
1034 
1035 	return false;
1036 }
1037 
1038 static int kfd_ioctl_get_available_memory(struct file *filep,
1039 					  struct kfd_process *p, void *data)
1040 {
1041 	struct kfd_ioctl_get_available_memory_args *args = data;
1042 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1043 
1044 	if (!pdd)
1045 		return -EINVAL;
1046 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1047 							pdd->dev->node_id);
1048 	kfd_unlock_pdd(pdd);
1049 	return 0;
1050 }
1051 
1052 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1053 					struct kfd_process *p, void *data)
1054 {
1055 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1056 	struct kfd_process_device *pdd;
1057 	void *mem;
1058 	struct kfd_node *dev;
1059 	int idr_handle;
1060 	long err;
1061 	uint64_t offset = args->mmap_offset;
1062 	uint32_t flags = args->flags;
1063 
1064 	if (args->size == 0)
1065 		return -EINVAL;
1066 
1067 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1068 	/* Flush pending deferred work to avoid racing with deferred actions
1069 	 * from previous memory map changes (e.g. munmap).
1070 	 */
1071 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1072 	mutex_lock(&p->svms.lock);
1073 	mmap_write_unlock(current->mm);
1074 	if (interval_tree_iter_first(&p->svms.objects,
1075 				     args->va_addr >> PAGE_SHIFT,
1076 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1077 		pr_err("Address: 0x%llx already allocated by SVM\n",
1078 			args->va_addr);
1079 		mutex_unlock(&p->svms.lock);
1080 		return -EADDRINUSE;
1081 	}
1082 
1083 	/* When register user buffer check if it has been registered by svm by
1084 	 * buffer cpu virtual address.
1085 	 */
1086 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1087 	    interval_tree_iter_first(&p->svms.objects,
1088 				     args->mmap_offset >> PAGE_SHIFT,
1089 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1090 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1091 			args->mmap_offset);
1092 		mutex_unlock(&p->svms.lock);
1093 		return -EADDRINUSE;
1094 	}
1095 
1096 	mutex_unlock(&p->svms.lock);
1097 #endif
1098 	mutex_lock(&p->mutex);
1099 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1100 	if (!pdd) {
1101 		err = -EINVAL;
1102 		goto err_pdd;
1103 	}
1104 
1105 	dev = pdd->dev;
1106 
1107 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1108 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1109 		!kfd_dev_is_large_bar(dev)) {
1110 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1111 		err = -EINVAL;
1112 		goto err_large_bar;
1113 	}
1114 
1115 	pdd = kfd_bind_process_to_device(dev, p);
1116 	if (IS_ERR(pdd)) {
1117 		err = PTR_ERR(pdd);
1118 		goto err_unlock;
1119 	}
1120 
1121 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1122 		if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1123 			err = -EINVAL;
1124 			goto err_unlock;
1125 		}
1126 		offset = kfd_get_process_doorbells(pdd);
1127 		if (!offset) {
1128 			err = -ENOMEM;
1129 			goto err_unlock;
1130 		}
1131 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1132 		if (args->size != PAGE_SIZE) {
1133 			err = -EINVAL;
1134 			goto err_unlock;
1135 		}
1136 		offset = dev->adev->rmmio_remap.bus_addr;
1137 		if (!offset || (PAGE_SIZE > 4096)) {
1138 			err = -ENOMEM;
1139 			goto err_unlock;
1140 		}
1141 	}
1142 
1143 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1144 		dev->adev, args->va_addr, args->size,
1145 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1146 		flags, false);
1147 
1148 	if (err)
1149 		goto err_unlock;
1150 
1151 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1152 	if (idr_handle < 0) {
1153 		err = -EFAULT;
1154 		goto err_free;
1155 	}
1156 
1157 	/* Update the VRAM usage count */
1158 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1159 		uint64_t size = args->size;
1160 
1161 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1162 			size >>= 1;
1163 		atomic64_add(PAGE_ALIGN(size), &pdd->vram_usage);
1164 	}
1165 
1166 	mutex_unlock(&p->mutex);
1167 
1168 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1169 	args->mmap_offset = offset;
1170 
1171 	/* MMIO is mapped through kfd device
1172 	 * Generate a kfd mmap offset
1173 	 */
1174 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1175 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1176 					| KFD_MMAP_GPU_ID(args->gpu_id);
1177 
1178 	return 0;
1179 
1180 err_free:
1181 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1182 					       pdd->drm_priv, NULL);
1183 err_unlock:
1184 err_pdd:
1185 err_large_bar:
1186 	mutex_unlock(&p->mutex);
1187 	return err;
1188 }
1189 
1190 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1191 					struct kfd_process *p, void *data)
1192 {
1193 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1194 	struct kfd_process_device *pdd;
1195 	void *mem;
1196 	int ret;
1197 	uint64_t size = 0;
1198 
1199 	mutex_lock(&p->mutex);
1200 	/*
1201 	 * Safeguard to prevent user space from freeing signal BO.
1202 	 * It will be freed at process termination.
1203 	 */
1204 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1205 		pr_err("Free signal BO is not allowed\n");
1206 		ret = -EPERM;
1207 		goto err_unlock;
1208 	}
1209 
1210 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1211 	if (!pdd) {
1212 		pr_err("Process device data doesn't exist\n");
1213 		ret = -EINVAL;
1214 		goto err_pdd;
1215 	}
1216 
1217 	mem = kfd_process_device_translate_handle(
1218 		pdd, GET_IDR_HANDLE(args->handle));
1219 	if (!mem) {
1220 		ret = -EINVAL;
1221 		goto err_unlock;
1222 	}
1223 
1224 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1225 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1226 
1227 	/* If freeing the buffer failed, leave the handle in place for
1228 	 * clean-up during process tear-down.
1229 	 */
1230 	if (!ret)
1231 		kfd_process_device_remove_obj_handle(
1232 			pdd, GET_IDR_HANDLE(args->handle));
1233 
1234 	atomic64_sub(size, &pdd->vram_usage);
1235 
1236 err_unlock:
1237 err_pdd:
1238 	mutex_unlock(&p->mutex);
1239 	return ret;
1240 }
1241 
1242 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1243 					struct kfd_process *p, void *data)
1244 {
1245 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1246 	struct kfd_process_device *pdd, *peer_pdd;
1247 	void *mem;
1248 	struct kfd_node *dev;
1249 	long err = 0;
1250 	int i;
1251 	uint32_t *devices_arr = NULL;
1252 
1253 	if (!args->n_devices) {
1254 		pr_debug("Device IDs array empty\n");
1255 		return -EINVAL;
1256 	}
1257 	if (args->n_success > args->n_devices) {
1258 		pr_debug("n_success exceeds n_devices\n");
1259 		return -EINVAL;
1260 	}
1261 
1262 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1263 				    GFP_KERNEL);
1264 	if (!devices_arr)
1265 		return -ENOMEM;
1266 
1267 	err = copy_from_user(devices_arr,
1268 			     (void __user *)args->device_ids_array_ptr,
1269 			     args->n_devices * sizeof(*devices_arr));
1270 	if (err != 0) {
1271 		err = -EFAULT;
1272 		goto copy_from_user_failed;
1273 	}
1274 
1275 	mutex_lock(&p->mutex);
1276 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1277 	if (!pdd) {
1278 		err = -EINVAL;
1279 		goto get_process_device_data_failed;
1280 	}
1281 	dev = pdd->dev;
1282 
1283 	pdd = kfd_bind_process_to_device(dev, p);
1284 	if (IS_ERR(pdd)) {
1285 		err = PTR_ERR(pdd);
1286 		goto bind_process_to_device_failed;
1287 	}
1288 
1289 	mem = kfd_process_device_translate_handle(pdd,
1290 						GET_IDR_HANDLE(args->handle));
1291 	if (!mem) {
1292 		err = -ENOMEM;
1293 		goto get_mem_obj_from_handle_failed;
1294 	}
1295 
1296 	for (i = args->n_success; i < args->n_devices; i++) {
1297 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1298 		if (!peer_pdd) {
1299 			pr_debug("Getting device by id failed for 0x%x\n",
1300 				 devices_arr[i]);
1301 			err = -EINVAL;
1302 			goto get_mem_obj_from_handle_failed;
1303 		}
1304 
1305 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1306 		if (IS_ERR(peer_pdd)) {
1307 			err = PTR_ERR(peer_pdd);
1308 			goto get_mem_obj_from_handle_failed;
1309 		}
1310 
1311 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1312 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1313 			peer_pdd->drm_priv);
1314 		if (err) {
1315 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1316 
1317 			dev_err(dev->adev->dev,
1318 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1319 			       pci_domain_nr(pdev->bus),
1320 			       pdev->bus->number,
1321 			       PCI_SLOT(pdev->devfn),
1322 			       PCI_FUNC(pdev->devfn),
1323 			       ((struct kgd_mem *)mem)->domain);
1324 			goto map_memory_to_gpu_failed;
1325 		}
1326 		args->n_success = i+1;
1327 	}
1328 
1329 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1330 	if (err) {
