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 static struct class *kfd_class;
67 struct device *kfd_device;
68 
69 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
70 {
71 	struct kfd_process_device *pdd;
72 
73 	mutex_lock(&p->mutex);
74 	pdd = kfd_process_device_data_by_id(p, gpu_id);
75 
76 	if (pdd)
77 		return pdd;
78 
79 	mutex_unlock(&p->mutex);
80 	return NULL;
81 }
82 
83 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
84 {
85 	mutex_unlock(&pdd->process->mutex);
86 }
87 
88 int kfd_chardev_init(void)
89 {
90 	int err = 0;
91 
92 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
93 	err = kfd_char_dev_major;
94 	if (err < 0)
95 		goto err_register_chrdev;
96 
97 	kfd_class = class_create(kfd_dev_name);
98 	err = PTR_ERR(kfd_class);
99 	if (IS_ERR(kfd_class))
100 		goto err_class_create;
101 
102 	kfd_device = device_create(kfd_class, NULL,
103 					MKDEV(kfd_char_dev_major, 0),
104 					NULL, kfd_dev_name);
105 	err = PTR_ERR(kfd_device);
106 	if (IS_ERR(kfd_device))
107 		goto err_device_create;
108 
109 	return 0;
110 
111 err_device_create:
112 	class_destroy(kfd_class);
113 err_class_create:
114 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
115 err_register_chrdev:
116 	return err;
117 }
118 
119 void kfd_chardev_exit(void)
120 {
121 	device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
122 	class_destroy(kfd_class);
123 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
124 	kfd_device = NULL;
125 }
126 
127 
128 static int kfd_open(struct inode *inode, struct file *filep)
129 {
130 	struct kfd_process *process;
131 	bool is_32bit_user_mode;
132 
133 	if (iminor(inode) != 0)
134 		return -ENODEV;
135 
136 	is_32bit_user_mode = in_compat_syscall();
137 
138 	if (is_32bit_user_mode) {
139 		dev_warn(kfd_device,
140 			"Process %d (32-bit) failed to open /dev/kfd\n"
141 			"32-bit processes are not supported by amdkfd\n",
142 			current->pid);
143 		return -EPERM;
144 	}
145 
146 	process = kfd_create_process(current);
147 	if (IS_ERR(process))
148 		return PTR_ERR(process);
149 
150 	if (kfd_process_init_cwsr_apu(process, filep)) {
151 		kfd_unref_process(process);
152 		return -EFAULT;
153 	}
154 
155 	/* filep now owns the reference returned by kfd_create_process */
156 	filep->private_data = process;
157 
158 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
159 		process->pasid, process->is_32bit_user_mode);
160 
161 	return 0;
162 }
163 
164 static int kfd_release(struct inode *inode, struct file *filep)
165 {
166 	struct kfd_process *process = filep->private_data;
167 
168 	if (process)
169 		kfd_unref_process(process);
170 
171 	return 0;
172 }
173 
174 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
175 					void *data)
176 {
177 	struct kfd_ioctl_get_version_args *args = data;
178 
179 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
180 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
181 
182 	return 0;
183 }
184 
185 static int set_queue_properties_from_user(struct queue_properties *q_properties,
186 				struct kfd_ioctl_create_queue_args *args)
187 {
188 	/*
189 	 * Repurpose queue percentage to accommodate new features:
190 	 * bit 0-7: queue percentage
191 	 * bit 8-15: pm4_target_xcc
192 	 */
193 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
194 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
195 		return -EINVAL;
196 	}
197 
198 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
199 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
200 		return -EINVAL;
201 	}
202 
203 	if ((args->ring_base_address) &&
204 		(!access_ok((const void __user *) args->ring_base_address,
205 			sizeof(uint64_t)))) {
206 		pr_err("Can't access ring base address\n");
207 		return -EFAULT;
208 	}
209 
210 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
211 		pr_err("Ring size must be a power of 2 or 0\n");
212 		return -EINVAL;
213 	}
214 
215 	if (!access_ok((const void __user *) args->read_pointer_address,
216 			sizeof(uint32_t))) {
217 		pr_err("Can't access read pointer\n");
218 		return -EFAULT;
219 	}
220 
221 	if (!access_ok((const void __user *) args->write_pointer_address,
222 			sizeof(uint32_t))) {
223 		pr_err("Can't access write pointer\n");
224 		return -EFAULT;
225 	}
226 
227 	if (args->eop_buffer_address &&
228 		!access_ok((const void __user *) args->eop_buffer_address,
229 			sizeof(uint32_t))) {
230 		pr_debug("Can't access eop buffer");
231 		return -EFAULT;
232 	}
233 
234 	if (args->ctx_save_restore_address &&
235 		!access_ok((const void __user *) args->ctx_save_restore_address,
236 			sizeof(uint32_t))) {
237 		pr_debug("Can't access ctx save restore buffer");
