xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c (revision 6fa4bf3dce0668a96faca0024e382f4489a9cc9b)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include <linux/device.h>
25 #include <linux/export.h>
26 #include <linux/err.h>
27 #include <linux/fs.h>
28 #include <linux/file.h>
29 #include <linux/sched.h>
30 #include <linux/slab.h>
31 #include <linux/uaccess.h>
32 #include <linux/compat.h>
33 #include <uapi/linux/kfd_ioctl.h>
34 #include <linux/time.h>
35 #include <linux/mm.h>
36 #include <linux/mman.h>
37 #include <linux/ptrace.h>
38 #include <linux/dma-buf.h>
39 #include <linux/fdtable.h>
40 #include <linux/processor.h>
41 #include "kfd_priv.h"
42 #include "kfd_device_queue_manager.h"
43 #include "kfd_svm.h"
44 #include "amdgpu_amdkfd.h"
45 #include "kfd_smi_events.h"
46 #include "amdgpu_dma_buf.h"
47 #include "kfd_debug.h"
48 
49 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
50 static int kfd_open(struct inode *, struct file *);
51 static int kfd_release(struct inode *, struct file *);
52 static int kfd_mmap(struct file *, struct vm_area_struct *);
53 
54 static const char kfd_dev_name[] = "kfd";
55 
56 static const struct file_operations kfd_fops = {
57 	.owner = THIS_MODULE,
58 	.unlocked_ioctl = kfd_ioctl,
59 	.compat_ioctl = compat_ptr_ioctl,
60 	.open = kfd_open,
61 	.release = kfd_release,
62 	.mmap = kfd_mmap,
63 };
64 
65 static int kfd_char_dev_major = -1;
66 struct device *kfd_device;
67 static const struct class kfd_class = {
68 	.name = kfd_dev_name,
69 };
70 
71 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
72 {
73 	struct kfd_process_device *pdd;
74 
75 	mutex_lock(&p->mutex);
76 	pdd = kfd_process_device_data_by_id(p, gpu_id);
77 
78 	if (pdd)
79 		return pdd;
80 
81 	mutex_unlock(&p->mutex);
82 	return NULL;
83 }
84 
85 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
86 {
87 	mutex_unlock(&pdd->process->mutex);
88 }
89 
90 int kfd_chardev_init(void)
91 {
92 	int err = 0;
93 
94 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
95 	err = kfd_char_dev_major;
96 	if (err < 0)
97 		goto err_register_chrdev;
98 
99 	err = class_register(&kfd_class);
100 	if (err)
101 		goto err_class_create;
102 
103 	kfd_device = device_create(&kfd_class, NULL,
104 				   MKDEV(kfd_char_dev_major, 0),
105 				   NULL, kfd_dev_name);
106 	err = PTR_ERR(kfd_device);
107 	if (IS_ERR(kfd_device))
108 		goto err_device_create;
109 
110 	return 0;
111 
112 err_device_create:
113 	class_unregister(&kfd_class);
114 err_class_create:
115 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
116 err_register_chrdev:
117 	return err;
118 }
119 
120 void kfd_chardev_exit(void)
121 {
122 	device_destroy(&kfd_class, MKDEV(kfd_char_dev_major, 0));
123 	class_unregister(&kfd_class);
124 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
125 	kfd_device = NULL;
126 }
127 
128 
129 static int kfd_open(struct inode *inode, struct file *filep)
130 {
131 	struct kfd_process *process;
132 	bool is_32bit_user_mode;
133 
134 	if (iminor(inode) != 0)
135 		return -ENODEV;
136 
137 	is_32bit_user_mode = in_compat_syscall();
138 
139 	if (is_32bit_user_mode) {
140 		dev_warn(kfd_device,
141 			"Process %d (32-bit) failed to open /dev/kfd\n"
142 			"32-bit processes are not supported by amdkfd\n",
143 			current->pid);
144 		return -EPERM;
145 	}
146 
147 	process = kfd_create_process(current);
148 	if (IS_ERR(process))
149 		return PTR_ERR(process);
150 
151 	if (kfd_process_init_cwsr_apu(process, filep)) {
152 		kfd_unref_process(process);
153 		return -EFAULT;
154 	}
155 
156 	/* filep now owns the reference returned by kfd_create_process */
157 	filep->private_data = process;
158 
159 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
160 		process->pasid, process->is_32bit_user_mode);
161 
162 	return 0;
163 }
164 
165 static int kfd_release(struct inode *inode, struct file *filep)
166 {
167 	struct kfd_process *process = filep->private_data;
168 
169 	if (process)
170 		kfd_unref_process(process);
171 
172 	return 0;
173 }
174 
175 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
176 					void *data)
177 {
178 	struct kfd_ioctl_get_version_args *args = data;
179 
180 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
181 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
182 
183 	return 0;
184 }
185 
186 static int set_queue_properties_from_user(struct queue_properties *q_properties,
187 				struct kfd_ioctl_create_queue_args *args)
188 {
189 	/*
190 	 * Repurpose queue percentage to accommodate new features:
191 	 * bit 0-7: queue percentage
192 	 * bit 8-15: pm4_target_xcc
193 	 */
194 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
195 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
196 		return -EINVAL;
197 	}
198 
199 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
200 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
201 		return -EINVAL;
202 	}
203 
204 	if ((args->ring_base_address) &&
205 		(!access_ok((const void __user *) args->ring_base_address,
206 			sizeof(uint64_t)))) {
207 		pr_err("Can't access ring base address\n");
208 		return -EFAULT;
209 	}
210 
211 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
212 		pr_err("Ring size must be a power of 2 or 0\n");
213 		return -EINVAL;
214 	}
215 
216 	if (!access_ok((const void __user *) args->read_pointer_address,
217 			sizeof(uint32_t))) {
218 		pr_err("Can't access read pointer\n");
219 		return -EFAULT;
220 	}
221 
222 	if (!access_ok((const void __user *) args->write_pointer_address,
223 			sizeof(uint32_t))) {
224 		pr_err("Can't access write pointer\n");
225 		return -EFAULT;
226 	}
227 
228 	if (args->eop_buffer_address &&
229 		!access_ok((const void __user *) args->eop_buffer_address,
230 			sizeof(uint32_t))) {
231 		pr_debug("Can't access eop buffer");
232 		return -EFAULT;
233 	}
234 
235 	if (args->ctx_save_restore_address &&
236 		!access_ok((const void __user *) args->ctx_save_restore_address,
237 			sizeof(uint32_t))) {
238 		pr_debug("Can't access ctx save restore buffer");
239 		return -EFAULT;
240 	}
241 
242 	q_properties->is_interop = false;
243 	q_properties->is_gws = false;
244 	q_properties->queue_percent = args->queue_percentage & 0xFF;
245 	/* bit 8-15 are repurposed to be PM4 target XCC */
246 	q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
247 	q_properties->priority = args->queue_priority;
248 	q_properties->queue_address = args->ring_base_address;
249 	q_properties->queue_size = args->ring_size;
250 	q_properties->read_ptr = (void __user *)args->read_pointer_address;
251 	q_properties->write_ptr = (void __user *)args->write_pointer_address;
252 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
253 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
254 	q_properties->ctx_save_restore_area_address =
255 			args->ctx_save_restore_address;
256 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
257 	q_properties->ctl_stack_size = args->ctl_stack_size;
258 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
259 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
260 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
261 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
262 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
263 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
264 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
265 	else
266 		return -ENOTSUPP;
267 
268 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
269 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
270 	else
271 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
272 
273 	pr_debug("Queue Percentage: %d, %d\n",
274 			q_properties->queue_percent, args->queue_percentage);
275 
276 	pr_debug("Queue Priority: %d, %d\n",
277 			q_properties->priority, args->queue_priority);
278 
279 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
280 			q_properties->queue_address, args->ring_base_address);
281 
282 	pr_debug("Queue Size: 0x%llX, %u\n",
283 			q_properties->queue_size, args->ring_size);
284 
285 	pr_debug("Queue r/w Pointers: %px, %px\n",
286 			q_properties->read_ptr,
287 			q_properties->write_ptr);
288 
289 	pr_debug("Queue Format: %d\n", q_properties->format);
290 
291 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
292 
293 	pr_debug("Queue CTX save area: 0x%llX\n",
294 			q_properties->ctx_save_restore_area_address);
295 
296 	return 0;
297 }
298 
299 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
300 					void *data)
301 {
302 	struct kfd_ioctl_create_queue_args *args = data;
303 	struct kfd_node *dev;
304 	int err = 0;
305 	unsigned int queue_id;
306 	struct kfd_process_device *pdd;
307 	struct queue_properties q_properties;
308 	uint32_t doorbell_offset_in_process = 0;
309 
310 	memset(&q_properties, 0, sizeof(struct queue_properties));
311 
312 	pr_debug("Creating queue ioctl\n");
313 
314 	err = set_queue_properties_from_user(&q_properties, args);
315 	if (err)
316 		return err;
317 
318 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
319 
320 	mutex_lock(&p->mutex);
321 
322 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
323 	if (!pdd) {
324 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
325 		err = -EINVAL;
326 		goto err_pdd;
327 	}
328 	dev = pdd->dev;
329 
330 	pdd = kfd_bind_process_to_device(dev, p);
331 	if (IS_ERR(pdd)) {
332 		err = -ESRCH;
333 		goto err_bind_process;
334 	}
335 
336 	if (!pdd->qpd.proc_doorbells) {
337 		err = kfd_alloc_process_doorbells(dev->kfd, pdd);
338 		if (err) {
339 			pr_debug("failed to allocate process doorbells\n");
340 			goto err_bind_process;
341 		}
342 	}
343 
344 	err = kfd_queue_acquire_buffers(pdd, &q_properties);
345 	if (err) {
346 		pr_debug("failed to acquire user queue buffers\n");
347 		goto err_acquire_queue_buf;
348 	}
349 
350 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
351 			p->pasid,
352 			dev->id);
353 
354 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id,
355 			NULL, NULL, NULL, &doorbell_offset_in_process);
356 	if (err != 0)
357 		goto err_create_queue;
358 
359 	args->queue_id = queue_id;
360 
361 
362 	/* Return gpu_id as doorbell offset for mmap usage */
363 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
364 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
365 	if (KFD_IS_SOC15(dev))
366 		/* On SOC15 ASICs, include the doorbell offset within the
367 		 * process doorbell frame, which is 2 pages.
368 		 */
369 		args->doorbell_offset |= doorbell_offset_in_process;
370 
371 	mutex_unlock(&p->mutex);
372 
373 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
374 
375 	pr_debug("Ring buffer address == 0x%016llX\n",
376 			args->ring_base_address);
377 
378 	pr_debug("Read ptr address    == 0x%016llX\n",
379 			args->read_pointer_address);
380 
381 	pr_debug("Write ptr address   == 0x%016llX\n",
382 			args->write_pointer_address);
383 
384 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
385 	return 0;
386 
387 err_create_queue:
388 	kfd_queue_release_buffers(pdd, &q_properties);
389 err_acquire_queue_buf:
390 err_bind_process:
391 err_pdd:
392 	mutex_unlock(&p->mutex);
393 	return err;
394 }
395 
396 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
397 					void *data)
398 {
399 	int retval;
400 	struct kfd_ioctl_destroy_queue_args *args = data;
401 
402 	pr_debug("Destroying queue id %d for pasid 0x%x\n",
403 				args->queue_id,
404 				p->pasid);
405 
406 	mutex_lock(&p->mutex);
407 
408 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
409 
410 	mutex_unlock(&p->mutex);
411 	return retval;
412 }
413 
414 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
415 					void *data)
416 {
417 	int retval;
418 	struct kfd_ioctl_update_queue_args *args = data;
419 	struct queue_properties properties;
420 
421 	/*
422 	 * Repurpose queue percentage to accommodate new features:
423 	 * bit 0-7: queue percentage
424 	 * bit 8-15: pm4_target_xcc
425 	 */
426 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
427 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
428 		return -EINVAL;
429 	}
430 
431 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
432 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
433 		return -EINVAL;
434 	}
435 
436 	if ((args->ring_base_address) &&
437 		(!access_ok((const void __user *) args->ring_base_address,
438 			sizeof(uint64_t)))) {
439 		pr_err("Can't access ring base address\n");
440 		return -EFAULT;
441 	}
442 
443 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
444 		pr_err("Ring size must be a power of 2 or 0\n");
445 		return -EINVAL;
446 	}
447 
448 	properties.queue_address = args->ring_base_address;
449 	properties.queue_size = args->ring_size;
450 	properties.queue_percent = args->queue_percentage & 0xFF;
451 	/* bit 8-15 are repurposed to be PM4 target XCC */
452 	properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
453 	properties.priority = args->queue_priority;
454 
455 	pr_debug("Updating queue id %d for pasid 0x%x\n",
456 			args->queue_id, p->pasid);
457 
458 	mutex_lock(&p->mutex);
459 
460 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
461 
462 	mutex_unlock(&p->mutex);
463 
464 	return retval;
465 }
466 
467 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
468 					void *data)
469 {
470 	int retval;
471 	const int max_num_cus = 1024;
472 	struct kfd_ioctl_set_cu_mask_args *args = data;
473 	struct mqd_update_info minfo = {0};
474 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
475 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
476 
477 	if ((args->num_cu_mask % 32) != 0) {
478 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
479 				args->num_cu_mask);
480 		return -EINVAL;
481 	}
482 
483 	minfo.cu_mask.count = args->num_cu_mask;
484 	if (minfo.cu_mask.count == 0) {
485 		pr_debug("CU mask cannot be 0");
486 		return -EINVAL;
487 	}
488 
489 	/* To prevent an unreasonably large CU mask size, set an arbitrary
490 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
491 	 * past max_num_cus bits and just use the first max_num_cus bits.