1331 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1332 		goto sync_memory_failed;
1333 	}
1334 
1335 	mutex_unlock(&p->mutex);
1336 
1337 	/* Flush TLBs after waiting for the page table updates to complete */
1338 	for (i = 0; i < args->n_devices; i++) {
1339 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1340 		if (WARN_ON_ONCE(!peer_pdd))
1341 			continue;
1342 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1343 	}
1344 	kfree(devices_arr);
1345 
1346 	return err;
1347 
1348 get_process_device_data_failed:
1349 bind_process_to_device_failed:
1350 get_mem_obj_from_handle_failed:
1351 map_memory_to_gpu_failed:
1352 sync_memory_failed:
1353 	mutex_unlock(&p->mutex);
1354 copy_from_user_failed:
1355 	kfree(devices_arr);
1356 
1357 	return err;
1358 }
1359 
1360 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1361 					struct kfd_process *p, void *data)
1362 {
1363 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1364 	struct kfd_process_device *pdd, *peer_pdd;
1365 	void *mem;
1366 	long err = 0;
1367 	uint32_t *devices_arr = NULL, i;
1368 	bool flush_tlb;
1369 
1370 	if (!args->n_devices) {
1371 		pr_debug("Device IDs array empty\n");
1372 		return -EINVAL;
1373 	}
1374 	if (args->n_success > args->n_devices) {
1375 		pr_debug("n_success exceeds n_devices\n");
1376 		return -EINVAL;
1377 	}
1378 
1379 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1380 				    GFP_KERNEL);
1381 	if (!devices_arr)
1382 		return -ENOMEM;
1383 
1384 	err = copy_from_user(devices_arr,
1385 			     (void __user *)args->device_ids_array_ptr,
1386 			     args->n_devices * sizeof(*devices_arr));
1387 	if (err != 0) {
1388 		err = -EFAULT;
1389 		goto copy_from_user_failed;
1390 	}
1391 
1392 	mutex_lock(&p->mutex);
1393 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1394 	if (!pdd) {
1395 		err = -EINVAL;
1396 		goto bind_process_to_device_failed;
1397 	}
1398 
1399 	mem = kfd_process_device_translate_handle(pdd,
1400 						GET_IDR_HANDLE(args->handle));
1401 	if (!mem) {
1402 		err = -ENOMEM;
1403 		goto get_mem_obj_from_handle_failed;
1404 	}
1405 
1406 	for (i = args->n_success; i < args->n_devices; i++) {
1407 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1408 		if (!peer_pdd) {
1409 			err = -EINVAL;
1410 			goto get_mem_obj_from_handle_failed;
1411 		}
1412 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1413 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1414 		if (err) {
1415 			pr_debug("Failed to unmap from gpu %d/%d\n", i, args->n_devices);
1416 			goto unmap_memory_from_gpu_failed;
1417 		}
1418 		args->n_success = i+1;
1419 	}
1420 
1421 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1422 	if (flush_tlb) {
1423 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1424 				(struct kgd_mem *) mem, true);
1425 		if (err) {
1426 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1427 			goto sync_memory_failed;
1428 		}
1429 	}
1430 
1431 	/* Flush TLBs after waiting for the page table updates to complete */
1432 	for (i = 0; i < args->n_devices; i++) {
1433 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1434 		if (WARN_ON_ONCE(!peer_pdd))
1435 			continue;
1436 		if (flush_tlb)
1437 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1438 
1439 		/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */
1440 		err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd->drm_priv);
1441 		if (err)
1442 			goto sync_memory_failed;
1443 	}
1444 
1445 	mutex_unlock(&p->mutex);
1446 
1447 	kfree(devices_arr);
1448 
1449 	return 0;
1450 
1451 bind_process_to_device_failed:
1452 get_mem_obj_from_handle_failed:
1453 unmap_memory_from_gpu_failed:
1454 sync_memory_failed:
1455 	mutex_unlock(&p->mutex);
1456 copy_from_user_failed:
1457 	kfree(devices_arr);
1458 	return err;
1459 }
1460 
1461 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1462 		struct kfd_process *p, void *data)
1463 {
1464 	int retval;
1465 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1466 	struct queue *q;
1467 	struct kfd_node *dev;
1468 
1469 	mutex_lock(&p->mutex);
1470 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1471 
1472 	if (q) {
1473 		dev = q->device;
1474 	} else {
1475 		retval = -EINVAL;
1476 		goto out_unlock;
1477 	}
1478 
1479 	if (!dev->gws) {
1480 		retval = -ENODEV;
1481 		goto out_unlock;
1482 	}
1483 
1484 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1485 		retval = -ENODEV;
1486 		goto out_unlock;
1487 	}
1488 
1489 	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1490 				      kfd_dbg_has_cwsr_workaround(dev))) {
1491 		retval = -EBUSY;
1492 		goto out_unlock;
1493 	}
1494 
1495 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1496 	mutex_unlock(&p->mutex);
1497 
1498 	args->first_gws = 0;
1499 	return retval;
1500 
1501 out_unlock:
1502 	mutex_unlock(&p->mutex);
1503 	return retval;
1504 }
1505 
1506 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1507 		struct kfd_process *p, void *data)
1508 {
1509 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1510 	struct kfd_node *dev = NULL;
1511 	struct amdgpu_device *dmabuf_adev;
1512 	void *metadata_buffer = NULL;
1513 	uint32_t flags;
1514 	int8_t xcp_id;
1515 	unsigned int i;
1516 	int r;
1517 
1518 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1519 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1520 		if (dev && !kfd_devcgroup_check_permission(dev))
1521 			break;
1522 	if (!dev)
1523 		return -EINVAL;
1524 
1525 	if (args->metadata_ptr) {
1526 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1527 		if (!metadata_buffer)
1528 			return -ENOMEM;
1529 	}
1530 
1531 	/* Get dmabuf info from KGD */
1532 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1533 					  &dmabuf_adev, &args->size,
1534 					  metadata_buffer, args->metadata_size,
1535 					  &args->metadata_size, &flags, &xcp_id);
1536 	if (r)
1537 		goto exit;
1538 
1539 	if (xcp_id >= 0)
1540 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1541 	else
1542 		args->gpu_id = dev->id;
1543 	args->flags = flags;
1544 
1545 	/* Copy metadata buffer to user mode */
1546 	if (metadata_buffer) {
1547 		r = copy_to_user((void __user *)args->metadata_ptr,
1548 				 metadata_buffer, args->metadata_size);
1549 		if (r != 0)
1550 			r = -EFAULT;
1551 	}
1552 
1553 exit:
1554 	kfree(metadata_buffer);
1555 
1556 	return r;
1557 }
1558 
1559 static int kfd_ioctl_import_dmabuf(struct file *filep,
1560 				   struct kfd_process *p, void *data)
1561 {
1562 	struct kfd_ioctl_import_dmabuf_args *args = data;
1563 	struct kfd_process_device *pdd;
1564 	int idr_handle;
1565 	uint64_t size;
1566 	void *mem;
1567 	int r;
1568 
1569 	mutex_lock(&p->mutex);
1570 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1571 	if (!pdd) {
1572 		r = -EINVAL;
1573 		goto err_unlock;
1574 	}
1575 
1576 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1577 	if (IS_ERR(pdd)) {
1578 		r = PTR_ERR(pdd);
1579 		goto err_unlock;
1580 	}
1581 
1582 	r = amdgpu_amdkfd_gpuvm_import_dmabuf_fd(pdd->dev->adev, args->dmabuf_fd,
1583 						 args->va_addr, pdd->drm_priv,
1584 						 (struct kgd_mem **)&mem, &size,
1585 						 NULL);
1586 	if (r)
1587 		goto err_unlock;
1588 
1589 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1590 	if (idr_handle < 0) {
1591 		r = -EFAULT;
1592 		goto err_free;
1593 	}
1594 
1595 	mutex_unlock(&p->mutex);
1596 
1597 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1598 
1599 	return 0;
1600 
1601 err_free:
1602 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1603 					       pdd->drm_priv, NULL);
1604 err_unlock:
1605 	mutex_unlock(&p->mutex);
1606 	return r;
1607 }
1608 
1609 static int kfd_ioctl_export_dmabuf(struct file *filep,
1610 				   struct kfd_process *p, void *data)
1611 {
1612 	struct kfd_ioctl_export_dmabuf_args *args = data;
1613 	struct kfd_process_device *pdd;
1614 	struct dma_buf *dmabuf;
1615 	struct kfd_node *dev;
1616 	void *mem;
1617 	int ret = 0;
1618 
1619 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1620 	if (!dev)
1621 		return -EINVAL;
1622 
1623 	mutex_lock(&p->mutex);
1624 
1625 	pdd = kfd_get_process_device_data(dev, p);
1626 	if (!pdd) {
1627 		ret = -EINVAL;
1628 		goto err_unlock;
1629 	}
1630 
1631 	mem = kfd_process_device_translate_handle(pdd,
1632 						GET_IDR_HANDLE(args->handle));
1633 	if (!mem) {
1634 		ret = -EINVAL;
1635 		goto err_unlock;
1636 	}
1637 
1638 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1639 	mutex_unlock(&p->mutex);
1640 	if (ret)
1641 		goto err_out;
1642 
1643 	ret = dma_buf_fd(dmabuf, args->flags);
1644 	if (ret < 0) {
1645 		dma_buf_put(dmabuf);
1646 		goto err_out;
1647 	}
1648 	/* dma_buf_fd assigns the reference count to the fd, no need to
1649 	 * put the reference here.