238 		return -EFAULT;
239 	}
240 
241 	q_properties->is_interop = false;
242 	q_properties->is_gws = false;
243 	q_properties->queue_percent = args->queue_percentage & 0xFF;
244 	/* bit 8-15 are repurposed to be PM4 target XCC */
245 	q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
246 	q_properties->priority = args->queue_priority;
247 	q_properties->queue_address = args->ring_base_address;
248 	q_properties->queue_size = args->ring_size;
249 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
250 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
251 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
252 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
253 	q_properties->ctx_save_restore_area_address =
254 			args->ctx_save_restore_address;
255 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
256 	q_properties->ctl_stack_size = args->ctl_stack_size;
257 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
258 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
259 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
260 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
261 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
262 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
263 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
264 	else
265 		return -ENOTSUPP;
266 
267 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
268 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
269 	else
270 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
271 
272 	pr_debug("Queue Percentage: %d, %d\n",
273 			q_properties->queue_percent, args->queue_percentage);
274 
275 	pr_debug("Queue Priority: %d, %d\n",
276 			q_properties->priority, args->queue_priority);
277 
278 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
279 			q_properties->queue_address, args->ring_base_address);
280 
281 	pr_debug("Queue Size: 0x%llX, %u\n",
282 			q_properties->queue_size, args->ring_size);
283 
284 	pr_debug("Queue r/w Pointers: %px, %px\n",
285 			q_properties->read_ptr,
286 			q_properties->write_ptr);
287 
288 	pr_debug("Queue Format: %d\n", q_properties->format);
289 
290 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
291 
292 	pr_debug("Queue CTX save area: 0x%llX\n",
293 			q_properties->ctx_save_restore_area_address);
294 
295 	return 0;
296 }
297 
298 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
299 					void *data)
300 {
301 	struct kfd_ioctl_create_queue_args *args = data;
302 	struct kfd_node *dev;
303 	int err = 0;
304 	unsigned int queue_id;
305 	struct kfd_process_device *pdd;
306 	struct queue_properties q_properties;
307 	uint32_t doorbell_offset_in_process = 0;
308 	struct amdgpu_bo *wptr_bo = NULL;
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 	/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
345 	 * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
346 	 */
347 	if (dev->kfd->shared_resources.enable_mes &&
348 			((dev->adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK)
349 			>> AMDGPU_MES_API_VERSION_SHIFT) >= 2) {
350 		struct amdgpu_bo_va_mapping *wptr_mapping;
351 		struct amdgpu_vm *wptr_vm;
352 
353 		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
354 		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
355 		if (err)
356 			goto err_wptr_map_gart;
357 
358 		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
359 				wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
360 		amdgpu_bo_unreserve(wptr_vm->root.bo);
361 		if (!wptr_mapping) {
362 			pr_err("Failed to lookup wptr bo\n");
363 			err = -EINVAL;
364 			goto err_wptr_map_gart;
365 		}
366 
367 		wptr_bo = wptr_mapping->bo_va->base.bo;
368 		if (wptr_bo->tbo.base.size > PAGE_SIZE) {
369 			pr_err("Requested GART mapping for wptr bo larger than one page\n");
370 			err = -EINVAL;
371 			goto err_wptr_map_gart;
372 		}
373 
374 		err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
375 		if (err) {
376 			pr_err("Failed to map wptr bo to GART\n");
377 			goto err_wptr_map_gart;
378 		}
379 	}
380 
381 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
382 			p->pasid,
383 			dev->id);
384 
385 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
386 			NULL, NULL, NULL, &doorbell_offset_in_process);
387 	if (err != 0)
388 		goto err_create_queue;
389 
390 	args->queue_id = queue_id;
391 
392 
393 	/* Return gpu_id as doorbell offset for mmap usage */
394 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
395 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
396 	if (KFD_IS_SOC15(dev))
397 		/* On SOC15 ASICs, include the doorbell offset within the
398 		 * process doorbell frame, which is 2 pages.