492 	 */
493 	if (minfo.cu_mask.count > max_num_cus) {
494 		pr_debug("CU mask cannot be greater than 1024 bits");
495 		minfo.cu_mask.count = max_num_cus;
496 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
497 	}
498 
499 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
500 	if (!minfo.cu_mask.ptr)
501 		return -ENOMEM;
502 
503 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
504 	if (retval) {
505 		pr_debug("Could not copy CU mask from userspace");
506 		retval = -EFAULT;
507 		goto out;
508 	}
509 
510 	mutex_lock(&p->mutex);
511 
512 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
513 
514 	mutex_unlock(&p->mutex);
515 
516 out:
517 	kfree(minfo.cu_mask.ptr);
518 	return retval;
519 }
520 
521 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
522 					  struct kfd_process *p, void *data)
523 {
524 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
525 	int r;
526 
527 	mutex_lock(&p->mutex);
528 
529 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
530 			       (void __user *)args->ctl_stack_address,
531 			       &args->ctl_stack_used_size,
532 			       &args->save_area_used_size);
533 
534 	mutex_unlock(&p->mutex);
535 
536 	return r;
537 }
538 
539 static int kfd_ioctl_set_memory_policy(struct file *filep,
540 					struct kfd_process *p, void *data)
541 {
542 	struct kfd_ioctl_set_memory_policy_args *args = data;
543 	int err = 0;
544 	struct kfd_process_device *pdd;
545 	enum cache_policy default_policy, alternate_policy;
546 
547 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
548 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
549 		return -EINVAL;
550 	}
551 
552 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
553 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
554 		return -EINVAL;
555 	}
556 
557 	mutex_lock(&p->mutex);
558 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
559 	if (!pdd) {
560 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
561 		err = -EINVAL;
562 		goto err_pdd;
563 	}
564 
565 	pdd = kfd_bind_process_to_device(pdd->dev, p);
566 	if (IS_ERR(pdd)) {
567 		err = -ESRCH;
568 		goto out;
569 	}
570 
571 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
572 			 ? cache_policy_coherent : cache_policy_noncoherent;
573 
574 	alternate_policy =
575 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
576 		   ? cache_policy_coherent : cache_policy_noncoherent;
577 
578 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
579 				&pdd->qpd,
580 				default_policy,
581 				alternate_policy,
582 				(void __user *)args->alternate_aperture_base,
583 				args->alternate_aperture_size))
584 		err = -EINVAL;
585 
586 out:
587 err_pdd:
588 	mutex_unlock(&p->mutex);
589 
590 	return err;
591 }
592 
593 static int kfd_ioctl_set_trap_handler(struct file *filep,
594 					struct kfd_process *p, void *data)
595 {
596 	struct kfd_ioctl_set_trap_handler_args *args = data;
597 	int err = 0;
598 	struct kfd_process_device *pdd;
599 
600 	mutex_lock(&p->mutex);
601 
602 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
603 	if (!pdd) {
604 		err = -EINVAL;
605 		goto err_pdd;
606 	}
607 
608 	pdd = kfd_bind_process_to_device(pdd->dev, p);
609 	if (IS_ERR(pdd)) {
610 		err = -ESRCH;
611 		goto out;
612 	}
613 
614 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
615 
616 out:
617 err_pdd:
618 	mutex_unlock(&p->mutex);
619 
620 	return err;
621 }
622 
623 static int kfd_ioctl_dbg_register(struct file *filep,
624 				struct kfd_process *p, void *data)
625 {
626 	return -EPERM;
627 }
628 
629 static int kfd_ioctl_dbg_unregister(struct file *filep,
630 				struct kfd_process *p, void *data)
631 {
632 	return -EPERM;
633 }
634 
635 static int kfd_ioctl_dbg_address_watch(struct file *filep,
636 					struct kfd_process *p, void *data)
637 {
638 	return -EPERM;
639 }
640 
641 /* Parse and generate fixed size data structure for wave control */
642 static int kfd_ioctl_dbg_wave_control(struct file *filep,
643 					struct kfd_process *p, void *data)
644 {
645 	return -EPERM;
646 }
647 
648 static int kfd_ioctl_get_clock_counters(struct file *filep,
649 				struct kfd_process *p, void *data)
650 {
651 	struct kfd_ioctl_get_clock_counters_args *args = data;
652 	struct kfd_process_device *pdd;
653 
654 	mutex_lock(&p->mutex);
655 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
656 	mutex_unlock(&p->mutex);
657 	if (pdd)
658 		/* Reading GPU clock counter from KGD */
659 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
660 	else
661 		/* Node without GPU resource */
662 		args->gpu_clock_counter = 0;
663 
664 	/* No access to rdtsc. Using raw monotonic time */
665 	args->cpu_clock_counter = ktime_get_raw_ns();
666 	args->system_clock_counter = ktime_get_boottime_ns();
667 
668 	/* Since the counter is in nano-seconds we use 1GHz frequency */
669 	args->system_clock_freq = 1000000000;
670 
671 	return 0;
672 }
673 
674 
675 static int kfd_ioctl_get_process_apertures(struct file *filp,
676 				struct kfd_process *p, void *data)
677 {
678 	struct kfd_ioctl_get_process_apertures_args *args = data;
679 	struct kfd_process_device_apertures *pAperture;
680 	int i;
681 
682 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
683 
684 	args->num_of_nodes = 0;
685 
686 	mutex_lock(&p->mutex);
687 	/* Run over all pdd of the process */
688 	for (i = 0; i < p->n_pdds; i++) {
689 		struct kfd_process_device *pdd = p->pdds[i];
690 
691 		pAperture =
692 			&args->process_apertures[args->num_of_nodes];
693 		pAperture->gpu_id = pdd->dev->id;
694 		pAperture->lds_base = pdd->lds_base;
695 		pAperture->lds_limit = pdd->lds_limit;
696 		pAperture->gpuvm_base = pdd->gpuvm_base;
697 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
698 		pAperture->scratch_base = pdd->scratch_base;
699 		pAperture->scratch_limit = pdd->scratch_limit;
700 
701 		dev_dbg(kfd_device,
702 			"node id %u\n", args->num_of_nodes);
703 		dev_dbg(kfd_device,
704 			"gpu id %u\n", pdd->dev->id);
705 		dev_dbg(kfd_device,
706 			"lds_base %llX\n", pdd->lds_base);
707 		dev_dbg(kfd_device,
708 			"lds_limit %llX\n", pdd->lds_limit);
709 		dev_dbg(kfd_device,
710 			"gpuvm_base %llX\n", pdd->gpuvm_base);
711 		dev_dbg(kfd_device,
712 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
713 		dev_dbg(kfd_device,
714 			"scratch_base %llX\n", pdd->scratch_base);
715 		dev_dbg(kfd_device,
716 			"scratch_limit %llX\n", pdd->scratch_limit);
717 
718 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
719 			break;
720 	}
721 	mutex_unlock(&p->mutex);
722 
723 	return 0;
724 }
725 
726 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
727 				struct kfd_process *p, void *data)
728 {
729 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
730 	struct kfd_process_device_apertures *pa;
731 	int ret;
732 	int i;
733 
734 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
735 
736 	if (args->num_of_nodes == 0) {
737 		/* Return number of nodes, so that user space can alloacate
738 		 * sufficient memory
739 		 */
740 		mutex_lock(&p->mutex);
741 		args->num_of_nodes = p->n_pdds;
742 		goto out_unlock;
743 	}
744 
745 	/* Fill in process-aperture information for all available
746 	 * nodes, but not more than args->num_of_nodes as that is
747 	 * the amount of memory allocated by user
748 	 */
749 	pa = kcalloc(args->num_of_nodes, sizeof(struct kfd_process_device_apertures),
750 		     GFP_KERNEL);
751 	if (!pa)
752 		return -ENOMEM;
753 
754 	mutex_lock(&p->mutex);
755 
756 	if (!p->n_pdds) {
757 		args->num_of_nodes = 0;
758 		kfree(pa);
759 		goto out_unlock;
760 	}
761 
762 	/* Run over all pdd of the process */
763 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
764 		struct kfd_process_device *pdd = p->pdds[i];
765 
766 		pa[i].gpu_id = pdd->dev->id;
767 		pa[i].lds_base = pdd->lds_base;
768 		pa[i].lds_limit = pdd->lds_limit;
769 		pa[i].gpuvm_base = pdd->gpuvm_base;
770 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
771 		pa[i].scratch_base = pdd->scratch_base;
772 		pa[i].scratch_limit = pdd->scratch_limit;
773 
774 		dev_dbg(kfd_device,
775 			"gpu id %u\n", pdd->dev->id);
776 		dev_dbg(kfd_device,
777 			"lds_base %llX\n", pdd->lds_base);
778 		dev_dbg(kfd_device,
779 			"lds_limit %llX\n", pdd->lds_limit);
780 		dev_dbg(kfd_device,
781 			"gpuvm_base %llX\n", pdd->gpuvm_base);
782 		dev_dbg(kfd_device,
783 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
784 		dev_dbg(kfd_device,
785 			"scratch_base %llX\n", pdd->scratch_base);
786 		dev_dbg(kfd_device,
787 			"scratch_limit %llX\n", pdd->scratch_limit);
788 	}
789 	mutex_unlock(&p->mutex);
790 
791 	args->num_of_nodes = i;
792 	ret = copy_to_user(
793 			(void __user *)args->kfd_process_device_apertures_ptr,
794 			pa,
795 			(i * sizeof(struct kfd_process_device_apertures)));
796 	kfree(pa);
797 	return ret ? -EFAULT : 0;
798 
799 out_unlock:
800 	mutex_unlock(&p->mutex);
801 	return 0;
802 }
803 
804 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
805 					void *data)
806 {
807 	struct kfd_ioctl_create_event_args *args = data;
808 	int err;
809 
810 	/* For dGPUs the event page is allocated in user mode. The
811 	 * handle is passed to KFD with the first call to this IOCTL
812 	 * through the event_page_offset field.