1650 	 */
1651 	args->dmabuf_fd = ret;
1652 
1653 	return 0;
1654 
1655 err_unlock:
1656 	mutex_unlock(&p->mutex);
1657 err_out:
1658 	return ret;
1659 }
1660 
1661 /* Handle requests for watching SMI events */
1662 static int kfd_ioctl_smi_events(struct file *filep,
1663 				struct kfd_process *p, void *data)
1664 {
1665 	struct kfd_ioctl_smi_events_args *args = data;
1666 	struct kfd_process_device *pdd;
1667 
1668 	mutex_lock(&p->mutex);
1669 
1670 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1671 	mutex_unlock(&p->mutex);
1672 	if (!pdd)
1673 		return -EINVAL;
1674 
1675 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1676 }
1677 
1678 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1679 
1680 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1681 				    struct kfd_process *p, void *data)
1682 {
1683 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1684 	int r = 0;
1685 
1686 	mutex_lock(&p->mutex);
1687 	if (args->xnack_enabled >= 0) {
1688 		if (!list_empty(&p->pqm.queues)) {
1689 			pr_debug("Process has user queues running\n");
1690 			r = -EBUSY;
1691 			goto out_unlock;
1692 		}
1693 
1694 		if (p->xnack_enabled == args->xnack_enabled)
1695 			goto out_unlock;
1696 
1697 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1698 			r = -EPERM;
1699 			goto out_unlock;
1700 		}
1701 
1702 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1703 	} else {
1704 		args->xnack_enabled = p->xnack_enabled;
1705 	}
1706 
1707 out_unlock:
1708 	mutex_unlock(&p->mutex);
1709 
1710 	return r;
1711 }
1712 
1713 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1714 {
1715 	struct kfd_ioctl_svm_args *args = data;
1716 	int r = 0;
1717 
1718 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1719 		 args->start_addr, args->size, args->op, args->nattr);
1720 
1721 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1722 		return -EINVAL;
1723 	if (!args->start_addr || !args->size)
1724 		return -EINVAL;
1725 
1726 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1727 		      args->attrs);
1728 
1729 	return r;
1730 }
1731 #else
1732 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1733 				    struct kfd_process *p, void *data)
1734 {
1735 	return -EPERM;
1736 }
1737 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1738 {
1739 	return -EPERM;
1740 }
1741 #endif
1742 
1743 static int criu_checkpoint_process(struct kfd_process *p,
1744 			     uint8_t __user *user_priv_data,
1745 			     uint64_t *priv_offset)
1746 {
1747 	struct kfd_criu_process_priv_data process_priv;
1748 	int ret;
1749 
1750 	memset(&process_priv, 0, sizeof(process_priv));
1751 
1752 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1753 	/* For CR, we don't consider negative xnack mode which is used for
1754 	 * querying without changing it, here 0 simply means disabled and 1
1755 	 * means enabled so retry for finding a valid PTE.
1756 	 */
1757 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1758 
1759 	ret = copy_to_user(user_priv_data + *priv_offset,
1760 				&process_priv, sizeof(process_priv));
1761 
1762 	if (ret) {
1763 		pr_err("Failed to copy process information to user\n");
1764 		ret = -EFAULT;
1765 	}
1766 
1767 	*priv_offset += sizeof(process_priv);
1768 	return ret;
1769 }
1770 
1771 static int criu_checkpoint_devices(struct kfd_process *p,
1772 			     uint32_t num_devices,
1773 			     uint8_t __user *user_addr,
1774 			     uint8_t __user *user_priv_data,
1775 			     uint64_t *priv_offset)
1776 {
1777 	struct kfd_criu_device_priv_data *device_priv = NULL;
1778 	struct kfd_criu_device_bucket *device_buckets = NULL;
1779 	int ret = 0, i;
1780 
1781 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1782 	if (!device_buckets) {
1783 		ret = -ENOMEM;
1784 		goto exit;
1785 	}
1786 
1787 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1788 	if (!device_priv) {
1789 		ret = -ENOMEM;
1790 		goto exit;
1791 	}
1792 
1793 	for (i = 0; i < num_devices; i++) {
1794 		struct kfd_process_device *pdd = p->pdds[i];
1795 
1796 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1797 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1798 
1799 		/*
1800 		 * priv_data does not contain useful information for now and is reserved for
1801 		 * future use, so we do not set its contents.
1802 		 */
1803 	}
1804 
1805 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1806 	if (ret) {
1807 		pr_err("Failed to copy device information to user\n");
1808 		ret = -EFAULT;
1809 		goto exit;
1810 	}
1811 
1812 	ret = copy_to_user(user_priv_data + *priv_offset,
1813 			   device_priv,
1814 			   num_devices * sizeof(*device_priv));
1815 	if (ret) {
1816 		pr_err("Failed to copy device information to user\n");
1817 		ret = -EFAULT;
1818 	}
1819 	*priv_offset += num_devices * sizeof(*device_priv);
1820 
1821 exit:
1822 	kvfree(device_buckets);
1823 	kvfree(device_priv);
1824 	return ret;
1825 }
1826 
1827 static uint32_t get_process_num_bos(struct kfd_process *p)
1828 {
1829 	uint32_t num_of_bos = 0;
1830 	int i;
1831 
1832 	/* Run over all PDDs of the process */
1833 	for (i = 0; i < p->n_pdds; i++) {
1834 		struct kfd_process_device *pdd = p->pdds[i];
1835 		void *mem;
1836 		int id;
1837 
1838 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1839 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1840 
1841 			if (!kgd_mem->va || kgd_mem->va > pdd->gpuvm_base)
1842 				num_of_bos++;
1843 		}
1844 	}
1845 	return num_of_bos;
1846 }
1847 
1848 static int criu_get_prime_handle(struct kgd_mem *mem,
1849 				 int flags, u32 *shared_fd,
1850 				 struct file **file)
1851 {
1852 	struct dma_buf *dmabuf;
1853 	int ret;
1854 
1855 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1856 	if (ret) {
1857 		pr_err("dmabuf export failed for the BO\n");
1858 		return ret;
1859 	}
1860 
1861 	ret = get_unused_fd_flags(flags);
1862 	if (ret < 0) {
1863 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1864 		goto out_free_dmabuf;
1865 	}
1866 
1867 	*shared_fd = ret;
1868 	*file = dmabuf->file;
1869 	return 0;
1870 
1871 out_free_dmabuf:
1872 	dma_buf_put(dmabuf);
1873 	return ret;
1874 }
1875 
1876 static void commit_files(struct file **files,
1877 			 struct kfd_criu_bo_bucket *bo_buckets,
1878 			 unsigned int count,
1879 			 int err)
1880 {
1881 	while (count--) {
1882 		struct file *file = files[count];
1883 
1884 		if (!file)
1885 			continue;
1886 		if (err) {
1887 			fput(file);
1888 			put_unused_fd(bo_buckets[count].dmabuf_fd);
1889 		} else {
1890 			fd_install(bo_buckets[count].dmabuf_fd, file);
1891 		}
1892 	}
1893 }
1894 
1895 static int criu_checkpoint_bos(struct kfd_process *p,
1896 			       uint32_t num_bos,
1897 			       uint8_t __user *user_bos,
1898 			       uint8_t __user *user_priv_data,
1899 			       uint64_t *priv_offset)
1900 {
1901 	struct kfd_criu_bo_bucket *bo_buckets;
1902 	struct kfd_criu_bo_priv_data *bo_privs;
1903 	struct file **files = NULL;
1904 	int ret = 0, pdd_index, bo_index = 0, id;
1905 	void *mem;
1906 
1907 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1908 	if (!bo_buckets)
1909 		return -ENOMEM;
1910 
1911 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1912 	if (!bo_privs) {
1913 		ret = -ENOMEM;
1914 		goto exit;
1915 	}
1916 
1917 	files = kvzalloc(num_bos * sizeof(struct file *), GFP_KERNEL);
1918 	if (!files) {
1919 		ret = -ENOMEM;
1920 		goto exit;
1921 	}
1922 
1923 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1924 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1925 		struct amdgpu_bo *dumper_bo;
1926 		struct kgd_mem *kgd_mem;
1927 
1928 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1929 			struct kfd_criu_bo_bucket *bo_bucket;
1930 			struct kfd_criu_bo_priv_data *bo_priv;
1931 			int i, dev_idx = 0;
1932 
1933 			kgd_mem = (struct kgd_mem *)mem;
1934 			dumper_bo = kgd_mem->bo;
1935 
1936 			/* Skip checkpointing BOs that are used for Trap handler
1937 			 * code and state. Currently, these BOs have a VA that
1938 			 * is less GPUVM Base
1939 			 */
1940 			if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base)
1941 				continue;
1942 
1943 			bo_bucket = &bo_buckets[bo_index];
1944 			bo_priv = &bo_privs[bo_index];
1945 
1946 			bo_bucket->gpu_id = pdd->user_gpu_id;
1947 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1948 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1949 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1950 			bo_priv->idr_handle = id;
1951 
1952 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1953 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1954 								&bo_priv->user_addr);
1955 				if (ret) {
1956 					pr_err("Failed to obtain user address for user-pointer bo\n");
1957 					goto exit;
1958 				}
1959 			}
1960 			if (bo_bucket->alloc_flags
1961 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1962 				ret = criu_get_prime_handle(kgd_mem,
1963 						bo_bucket->alloc_flags &
1964 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1965 						&bo_bucket->dmabuf_fd, &files[bo_index]);
1966 				if (ret)
1967 					goto exit;
1968 			} else {
1969 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1970 			}
1971 