399 		 */
400 		args->doorbell_offset |= doorbell_offset_in_process;
401 
402 	mutex_unlock(&p->mutex);
403 
404 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
405 
406 	pr_debug("Ring buffer address == 0x%016llX\n",
407 			args->ring_base_address);
408 
409 	pr_debug("Read ptr address    == 0x%016llX\n",
410 			args->read_pointer_address);
411 
412 	pr_debug("Write ptr address   == 0x%016llX\n",
413 			args->write_pointer_address);
414 
415 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
416 	return 0;
417 
418 err_create_queue:
419 	if (wptr_bo)
420 		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
421 err_wptr_map_gart:
422 err_bind_process:
423 err_pdd:
424 	mutex_unlock(&p->mutex);
425 	return err;
426 }
427 
428 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
429 					void *data)
430 {
431 	int retval;
432 	struct kfd_ioctl_destroy_queue_args *args = data;
433 
434 	pr_debug("Destroying queue id %d for pasid 0x%x\n",
435 				args->queue_id,
436 				p->pasid);
437 
438 	mutex_lock(&p->mutex);
439 
440 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
441 
442 	mutex_unlock(&p->mutex);
443 	return retval;
444 }
445 
446 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
447 					void *data)
448 {
449 	int retval;
450 	struct kfd_ioctl_update_queue_args *args = data;
451 	struct queue_properties properties;
452 
453 	/*
454 	 * Repurpose queue percentage to accommodate new features:
455 	 * bit 0-7: queue percentage
456 	 * bit 8-15: pm4_target_xcc
457 	 */
458 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
459 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
460 		return -EINVAL;
461 	}
462 
463 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
464 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
465 		return -EINVAL;
466 	}
467 
468 	if ((args->ring_base_address) &&
469 		(!access_ok((const void __user *) args->ring_base_address,
470 			sizeof(uint64_t)))) {
471 		pr_err("Can't access ring base address\n");
472 		return -EFAULT;
473 	}
474 
475 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
476 		pr_err("Ring size must be a power of 2 or 0\n");
477 		return -EINVAL;
478 	}
479 
480 	properties.queue_address = args->ring_base_address;
481 	properties.queue_size = args->ring_size;
482 	properties.queue_percent = args->queue_percentage & 0xFF;
483 	/* bit 8-15 are repurposed to be PM4 target XCC */
484 	properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
485 	properties.priority = args->queue_priority;
486 
487 	pr_debug("Updating queue id %d for pasid 0x%x\n",
488 			args->queue_id, p->pasid);
489 
490 	mutex_lock(&p->mutex);
491 
492 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
493 
494 	mutex_unlock(&p->mutex);
495 
496 	return retval;
497 }
498 
499 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
500 					void *data)
501 {
502 	int retval;
503 	const int max_num_cus = 1024;
504 	struct kfd_ioctl_set_cu_mask_args *args = data;
505 	struct mqd_update_info minfo = {0};
506 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
507 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
508 
509 	if ((args->num_cu_mask % 32) != 0) {
510 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
511 				args->num_cu_mask);
512 		return -EINVAL;
513 	}
514 
515 	minfo.cu_mask.count = args->num_cu_mask;
516 	if (minfo.cu_mask.count == 0) {
517 		pr_debug("CU mask cannot be 0");
518 		return -EINVAL;
519 	}
520 
521 	/* To prevent an unreasonably large CU mask size, set an arbitrary
522 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
523 	 * past max_num_cus bits and just use the first max_num_cus bits.