813 	 */
814 	if (args->event_page_offset) {
815 		mutex_lock(&p->mutex);
816 		err = kfd_kmap_event_page(p, args->event_page_offset);
817 		mutex_unlock(&p->mutex);
818 		if (err)
819 			return err;
820 	}
821 
822 	err = kfd_event_create(filp, p, args->event_type,
823 				args->auto_reset != 0, args->node_id,
824 				&args->event_id, &args->event_trigger_data,
825 				&args->event_page_offset,
826 				&args->event_slot_index);
827 
828 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
829 	return err;
830 }
831 
832 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
833 					void *data)
834 {
835 	struct kfd_ioctl_destroy_event_args *args = data;
836 
837 	return kfd_event_destroy(p, args->event_id);
838 }
839 
840 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
841 				void *data)
842 {
843 	struct kfd_ioctl_set_event_args *args = data;
844 
845 	return kfd_set_event(p, args->event_id);
846 }
847 
848 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
849 				void *data)
850 {
851 	struct kfd_ioctl_reset_event_args *args = data;
852 
853 	return kfd_reset_event(p, args->event_id);
854 }
855 
856 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
857 				void *data)
858 {
859 	struct kfd_ioctl_wait_events_args *args = data;
860 
861 	return kfd_wait_on_events(p, args->num_events,
862 			(void __user *)args->events_ptr,
863 			(args->wait_for_all != 0),
864 			&args->timeout, &args->wait_result);
865 }
866 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
867 					struct kfd_process *p, void *data)
868 {
869 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
870 	struct kfd_process_device *pdd;
871 	struct kfd_node *dev;
872 	long err;
873 
874 	mutex_lock(&p->mutex);
875 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
876 	if (!pdd) {
877 		err = -EINVAL;
878 		goto err_pdd;
879 	}
880 	dev = pdd->dev;
881 
882 	pdd = kfd_bind_process_to_device(dev, p);
883 	if (IS_ERR(pdd)) {
884 		err = PTR_ERR(pdd);
885 		goto bind_process_to_device_fail;
886 	}
887 
888 	pdd->qpd.sh_hidden_private_base = args->va_addr;
889 
890 	mutex_unlock(&p->mutex);
891 
892 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
893 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
894 		dev->kfd2kgd->set_scratch_backing_va(
895 			dev->adev, args->va_addr, pdd->qpd.vmid);
896 
897 	return 0;
898 
899 bind_process_to_device_fail:
900 err_pdd:
901 	mutex_unlock(&p->mutex);
902 	return err;
903 }
904 
905 static int kfd_ioctl_get_tile_config(struct file *filep,
906 		struct kfd_process *p, void *data)
907 {
908 	struct kfd_ioctl_get_tile_config_args *args = data;
909 	struct kfd_process_device *pdd;
910 	struct tile_config config;
911 	int err = 0;
912 
913 	mutex_lock(&p->mutex);
914 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
915 	mutex_unlock(&p->mutex);
916 	if (!pdd)
917 		return -EINVAL;
918 
919 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
920 
921 	args->gb_addr_config = config.gb_addr_config;
922 	args->num_banks = config.num_banks;
923 	args->num_ranks = config.num_ranks;
924 
925 	if (args->num_tile_configs > config.num_tile_configs)
926 		args->num_tile_configs = config.num_tile_configs;
927 	err = copy_to_user((void __user *)args->tile_config_ptr,
928 			config.tile_config_ptr,
929 			args->num_tile_configs * sizeof(uint32_t));
930 	if (err) {
931 		args->num_tile_configs = 0;
932 		return -EFAULT;
933 	}
934 
935 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
936 		args->num_macro_tile_configs =
937 				config.num_macro_tile_configs;
938 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
939 			config.macro_tile_config_ptr,
940 			args->num_macro_tile_configs * sizeof(uint32_t));
941 	if (err) {
942 		args->num_macro_tile_configs = 0;
943 		return -EFAULT;
944 	}
945 
946 	return 0;
947 }
948 
949 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
950 				void *data)
951 {
952 	struct kfd_ioctl_acquire_vm_args *args = data;
953 	struct kfd_process_device *pdd;
954 	struct file *drm_file;
955 	int ret;
956 
957 	drm_file = fget(args->drm_fd);
958 	if (!drm_file)
959 		return -EINVAL;
960 
961 	mutex_lock(&p->mutex);
962 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
963 	if (!pdd) {
964 		ret = -EINVAL;
965 		goto err_pdd;
966 	}
967 
968 	if (pdd->drm_file) {
969 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
970 		goto err_drm_file;
971 	}
972 
973 	ret = kfd_process_device_init_vm(pdd, drm_file);
974 	if (ret)
975 		goto err_unlock;
976 
977 	/* On success, the PDD keeps the drm_file reference */
978 	mutex_unlock(&p->mutex);
979 
980 	return 0;
981 
982 err_unlock:
983 err_pdd:
984 err_drm_file:
985 	mutex_unlock(&p->mutex);
986 	fput(drm_file);
987 	return ret;
988 }
989 
990 bool kfd_dev_is_large_bar(struct kfd_node *dev)
991 {
992 	if (dev->kfd->adev->debug_largebar) {
993 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
994 		return true;
995 	}
996 
997 	if (dev->local_mem_info.local_mem_size_private == 0 &&
998 	    dev->local_mem_info.local_mem_size_public > 0)
999 		return true;
1000 
1001 	if (dev->local_mem_info.local_mem_size_public == 0 &&
1002 	    dev->kfd->adev->gmc.is_app_apu) {
1003 		pr_debug("APP APU, Consider like a large bar system\n");
1004 		return true;
1005 	}
1006 
1007 	return false;
1008 }
1009 
1010 static int kfd_ioctl_get_available_memory(struct file *filep,
1011 					  struct kfd_process *p, void *data)
1012 {
1013 	struct kfd_ioctl_get_available_memory_args *args = data;
1014 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1015 
1016 	if (!pdd)
1017 		return -EINVAL;
1018 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1019 							pdd->dev->node_id);
1020 	kfd_unlock_pdd(pdd);
1021 	return 0;
1022 }
1023 
1024 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1025 					struct kfd_process *p, void *data)
1026 {
1027 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1028 	struct kfd_process_device *pdd;
1029 	void *mem;
1030 	struct kfd_node *dev;
1031 	int idr_handle;
1032 	long err;
1033 	uint64_t offset = args->mmap_offset;
1034 	uint32_t flags = args->flags;
1035 
1036 	if (args->size == 0)
1037 		return -EINVAL;
1038 
1039 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1040 	/* Flush pending deferred work to avoid racing with deferred actions
1041 	 * from previous memory map changes (e.g. munmap).
1042 	 */
1043 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1044 	mutex_lock(&p->svms.lock);
1045 	mmap_write_unlock(current->mm);
1046 	if (interval_tree_iter_first(&p->svms.objects,
1047 				     args->va_addr >> PAGE_SHIFT,
1048 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1049 		pr_err("Address: 0x%llx already allocated by SVM\n",
1050 			args->va_addr);
1051 		mutex_unlock(&p->svms.lock);
1052 		return -EADDRINUSE;
1053 	}
1054 
1055 	/* When register user buffer check if it has been registered by svm by
1056 	 * buffer cpu virtual address.
1057 	 */
1058 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1059 	    interval_tree_iter_first(&p->svms.objects,
1060 				     args->mmap_offset >> PAGE_SHIFT,
1061 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1062 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1063 			args->mmap_offset);
1064 		mutex_unlock(&p->svms.lock);
1065 		return -EADDRINUSE;
1066 	}
1067 
1068 	mutex_unlock(&p->svms.lock);
1069 #endif
1070 	mutex_lock(&p->mutex);
1071 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1072 	if (!pdd) {
1073 		err = -EINVAL;
1074 		goto err_pdd;
1075 	}
1076 
1077 	dev = pdd->dev;
1078 
1079 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1080 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1081 		!kfd_dev_is_large_bar(dev)) {
1082 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1083 		err = -EINVAL;
1084 		goto err_large_bar;
1085 	}
1086 
1087 	pdd = kfd_bind_process_to_device(dev, p);
1088 	if (IS_ERR(pdd)) {
1089 		err = PTR_ERR(pdd);
1090 		goto err_unlock;
1091 	}
1092 
1093 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1094 		if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1095 			err = -EINVAL;
1096 			goto err_unlock;
1097 		}
1098 		offset = kfd_get_process_doorbells(pdd);
1099 		if (!offset) {
1100 			err = -ENOMEM;
1101 			goto err_unlock;
1102 		}
1103 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1104 		if (args->size != PAGE_SIZE) {
1105 			err = -EINVAL;
1106 			goto err_unlock;
1107 		}
1108 		offset = dev->adev->rmmio_remap.bus_addr;
1109 		if (!offset || (PAGE_SIZE > 4096)) {
1110 			err = -ENOMEM;
1111 			goto err_unlock;
1112 		}
1113 	}
1114 
1115 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1116 		dev->adev, args->va_addr, args->size,
1117 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1118 		flags, false);
1119 
1120 	if (err)
1121 		goto err_unlock;
1122 
1123 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1124 	if (idr_handle < 0) {
1125 		err = -EFAULT;
1126 		goto err_free;
1127 	}
1128 
1129 	/* Update the VRAM usage count */
1130 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1131 		uint64_t size = args->size;
1132 
1133 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1134 			size >>= 1;
1135 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1136 	}
1137 
1138 	mutex_unlock(&p->mutex);
1139 
1140 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1141 	args->mmap_offset = offset;
1142 
1143 	/* MMIO is mapped through kfd device
1144 	 * Generate a kfd mmap offset
1145 	 */
1146 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1147 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1148 					| KFD_MMAP_GPU_ID(args->gpu_id);
1149 
1150 	return 0;
1151 
1152 err_free:
1153 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1154 					       pdd->drm_priv, NULL);
1155 err_unlock:
1156 err_pdd:
1157 err_large_bar:
1158 	mutex_unlock(&p->mutex);
1159 	return err;
1160 }
1161 
1162 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1163 					struct kfd_process *p, void *data)
1164 {
1165 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1166 	struct kfd_process_device *pdd;
1167 	void *mem;
1168 	int ret;
1169 	uint64_t size = 0;
1170 
1171 	mutex_lock(&p->mutex);
1172 	/*
1173 	 * Safeguard to prevent user space from freeing signal BO.
1174 	 * It will be freed at process termination.
1175 	 */
1176 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1177 		pr_err("Free signal BO is not allowed\n");
1178 		ret = -EPERM;
1179 		goto err_unlock;
1180 	}
1181 
1182 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1183 	if (!pdd) {
1184 		pr_err("Process device data doesn't exist\n");
1185 		ret = -EINVAL;
1186 		goto err_pdd;
1187 	}
1188 
1189 	mem = kfd_process_device_translate_handle(
1190 		pdd, GET_IDR_HANDLE(args->handle));
1191 	if (!mem) {
1192 		ret = -EINVAL;
1193 		goto err_unlock;
1194 	}
1195 
1196 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1197 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1198 
1199 	/* If freeing the buffer failed, leave the handle in place for
1200 	 * clean-up during process tear-down.