1972 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1973 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1974 					KFD_MMAP_GPU_ID(pdd->dev->id);
1975 			else if (bo_bucket->alloc_flags &
1976 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1977 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1978 					KFD_MMAP_GPU_ID(pdd->dev->id);
1979 			else
1980 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1981 
1982 			for (i = 0; i < p->n_pdds; i++) {
1983 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->drm_priv, kgd_mem))
1984 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1985 			}
1986 
1987 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1988 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1989 					bo_bucket->size,
1990 					bo_bucket->addr,
1991 					bo_bucket->offset,
1992 					bo_bucket->gpu_id,
1993 					bo_bucket->alloc_flags,
1994 					bo_priv->idr_handle);
1995 			bo_index++;
1996 		}
1997 	}
1998 
1999 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
2000 	if (ret) {
2001 		pr_err("Failed to copy BO information to user\n");
2002 		ret = -EFAULT;
2003 		goto exit;
2004 	}
2005 
2006 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
2007 	if (ret) {
2008 		pr_err("Failed to copy BO priv information to user\n");
2009 		ret = -EFAULT;
2010 		goto exit;
2011 	}
2012 
2013 	*priv_offset += num_bos * sizeof(*bo_privs);
2014 
2015 exit:
2016 	commit_files(files, bo_buckets, bo_index, ret);
2017 	kvfree(files);
2018 	kvfree(bo_buckets);
2019 	kvfree(bo_privs);
2020 	return ret;
2021 }
2022 
2023 static int criu_get_process_object_info(struct kfd_process *p,
2024 					uint32_t *num_devices,
2025 					uint32_t *num_bos,
2026 					uint32_t *num_objects,
2027 					uint64_t *objs_priv_size)
2028 {
2029 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2030 	uint32_t num_queues, num_events, num_svm_ranges;
2031 	int ret;
2032 
2033 	*num_devices = p->n_pdds;
2034 	*num_bos = get_process_num_bos(p);
2035 
2036 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2037 	if (ret)
2038 		return ret;
2039 
2040 	num_events = kfd_get_num_events(p);
2041 
2042 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2043 	if (ret)
2044 		return ret;
2045 
2046 	*num_objects = num_queues + num_events + num_svm_ranges;
2047 
2048 	if (objs_priv_size) {
2049 		priv_size = sizeof(struct kfd_criu_process_priv_data);
2050 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2051 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2052 		priv_size += queues_priv_data_size;
2053 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2054 		priv_size += svm_priv_data_size;
2055 		*objs_priv_size = priv_size;
2056 	}
2057 	return 0;
2058 }
2059 
2060 static int criu_checkpoint(struct file *filep,
2061 			   struct kfd_process *p,
2062 			   struct kfd_ioctl_criu_args *args)
2063 {
2064 	int ret;
2065 	uint32_t num_devices, num_bos, num_objects;
2066 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2067 
2068 	if (!args->devices || !args->bos || !args->priv_data)
2069 		return -EINVAL;
2070 
2071 	mutex_lock(&p->mutex);
2072 
2073 	if (!p->n_pdds) {
2074 		pr_err("No pdd for given process\n");
2075 		ret = -ENODEV;
2076 		goto exit_unlock;
2077 	}
2078 
2079 	/* Confirm all process queues are evicted */
2080 	if (!p->queues_paused) {
2081 		pr_err("Cannot dump process when queues are not in evicted state\n");
2082 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2083 		ret = -EINVAL;
2084 		goto exit_unlock;
2085 	}
2086 
2087 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2088 	if (ret)
2089 		goto exit_unlock;
2090 
2091 	if (num_devices != args->num_devices ||
2092 	    num_bos != args->num_bos ||
2093 	    num_objects != args->num_objects ||
2094 	    priv_size != args->priv_data_size) {
2095 
2096 		ret = -EINVAL;
2097 		goto exit_unlock;
2098 	}
2099 
2100 	/* each function will store private data inside priv_data and adjust priv_offset */
2101 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2102 	if (ret)
2103 		goto exit_unlock;
2104 
2105 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2106 				(uint8_t __user *)args->priv_data, &priv_offset);
2107 	if (ret)
2108 		goto exit_unlock;
2109 
2110 	/* Leave room for BOs in the private data. They need to be restored
2111 	 * before events, but we checkpoint them last to simplify the error
2112 	 * handling.
2113 	 */
2114 	bo_priv_offset = priv_offset;
2115 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2116 
2117 	if (num_objects) {
2118 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2119 						 &priv_offset);
2120 		if (ret)
2121 			goto exit_unlock;
2122 
2123 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2124 						 &priv_offset);
2125 		if (ret)
2126 			goto exit_unlock;
2127 
2128 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2129 		if (ret)
2130 			goto exit_unlock;
2131 	}
2132 
2133 	/* This must be the last thing in this function that can fail.
2134 	 * Otherwise we leak dmabuf file descriptors.
2135 	 */
2136 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2137 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2138 
2139 exit_unlock:
2140 	mutex_unlock(&p->mutex);
2141 	if (ret)
2142 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2143 	else
2144 		pr_debug("CRIU dump ret:%d\n", ret);
2145 
2146 	return ret;
2147 }
2148 
2149 static int criu_restore_process(struct kfd_process *p,
2150 				struct kfd_ioctl_criu_args *args,
2151 				uint64_t *priv_offset,
2152 				uint64_t max_priv_data_size)
2153 {
2154 	int ret = 0;
2155 	struct kfd_criu_process_priv_data process_priv;
2156 
2157 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2158 		return -EINVAL;
2159 
2160 	ret = copy_from_user(&process_priv,
2161 				(void __user *)(args->priv_data + *priv_offset),
2162 				sizeof(process_priv));
2163 	if (ret) {
2164 		pr_err("Failed to copy process private information from user\n");
2165 		ret = -EFAULT;
2166 		goto exit;
2167 	}
2168 	*priv_offset += sizeof(process_priv);
2169 
2170 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2171 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2172 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2173 		return -EINVAL;
2174 	}
2175 
2176 	pr_debug("Setting XNACK mode\n");
2177 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2178 		pr_err("xnack mode cannot be set\n");
2179 		ret = -EPERM;
2180 		goto exit;
2181 	} else {
2182 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2183 		p->xnack_enabled = process_priv.xnack_mode;
2184 	}
2185 
2186 exit:
2187 	return ret;
2188 }
2189 
2190 static int criu_restore_devices(struct kfd_process *p,
2191 				struct kfd_ioctl_criu_args *args,
2192 				uint64_t *priv_offset,
2193 				uint64_t max_priv_data_size)
2194 {
2195 	struct kfd_criu_device_bucket *device_buckets;
2196 	struct kfd_criu_device_priv_data *device_privs;
2197 	int ret = 0;
2198 	uint32_t i;
2199 
2200 	if (args->num_devices != p->n_pdds)
2201 		return -EINVAL;
2202 
2203 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2204 		return -EINVAL;
2205 
2206 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2207 	if (!device_buckets)
2208 		return -ENOMEM;
2209 
2210 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2211 				args->num_devices * sizeof(*device_buckets));
2212 	if (ret) {
2213 		pr_err("Failed to copy devices buckets from user\n");
2214 		ret = -EFAULT;
2215 		goto exit;
2216 	}
2217 
2218 	for (i = 0; i < args->num_devices; i++) {
2219 		struct kfd_node *dev;
2220 		struct kfd_process_device *pdd;
2221 		struct file *drm_file;
2222 
2223 		/* device private data is not currently used */
2224 
2225 		if (!device_buckets[i].user_gpu_id) {
2226 			pr_err("Invalid user gpu_id\n");
2227 			ret = -EINVAL;
2228 			goto exit;
2229 		}
2230 
2231 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2232 		if (!dev) {
2233 			pr_err("Failed to find device with gpu_id = %x\n",
2234 				device_buckets[i].actual_gpu_id);
2235 			ret = -EINVAL;
2236 			goto exit;
2237 		}
2238 
2239 		pdd = kfd_get_process_device_data(dev, p);
2240 		if (!pdd) {
2241 			pr_err("Failed to get pdd for gpu_id = %x\n",
2242 					device_buckets[i].actual_gpu_id);
2243 			ret = -EINVAL;
2244 			goto exit;
2245 		}
2246 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2247 
2248 		drm_file = fget(device_buckets[i].drm_fd);
2249 		if (!drm_file) {
2250 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2251 				device_buckets[i].drm_fd);
2252 			ret = -EINVAL;
2253 			goto exit;
2254 		}
2255 
2256 		if (pdd->drm_file) {
2257 			ret = -EINVAL;
2258 			goto exit;
2259 		}
2260 
2261 		/* create the vm using render nodes for kfd pdd */
2262 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2263 			pr_err("could not init vm for given pdd\n");
2264 			/* On success, the PDD keeps the drm_file reference */
2265 			fput(drm_file);
2266 			ret = -EINVAL;
2267 			goto exit;
2268 		}
2269 		/*
2270 		 * pdd now already has the vm bound to render node so below api won't create a new
2271 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2272 		 * for iommu v2 binding  and runtime pm.