524 	 */
525 	if (minfo.cu_mask.count > max_num_cus) {
526 		pr_debug("CU mask cannot be greater than 1024 bits");
527 		minfo.cu_mask.count = max_num_cus;
528 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
529 	}
530 
531 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
532 	if (!minfo.cu_mask.ptr)
533 		return -ENOMEM;
534 
535 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
536 	if (retval) {
537 		pr_debug("Could not copy CU mask from userspace");
538 		retval = -EFAULT;
539 		goto out;
540 	}
541 
542 	mutex_lock(&p->mutex);
543 
544 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
545 
546 	mutex_unlock(&p->mutex);
547 
548 out:
549 	kfree(minfo.cu_mask.ptr);
550 	return retval;
551 }
552 
553 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
554 					  struct kfd_process *p, void *data)
555 {
556 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
557 	int r;
558 
559 	mutex_lock(&p->mutex);
560 
561 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
562 			       (void __user *)args->ctl_stack_address,
563 			       &args->ctl_stack_used_size,
564 			       &args->save_area_used_size);
565 
566 	mutex_unlock(&p->mutex);
567 
568 	return r;
569 }
570 
571 static int kfd_ioctl_set_memory_policy(struct file *filep,
572 					struct kfd_process *p, void *data)
573 {
574 	struct kfd_ioctl_set_memory_policy_args *args = data;
575 	int err = 0;
576 	struct kfd_process_device *pdd;
577 	enum cache_policy default_policy, alternate_policy;
578 
579 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
580 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
581 		return -EINVAL;
582 	}
583 
584 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
585 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
586 		return -EINVAL;
587 	}
588 
589 	mutex_lock(&p->mutex);
590 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
591 	if (!pdd) {
592 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
593 		err = -EINVAL;
594 		goto err_pdd;
595 	}
596 
597 	pdd = kfd_bind_process_to_device(pdd->dev, p);
598 	if (IS_ERR(pdd)) {
599 		err = -ESRCH;
600 		goto out;
601 	}
602 
603 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
604 			 ? cache_policy_coherent : cache_policy_noncoherent;
605 
606 	alternate_policy =
607 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
608 		   ? cache_policy_coherent : cache_policy_noncoherent;
609 
610 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
611 				&pdd->qpd,
612 				default_policy,
613 				alternate_policy,
614 				(void __user *)args->alternate_aperture_base,
615 				args->alternate_aperture_size))
616 		err = -EINVAL;
617 
618 out:
619 err_pdd:
620 	mutex_unlock(&p->mutex);
621 
622 	return err;
623 }
624 
625 static int kfd_ioctl_set_trap_handler(struct file *filep,
626 					struct kfd_process *p, void *data)
627 {
628 	struct kfd_ioctl_set_trap_handler_args *args = data;
629 	int err = 0;
630 	struct kfd_process_device *pdd;
631 
632 	mutex_lock(&p->mutex);
633 
634 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
635 	if (!pdd) {
636 		err = -EINVAL;
637 		goto err_pdd;
638 	}
639 
640 	pdd = kfd_bind_process_to_device(pdd->dev, p);
641 	if (IS_ERR(pdd)) {
642 		err = -ESRCH;
643 		goto out;
644 	}
645 
646 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
647 
648 out:
649 err_pdd:
650 	mutex_unlock(&p->mutex);
651 
652 	return err;
653 }
654 
655 static int kfd_ioctl_dbg_register(struct file *filep,
656 				struct kfd_process *p, void *data)
657 {
658 	return -EPERM;
659 }
660 
661 static int kfd_ioctl_dbg_unregister(struct file *filep,
662 				struct kfd_process *p, void *data)
663 {
664 	return -EPERM;
665 }
666 
667 static int kfd_ioctl_dbg_address_watch(struct file *filep,
668 					struct kfd_process *p, void *data)
669 {
670 	return -EPERM;
671 }
672 
673 /* Parse and generate fixed size data structure for wave control */
674 static int kfd_ioctl_dbg_wave_control(struct file *filep,
675 					struct kfd_process *p, void *data)
676 {
677 	return -EPERM;
678 }
679 
680 static int kfd_ioctl_get_clock_counters(struct file *filep,
681 				struct kfd_process *p, void *data)
682 {
683 	struct kfd_ioctl_get_clock_counters_args *args = data;
684 	struct kfd_process_device *pdd;
685 
686 	mutex_lock(&p->mutex);
687 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
688 	mutex_unlock(&p->mutex);
689 	if (pdd)
690 		/* Reading GPU clock counter from KGD */
691 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
692 	else
693 		/* Node without GPU resource */
694 		args->gpu_clock_counter = 0;
695 
696 	/* No access to rdtsc. Using raw monotonic time */