1201 	 */
1202 	if (!ret)
1203 		kfd_process_device_remove_obj_handle(
1204 			pdd, GET_IDR_HANDLE(args->handle));
1205 
1206 	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1207 
1208 err_unlock:
1209 err_pdd:
1210 	mutex_unlock(&p->mutex);
1211 	return ret;
1212 }
1213 
1214 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1215 					struct kfd_process *p, void *data)
1216 {
1217 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1218 	struct kfd_process_device *pdd, *peer_pdd;
1219 	void *mem;
1220 	struct kfd_node *dev;
1221 	long err = 0;
1222 	int i;
1223 	uint32_t *devices_arr = NULL;
1224 
1225 	if (!args->n_devices) {
1226 		pr_debug("Device IDs array empty\n");
1227 		return -EINVAL;
1228 	}
1229 	if (args->n_success > args->n_devices) {
1230 		pr_debug("n_success exceeds n_devices\n");
1231 		return -EINVAL;
1232 	}
1233 
1234 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1235 				    GFP_KERNEL);
1236 	if (!devices_arr)
1237 		return -ENOMEM;
1238 
1239 	err = copy_from_user(devices_arr,
1240 			     (void __user *)args->device_ids_array_ptr,
1241 			     args->n_devices * sizeof(*devices_arr));
1242 	if (err != 0) {
1243 		err = -EFAULT;
1244 		goto copy_from_user_failed;
1245 	}
1246 
1247 	mutex_lock(&p->mutex);
1248 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1249 	if (!pdd) {
1250 		err = -EINVAL;
1251 		goto get_process_device_data_failed;
1252 	}
1253 	dev = pdd->dev;
1254 
1255 	pdd = kfd_bind_process_to_device(dev, p);
1256 	if (IS_ERR(pdd)) {
1257 		err = PTR_ERR(pdd);
1258 		goto bind_process_to_device_failed;
1259 	}
1260 
1261 	mem = kfd_process_device_translate_handle(pdd,
1262 						GET_IDR_HANDLE(args->handle));
1263 	if (!mem) {
1264 		err = -ENOMEM;
1265 		goto get_mem_obj_from_handle_failed;
1266 	}
1267 
1268 	for (i = args->n_success; i < args->n_devices; i++) {
1269 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1270 		if (!peer_pdd) {
1271 			pr_debug("Getting device by id failed for 0x%x\n",
1272 				 devices_arr[i]);
1273 			err = -EINVAL;
1274 			goto get_mem_obj_from_handle_failed;
1275 		}
1276 
1277 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1278 		if (IS_ERR(peer_pdd)) {
1279 			err = PTR_ERR(peer_pdd);
1280 			goto get_mem_obj_from_handle_failed;
1281 		}
1282 
1283 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1284 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1285 			peer_pdd->drm_priv);
1286 		if (err) {
1287 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1288 
1289 			dev_err(dev->adev->dev,
1290 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1291 			       pci_domain_nr(pdev->bus),
1292 			       pdev->bus->number,
1293 			       PCI_SLOT(pdev->devfn),
1294 			       PCI_FUNC(pdev->devfn),
1295 			       ((struct kgd_mem *)mem)->domain);
1296 			goto map_memory_to_gpu_failed;
1297 		}
1298 		args->n_success = i+1;
1299 	}
1300 
1301 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1302 	if (err) {
1303 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1304 		goto sync_memory_failed;
1305 	}
1306 
1307 	mutex_unlock(&p->mutex);
1308 
1309 	/* Flush TLBs after waiting for the page table updates to complete */
1310 	for (i = 0; i < args->n_devices; i++) {
1311 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1312 		if (WARN_ON_ONCE(!peer_pdd))
1313 			continue;
1314 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1315 	}
1316 	kfree(devices_arr);
1317 
1318 	return err;
1319 
1320 get_process_device_data_failed:
1321 bind_process_to_device_failed:
1322 get_mem_obj_from_handle_failed:
1323 map_memory_to_gpu_failed:
1324 sync_memory_failed:
1325 	mutex_unlock(&p->mutex);
1326 copy_from_user_failed:
1327 	kfree(devices_arr);
1328 
1329 	return err;
1330 }
1331 
1332 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1333 					struct kfd_process *p, void *data)
1334 {
1335 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1336 	struct kfd_process_device *pdd, *peer_pdd;
1337 	void *mem;
1338 	long err = 0;
1339 	uint32_t *devices_arr = NULL, i;
1340 	bool flush_tlb;
1341 
1342 	if (!args->n_devices) {
1343 		pr_debug("Device IDs array empty\n");
1344 		return -EINVAL;
1345 	}
1346 	if (args->n_success > args->n_devices) {
1347 		pr_debug("n_success exceeds n_devices\n");
1348 		return -EINVAL;
1349 	}
1350 
1351 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1352 				    GFP_KERNEL);
1353 	if (!devices_arr)
1354 		return -ENOMEM;
1355 
1356 	err = copy_from_user(devices_arr,
1357 			     (void __user *)args->device_ids_array_ptr,
1358 			     args->n_devices * sizeof(*devices_arr));
1359 	if (err != 0) {
1360 		err = -EFAULT;
1361 		goto copy_from_user_failed;
1362 	}
1363 
1364 	mutex_lock(&p->mutex);
1365 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1366 	if (!pdd) {
1367 		err = -EINVAL;
1368 		goto bind_process_to_device_failed;
1369 	}
1370 
1371 	mem = kfd_process_device_translate_handle(pdd,
1372 						GET_IDR_HANDLE(args->handle));
1373 	if (!mem) {
1374 		err = -ENOMEM;
1375 		goto get_mem_obj_from_handle_failed;
1376 	}
1377 
1378 	for (i = args->n_success; i < args->n_devices; i++) {
1379 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1380 		if (!peer_pdd) {
1381 			err = -EINVAL;
1382 			goto get_mem_obj_from_handle_failed;
1383 		}
1384 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1385 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1386 		if (err) {
1387 			pr_err("Failed to unmap from gpu %d/%d\n",
1388 			       i, args->n_devices);
1389 			goto unmap_memory_from_gpu_failed;
1390 		}
1391 		args->n_success = i+1;
1392 	}
1393 
1394 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1395 	if (flush_tlb) {
1396 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1397 				(struct kgd_mem *) mem, true);
1398 		if (err) {
1399 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1400 			goto sync_memory_failed;
1401 		}
1402 	}
1403 
1404 	/* Flush TLBs after waiting for the page table updates to complete */
1405 	for (i = 0; i < args->n_devices; i++) {
1406 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1407 		if (WARN_ON_ONCE(!peer_pdd))
1408 			continue;
1409 		if (flush_tlb)
1410 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1411 
1412 		/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */
1413 		err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd->drm_priv);
1414 		if (err)
1415 			goto sync_memory_failed;
1416 	}
1417 
1418 	mutex_unlock(&p->mutex);
1419 
1420 	kfree(devices_arr);
1421 
1422 	return 0;
1423 
1424 bind_process_to_device_failed:
1425 get_mem_obj_from_handle_failed:
1426 unmap_memory_from_gpu_failed:
1427 sync_memory_failed:
1428 	mutex_unlock(&p->mutex);
1429 copy_from_user_failed:
1430 	kfree(devices_arr);
1431 	return err;
1432 }
1433 
1434 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1435 		struct kfd_process *p, void *data)
1436 {
1437 	int retval;
1438 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1439 	struct queue *q;
1440 	struct kfd_node *dev;
1441 
1442 	mutex_lock(&p->mutex);
1443 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1444 
1445 	if (q) {
1446 		dev = q->device;
1447 	} else {
1448 		retval = -EINVAL;
1449 		goto out_unlock;
1450 	}
1451 
1452 	if (!dev->gws) {
1453 		retval = -ENODEV;
1454 		goto out_unlock;
1455 	}
1456 
1457 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1458 		retval = -ENODEV;
1459 		goto out_unlock;
1460 	}
1461 
1462 	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1463 				      kfd_dbg_has_cwsr_workaround(dev))) {
1464 		retval = -EBUSY;
1465 		goto out_unlock;
1466 	}
1467 
1468 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1469 	mutex_unlock(&p->mutex);
1470 
1471 	args->first_gws = 0;
1472 	return retval;
1473 
1474 out_unlock:
1475 	mutex_unlock(&p->mutex);
1476 	return retval;
1477 }
1478 
1479 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1480 		struct kfd_process *p, void *data)
1481 {
1482 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1483 	struct kfd_node *dev = NULL;
1484 	struct amdgpu_device *dmabuf_adev;
1485 	void *metadata_buffer = NULL;
1486 	uint32_t flags;
1487 	int8_t xcp_id;
1488 	unsigned int i;
1489 	int r;
1490 
1491 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1492 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1493 		if (dev && !kfd_devcgroup_check_permission(dev))
1494 			break;
1495 	if (!dev)
1496 		return -EINVAL;
1497 
1498 	if (args->metadata_ptr) {
1499 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1500 		if (!metadata_buffer)
1501 			return -ENOMEM;
1502 	}
1503 
1504 	/* Get dmabuf info from KGD */
1505 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1506 					  &dmabuf_adev, &args->size,
1507 					  metadata_buffer, args->metadata_size,
1508 					  &args->metadata_size, &flags, &xcp_id);
1509 	if (r)
1510 		goto exit;
1511 
1512 	if (xcp_id >= 0)
1513 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1514 	else
1515 		args->gpu_id = dev->id;
1516 	args->flags = flags;
1517 
1518 	/* Copy metadata buffer to user mode */
1519 	if (metadata_buffer) {
1520 		r = copy_to_user((void __user *)args->metadata_ptr,
1521 				 metadata_buffer, args->metadata_size);
1522 		if (r != 0)
1523 			r = -EFAULT;
1524 	}
1525 
1526 exit:
1527 	kfree(metadata_buffer);
1528 
1529 	return r;
1530 }
1531 
1532 static int kfd_ioctl_import_dmabuf(struct file *filep,
1533 				   struct kfd_process *p, void *data)
1534 {
1535 	struct kfd_ioctl_import_dmabuf_args *args = data;
1536 	struct kfd_process_device *pdd;
1537 	int idr_handle;
1538 	uint64_t size;
1539 	void *mem;
1540 	int r;
1541 
1542 	mutex_lock(&p->mutex);
1543 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1544 	if (!pdd) {
1545 		r = -EINVAL;
1546 		goto err_unlock;
1547 	}
1548 
1549 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1550 	if (IS_ERR(pdd)) {
1551 		r = PTR_ERR(pdd);
1552 		goto err_unlock;
1553 	}
1554 
1555 	r = amdgpu_amdkfd_gpuvm_import_dmabuf_fd(pdd->dev->adev, args->dmabuf_fd,
1556 						 args->va_addr, pdd->drm_priv,
1557 						 (struct kgd_mem **)&mem, &size,
1558 						 NULL);
1559 	if (r)
1560 		goto err_unlock;
1561 
1562 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1563 	if (idr_handle < 0) {
1564 		r = -EFAULT;
1565 		goto err_free;
1566 	}
1567 
1568 	mutex_unlock(&p->mutex);
1569 
1570 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1571 
1572 	return 0;
1573 
1574 err_free:
1575 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1576 					       pdd->drm_priv, NULL);
1577 err_unlock:
1578 	mutex_unlock(&p->mutex);
1579 	return r;
1580 }
1581 
1582 static int kfd_ioctl_export_dmabuf(struct file *filep,
1583 				   struct kfd_process *p, void *data)
1584 {
1585 	struct kfd_ioctl_export_dmabuf_args *args = data;
1586 	struct kfd_process_device *pdd;
1587 	struct dma_buf *dmabuf;
1588 	struct kfd_node *dev;
1589 	void *mem;
1590 	int ret = 0;
1591 
1592 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1593 	if (!dev)
1594 		return -EINVAL;
1595 
1596 	mutex_lock(&p->mutex);
1597 
1598 	pdd = kfd_get_process_device_data(dev, p);
1599 	if (!pdd) {
1600 		ret = -EINVAL;
1601 		goto err_unlock;
1602 	}
1603 
1604 	mem = kfd_process_device_translate_handle(pdd,
1605 						GET_IDR_HANDLE(args->handle));
1606 	if (!mem) {
1607 		ret = -EINVAL;
1608 		goto err_unlock;
1609 	}
1610 
1611 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1612 	mutex_unlock(&p->mutex);
1613 	if (ret)
1614 		goto err_out;
1615 
1616 	ret = dma_buf_fd(dmabuf, args->flags);
1617 	if (ret < 0) {
1618 		dma_buf_put(dmabuf);
1619 		goto err_out;
1620 	}
1621 	/* dma_buf_fd assigns the reference count to the fd, no need to
1622 	 * put the reference here.
1623 	 */
1624 	args->dmabuf_fd = ret;
1625 
1626 	return 0;
1627 
1628 err_unlock:
1629 	mutex_unlock(&p->mutex);
1630 err_out:
1631 	return ret;
1632 }
1633 
1634 /* Handle requests for watching SMI events */
1635 static int kfd_ioctl_smi_events(struct file *filep,
1636 				struct kfd_process *p, void *data)
1637 {
1638 	struct kfd_ioctl_smi_events_args *args = data;
1639 	struct kfd_process_device *pdd;
1640 
1641 	mutex_lock(&p->mutex);
1642 
1643 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1644 	mutex_unlock(&p->mutex);
1645 	if (!pdd)
1646 		return -EINVAL;
1647 
1648 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1649 }
1650 
1651 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1652 
1653 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1654 				    struct kfd_process *p, void *data)
1655 {
1656 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1657 	int r = 0;
1658 
1659 	mutex_lock(&p->mutex);
1660 	if (args->xnack_enabled >= 0) {
1661 		if (!list_empty(&p->pqm.queues)) {
1662 			pr_debug("Process has user queues running\n");
1663 			r = -EBUSY;
1664 			goto out_unlock;
1665 		}
1666 
1667 		if (p->xnack_enabled == args->xnack_enabled)
1668 			goto out_unlock;
1669 
1670 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1671 			r = -EPERM;
1672 			goto out_unlock;
1673 		}
1674 
1675 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1676 	} else {
1677 		args->xnack_enabled = p->xnack_enabled;
1678 	}
1679 
1680 out_unlock:
1681 	mutex_unlock(&p->mutex);
1682 
1683 	return r;
1684 }
1685 
1686 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1687 {
1688 	struct kfd_ioctl_svm_args *args = data;
1689 	int r = 0;
1690 
1691 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1692 		 args->start_addr, args->size, args->op, args->nattr);
1693 
1694 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1695 		return -EINVAL;
1696 	if (!args->start_addr || !args->size)
1697 		return -EINVAL;
1698 
1699 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1700 		      args->attrs);
1701 
1702 	return r;
1703 }
1704 #else
1705 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1706 				    struct kfd_process *p, void *data)
1707 {
1708 	return -EPERM;
1709 }
1710 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1711 {
1712 	return -EPERM;
1713 }
1714 #endif
1715 
1716 static int criu_checkpoint_process(struct kfd_process *p,
1717 			     uint8_t __user *user_priv_data,
1718 			     uint64_t *priv_offset)
1719 {
1720 	struct kfd_criu_process_priv_data process_priv;
1721 	int ret;
1722 
1723 	memset(&process_priv, 0, sizeof(process_priv));
1724 
1725 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1726 	/* For CR, we don't consider negative xnack mode which is used for
1727 	 * querying without changing it, here 0 simply means disabled and 1
1728 	 * means enabled so retry for finding a valid PTE.