2273 		 */
2274 		pdd = kfd_bind_process_to_device(dev, p);
2275 		if (IS_ERR(pdd)) {
2276 			ret = PTR_ERR(pdd);
2277 			goto exit;
2278 		}
2279 
2280 		if (!pdd->qpd.proc_doorbells) {
2281 			ret = kfd_alloc_process_doorbells(dev->kfd, pdd);
2282 			if (ret)
2283 				goto exit;
2284 		}
2285 	}
2286 
2287 	/*
2288 	 * We are not copying device private data from user as we are not using the data for now,
2289 	 * but we still adjust for its private data.
2290 	 */
2291 	*priv_offset += args->num_devices * sizeof(*device_privs);
2292 
2293 exit:
2294 	kfree(device_buckets);
2295 	return ret;
2296 }
2297 
2298 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2299 				      struct kfd_criu_bo_bucket *bo_bucket,
2300 				      struct kfd_criu_bo_priv_data *bo_priv,
2301 				      struct kgd_mem **kgd_mem)
2302 {
2303 	int idr_handle;
2304 	int ret;
2305 	const bool criu_resume = true;
2306 	u64 offset;
2307 
2308 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2309 		if (bo_bucket->size !=
2310 				kfd_doorbell_process_slice(pdd->dev->kfd))
2311 			return -EINVAL;
2312 
2313 		offset = kfd_get_process_doorbells(pdd);
2314 		if (!offset)
2315 			return -ENOMEM;
2316 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2317 		/* MMIO BOs need remapped bus address */
2318 		if (bo_bucket->size != PAGE_SIZE) {
2319 			pr_err("Invalid page size\n");
2320 			return -EINVAL;
2321 		}
2322 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2323 		if (!offset || (PAGE_SIZE > 4096)) {
2324 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2325 			return -ENOMEM;
2326 		}
2327 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2328 		offset = bo_priv->user_addr;
2329 	}
2330 	/* Create the BO */
2331 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2332 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2333 						      &offset, bo_bucket->alloc_flags, criu_resume);
2334 	if (ret) {
2335 		pr_err("Could not create the BO\n");
2336 		return ret;
2337 	}
2338 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2339 		 bo_bucket->size, bo_bucket->addr, offset);
2340 
2341 	/* Restore previous IDR handle */
2342 	pr_debug("Restoring old IDR handle for the BO");
2343 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2344 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2345 
2346 	if (idr_handle < 0) {
2347 		pr_err("Could not allocate idr\n");
2348 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2349 						       NULL);
2350 		return -ENOMEM;
2351 	}
2352 
2353 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2354 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2355 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2356 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2357 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2358 		bo_bucket->restored_offset = offset;
2359 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2360 		bo_bucket->restored_offset = offset;
2361 		/* Update the VRAM usage count */
2362 		atomic64_add(bo_bucket->size, &pdd->vram_usage);
2363 	}
2364 	return 0;
2365 }
2366 
2367 static int criu_restore_bo(struct kfd_process *p,
2368 			   struct kfd_criu_bo_bucket *bo_bucket,
2369 			   struct kfd_criu_bo_priv_data *bo_priv,
2370 			   struct file **file)
2371 {
2372 	struct kfd_process_device *pdd;
2373 	struct kgd_mem *kgd_mem;
2374 	int ret;
2375 	int j;
2376 
2377 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2378 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2379 		 bo_priv->idr_handle);
2380 
2381 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2382 	if (!pdd) {
2383 		pr_err("Failed to get pdd\n");
2384 		return -ENODEV;
2385 	}
2386 
2387 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2388 	if (ret)
2389 		return ret;
2390 
2391 	/* now map these BOs to GPU/s */
2392 	for (j = 0; j < p->n_pdds; j++) {
2393 		struct kfd_node *peer;
2394 		struct kfd_process_device *peer_pdd;
2395 
2396 		if (!bo_priv->mapped_gpuids[j])
2397 			break;
2398 
2399 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2400 		if (!peer_pdd)
2401 			return -EINVAL;
2402 
2403 		peer = peer_pdd->dev;
2404 
2405 		peer_pdd = kfd_bind_process_to_device(peer, p);
2406 		if (IS_ERR(peer_pdd))
2407 			return PTR_ERR(peer_pdd);
2408 
2409 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2410 							    peer_pdd->drm_priv);
2411 		if (ret) {
2412 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2413 			return ret;
2414 		}
2415 	}
2416 
2417 	pr_debug("map memory was successful for the BO\n");
2418 	/* create the dmabuf object and export the bo */
2419 	if (bo_bucket->alloc_flags
2420 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2421 		ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2422 					    &bo_bucket->dmabuf_fd, file);
2423 		if (ret)
2424 			return ret;
2425 	} else {
2426 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2427 	}
2428 
2429 	return 0;
2430 }
2431 
2432 static int criu_restore_bos(struct kfd_process *p,
2433 			    struct kfd_ioctl_criu_args *args,
2434 			    uint64_t *priv_offset,
2435 			    uint64_t max_priv_data_size)
2436 {
2437 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2438 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2439 	struct file **files = NULL;
2440 	int ret = 0;
2441 	uint32_t i = 0;
2442 
2443 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2444 		return -EINVAL;
2445 
2446 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2447 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2448 
2449 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2450 	if (!bo_buckets)
2451 		return -ENOMEM;
2452 
2453 	files = kvzalloc(args->num_bos * sizeof(struct file *), GFP_KERNEL);
2454 	if (!files) {
2455 		ret = -ENOMEM;
2456 		goto exit;
2457 	}
2458 
2459 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2460 			     args->num_bos * sizeof(*bo_buckets));
2461 	if (ret) {
2462 		pr_err("Failed to copy BOs information from user\n");
2463 		ret = -EFAULT;
2464 		goto exit;
2465 	}
2466 
2467 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2468 	if (!bo_privs) {
2469 		ret = -ENOMEM;
2470 		goto exit;
2471 	}
2472 
2473 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2474 			     args->num_bos * sizeof(*bo_privs));
2475 	if (ret) {
2476 		pr_err("Failed to copy BOs information from user\n");
2477 		ret = -EFAULT;
2478 		goto exit;
2479 	}
2480 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2481 
2482 	/* Create and map new BOs */
2483 	for (; i < args->num_bos; i++) {
2484 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i], &files[i]);
2485 		if (ret) {
2486 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2487 			goto exit;
2488 		}
2489 	} /* done */
2490 
2491 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2492 	ret = copy_to_user((void __user *)args->bos,
2493 				bo_buckets,
2494 				(args->num_bos * sizeof(*bo_buckets)));
2495 	if (ret)
2496 		ret = -EFAULT;
2497 
2498 exit:
2499 	commit_files(files, bo_buckets, i, ret);
2500 	kvfree(files);
2501 	kvfree(bo_buckets);
2502 	kvfree(bo_privs);
2503 	return ret;
2504 }
2505 
2506 static int criu_restore_objects(struct file *filep,
2507 				struct kfd_process *p,
2508 				struct kfd_ioctl_criu_args *args,
2509 				uint64_t *priv_offset,
2510 				uint64_t max_priv_data_size)
2511 {
2512 	int ret = 0;
2513 	uint32_t i;
2514 
2515 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2516 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2517 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2518 
2519 	for (i = 0; i < args->num_objects; i++) {
2520 		uint32_t object_type;
2521 
2522 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2523 			pr_err("Invalid private data size\n");
2524 			return -EINVAL;
2525 		}
2526 
2527 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2528 		if (ret) {
2529 			pr_err("Failed to copy private information from user\n");
2530 			goto exit;
2531 		}
2532 
2533 		switch (object_type) {
2534 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2535 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2536 						     priv_offset, max_priv_data_size);
2537 			if (ret)
2538 				goto exit;
2539 			break;
2540 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2541 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2542 						     priv_offset, max_priv_data_size);
2543 			if (ret)
2544 				goto exit;
2545 			break;
2546 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2547 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2548 						     priv_offset, max_priv_data_size);
2549 			if (ret)
2550 				goto exit;
2551 			break;
2552 		default:
2553 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2554 			ret = -EINVAL;
2555 			goto exit;
2556 		}
2557 	}
2558 exit:
2559 	return ret;
2560 }
2561 
2562 static int criu_restore(struct file *filep,
2563 			struct kfd_process *p,
2564 			struct kfd_ioctl_criu_args *args)
2565 {
2566 	uint64_t priv_offset = 0;
2567 	int ret = 0;
2568 
2569 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2570 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2571 
2572 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2573 	    !args->num_devices || !args->num_bos)
2574 		return -EINVAL;
2575 
2576 	mutex_lock(&p->mutex);
2577 
2578 	/*
2579 	 * Set the process to evicted state to avoid running any new queues before all the memory
2580 	 * mappings are ready.