697 	args->cpu_clock_counter = ktime_get_raw_ns();
698 	args->system_clock_counter = ktime_get_boottime_ns();
699 
700 	/* Since the counter is in nano-seconds we use 1GHz frequency */
701 	args->system_clock_freq = 1000000000;
702 
703 	return 0;
704 }
705 
706 
707 static int kfd_ioctl_get_process_apertures(struct file *filp,
708 				struct kfd_process *p, void *data)
709 {
710 	struct kfd_ioctl_get_process_apertures_args *args = data;
711 	struct kfd_process_device_apertures *pAperture;
712 	int i;
713 
714 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
715 
716 	args->num_of_nodes = 0;
717 
718 	mutex_lock(&p->mutex);
719 	/* Run over all pdd of the process */
720 	for (i = 0; i < p->n_pdds; i++) {
721 		struct kfd_process_device *pdd = p->pdds[i];
722 
723 		pAperture =
724 			&args->process_apertures[args->num_of_nodes];
725 		pAperture->gpu_id = pdd->dev->id;
726 		pAperture->lds_base = pdd->lds_base;
727 		pAperture->lds_limit = pdd->lds_limit;
728 		pAperture->gpuvm_base = pdd->gpuvm_base;
729 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
730 		pAperture->scratch_base = pdd->scratch_base;
731 		pAperture->scratch_limit = pdd->scratch_limit;
732 
733 		dev_dbg(kfd_device,
734 			"node id %u\n", args->num_of_nodes);
735 		dev_dbg(kfd_device,
736 			"gpu id %u\n", pdd->dev->id);
737 		dev_dbg(kfd_device,
738 			"lds_base %llX\n", pdd->lds_base);
739 		dev_dbg(kfd_device,
740 			"lds_limit %llX\n", pdd->lds_limit);
741 		dev_dbg(kfd_device,
742 			"gpuvm_base %llX\n", pdd->gpuvm_base);
743 		dev_dbg(kfd_device,
744 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
745 		dev_dbg(kfd_device,
746 			"scratch_base %llX\n", pdd->scratch_base);
747 		dev_dbg(kfd_device,
748 			"scratch_limit %llX\n", pdd->scratch_limit);
749 
750 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
751 			break;
752 	}
753 	mutex_unlock(&p->mutex);
754 
755 	return 0;
756 }
757 
758 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
759 				struct kfd_process *p, void *data)
760 {
761 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
762 	struct kfd_process_device_apertures *pa;
763 	int ret;
764 	int i;
765 
766 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
767 
768 	if (args->num_of_nodes == 0) {
769 		/* Return number of nodes, so that user space can alloacate
770 		 * sufficient memory
771 		 */
772 		mutex_lock(&p->mutex);
773 		args->num_of_nodes = p->n_pdds;
774 		goto out_unlock;
775 	}
776 
777 	/* Fill in process-aperture information for all available
778 	 * nodes, but not more than args->num_of_nodes as that is
779 	 * the amount of memory allocated by user
780 	 */
781 	pa = kcalloc(args->num_of_nodes, sizeof(struct kfd_process_device_apertures),
782 		     GFP_KERNEL);
783 	if (!pa)
784 		return -ENOMEM;
785 
786 	mutex_lock(&p->mutex);
787 
788 	if (!p->n_pdds) {
789 		args->num_of_nodes = 0;
790 		kfree(pa);
791 		goto out_unlock;
792 	}
793 
794 	/* Run over all pdd of the process */
795 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
796 		struct kfd_process_device *pdd = p->pdds[i];
797 
798 		pa[i].gpu_id = pdd->dev->id;
799 		pa[i].lds_base = pdd->lds_base;
800 		pa[i].lds_limit = pdd->lds_limit;
801 		pa[i].gpuvm_base = pdd->gpuvm_base;
802 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
803 		pa[i].scratch_base = pdd->scratch_base;
804 		pa[i].scratch_limit = pdd->scratch_limit;
805 
806 		dev_dbg(kfd_device,
807 			"gpu id %u\n", pdd->dev->id);
808 		dev_dbg(kfd_device,
809 			"lds_base %llX\n", pdd->lds_base);
810 		dev_dbg(kfd_device,
811 			"lds_limit %llX\n", pdd->lds_limit);
812 		dev_dbg(kfd_device,
813 			"gpuvm_base %llX\n", pdd->gpuvm_base);
814 		dev_dbg(kfd_device,
815 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
816 		dev_dbg(kfd_device,
817 			"scratch_base %llX\n", pdd->scratch_base);
818 		dev_dbg(kfd_device,
819 			"scratch_limit %llX\n", pdd->scratch_limit);
820 	}
821 	mutex_unlock(&p->mutex);
822 
823 	args->num_of_nodes = i;
824 	ret = copy_to_user(
825 			(void __user *)args->kfd_process_device_apertures_ptr,
826 			pa,
827 			(i * sizeof(struct kfd_process_device_apertures)));
828 	kfree(pa);
829 	return ret ? -EFAULT : 0;
830 
831 out_unlock:
832 	mutex_unlock(&p->mutex);
833 	return 0;
834 }
835 
836 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
837 					void *data)
838 {
839 	struct kfd_ioctl_create_event_args *args = data;
840 	int err;
841 
842 	/* For dGPUs the event page is allocated in user mode. The
843 	 * handle is passed to KFD with the first call to this IOCTL
844 	 * through the event_page_offset field.