1729 	 */
1730 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1731 
1732 	ret = copy_to_user(user_priv_data + *priv_offset,
1733 				&process_priv, sizeof(process_priv));
1734 
1735 	if (ret) {
1736 		pr_err("Failed to copy process information to user\n");
1737 		ret = -EFAULT;
1738 	}
1739 
1740 	*priv_offset += sizeof(process_priv);
1741 	return ret;
1742 }
1743 
1744 static int criu_checkpoint_devices(struct kfd_process *p,
1745 			     uint32_t num_devices,
1746 			     uint8_t __user *user_addr,
1747 			     uint8_t __user *user_priv_data,
1748 			     uint64_t *priv_offset)
1749 {
1750 	struct kfd_criu_device_priv_data *device_priv = NULL;
1751 	struct kfd_criu_device_bucket *device_buckets = NULL;
1752 	int ret = 0, i;
1753 
1754 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1755 	if (!device_buckets) {
1756 		ret = -ENOMEM;
1757 		goto exit;
1758 	}
1759 
1760 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1761 	if (!device_priv) {
1762 		ret = -ENOMEM;
1763 		goto exit;
1764 	}
1765 
1766 	for (i = 0; i < num_devices; i++) {
1767 		struct kfd_process_device *pdd = p->pdds[i];
1768 
1769 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1770 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1771 
1772 		/*
1773 		 * priv_data does not contain useful information for now and is reserved for
1774 		 * future use, so we do not set its contents.
1775 		 */
1776 	}
1777 
1778 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1779 	if (ret) {
1780 		pr_err("Failed to copy device information to user\n");
1781 		ret = -EFAULT;
1782 		goto exit;
1783 	}
1784 
1785 	ret = copy_to_user(user_priv_data + *priv_offset,
1786 			   device_priv,
1787 			   num_devices * sizeof(*device_priv));
1788 	if (ret) {
1789 		pr_err("Failed to copy device information to user\n");
1790 		ret = -EFAULT;
1791 	}
1792 	*priv_offset += num_devices * sizeof(*device_priv);
1793 
1794 exit:
1795 	kvfree(device_buckets);
1796 	kvfree(device_priv);
1797 	return ret;
1798 }
1799 
1800 static uint32_t get_process_num_bos(struct kfd_process *p)
1801 {
1802 	uint32_t num_of_bos = 0;
1803 	int i;
1804 
1805 	/* Run over all PDDs of the process */
1806 	for (i = 0; i < p->n_pdds; i++) {
1807 		struct kfd_process_device *pdd = p->pdds[i];
1808 		void *mem;
1809 		int id;
1810 
1811 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1812 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1813 
1814 			if (!kgd_mem->va || kgd_mem->va > pdd->gpuvm_base)
1815 				num_of_bos++;
1816 		}
1817 	}
1818 	return num_of_bos;
1819 }
1820 
1821 static int criu_get_prime_handle(struct kgd_mem *mem,
1822 				 int flags, u32 *shared_fd)
1823 {
1824 	struct dma_buf *dmabuf;
1825 	int ret;
1826 
1827 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1828 	if (ret) {
1829 		pr_err("dmabuf export failed for the BO\n");
1830 		return ret;
1831 	}
1832 
1833 	ret = dma_buf_fd(dmabuf, flags);
1834 	if (ret < 0) {
1835 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1836 		goto out_free_dmabuf;
1837 	}
1838 
1839 	*shared_fd = ret;
1840 	return 0;
1841 
1842 out_free_dmabuf:
1843 	dma_buf_put(dmabuf);
1844 	return ret;
1845 }
1846 
1847 static int criu_checkpoint_bos(struct kfd_process *p,
1848 			       uint32_t num_bos,
1849 			       uint8_t __user *user_bos,
1850 			       uint8_t __user *user_priv_data,
1851 			       uint64_t *priv_offset)
1852 {
1853 	struct kfd_criu_bo_bucket *bo_buckets;
1854 	struct kfd_criu_bo_priv_data *bo_privs;
1855 	int ret = 0, pdd_index, bo_index = 0, id;
1856 	void *mem;
1857 
1858 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1859 	if (!bo_buckets)
1860 		return -ENOMEM;
1861 
1862 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1863 	if (!bo_privs) {
1864 		ret = -ENOMEM;
1865 		goto exit;
1866 	}
1867 
1868 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1869 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1870 		struct amdgpu_bo *dumper_bo;
1871 		struct kgd_mem *kgd_mem;
1872 
1873 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1874 			struct kfd_criu_bo_bucket *bo_bucket;
1875 			struct kfd_criu_bo_priv_data *bo_priv;
1876 			int i, dev_idx = 0;
1877 
1878 			kgd_mem = (struct kgd_mem *)mem;
1879 			dumper_bo = kgd_mem->bo;
1880 
1881 			/* Skip checkpointing BOs that are used for Trap handler
1882 			 * code and state. Currently, these BOs have a VA that
1883 			 * is less GPUVM Base
1884 			 */
1885 			if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base)
1886 				continue;
1887 
1888 			bo_bucket = &bo_buckets[bo_index];
1889 			bo_priv = &bo_privs[bo_index];
1890 
1891 			bo_bucket->gpu_id = pdd->user_gpu_id;
1892 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1893 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1894 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1895 			bo_priv->idr_handle = id;
1896 
1897 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1898 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1899 								&bo_priv->user_addr);
1900 				if (ret) {
1901 					pr_err("Failed to obtain user address for user-pointer bo\n");
1902 					goto exit;
1903 				}
1904 			}
1905 			if (bo_bucket->alloc_flags
1906 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1907 				ret = criu_get_prime_handle(kgd_mem,
1908 						bo_bucket->alloc_flags &
1909 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1910 						&bo_bucket->dmabuf_fd);
1911 				if (ret)
1912 					goto exit;
1913 			} else {
1914 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1915 			}
1916 
1917 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1918 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1919 					KFD_MMAP_GPU_ID(pdd->dev->id);
1920 			else if (bo_bucket->alloc_flags &
1921 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1922 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1923 					KFD_MMAP_GPU_ID(pdd->dev->id);
1924 			else
1925 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1926 
1927 			for (i = 0; i < p->n_pdds; i++) {
1928 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->drm_priv, kgd_mem))
1929 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1930 			}
1931 
1932 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1933 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1934 					bo_bucket->size,
1935 					bo_bucket->addr,
1936 					bo_bucket->offset,
1937 					bo_bucket->gpu_id,
1938 					bo_bucket->alloc_flags,
1939 					bo_priv->idr_handle);
1940 			bo_index++;
1941 		}
1942 	}
1943 
1944 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1945 	if (ret) {
1946 		pr_err("Failed to copy BO information to user\n");
1947 		ret = -EFAULT;
1948 		goto exit;
1949 	}
1950 
1951 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1952 	if (ret) {
1953 		pr_err("Failed to copy BO priv information to user\n");
1954 		ret = -EFAULT;
1955 		goto exit;
1956 	}
1957 
1958 	*priv_offset += num_bos * sizeof(*bo_privs);
1959 
1960 exit:
1961 	while (ret && bo_index--) {
1962 		if (bo_buckets[bo_index].alloc_flags
1963 		    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
1964 			close_fd(bo_buckets[bo_index].dmabuf_fd);
1965 	}
1966 
1967 	kvfree(bo_buckets);
1968 	kvfree(bo_privs);
1969 	return ret;
1970 }
1971 
1972 static int criu_get_process_object_info(struct kfd_process *p,
1973 					uint32_t *num_devices,
1974 					uint32_t *num_bos,
1975 					uint32_t *num_objects,
1976 					uint64_t *objs_priv_size)
1977 {
1978 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
1979 	uint32_t num_queues, num_events, num_svm_ranges;
1980 	int ret;
1981 
1982 	*num_devices = p->n_pdds;
1983 	*num_bos = get_process_num_bos(p);
1984 
1985 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
1986 	if (ret)
1987 		return ret;
1988 
1989 	num_events = kfd_get_num_events(p);
1990 
1991 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
1992 	if (ret)
1993 		return ret;
1994 
1995 	*num_objects = num_queues + num_events + num_svm_ranges;
1996 
1997 	if (objs_priv_size) {
1998 		priv_size = sizeof(struct kfd_criu_process_priv_data);
1999 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2000 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2001 		priv_size += queues_priv_data_size;
2002 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2003 		priv_size += svm_priv_data_size;
2004 		*objs_priv_size = priv_size;
2005 	}
2006 	return 0;
2007 }
2008 
2009 static int criu_checkpoint(struct file *filep,
2010 			   struct kfd_process *p,
2011 			   struct kfd_ioctl_criu_args *args)
2012 {
2013 	int ret;
2014 	uint32_t num_devices, num_bos, num_objects;
2015 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2016 
2017 	if (!args->devices || !args->bos || !args->priv_data)
2018 		return -EINVAL;
2019 
2020 	mutex_lock(&p->mutex);
2021 
2022 	if (!p->n_pdds) {
2023 		pr_err("No pdd for given process\n");
2024 		ret = -ENODEV;
2025 		goto exit_unlock;
2026 	}
2027 
2028 	/* Confirm all process queues are evicted */
2029 	if (!p->queues_paused) {
2030 		pr_err("Cannot dump process when queues are not in evicted state\n");
2031 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2032 		ret = -EINVAL;
2033 		goto exit_unlock;
2034 	}
2035 
2036 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2037 	if (ret)
2038 		goto exit_unlock;
2039 
2040 	if (num_devices != args->num_devices ||
2041 	    num_bos != args->num_bos ||
2042 	    num_objects != args->num_objects ||
2043 	    priv_size != args->priv_data_size) {
2044 
2045 		ret = -EINVAL;
2046 		goto exit_unlock;
2047 	}
2048 
2049 	/* each function will store private data inside priv_data and adjust priv_offset */
2050 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2051 	if (ret)
2052 		goto exit_unlock;
2053 
2054 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2055 				(uint8_t __user *)args->priv_data, &priv_offset);
2056 	if (ret)
2057 		goto exit_unlock;
2058 
2059 	/* Leave room for BOs in the private data. They need to be restored
2060 	 * before events, but we checkpoint them last to simplify the error
2061 	 * handling.
2062 	 */
2063 	bo_priv_offset = priv_offset;
2064 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2065 
2066 	if (num_objects) {
2067 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2068 						 &priv_offset);
2069 		if (ret)
2070 			goto exit_unlock;
2071 
2072 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2073 						 &priv_offset);
2074 		if (ret)
2075 			goto exit_unlock;
2076 
2077 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2078 		if (ret)
2079 			goto exit_unlock;
2080 	}
2081 
2082 	/* This must be the last thing in this function that can fail.
2083 	 * Otherwise we leak dmabuf file descriptors.