2581 	 */
2582 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2583 	if (ret)
2584 		goto exit_unlock;
2585 
2586 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2587 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2588 	if (ret)
2589 		goto exit_unlock;
2590 
2591 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2592 	if (ret)
2593 		goto exit_unlock;
2594 
2595 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2596 	if (ret)
2597 		goto exit_unlock;
2598 
2599 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2600 	if (ret)
2601 		goto exit_unlock;
2602 
2603 	if (priv_offset != args->priv_data_size) {
2604 		pr_err("Invalid private data size\n");
2605 		ret = -EINVAL;
2606 	}
2607 
2608 exit_unlock:
2609 	mutex_unlock(&p->mutex);
2610 	if (ret)
2611 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2612 	else
2613 		pr_debug("CRIU restore successful\n");
2614 
2615 	return ret;
2616 }
2617 
2618 static int criu_unpause(struct file *filep,
2619 			struct kfd_process *p,
2620 			struct kfd_ioctl_criu_args *args)
2621 {
2622 	int ret;
2623 
2624 	mutex_lock(&p->mutex);
2625 
2626 	if (!p->queues_paused) {
2627 		mutex_unlock(&p->mutex);
2628 		return -EINVAL;
2629 	}
2630 
2631 	ret = kfd_process_restore_queues(p);
2632 	if (ret)
2633 		pr_err("Failed to unpause queues ret:%d\n", ret);
2634 	else
2635 		p->queues_paused = false;
2636 
2637 	mutex_unlock(&p->mutex);
2638 
2639 	return ret;
2640 }
2641 
2642 static int criu_resume(struct file *filep,
2643 			struct kfd_process *p,
2644 			struct kfd_ioctl_criu_args *args)
2645 {
2646 	struct kfd_process *target = NULL;
2647 	struct pid *pid = NULL;
2648 	int ret = 0;
2649 
2650 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2651 		 args->pid);
2652 
2653 	pid = find_get_pid(args->pid);
2654 	if (!pid) {
2655 		pr_err("Cannot find pid info for %i\n", args->pid);
2656 		return -ESRCH;
2657 	}
2658 
2659 	pr_debug("calling kfd_lookup_process_by_pid\n");
2660 	target = kfd_lookup_process_by_pid(pid);
2661 
2662 	put_pid(pid);
2663 
2664 	if (!target) {
2665 		pr_debug("Cannot find process info for %i\n", args->pid);
2666 		return -ESRCH;
2667 	}
2668 
2669 	mutex_lock(&target->mutex);
2670 	ret = kfd_criu_resume_svm(target);
2671 	if (ret) {
2672 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2673 		goto exit;
2674 	}
2675 
2676 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2677 	if (ret)
2678 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2679 
2680 exit:
2681 	mutex_unlock(&target->mutex);
2682 
2683 	kfd_unref_process(target);
2684 	return ret;
2685 }
2686 
2687 static int criu_process_info(struct file *filep,
2688 				struct kfd_process *p,
2689 				struct kfd_ioctl_criu_args *args)
2690 {
2691 	int ret = 0;
2692 
2693 	mutex_lock(&p->mutex);
2694 
2695 	if (!p->n_pdds) {
2696 		pr_err("No pdd for given process\n");
2697 		ret = -ENODEV;
2698 		goto err_unlock;
2699 	}
2700 
2701 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2702 	if (ret)
2703 		goto err_unlock;
2704 
2705 	p->queues_paused = true;
2706 
2707 	args->pid = task_pid_nr_ns(p->lead_thread,
2708 					task_active_pid_ns(p->lead_thread));
2709 
2710 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2711 					   &args->num_objects, &args->priv_data_size);
2712 	if (ret)
2713 		goto err_unlock;
2714 
2715 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2716 				args->num_devices, args->num_bos, args->num_objects,
2717 				args->priv_data_size);
2718 
2719 err_unlock:
2720 	if (ret) {
2721 		kfd_process_restore_queues(p);
2722 		p->queues_paused = false;
2723 	}
2724 	mutex_unlock(&p->mutex);
2725 	return ret;
2726 }
2727 
2728 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2729 {
2730 	struct kfd_ioctl_criu_args *args = data;
2731 	int ret;
2732 
2733 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2734 	switch (args->op) {
2735 	case KFD_CRIU_OP_PROCESS_INFO:
2736 		ret = criu_process_info(filep, p, args);
2737 		break;
2738 	case KFD_CRIU_OP_CHECKPOINT:
2739 		ret = criu_checkpoint(filep, p, args);
2740 		break;
2741 	case KFD_CRIU_OP_UNPAUSE:
2742 		ret = criu_unpause(filep, p, args);
2743 		break;
2744 	case KFD_CRIU_OP_RESTORE:
2745 		ret = criu_restore(filep, p, args);
2746 		break;
2747 	case KFD_CRIU_OP_RESUME:
2748 		ret = criu_resume(filep, p, args);
2749 		break;
2750 	default:
2751 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2752 		ret = -EINVAL;
2753 		break;
2754 	}
2755 
2756 	if (ret)
2757 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2758 
2759 	return ret;
2760 }
2761 
2762 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2763 			bool enable_ttmp_setup)
2764 {
2765 	int i = 0, ret = 0;
2766 
2767 	if (p->is_runtime_retry)
2768 		goto retry;
2769 
2770 	if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2771 		return -EBUSY;
2772 
2773 	for (i = 0; i < p->n_pdds; i++) {
2774 		struct kfd_process_device *pdd = p->pdds[i];
2775 
2776 		if (pdd->qpd.queue_count)
2777 			return -EEXIST;
2778 
2779 		/*
2780 		 * Setup TTMPs by default.
2781 		 * Note that this call must remain here for MES ADD QUEUE to
2782 		 * skip_process_ctx_clear unconditionally as the first call to
2783 		 * SET_SHADER_DEBUGGER clears any stale process context data
2784 		 * saved in MES.