845 	 */
846 	if (args->event_page_offset) {
847 		mutex_lock(&p->mutex);
848 		err = kfd_kmap_event_page(p, args->event_page_offset);
849 		mutex_unlock(&p->mutex);
850 		if (err)
851 			return err;
852 	}
853 
854 	err = kfd_event_create(filp, p, args->event_type,
855 				args->auto_reset != 0, args->node_id,
856 				&args->event_id, &args->event_trigger_data,
857 				&args->event_page_offset,
858 				&args->event_slot_index);
859 
860 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
861 	return err;
862 }
863 
864 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
865 					void *data)
866 {
867 	struct kfd_ioctl_destroy_event_args *args = data;
868 
869 	return kfd_event_destroy(p, args->event_id);
870 }
871 
872 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
873 				void *data)
874 {
875 	struct kfd_ioctl_set_event_args *args = data;
876 
877 	return kfd_set_event(p, args->event_id);
878 }
879 
880 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
881 				void *data)
882 {
883 	struct kfd_ioctl_reset_event_args *args = data;
884 
885 	return kfd_reset_event(p, args->event_id);
886 }
887 
888 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
889 				void *data)
890 {
891 	struct kfd_ioctl_wait_events_args *args = data;
892 
893 	return kfd_wait_on_events(p, args->num_events,
894 			(void __user *)args->events_ptr,
895 			(args->wait_for_all != 0),
896 			&args->timeout, &args->wait_result);
897 }
898 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
899 					struct kfd_process *p, void *data)
900 {
901 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
902 	struct kfd_process_device *pdd;
903 	struct kfd_node *dev;
904 	long err;
905 
906 	mutex_lock(&p->mutex);
907 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
908 	if (!pdd) {
909 		err = -EINVAL;
910 		goto err_pdd;
911 	}
912 	dev = pdd->dev;
913 
914 	pdd = kfd_bind_process_to_device(dev, p);
915 	if (IS_ERR(pdd)) {
916 		err = PTR_ERR(pdd);
917 		goto bind_process_to_device_fail;
918 	}
919 
920 	pdd->qpd.sh_hidden_private_base = args->va_addr;
921 
922 	mutex_unlock(&p->mutex);
923 
924 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
925 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
926 		dev->kfd2kgd->set_scratch_backing_va(
927 			dev->adev, args->va_addr, pdd->qpd.vmid);
928 
929 	return 0;
930 
931 bind_process_to_device_fail:
932 err_pdd:
933 	mutex_unlock(&p->mutex);
934 	return err;
935 }
936 
937 static int kfd_ioctl_get_tile_config(struct file *filep,
938 		struct kfd_process *p, void *data)
939 {
940 	struct kfd_ioctl_get_tile_config_args *args = data;
941 	struct kfd_process_device *pdd;
942 	struct tile_config config;
943 	int err = 0;
944 
945 	mutex_lock(&p->mutex);
946 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
947 	mutex_unlock(&p->mutex);
948 	if (!pdd)
949 		return -EINVAL;
950 
951 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
952 
953 	args->gb_addr_config = config.gb_addr_config;
954 	args->num_banks = config.num_banks;
955 	args->num_ranks = config.num_ranks;
956 
957 	if (args->num_tile_configs > config.num_tile_configs)
958 		args->num_tile_configs = config.num_tile_configs;
959 	err = copy_to_user((void __user *)args->tile_config_ptr,
960 			config.tile_config_ptr,
961 			args->num_tile_configs * sizeof(uint32_t));
962 	if (err) {
963 		args->num_tile_configs = 0;
964 		return -EFAULT;
965 	}
966 
967 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
968 		args->num_macro_tile_configs =
969 				config.num_macro_tile_configs;
970 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
971 			config.macro_tile_config_ptr,
972 			args->num_macro_tile_configs * sizeof(uint32_t));
973 	if (err) {
974 		args->num_macro_tile_configs = 0;
975 		return -EFAULT;
976 	}
977 
978 	return 0;
979 }
980 
981 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
982 				void *data)
983 {
984 	struct kfd_ioctl_acquire_vm_args *args = data;
985 	struct kfd_process_device *pdd;
986 	struct file *drm_file;
987 	int ret;
988 
989 	drm_file = fget(args->drm_fd);
990 	if (!drm_file)
991 		return -EINVAL;
992 
993 	mutex_lock(&p->mutex);
994 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
995 	if (!pdd) {
996 		ret = -EINVAL;
997 		goto err_pdd;
998 	}
999 
1000 	if (pdd->drm_file) {