2084 	 */
2085 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2086 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2087 
2088 exit_unlock:
2089 	mutex_unlock(&p->mutex);
2090 	if (ret)
2091 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2092 	else
2093 		pr_debug("CRIU dump ret:%d\n", ret);
2094 
2095 	return ret;
2096 }
2097 
2098 static int criu_restore_process(struct kfd_process *p,
2099 				struct kfd_ioctl_criu_args *args,
2100 				uint64_t *priv_offset,
2101 				uint64_t max_priv_data_size)
2102 {
2103 	int ret = 0;
2104 	struct kfd_criu_process_priv_data process_priv;
2105 
2106 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2107 		return -EINVAL;
2108 
2109 	ret = copy_from_user(&process_priv,
2110 				(void __user *)(args->priv_data + *priv_offset),
2111 				sizeof(process_priv));
2112 	if (ret) {
2113 		pr_err("Failed to copy process private information from user\n");
2114 		ret = -EFAULT;
2115 		goto exit;
2116 	}
2117 	*priv_offset += sizeof(process_priv);
2118 
2119 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2120 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2121 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2122 		return -EINVAL;
2123 	}
2124 
2125 	pr_debug("Setting XNACK mode\n");
2126 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2127 		pr_err("xnack mode cannot be set\n");
2128 		ret = -EPERM;
2129 		goto exit;
2130 	} else {
2131 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2132 		p->xnack_enabled = process_priv.xnack_mode;
2133 	}
2134 
2135 exit:
2136 	return ret;
2137 }
2138 
2139 static int criu_restore_devices(struct kfd_process *p,
2140 				struct kfd_ioctl_criu_args *args,
2141 				uint64_t *priv_offset,
2142 				uint64_t max_priv_data_size)
2143 {
2144 	struct kfd_criu_device_bucket *device_buckets;
2145 	struct kfd_criu_device_priv_data *device_privs;
2146 	int ret = 0;
2147 	uint32_t i;
2148 
2149 	if (args->num_devices != p->n_pdds)
2150 		return -EINVAL;
2151 
2152 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2153 		return -EINVAL;
2154 
2155 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2156 	if (!device_buckets)
2157 		return -ENOMEM;
2158 
2159 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2160 				args->num_devices * sizeof(*device_buckets));
2161 	if (ret) {
2162 		pr_err("Failed to copy devices buckets from user\n");
2163 		ret = -EFAULT;
2164 		goto exit;
2165 	}
2166 
2167 	for (i = 0; i < args->num_devices; i++) {
2168 		struct kfd_node *dev;
2169 		struct kfd_process_device *pdd;
2170 		struct file *drm_file;
2171 
2172 		/* device private data is not currently used */
2173 
2174 		if (!device_buckets[i].user_gpu_id) {
2175 			pr_err("Invalid user gpu_id\n");
2176 			ret = -EINVAL;
2177 			goto exit;
2178 		}
2179 
2180 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2181 		if (!dev) {
2182 			pr_err("Failed to find device with gpu_id = %x\n",
2183 				device_buckets[i].actual_gpu_id);
2184 			ret = -EINVAL;
2185 			goto exit;
2186 		}
2187 
2188 		pdd = kfd_get_process_device_data(dev, p);
2189 		if (!pdd) {
2190 			pr_err("Failed to get pdd for gpu_id = %x\n",
2191 					device_buckets[i].actual_gpu_id);
2192 			ret = -EINVAL;
2193 			goto exit;
2194 		}
2195 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2196 
2197 		drm_file = fget(device_buckets[i].drm_fd);
2198 		if (!drm_file) {
2199 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2200 				device_buckets[i].drm_fd);
2201 			ret = -EINVAL;
2202 			goto exit;
2203 		}
2204 
2205 		if (pdd->drm_file) {
2206 			ret = -EINVAL;
2207 			goto exit;
2208 		}
2209 
2210 		/* create the vm using render nodes for kfd pdd */
2211 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2212 			pr_err("could not init vm for given pdd\n");
2213 			/* On success, the PDD keeps the drm_file reference */
2214 			fput(drm_file);
2215 			ret = -EINVAL;
2216 			goto exit;
2217 		}
2218 		/*
2219 		 * pdd now already has the vm bound to render node so below api won't create a new
2220 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2221 		 * for iommu v2 binding  and runtime pm.
2222 		 */
2223 		pdd = kfd_bind_process_to_device(dev, p);
2224 		if (IS_ERR(pdd)) {
2225 			ret = PTR_ERR(pdd);
2226 			goto exit;
2227 		}
2228 
2229 		if (!pdd->qpd.proc_doorbells) {
2230 			ret = kfd_alloc_process_doorbells(dev->kfd, pdd);
2231 			if (ret)
2232 				goto exit;
2233 		}
2234 	}
2235 
2236 	/*
2237 	 * We are not copying device private data from user as we are not using the data for now,
2238 	 * but we still adjust for its private data.
2239 	 */
2240 	*priv_offset += args->num_devices * sizeof(*device_privs);
2241 
2242 exit:
2243 	kfree(device_buckets);
2244 	return ret;
2245 }
2246 
2247 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2248 				      struct kfd_criu_bo_bucket *bo_bucket,
2249 				      struct kfd_criu_bo_priv_data *bo_priv,
2250 				      struct kgd_mem **kgd_mem)
2251 {
2252 	int idr_handle;
2253 	int ret;
2254 	const bool criu_resume = true;
2255 	u64 offset;
2256 
2257 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2258 		if (bo_bucket->size !=
2259 				kfd_doorbell_process_slice(pdd->dev->kfd))
2260 			return -EINVAL;
2261 
2262 		offset = kfd_get_process_doorbells(pdd);
2263 		if (!offset)
2264 			return -ENOMEM;
2265 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2266 		/* MMIO BOs need remapped bus address */
2267 		if (bo_bucket->size != PAGE_SIZE) {
2268 			pr_err("Invalid page size\n");
2269 			return -EINVAL;
2270 		}
2271 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2272 		if (!offset || (PAGE_SIZE > 4096)) {
2273 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2274 			return -ENOMEM;
2275 		}
2276 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2277 		offset = bo_priv->user_addr;
2278 	}
2279 	/* Create the BO */
2280 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2281 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2282 						      &offset, bo_bucket->alloc_flags, criu_resume);
2283 	if (ret) {
2284 		pr_err("Could not create the BO\n");
2285 		return ret;
2286 	}
2287 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2288 		 bo_bucket->size, bo_bucket->addr, offset);
2289 
2290 	/* Restore previous IDR handle */
2291 	pr_debug("Restoring old IDR handle for the BO");
2292 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2293 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2294 
2295 	if (idr_handle < 0) {
2296 		pr_err("Could not allocate idr\n");
2297 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2298 						       NULL);
2299 		return -ENOMEM;
2300 	}
2301 
2302 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2303 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2304 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2305 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2306 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2307 		bo_bucket->restored_offset = offset;
2308 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2309 		bo_bucket->restored_offset = offset;
2310 		/* Update the VRAM usage count */
2311 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2312 	}
2313 	return 0;
2314 }
2315 
2316 static int criu_restore_bo(struct kfd_process *p,
2317 			   struct kfd_criu_bo_bucket *bo_bucket,
2318 			   struct kfd_criu_bo_priv_data *bo_priv)
2319 {
2320 	struct kfd_process_device *pdd;
2321 	struct kgd_mem *kgd_mem;
2322 	int ret;
2323 	int j;
2324 
2325 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2326 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2327 		 bo_priv->idr_handle);
2328 
2329 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2330 	if (!pdd) {
2331 		pr_err("Failed to get pdd\n");
2332 		return -ENODEV;
2333 	}
2334 
2335 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2336 	if (ret)
2337 		return ret;
2338 
2339 	/* now map these BOs to GPU/s */
2340 	for (j = 0; j < p->n_pdds; j++) {
2341 		struct kfd_node *peer;
2342 		struct kfd_process_device *peer_pdd;
2343 
2344 		if (!bo_priv->mapped_gpuids[j])
2345 			break;
2346 
2347 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2348 		if (!peer_pdd)
2349 			return -EINVAL;
2350 
2351 		peer = peer_pdd->dev;
2352 
2353 		peer_pdd = kfd_bind_process_to_device(peer, p);
2354 		if (IS_ERR(peer_pdd))
2355 			return PTR_ERR(peer_pdd);
2356 
2357 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2358 							    peer_pdd->drm_priv);
2359 		if (ret) {
2360 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2361 			return ret;
2362 		}
2363 	}
2364 
2365 	pr_debug("map memory was successful for the BO\n");
2366 	/* create the dmabuf object and export the bo */
2367 	if (bo_bucket->alloc_flags
2368 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2369 		ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2370 					    &bo_bucket->dmabuf_fd);
2371 		if (ret)
2372 			return ret;
2373 	} else {
2374 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2375 	}
2376 
2377 	return 0;
2378 }
2379 
2380 static int criu_restore_bos(struct kfd_process *p,
2381 			    struct kfd_ioctl_criu_args *args,
2382 			    uint64_t *priv_offset,
2383 			    uint64_t max_priv_data_size)
2384 {
2385 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2386 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2387 	int ret = 0;
2388 	uint32_t i = 0;
2389 
2390 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2391 		return -EINVAL;
2392 
2393 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2394 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2395 
2396 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2397 	if (!bo_buckets)
2398 		return -ENOMEM;
2399 
2400 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2401 			     args->num_bos * sizeof(*bo_buckets));
2402 	if (ret) {
2403 		pr_err("Failed to copy BOs information from user\n");
2404 		ret = -EFAULT;
2405 		goto exit;
2406 	}
2407 
2408 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2409 	if (!bo_privs) {
2410 		ret = -ENOMEM;
2411 		goto exit;
2412 	}
2413 
2414 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2415 			     args->num_bos * sizeof(*bo_privs));
2416 	if (ret) {
2417 		pr_err("Failed to copy BOs information from user\n");
2418 		ret = -EFAULT;
2419 		goto exit;
2420 	}
2421 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2422 
2423 	/* Create and map new BOs */
2424 	for (; i < args->num_bos; i++) {
2425 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2426 		if (ret) {
2427 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2428 			goto exit;
2429 		}
2430 	} /* done */
2431 
2432 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2433 	ret = copy_to_user((void __user *)args->bos,
2434 				bo_buckets,
2435 				(args->num_bos * sizeof(*bo_buckets)));
2436 	if (ret)
2437 		ret = -EFAULT;
2438 
2439 exit:
2440 	while (ret && i--) {
2441 		if (bo_buckets[i].alloc_flags
2442 		   & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2443 			close_fd(bo_buckets[i].dmabuf_fd);
2444 	}
2445 	kvfree(bo_buckets);
2446 	kvfree(bo_privs);
2447 	return ret;
2448 }
2449 
2450 static int criu_restore_objects(struct file *filep,
2451 				struct kfd_process *p,
2452 				struct kfd_ioctl_criu_args *args,
2453 				uint64_t *priv_offset,
2454 				uint64_t max_priv_data_size)
2455 {
2456 	int ret = 0;
2457 	uint32_t i;
2458 
2459 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2460 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2461 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2462 
2463 	for (i = 0; i < args->num_objects; i++) {
2464 		uint32_t object_type;
2465 
2466 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2467 			pr_err("Invalid private data size\n");
2468 			return -EINVAL;
2469 		}
2470 
2471 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2472 		if (ret) {
2473 			pr_err("Failed to copy private information from user\n");
2474 			goto exit;
2475 		}
2476 
2477 		switch (object_type) {
2478 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2479 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2480 						     priv_offset, max_priv_data_size);
2481 			if (ret)
2482 				goto exit;
2483 			break;
2484 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2485 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2486 						     priv_offset, max_priv_data_size);
2487 			if (ret)
2488 				goto exit;
2489 			break;
2490 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2491 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2492 						     priv_offset, max_priv_data_size);
2493 			if (ret)
2494 				goto exit;
2495 			break;
2496 		default:
2497 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2498 			ret = -EINVAL;
2499 			goto exit;
2500 		}
2501 	}
2502 exit:
2503 	return ret;
2504 }
2505 
2506 static int criu_restore(struct file *filep,
2507 			struct kfd_process *p,
2508 			struct kfd_ioctl_criu_args *args)
2509 {
2510 	uint64_t priv_offset = 0;
2511 	int ret = 0;
2512 
2513 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2514 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2515 
2516 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2517 	    !args->num_devices || !args->num_bos)
2518 		return -EINVAL;
2519 
2520 	mutex_lock(&p->mutex);
2521 
2522 	/*
2523 	 * Set the process to evicted state to avoid running any new queues before all the memory
2524 	 * mappings are ready.