2785 		 */
2786 		if (pdd->dev->kfd->shared_resources.enable_mes)
2787 			kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
2788 	}
2789 
2790 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2791 	p->runtime_info.r_debug = r_debug;
2792 	p->runtime_info.ttmp_setup = enable_ttmp_setup;
2793 
2794 	if (p->runtime_info.ttmp_setup) {
2795 		for (i = 0; i < p->n_pdds; i++) {
2796 			struct kfd_process_device *pdd = p->pdds[i];
2797 
2798 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2799 				amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2800 				pdd->dev->kfd2kgd->enable_debug_trap(
2801 						pdd->dev->adev,
2802 						true,
2803 						pdd->dev->vm_info.last_vmid_kfd);
2804 			} else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2805 				pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2806 						pdd->dev->adev,
2807 						false,
2808 						0);
2809 			}
2810 		}
2811 	}
2812 
2813 retry:
2814 	if (p->debug_trap_enabled) {
2815 		if (!p->is_runtime_retry) {
2816 			kfd_dbg_trap_activate(p);
2817 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2818 					p, NULL, 0, false, NULL, 0);
2819 		}
2820 
2821 		mutex_unlock(&p->mutex);
2822 		ret = down_interruptible(&p->runtime_enable_sema);
2823 		mutex_lock(&p->mutex);
2824 
2825 		p->is_runtime_retry = !!ret;
2826 	}
2827 
2828 	return ret;
2829 }
2830 
2831 static int runtime_disable(struct kfd_process *p)
2832 {
2833 	int i = 0, ret;
2834 	bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2835 
2836 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2837 	p->runtime_info.r_debug = 0;
2838 
2839 	if (p->debug_trap_enabled) {
2840 		if (was_enabled)
2841 			kfd_dbg_trap_deactivate(p, false, 0);
2842 
2843 		if (!p->is_runtime_retry)
2844 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2845 					p, NULL, 0, false, NULL, 0);
2846 
2847 		mutex_unlock(&p->mutex);
2848 		ret = down_interruptible(&p->runtime_enable_sema);
2849 		mutex_lock(&p->mutex);
2850 
2851 		p->is_runtime_retry = !!ret;
2852 		if (ret)
2853 			return ret;
2854 	}
2855 
2856 	if (was_enabled && p->runtime_info.ttmp_setup) {
2857 		for (i = 0; i < p->n_pdds; i++) {
2858 			struct kfd_process_device *pdd = p->pdds[i];
2859 
2860 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2861 				amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2862 		}
2863 	}
2864 
2865 	p->runtime_info.ttmp_setup = false;
2866 
2867 	/* disable ttmp setup */
2868 	for (i = 0; i < p->n_pdds; i++) {
2869 		struct kfd_process_device *pdd = p->pdds[i];
2870 
2871 		if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2872 			pdd->spi_dbg_override =
2873 					pdd->dev->kfd2kgd->disable_debug_trap(
2874 					pdd->dev->adev,
2875 					false,
2876 					pdd->dev->vm_info.last_vmid_kfd);
2877 
2878 			if (!pdd->dev->kfd->shared_resources.enable_mes)
2879 				debug_refresh_runlist(pdd->dev->dqm);
2880 			else
2881 				kfd_dbg_set_mes_debug_mode(pdd,
2882 							   !kfd_dbg_has_cwsr_workaround(pdd->dev));
2883 		}
2884 	}
2885 
2886 	return 0;
2887 }
2888 
2889 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2890 {
2891 	struct kfd_ioctl_runtime_enable_args *args = data;
2892 	int r;
2893 
2894 	mutex_lock(&p->mutex);
2895 
2896 	if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2897 		r = runtime_enable(p, args->r_debug,
2898 				!!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2899 	else
2900 		r = runtime_disable(p);
2901 
2902 	mutex_unlock(&p->mutex);
2903 
2904 	return r;
2905 }
2906 
2907 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2908 {
2909 	struct kfd_ioctl_dbg_trap_args *args = data;
2910 	struct task_struct *thread = NULL;
2911 	struct mm_struct *mm = NULL;
2912 	struct pid *pid = NULL;
2913 	struct kfd_process *target = NULL;
2914 	struct kfd_process_device *pdd = NULL;
2915 	int r = 0;
2916 
2917 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2918 		pr_err("Debugging does not support sched_policy %i", sched_policy);
2919 		return -EINVAL;
2920 	}
2921 
2922 	pid = find_get_pid(args->pid);
2923 	if (!pid) {
2924 		pr_debug("Cannot find pid info for %i\n", args->pid);
2925 		r = -ESRCH;
2926 		goto out;
2927 	}
2928 
2929 	thread = get_pid_task(pid, PIDTYPE_PID);
2930 	if (!thread) {
2931 		r = -ESRCH;
2932 		goto out;
2933 	}
2934 
2935 	mm = get_task_mm(thread);
2936 	if (!mm) {
2937 		r = -ESRCH;
2938 		goto out;
2939 	}
2940 
2941 	if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2942 		bool create_process;
2943 
2944 		rcu_read_lock();
2945 		create_process = thread && thread != current && ptrace_parent(thread) == current;
2946 		rcu_read_unlock();
2947 
2948 		target = create_process ? kfd_create_process(thread) :
2949 					kfd_lookup_process_by_pid(pid);
2950 	} else {
2951 		target = kfd_lookup_process_by_pid(pid);
2952 	}
2953 
2954 	if (IS_ERR_OR_NULL(target)) {
2955 		pr_debug("Cannot find process PID %i to debug\n", args->pid);
2956 		r = target ? PTR_ERR(target) : -ESRCH;
2957 		target = NULL;
2958 		goto out;
2959 	}
2960 
2961 	/* Check if target is still PTRACED. */
2962 	rcu_read_lock();
2963 	if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2964 				&& ptrace_parent(target->lead_thread) != current) {
2965 		pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2966 		r = -EPERM;
2967 	}
2968 	rcu_read_unlock();
2969 
2970 	if (r)
2971 		goto out;
2972 
2973 	mutex_lock(&target->mutex);
2974 
2975 	if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2976 		pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2977 		r = -EINVAL;
2978 		goto unlock_out;
2979 	}
2980 
2981 	if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2982 			(args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2983 			 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2984 			 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2985 			 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2986 			 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2987 			 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2988 			 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2989 		r = -EPERM;
2990 		goto unlock_out;
2991 	}
2992 
2993 	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2994 	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2995 		int user_gpu_id = kfd_process_get_user_gpu_id(target,
2996 				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2997 					args->set_node_address_watch.gpu_id :
2998 					args->clear_node_address_watch.gpu_id);
2999 
3000 		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
3001 		if (user_gpu_id == -EINVAL || !pdd) {
3002 			r = -ENODEV;
3003 			goto unlock_out;
3004 		}
3005 	}
3006 
3007 	switch (args->op) {
3008 	case KFD_IOC_DBG_TRAP_ENABLE:
3009 		if (target != p)
3010 			target->debugger_process = p;
3011 
3012 		r = kfd_dbg_trap_enable(target,
3013 					args->enable.dbg_fd,
3014 					(void __user *)args->enable.rinfo_ptr,
3015 					&args->enable.rinfo_size);
3016 		if (!r)
3017 			target->exception_enable_mask = args->enable.exception_mask;
3018 
3019 		break;
3020 	case KFD_IOC_DBG_TRAP_DISABLE:
3021 		r = kfd_dbg_trap_disable(target);
3022 		break;
3023 	case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
3024 		r = kfd_dbg_send_exception_to_runtime(target,
3025 				args->send_runtime_event.gpu_id,
3026 				args->send_runtime_event.queue_id,
3027 				args->send_runtime_event.exception_mask);
3028 		break;
3029 	case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
3030 		kfd_dbg_set_enabled_debug_exception_mask(target,
3031 				args->set_exceptions_enabled.exception_mask);
3032 		break;
3033 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
3034 		r = kfd_dbg_trap_set_wave_launch_override(target,
3035 				args->launch_override.override_mode,
3036 				args->launch_override.enable_mask,
3037 				args->launch_override.support_request_mask,
3038 				&args->launch_override.enable_mask,
3039 				&args->launch_override.support_request_mask);
3040 		break;
3041 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
3042 		r = kfd_dbg_trap_set_wave_launch_mode(target,
3043 				args->launch_mode.launch_mode);
3044 		break;
3045 	case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
3046 		r = suspend_queues(target,
3047 				args->suspend_queues.num_queues,
3048 				args->suspend_queues.grace_period,
3049 				args->suspend_queues.exception_mask,
3050 				(uint32_t *)args->suspend_queues.queue_array_ptr);
3051 
3052 		break;
3053 	case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
3054 		r = resume_queues(target, args->resume_queues.num_queues,
3055 				(uint32_t *)args->resume_queues.queue_array_ptr);
3056 		break;
3057 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3058 		r = kfd_dbg_trap_set_dev_address_watch(pdd,
3059 				args->set_node_address_watch.address,
3060 				args->set_node_address_watch.mask,
3061 				&args->set_node_address_watch.id,
3062 				args->set_node_address_watch.mode);
3063 		break;
3064 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3065 		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3066 				args->clear_node_address_watch.id);
3067 		break;
3068 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
3069 		r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3070 		break;
3071 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3072 		r = kfd_dbg_ev_query_debug_event(target,
3073 				&args->query_debug_event.queue_id,
3074 				&args->query_debug_event.gpu_id,
3075 				args->query_debug_event.exception_mask,
3076 				&args->query_debug_event.exception_mask);
3077 		break;
3078 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3079 		r = kfd_dbg_trap_query_exception_info(target,
3080 				args->query_exception_info.source_id,
3081 				args->query_exception_info.exception_code,
3082 				args->query_exception_info.clear_exception,
3083 				(void __user *)args->query_exception_info.info_ptr,
3084 				&args->query_exception_info.info_size);
3085 		break;
3086 	case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3087 		r = pqm_get_queue_snapshot(&target->pqm,
3088 				args->queue_snapshot.exception_mask,
3089 				(void __user *)args->queue_snapshot.snapshot_buf_ptr,
3090 				&args->queue_snapshot.num_queues,
3091 				&args->queue_snapshot.entry_size);
3092 		break;