1001 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
1002 		goto err_drm_file;
1003 	}
1004 
1005 	ret = kfd_process_device_init_vm(pdd, drm_file);
1006 	if (ret)
1007 		goto err_unlock;
1008 
1009 	/* On success, the PDD keeps the drm_file reference */
1010 	mutex_unlock(&p->mutex);
1011 
1012 	return 0;
1013 
1014 err_unlock:
1015 err_pdd:
1016 err_drm_file:
1017 	mutex_unlock(&p->mutex);
1018 	fput(drm_file);
1019 	return ret;
1020 }
1021 
1022 bool kfd_dev_is_large_bar(struct kfd_node *dev)
1023 {
1024 	if (debug_largebar) {
1025 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1026 		return true;
1027 	}
1028 
1029 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1030 	    dev->local_mem_info.local_mem_size_public > 0)
1031 		return true;
1032 
1033 	if (dev->local_mem_info.local_mem_size_public == 0 &&
1034 	    dev->kfd->adev->gmc.is_app_apu) {
1035 		pr_debug("APP APU, Consider like a large bar system\n");
1036 		return true;
1037 	}
1038 
1039 	return false;
1040 }
1041 
1042 static int kfd_ioctl_get_available_memory(struct file *filep,
1043 					  struct kfd_process *p, void *data)
1044 {
1045 	struct kfd_ioctl_get_available_memory_args *args = data;
1046 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1047 
1048 	if (!pdd)
1049 		return -EINVAL;
1050 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1051 							pdd->dev->node_id);
1052 	kfd_unlock_pdd(pdd);
1053 	return 0;
1054 }
1055 
1056 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1057 					struct kfd_process *p, void *data)
1058 {
1059 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1060 	struct kfd_process_device *pdd;
1061 	void *mem;
1062 	struct kfd_node *dev;
1063 	int idr_handle;
1064 	long err;
1065 	uint64_t offset = args->mmap_offset;
1066 	uint32_t flags = args->flags;
1067 
1068 	if (args->size == 0)
1069 		return -EINVAL;
1070 
1071 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1072 	/* Flush pending deferred work to avoid racing with deferred actions
1073 	 * from previous memory map changes (e.g. munmap).
1074 	 */
1075 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1076 	mutex_lock(&p->svms.lock);
1077 	mmap_write_unlock(current->mm);
1078 	if (interval_tree_iter_first(&p->svms.objects,
1079 				     args->va_addr >> PAGE_SHIFT,
1080 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1081 		pr_err("Address: 0x%llx already allocated by SVM\n",
1082 			args->va_addr);
1083 		mutex_unlock(&p->svms.lock);
1084 		return -EADDRINUSE;
1085 	}
1086 
1087 	/* When register user buffer check if it has been registered by svm by
1088 	 * buffer cpu virtual address.
1089 	 */
1090 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1091 	    interval_tree_iter_first(&p->svms.objects,
1092 				     args->mmap_offset >> PAGE_SHIFT,
1093 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1094 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1095 			args->mmap_offset);
1096 		mutex_unlock(&p->svms.lock);
1097 		return -EADDRINUSE;
1098 	}
1099 
1100 	mutex_unlock(&p->svms.lock);
1101 #endif
1102 	mutex_lock(&p->mutex);
1103 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1104 	if (!pdd) {
1105 		err = -EINVAL;
1106 		goto err_pdd;
1107 	}
1108 
1109 	dev = pdd->dev;
1110 
1111 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1112 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1113 		!kfd_dev_is_large_bar(dev)) {
1114 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1115 		err = -EINVAL;
1116 		goto err_large_bar;
1117 	}
1118 
1119 	pdd = kfd_bind_process_to_device(dev, p);
1120 	if (IS_ERR(pdd)) {
1121 		err = PTR_ERR(pdd);
1122 		goto err_unlock;
1123 	}
1124 
1125 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1126 		if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1127 			err = -EINVAL;
1128 			goto err_unlock;
1129 		}
1130 		offset = kfd_get_process_doorbells(pdd);
1131 		if (!offset) {
1132 			err = -ENOMEM;
1133 			goto err_unlock;
1134 		}
1135 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1136 		if (args->size != PAGE_SIZE) {
1137 			err = -EINVAL;
1138 			goto err_unlock;
1139 		}
1140 		offset = dev->adev->rmmio_remap.bus_addr;
1141 		if (!offset || (PAGE_SIZE > 4096)) {
1142 			err = -ENOMEM;
1143 			goto err_unlock;
1144 		}
1145 	}
1146 