2525 	 */
2526 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2527 	if (ret)
2528 		goto exit_unlock;
2529 
2530 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2531 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2532 	if (ret)
2533 		goto exit_unlock;
2534 
2535 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2536 	if (ret)
2537 		goto exit_unlock;
2538 
2539 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2540 	if (ret)
2541 		goto exit_unlock;
2542 
2543 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2544 	if (ret)
2545 		goto exit_unlock;
2546 
2547 	if (priv_offset != args->priv_data_size) {
2548 		pr_err("Invalid private data size\n");
2549 		ret = -EINVAL;
2550 	}
2551 
2552 exit_unlock:
2553 	mutex_unlock(&p->mutex);
2554 	if (ret)
2555 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2556 	else
2557 		pr_debug("CRIU restore successful\n");
2558 
2559 	return ret;
2560 }
2561 
2562 static int criu_unpause(struct file *filep,
2563 			struct kfd_process *p,
2564 			struct kfd_ioctl_criu_args *args)
2565 {
2566 	int ret;
2567 
2568 	mutex_lock(&p->mutex);
2569 
2570 	if (!p->queues_paused) {
2571 		mutex_unlock(&p->mutex);
2572 		return -EINVAL;
2573 	}
2574 
2575 	ret = kfd_process_restore_queues(p);
2576 	if (ret)
2577 		pr_err("Failed to unpause queues ret:%d\n", ret);
2578 	else
2579 		p->queues_paused = false;
2580 
2581 	mutex_unlock(&p->mutex);
2582 
2583 	return ret;
2584 }
2585 
2586 static int criu_resume(struct file *filep,
2587 			struct kfd_process *p,
2588 			struct kfd_ioctl_criu_args *args)
2589 {
2590 	struct kfd_process *target = NULL;
2591 	struct pid *pid = NULL;
2592 	int ret = 0;
2593 
2594 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2595 		 args->pid);
2596 
2597 	pid = find_get_pid(args->pid);
2598 	if (!pid) {
2599 		pr_err("Cannot find pid info for %i\n", args->pid);
2600 		return -ESRCH;
2601 	}
2602 
2603 	pr_debug("calling kfd_lookup_process_by_pid\n");
2604 	target = kfd_lookup_process_by_pid(pid);
2605 
2606 	put_pid(pid);
2607 
2608 	if (!target) {
2609 		pr_debug("Cannot find process info for %i\n", args->pid);
2610 		return -ESRCH;
2611 	}
2612 
2613 	mutex_lock(&target->mutex);
2614 	ret = kfd_criu_resume_svm(target);
2615 	if (ret) {
2616 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2617 		goto exit;
2618 	}
2619 
2620 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2621 	if (ret)
2622 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2623 
2624 exit:
2625 	mutex_unlock(&target->mutex);
2626 
2627 	kfd_unref_process(target);
2628 	return ret;
2629 }
2630 
2631 static int criu_process_info(struct file *filep,
2632 				struct kfd_process *p,
2633 				struct kfd_ioctl_criu_args *args)
2634 {
2635 	int ret = 0;
2636 
2637 	mutex_lock(&p->mutex);
2638 
2639 	if (!p->n_pdds) {
2640 		pr_err("No pdd for given process\n");
2641 		ret = -ENODEV;
2642 		goto err_unlock;
2643 	}
2644 
2645 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2646 	if (ret)
2647 		goto err_unlock;
2648 
2649 	p->queues_paused = true;
2650 
2651 	args->pid = task_pid_nr_ns(p->lead_thread,
2652 					task_active_pid_ns(p->lead_thread));
2653 
2654 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2655 					   &args->num_objects, &args->priv_data_size);
2656 	if (ret)
2657 		goto err_unlock;
2658 
2659 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2660 				args->num_devices, args->num_bos, args->num_objects,
2661 				args->priv_data_size);
2662 
2663 err_unlock:
2664 	if (ret) {
2665 		kfd_process_restore_queues(p);
2666 		p->queues_paused = false;
2667 	}
2668 	mutex_unlock(&p->mutex);
2669 	return ret;
2670 }
2671 
2672 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2673 {
2674 	struct kfd_ioctl_criu_args *args = data;
2675 	int ret;
2676 
2677 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2678 	switch (args->op) {
2679 	case KFD_CRIU_OP_PROCESS_INFO:
2680 		ret = criu_process_info(filep, p, args);
2681 		break;
2682 	case KFD_CRIU_OP_CHECKPOINT:
2683 		ret = criu_checkpoint(filep, p, args);
2684 		break;
2685 	case KFD_CRIU_OP_UNPAUSE:
2686 		ret = criu_unpause(filep, p, args);
2687 		break;
2688 	case KFD_CRIU_OP_RESTORE:
2689 		ret = criu_restore(filep, p, args);
2690 		break;
2691 	case KFD_CRIU_OP_RESUME:
2692 		ret = criu_resume(filep, p, args);
2693 		break;
2694 	default:
2695 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2696 		ret = -EINVAL;
2697 		break;
2698 	}
2699 
2700 	if (ret)
2701 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2702 
2703 	return ret;
2704 }
2705 
2706 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2707 			bool enable_ttmp_setup)
2708 {
2709 	int i = 0, ret = 0;
2710 
2711 	if (p->is_runtime_retry)
2712 		goto retry;
2713 
2714 	if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2715 		return -EBUSY;
2716 
2717 	for (i = 0; i < p->n_pdds; i++) {
2718 		struct kfd_process_device *pdd = p->pdds[i];
2719 
2720 		if (pdd->qpd.queue_count)
2721 			return -EEXIST;
2722 
2723 		/*
2724 		 * Setup TTMPs by default.
2725 		 * Note that this call must remain here for MES ADD QUEUE to
2726 		 * skip_process_ctx_clear unconditionally as the first call to
2727 		 * SET_SHADER_DEBUGGER clears any stale process context data
2728 		 * saved in MES.
2729 		 */
2730 		if (pdd->dev->kfd->shared_resources.enable_mes)
2731 			kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
2732 	}
2733 
2734 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2735 	p->runtime_info.r_debug = r_debug;
2736 	p->runtime_info.ttmp_setup = enable_ttmp_setup;
2737 
2738 	if (p->runtime_info.ttmp_setup) {
2739 		for (i = 0; i < p->n_pdds; i++) {
2740 			struct kfd_process_device *pdd = p->pdds[i];
2741 
2742 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2743 				amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2744 				pdd->dev->kfd2kgd->enable_debug_trap(
2745 						pdd->dev->adev,
2746 						true,
2747 						pdd->dev->vm_info.last_vmid_kfd);
2748 			} else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2749 				pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2750 						pdd->dev->adev,
2751 						false,
2752 						0);
2753 			}
2754 		}
2755 	}
2756 
2757 retry:
2758 	if (p->debug_trap_enabled) {
2759 		if (!p->is_runtime_retry) {
2760 			kfd_dbg_trap_activate(p);
2761 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2762 					p, NULL, 0, false, NULL, 0);
2763 		}
2764 
2765 		mutex_unlock(&p->mutex);
2766 		ret = down_interruptible(&p->runtime_enable_sema);
2767 		mutex_lock(&p->mutex);
2768 
2769 		p->is_runtime_retry = !!ret;
2770 	}
2771 
2772 	return ret;
2773 }
2774 
2775 static int runtime_disable(struct kfd_process *p)
2776 {
2777 	int i = 0, ret;
2778 	bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2779 
2780 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2781 	p->runtime_info.r_debug = 0;
2782 
2783 	if (p->debug_trap_enabled) {
2784 		if (was_enabled)
2785 			kfd_dbg_trap_deactivate(p, false, 0);
2786 
2787 		if (!p->is_runtime_retry)
2788 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2789 					p, NULL, 0, false, NULL, 0);
2790 
2791 		mutex_unlock(&p->mutex);
2792 		ret = down_interruptible(&p->runtime_enable_sema);
2793 		mutex_lock(&p->mutex);
2794 
2795 		p->is_runtime_retry = !!ret;
2796 		if (ret)
2797 			return ret;
2798 	}
2799 
2800 	if (was_enabled && p->runtime_info.ttmp_setup) {
2801 		for (i = 0; i < p->n_pdds; i++) {
2802 			struct kfd_process_device *pdd = p->pdds[i];
2803 
2804 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2805 				amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2806 		}
2807 	}
2808 
2809 	p->runtime_info.ttmp_setup = false;
2810 
2811 	/* disable ttmp setup */
2812 	for (i = 0; i < p->n_pdds; i++) {
2813 		struct kfd_process_device *pdd = p->pdds[i];
2814 
2815 		if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2816 			pdd->spi_dbg_override =
2817 					pdd->dev->kfd2kgd->disable_debug_trap(
2818 					pdd->dev->adev,
2819 					false,
2820 					pdd->dev->vm_info.last_vmid_kfd);
2821 
2822 			if (!pdd->dev->kfd->shared_resources.enable_mes)
2823 				debug_refresh_runlist(pdd->dev->dqm);
2824 			else
2825 				kfd_dbg_set_mes_debug_mode(pdd,
2826 							   !kfd_dbg_has_cwsr_workaround(pdd->dev));
2827 		}
2828 	}
2829 
2830 	return 0;
2831 }
2832 
2833 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2834 {
2835 	struct kfd_ioctl_runtime_enable_args *args = data;
2836 	int r;
2837 
2838 	mutex_lock(&p->mutex);
2839 
2840 	if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2841 		r = runtime_enable(p, args->r_debug,
2842 				!!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2843 	else
2844 		r = runtime_disable(p);
2845 
2846 	mutex_unlock(&p->mutex);
2847 
2848 	return r;
2849 }
2850 
2851 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2852 {
2853 	struct kfd_ioctl_dbg_trap_args *args = data;
2854 	struct task_struct *thread = NULL;
2855 	struct mm_struct *mm = NULL;
2856 	struct pid *pid = NULL;
2857 	struct kfd_process *target = NULL;
2858 	struct kfd_process_device *pdd = NULL;
2859 	int r = 0;
2860 
2861 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2862 		pr_err("Debugging does not support sched_policy %i", sched_policy);
2863 		return -EINVAL;
2864 	}
2865 
2866 	pid = find_get_pid(args->pid);
2867 	if (!pid) {
2868 		pr_debug("Cannot find pid info for %i\n", args->pid);
2869 		r = -ESRCH;
2870 		goto out;
2871 	}
2872 
2873 	thread = get_pid_task(pid, PIDTYPE_PID);
2874 	if (!thread) {
2875 		r = -ESRCH;
2876 		goto out;
2877 	}
2878 
2879 	mm = get_task_mm(thread);
2880 	if (!mm) {
2881 		r = -ESRCH;
2882 		goto out;
2883 	}
2884 
2885 	if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2886 		bool create_process;
2887 
2888 		rcu_read_lock();
2889 		create_process = thread && thread != current && ptrace_parent(thread) == current;
2890 		rcu_read_unlock();
2891 
2892 		target = create_process ? kfd_create_process(thread) :
2893 					kfd_lookup_process_by_pid(pid);
2894 	} else {
2895 		target = kfd_lookup_process_by_pid(pid);
2896 	}
2897 
2898 	if (IS_ERR_OR_NULL(target)) {
2899 		pr_debug("Cannot find process PID %i to debug\n", args->pid);
2900 		r = target ? PTR_ERR(target) : -ESRCH;
2901 		target = NULL;
2902 		goto out;
2903 	}
2904 
2905 	/* Check if target is still PTRACED. */
2906 	rcu_read_lock();
2907 	if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2908 				&& ptrace_parent(target->lead_thread) != current) {
2909 		pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2910 		r = -EPERM;
2911 	}
2912 	rcu_read_unlock();
2913 
2914 	if (r)
2915 		goto out;
2916 
2917 	mutex_lock(&target->mutex);
2918 
2919 	if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2920 		pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2921 		r = -EINVAL;
2922 		goto unlock_out;
2923 	}
2924 
2925 	if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2926 			(args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2927 			 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2928 			 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2929 			 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2930 			 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2931 			 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2932 			 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2933 		r = -EPERM;
2934 		goto unlock_out;
2935 	}
2936 
2937 	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2938 	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2939 		int user_gpu_id = kfd_process_get_user_gpu_id(target,
2940 				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2941 					args->set_node_address_watch.gpu_id :
2942 					args->clear_node_address_watch.gpu_id);
2943 
2944 		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
2945 		if (user_gpu_id == -EINVAL || !pdd) {
2946 			r = -ENODEV;
2947 			goto unlock_out;
2948 		}
2949 	}
2950 
2951 	switch (args->op) {
2952 	case KFD_IOC_DBG_TRAP_ENABLE:
2953 		if (target != p)
2954 			target->debugger_process = p;
2955 
2956 		r = kfd_dbg_trap_enable(target,
2957 					args->enable.dbg_fd,
2958 					(void __user *)args->enable.rinfo_ptr,
2959 					&args->enable.rinfo_size);
2960 		if (!r)
2961 			target->exception_enable_mask = args->enable.exception_mask;
2962 
2963 		break;
2964 	case KFD_IOC_DBG_TRAP_DISABLE:
2965 		r = kfd_dbg_trap_disable(target);
2966 		break;
2967 	case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
2968 		r = kfd_dbg_send_exception_to_runtime(target,
2969 				args->send_runtime_event.gpu_id,
2970 				args->send_runtime_event.queue_id,
2971 				args->send_runtime_event.exception_mask);
2972 		break;
2973 	case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
2974 		kfd_dbg_set_enabled_debug_exception_mask(target,
2975 				args->set_exceptions_enabled.exception_mask);
2976 		break;