3093 	case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3094 		r = kfd_dbg_trap_device_snapshot(target,
3095 				args->device_snapshot.exception_mask,
3096 				(void __user *)args->device_snapshot.snapshot_buf_ptr,
3097 				&args->device_snapshot.num_devices,
3098 				&args->device_snapshot.entry_size);
3099 		break;
3100 	default:
3101 		pr_err("Invalid option: %i\n", args->op);
3102 		r = -EINVAL;
3103 	}
3104 
3105 unlock_out:
3106 	mutex_unlock(&target->mutex);
3107 
3108 out:
3109 	if (thread)
3110 		put_task_struct(thread);
3111 
3112 	if (mm)
3113 		mmput(mm);
3114 
3115 	if (pid)
3116 		put_pid(pid);
3117 
3118 	if (target)
3119 		kfd_unref_process(target);
3120 
3121 	return r;
3122 }
3123 
3124 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3125 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3126 			    .cmd_drv = 0, .name = #ioctl}
3127 
3128 /** Ioctl table */
3129 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3130 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3131 			kfd_ioctl_get_version, 0),
3132 
3133 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3134 			kfd_ioctl_create_queue, 0),
3135 
3136 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3137 			kfd_ioctl_destroy_queue, 0),
3138 
3139 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3140 			kfd_ioctl_set_memory_policy, 0),
3141 
3142 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3143 			kfd_ioctl_get_clock_counters, 0),
3144 
3145 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3146 			kfd_ioctl_get_process_apertures, 0),
3147 
3148 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3149 			kfd_ioctl_update_queue, 0),
3150 
3151 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3152 			kfd_ioctl_create_event, 0),
3153 
3154 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3155 			kfd_ioctl_destroy_event, 0),
3156 
3157 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3158 			kfd_ioctl_set_event, 0),
3159 
3160 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3161 			kfd_ioctl_reset_event, 0),
3162 
3163 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3164 			kfd_ioctl_wait_events, 0),
3165 
3166 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3167 			kfd_ioctl_dbg_register, 0),
3168 
3169 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3170 			kfd_ioctl_dbg_unregister, 0),
3171 
3172 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3173 			kfd_ioctl_dbg_address_watch, 0),
3174 
3175 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3176 			kfd_ioctl_dbg_wave_control, 0),
3177 
3178 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3179 			kfd_ioctl_set_scratch_backing_va, 0),
3180 
3181 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3182 			kfd_ioctl_get_tile_config, 0),
3183 
3184 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3185 			kfd_ioctl_set_trap_handler, 0),
3186 
3187 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3188 			kfd_ioctl_get_process_apertures_new, 0),
3189 
3190 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3191 			kfd_ioctl_acquire_vm, 0),
3192 
3193 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3194 			kfd_ioctl_alloc_memory_of_gpu, 0),
3195 
3196 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3197 			kfd_ioctl_free_memory_of_gpu, 0),
3198 
3199 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3200 			kfd_ioctl_map_memory_to_gpu, 0),
3201 
3202 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3203 			kfd_ioctl_unmap_memory_from_gpu, 0),
3204 
3205 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3206 			kfd_ioctl_set_cu_mask, 0),
3207 
3208 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3209 			kfd_ioctl_get_queue_wave_state, 0),
3210 
3211 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3212 				kfd_ioctl_get_dmabuf_info, 0),
3213 
3214 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3215 				kfd_ioctl_import_dmabuf, 0),
3216 
3217 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3218 			kfd_ioctl_alloc_queue_gws, 0),
3219 
3220 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3221 			kfd_ioctl_smi_events, 0),
3222 
3223 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3224 
3225 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3226 			kfd_ioctl_set_xnack_mode, 0),
3227 
3228 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3229 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3230 
3231 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3232 			kfd_ioctl_get_available_memory, 0),
3233 
3234 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3235 				kfd_ioctl_export_dmabuf, 0),
3236 
3237 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3238 			kfd_ioctl_runtime_enable, 0),
3239 
3240 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3241 			kfd_ioctl_set_debug_trap, 0),
3242 };
3243 
3244 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
3245 
3246 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3247 {
3248 	struct kfd_process *process;
3249 	amdkfd_ioctl_t *func;
3250 	const struct amdkfd_ioctl_desc *ioctl = NULL;
3251 	unsigned int nr = _IOC_NR(cmd);
3252 	char stack_kdata[128];
3253 	char *kdata = NULL;
3254 	unsigned int usize, asize;
3255 	int retcode = -EINVAL;
3256 	bool ptrace_attached = false;
3257 
3258 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3259 		goto err_i1;
3260 
3261 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3262 		u32 amdkfd_size;
3263 
3264 		ioctl = &amdkfd_ioctls[nr];
3265 
3266 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
3267 		usize = asize = _IOC_SIZE(cmd);
3268 		if (amdkfd_size > asize)
3269 			asize = amdkfd_size;
3270 
3271 		cmd = ioctl->cmd;
3272 	} else
3273 		goto err_i1;
3274 
3275 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3276 
3277 	/* Get the process struct from the filep. Only the process
3278 	 * that opened /dev/kfd can use the file descriptor. Child
3279 	 * processes need to create their own KFD device context.
3280 	 */
3281 	process = filep->private_data;
3282 
3283 	rcu_read_lock();
3284 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3285 	    ptrace_parent(process->lead_thread) == current)
3286 		ptrace_attached = true;
3287 	rcu_read_unlock();
3288 
3289 	if (process->lead_thread != current->group_leader
3290 	    && !ptrace_attached) {
3291 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3292 		retcode = -EBADF;
3293 		goto err_i1;
3294 	}
3295 
3296 	/* Do not trust userspace, use our own definition */
3297 	func = ioctl->func;
3298 
3299 	if (unlikely(!func)) {
3300 		dev_dbg(kfd_device, "no function\n");
3301 		retcode = -EINVAL;
3302 		goto err_i1;
3303 	}
3304 
3305 	/*
3306 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3307 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3308 	 * more priviledged access.
3309 	 */
3310 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3311 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
3312 						!capable(CAP_SYS_ADMIN)) {
3313 			retcode = -EACCES;
3314 			goto err_i1;
3315 		}
3316 	}
3317 
3318 	if (cmd & (IOC_IN | IOC_OUT)) {
3319 		if (asize <= sizeof(stack_kdata)) {
3320 			kdata = stack_kdata;
3321 		} else {
3322 			kdata = kmalloc(asize, GFP_KERNEL);
3323 			if (!kdata) {
3324 				retcode = -ENOMEM;
3325 				goto err_i1;
3326 			}
3327 		}
3328 		if (asize > usize)
3329 			memset(kdata + usize, 0, asize - usize);
3330 	}
3331 
3332 	if (cmd & IOC_IN) {
3333 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3334 			retcode = -EFAULT;
3335 			goto err_i1;
3336 		}
3337 	} else if (cmd & IOC_OUT) {
3338 		memset(kdata, 0, usize);
3339 	}
3340 
3341 	retcode = func(filep, process, kdata);
3342 
3343 	if (cmd & IOC_OUT)
3344 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3345 			retcode = -EFAULT;
3346 
3347 err_i1:
3348 	if (!ioctl)
3349 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3350 			  task_pid_nr(current), cmd, nr);
3351 
3352 	if (kdata != stack_kdata)
3353 		kfree(kdata);
3354 
3355 	if (retcode)
3356 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3357 				nr, arg, retcode);
3358 
3359 	return retcode;
3360 }
3361 
3362 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3363 		      struct vm_area_struct *vma)
3364 {
3365 	phys_addr_t address;
3366 
3367 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3368 		return -EINVAL;
3369 
3370 	if (PAGE_SIZE > 4096)
3371 		return -EINVAL;
3372 
3373 	address = dev->adev->rmmio_remap.bus_addr;
3374 
3375 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3376 				VM_DONTDUMP | VM_PFNMAP);
3377 
3378 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3379 
3380 	pr_debug("process pid %d mapping mmio page\n"
3381 		 "     target user address == 0x%08llX\n"
3382 		 "     physical address    == 0x%08llX\n"
3383 		 "     vm_flags            == 0x%04lX\n"
3384 		 "     size                == 0x%04lX\n",
3385 		 process->lead_thread->pid, (unsigned long long) vma->vm_start,
3386 		 address, vma->vm_flags, PAGE_SIZE);
3387 
3388 	return io_remap_pfn_range(vma,
3389 				vma->vm_start,
3390 				address >> PAGE_SHIFT,
3391 				PAGE_SIZE,
3392 				vma->vm_page_prot);
3393 }
3394 
3395 
3396 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3397 {
3398 	struct kfd_process *process;
3399 	struct kfd_node *dev = NULL;
3400 	unsigned long mmap_offset;
3401 	unsigned int gpu_id;
3402 
3403 	process = kfd_get_process(current);
3404 	if (IS_ERR(process))
3405 		return PTR_ERR(process);
3406 
3407 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3408 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3409 	if (gpu_id)
3410 		dev = kfd_device_by_id(gpu_id);
3411 
3412 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3413 	case KFD_MMAP_TYPE_DOORBELL:
3414 		if (!dev)
3415 			return -ENODEV;
3416 		return kfd_doorbell_mmap(dev, process, vma);
3417 
3418 	case KFD_MMAP_TYPE_EVENTS:
3419 		return kfd_event_mmap(process, vma);
3420 
3421 	case KFD_MMAP_TYPE_RESERVED_MEM:
3422 		if (!dev)
3423 			return -ENODEV;
3424 		return kfd_reserved_mem_mmap(dev, process, vma);
3425 	case KFD_MMAP_TYPE_MMIO:
3426 		if (!dev)
3427 			return -ENODEV;
3428 		return kfd_mmio_mmap(dev, process, vma);
3429 	}
3430 
3431 	return -EFAULT;
3432 }
3433