2977 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
2978 		r = kfd_dbg_trap_set_wave_launch_override(target,
2979 				args->launch_override.override_mode,
2980 				args->launch_override.enable_mask,
2981 				args->launch_override.support_request_mask,
2982 				&args->launch_override.enable_mask,
2983 				&args->launch_override.support_request_mask);
2984 		break;
2985 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
2986 		r = kfd_dbg_trap_set_wave_launch_mode(target,
2987 				args->launch_mode.launch_mode);
2988 		break;
2989 	case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
2990 		r = suspend_queues(target,
2991 				args->suspend_queues.num_queues,
2992 				args->suspend_queues.grace_period,
2993 				args->suspend_queues.exception_mask,
2994 				(uint32_t *)args->suspend_queues.queue_array_ptr);
2995 
2996 		break;
2997 	case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
2998 		r = resume_queues(target, args->resume_queues.num_queues,
2999 				(uint32_t *)args->resume_queues.queue_array_ptr);
3000 		break;
3001 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3002 		r = kfd_dbg_trap_set_dev_address_watch(pdd,
3003 				args->set_node_address_watch.address,
3004 				args->set_node_address_watch.mask,
3005 				&args->set_node_address_watch.id,
3006 				args->set_node_address_watch.mode);
3007 		break;
3008 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3009 		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3010 				args->clear_node_address_watch.id);
3011 		break;
3012 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
3013 		r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3014 		break;
3015 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3016 		r = kfd_dbg_ev_query_debug_event(target,
3017 				&args->query_debug_event.queue_id,
3018 				&args->query_debug_event.gpu_id,
3019 				args->query_debug_event.exception_mask,
3020 				&args->query_debug_event.exception_mask);
3021 		break;
3022 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3023 		r = kfd_dbg_trap_query_exception_info(target,
3024 				args->query_exception_info.source_id,
3025 				args->query_exception_info.exception_code,
3026 				args->query_exception_info.clear_exception,
3027 				(void __user *)args->query_exception_info.info_ptr,
3028 				&args->query_exception_info.info_size);
3029 		break;
3030 	case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3031 		r = pqm_get_queue_snapshot(&target->pqm,
3032 				args->queue_snapshot.exception_mask,
3033 				(void __user *)args->queue_snapshot.snapshot_buf_ptr,
3034 				&args->queue_snapshot.num_queues,
3035 				&args->queue_snapshot.entry_size);
3036 		break;
3037 	case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3038 		r = kfd_dbg_trap_device_snapshot(target,
3039 				args->device_snapshot.exception_mask,
3040 				(void __user *)args->device_snapshot.snapshot_buf_ptr,
3041 				&args->device_snapshot.num_devices,
3042 				&args->device_snapshot.entry_size);
3043 		break;
3044 	default:
3045 		pr_err("Invalid option: %i\n", args->op);
3046 		r = -EINVAL;
3047 	}
3048 
3049 unlock_out:
3050 	mutex_unlock(&target->mutex);
3051 
3052 out:
3053 	if (thread)
3054 		put_task_struct(thread);
3055 
3056 	if (mm)
3057 		mmput(mm);
3058 
3059 	if (pid)
3060 		put_pid(pid);
3061 
3062 	if (target)
3063 		kfd_unref_process(target);
3064 
3065 	return r;
3066 }
3067 
3068 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3069 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3070 			    .cmd_drv = 0, .name = #ioctl}
3071 
3072 /** Ioctl table */
3073 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3074 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3075 			kfd_ioctl_get_version, 0),
3076 
3077 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3078 			kfd_ioctl_create_queue, 0),
3079 
3080 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3081 			kfd_ioctl_destroy_queue, 0),
3082 
3083 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3084 			kfd_ioctl_set_memory_policy, 0),
3085 
3086 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3087 			kfd_ioctl_get_clock_counters, 0),
3088 
3089 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3090 			kfd_ioctl_get_process_apertures, 0),
3091 
3092 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3093 			kfd_ioctl_update_queue, 0),
3094 
3095 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3096 			kfd_ioctl_create_event, 0),
3097 
3098 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3099 			kfd_ioctl_destroy_event, 0),
3100 
3101 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3102 			kfd_ioctl_set_event, 0),
3103 
3104 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3105 			kfd_ioctl_reset_event, 0),
3106 
3107 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3108 			kfd_ioctl_wait_events, 0),
3109 
3110 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3111 			kfd_ioctl_dbg_register, 0),
3112 
3113 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3114 			kfd_ioctl_dbg_unregister, 0),
3115 
3116 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3117 			kfd_ioctl_dbg_address_watch, 0),
3118 
3119 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3120 			kfd_ioctl_dbg_wave_control, 0),
3121 
3122 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3123 			kfd_ioctl_set_scratch_backing_va, 0),
3124 
3125 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3126 			kfd_ioctl_get_tile_config, 0),
3127 
3128 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3129 			kfd_ioctl_set_trap_handler, 0),
3130 
3131 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3132 			kfd_ioctl_get_process_apertures_new, 0),
3133 
3134 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3135 			kfd_ioctl_acquire_vm, 0),
3136 
3137 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3138 			kfd_ioctl_alloc_memory_of_gpu, 0),
3139 
3140 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3141 			kfd_ioctl_free_memory_of_gpu, 0),
3142 
3143 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3144 			kfd_ioctl_map_memory_to_gpu, 0),
3145 
3146 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3147 			kfd_ioctl_unmap_memory_from_gpu, 0),
3148 
3149 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3150 			kfd_ioctl_set_cu_mask, 0),
3151 
3152 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3153 			kfd_ioctl_get_queue_wave_state, 0),
3154 
3155 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3156 				kfd_ioctl_get_dmabuf_info, 0),
3157 
3158 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3159 				kfd_ioctl_import_dmabuf, 0),
3160 
3161 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3162 			kfd_ioctl_alloc_queue_gws, 0),
3163 
3164 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3165 			kfd_ioctl_smi_events, 0),
3166 
3167 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3168 
3169 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3170 			kfd_ioctl_set_xnack_mode, 0),
3171 
3172 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3173 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3174 
3175 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3176 			kfd_ioctl_get_available_memory, 0),
3177 
3178 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3179 				kfd_ioctl_export_dmabuf, 0),
3180 
3181 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3182 			kfd_ioctl_runtime_enable, 0),
3183 
3184 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3185 			kfd_ioctl_set_debug_trap, 0),
3186 };
3187 
3188 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
3189 
3190 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3191 {
3192 	struct kfd_process *process;
3193 	amdkfd_ioctl_t *func;
3194 	const struct amdkfd_ioctl_desc *ioctl = NULL;
3195 	unsigned int nr = _IOC_NR(cmd);
3196 	char stack_kdata[128];
3197 	char *kdata = NULL;
3198 	unsigned int usize, asize;
3199 	int retcode = -EINVAL;
3200 	bool ptrace_attached = false;
3201 
3202 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3203 		goto err_i1;
3204 
3205 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3206 		u32 amdkfd_size;
3207 
3208 		ioctl = &amdkfd_ioctls[nr];
3209 
3210 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
3211 		usize = asize = _IOC_SIZE(cmd);
3212 		if (amdkfd_size > asize)
3213 			asize = amdkfd_size;
3214 
3215 		cmd = ioctl->cmd;
3216 	} else
3217 		goto err_i1;
3218 
3219 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3220 
3221 	/* Get the process struct from the filep. Only the process
3222 	 * that opened /dev/kfd can use the file descriptor. Child
3223 	 * processes need to create their own KFD device context.
3224 	 */
3225 	process = filep->private_data;
3226 
3227 	rcu_read_lock();
3228 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3229 	    ptrace_parent(process->lead_thread) == current)
3230 		ptrace_attached = true;
3231 	rcu_read_unlock();
3232 
3233 	if (process->lead_thread != current->group_leader
3234 	    && !ptrace_attached) {
3235 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3236 		retcode = -EBADF;
3237 		goto err_i1;
3238 	}
3239 
3240 	/* Do not trust userspace, use our own definition */
3241 	func = ioctl->func;
3242 
3243 	if (unlikely(!func)) {
3244 		dev_dbg(kfd_device, "no function\n");
3245 		retcode = -EINVAL;
3246 		goto err_i1;
3247 	}
3248 
3249 	/*
3250 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3251 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3252 	 * more priviledged access.
3253 	 */
3254 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3255 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
3256 						!capable(CAP_SYS_ADMIN)) {
3257 			retcode = -EACCES;
3258 			goto err_i1;
3259 		}
3260 	}
3261 
3262 	if (cmd & (IOC_IN | IOC_OUT)) {
3263 		if (asize <= sizeof(stack_kdata)) {
3264 			kdata = stack_kdata;
3265 		} else {
3266 			kdata = kmalloc(asize, GFP_KERNEL);
3267 			if (!kdata) {
3268 				retcode = -ENOMEM;
3269 				goto err_i1;
3270 			}
3271 		}
3272 		if (asize > usize)
3273 			memset(kdata + usize, 0, asize - usize);
3274 	}
3275 
3276 	if (cmd & IOC_IN) {
3277 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3278 			retcode = -EFAULT;
3279 			goto err_i1;
3280 		}
3281 	} else if (cmd & IOC_OUT) {
3282 		memset(kdata, 0, usize);
3283 	}
3284 
3285 	retcode = func(filep, process, kdata);
3286 
3287 	if (cmd & IOC_OUT)
3288 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3289 			retcode = -EFAULT;
3290 
3291 err_i1:
3292 	if (!ioctl)
3293 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3294 			  task_pid_nr(current), cmd, nr);
3295 
3296 	if (kdata != stack_kdata)
3297 		kfree(kdata);
3298 
3299 	if (retcode)
3300 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3301 				nr, arg, retcode);
3302 
3303 	return retcode;
3304 }
3305 
3306 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3307 		      struct vm_area_struct *vma)
3308 {
3309 	phys_addr_t address;
3310 
3311 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3312 		return -EINVAL;
3313 
3314 	if (PAGE_SIZE > 4096)
3315 		return -EINVAL;
3316 
3317 	address = dev->adev->rmmio_remap.bus_addr;
3318 
3319 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3320 				VM_DONTDUMP | VM_PFNMAP);
3321 
3322 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3323 
3324 	pr_debug("pasid 0x%x mapping mmio page\n"
3325 		 "     target user address == 0x%08llX\n"
3326 		 "     physical address    == 0x%08llX\n"
3327 		 "     vm_flags            == 0x%04lX\n"
3328 		 "     size                == 0x%04lX\n",
3329 		 process->pasid, (unsigned long long) vma->vm_start,
3330 		 address, vma->vm_flags, PAGE_SIZE);
3331 
3332 	return io_remap_pfn_range(vma,
3333 				vma->vm_start,
3334 				address >> PAGE_SHIFT,
3335 				PAGE_SIZE,
3336 				vma->vm_page_prot);
3337 }
3338 
3339 
3340 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3341 {
3342 	struct kfd_process *process;
3343 	struct kfd_node *dev = NULL;
3344 	unsigned long mmap_offset;
3345 	unsigned int gpu_id;
3346 
3347 	process = kfd_get_process(current);
3348 	if (IS_ERR(process))
3349 		return PTR_ERR(process);
3350 
3351 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3352 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3353 	if (gpu_id)
3354 		dev = kfd_device_by_id(gpu_id);
3355 
3356 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3357 	case KFD_MMAP_TYPE_DOORBELL:
3358 		if (!dev)
3359 			return -ENODEV;
3360 		return kfd_doorbell_mmap(dev, process, vma);
3361 
3362 	case KFD_MMAP_TYPE_EVENTS:
3363 		return kfd_event_mmap(process, vma);
3364 
3365 	case KFD_MMAP_TYPE_RESERVED_MEM:
3366 		if (!dev)
3367 			return -ENODEV;
3368 		return kfd_reserved_mem_mmap(dev, process, vma);
3369 	case KFD_MMAP_TYPE_MMIO:
3370 		if (!dev)
3371 			return -ENODEV;
3372 		return kfd_mmio_mmap(dev, process, vma);
3373 	}
3374 
3375 	return -EFAULT;
3376 }
3377