1147 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1148 		dev->adev, args->va_addr, args->size,
1149 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1150 		flags, false);
1151 
1152 	if (err)
1153 		goto err_unlock;
1154 
1155 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1156 	if (idr_handle < 0) {
1157 		err = -EFAULT;
1158 		goto err_free;
1159 	}
1160 
1161 	/* Update the VRAM usage count */
1162 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1163 		uint64_t size = args->size;
1164 
1165 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1166 			size >>= 1;
1167 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1168 	}
1169 
1170 	mutex_unlock(&p->mutex);
1171 
1172 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1173 	args->mmap_offset = offset;
1174 
1175 	/* MMIO is mapped through kfd device
1176 	 * Generate a kfd mmap offset
1177 	 */
1178 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1179 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1180 					| KFD_MMAP_GPU_ID(args->gpu_id);
1181 
1182 	return 0;
1183 
1184 err_free:
1185 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1186 					       pdd->drm_priv, NULL);
1187 err_unlock:
1188 err_pdd:
1189 err_large_bar:
1190 	mutex_unlock(&p->mutex);
1191 	return err;
1192 }
1193 
1194 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1195 					struct kfd_process *p, void *data)
1196 {
1197 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1198 	struct kfd_process_device *pdd;
1199 	void *mem;
1200 	int ret;
1201 	uint64_t size = 0;
1202 
1203 	mutex_lock(&p->mutex);
1204 	/*
1205 	 * Safeguard to prevent user space from freeing signal BO.
1206 	 * It will be freed at process termination.
1207 	 */
1208 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1209 		pr_err("Free signal BO is not allowed\n");
1210 		ret = -EPERM;
1211 		goto err_unlock;
1212 	}
1213 
1214 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1215 	if (!pdd) {
1216 		pr_err("Process device data doesn't exist\n");
1217 		ret = -EINVAL;
1218 		goto err_pdd;
1219 	}
1220 
1221 	mem = kfd_process_device_translate_handle(
1222 		pdd, GET_IDR_HANDLE(args->handle));
1223 	if (!mem) {
1224 		ret = -EINVAL;
1225 		goto err_unlock;
1226 	}
1227 
1228 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1229 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1230 
1231 	/* If freeing the buffer failed, leave the handle in place for
1232 	 * clean-up during process tear-down.
1233 	 */
1234 	if (!ret)
1235 		kfd_process_device_remove_obj_handle(
1236 			pdd, GET_IDR_HANDLE(args->handle));
1237 
1238 	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1239 
1240 err_unlock:
1241 err_pdd:
1242 	mutex_unlock(&p->mutex);
1243 	return ret;
1244 }
1245 
1246 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1247 					struct kfd_process *p, void *data)
1248 {
1249 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1250 	struct kfd_process_device *pdd, *peer_pdd;
1251 	void *mem;
1252 	struct kfd_node *dev;
1253 	long err = 0;
1254 	int i;
1255 	uint32_t *devices_arr = NULL;
1256 
1257 	if (!args->n_devices) {
1258 		pr_debug("Device IDs array empty\n");
1259 		return -EINVAL;
1260 	}
1261 	if (args->n_success > args->n_devices) {
1262 		pr_debug("n_success exceeds n_devices\n");
1263 		return -EINVAL;
1264 	}
1265 
1266 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1267 				    GFP_KERNEL);
1268 	if (!devices_arr)
1269 		return -ENOMEM;
1270 
1271 	err = copy_from_user(devices_arr,
1272 			     (void __user *)args->device_ids_array_ptr,
1273 			     args->n_devices * sizeof(*devices_arr));
1274 	if (err != 0) {
1275 		err = -EFAULT;
1276 		goto copy_from_user_failed;
1277 	}
1278 
1279 	mutex_lock(&p->mutex);
1280 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1281 	if (!pdd) {
1282 		err = -EINVAL;
1283 		goto get_process_device_data_failed;
1284 	}
1285 	dev = pdd->dev;
1286 
1287 	pdd = kfd_bind_process_to_device(dev, p);
1288 	if (IS_ERR(pdd)) {
1289 		err = PTR_ERR(pdd);
1290 		goto bind_process_to_device_failed;
1291 	}
1292 
1293 	mem = kfd_process_device_translate_handle(pdd,
1294 						GET_IDR_HANDLE(args->handle));
1295 	if (!mem) {
1296 		err = -ENOMEM;
1297 		goto get_mem_obj_from_handle_failed;
1298 	}
1299 
1300 	for (i = args->n_success; i < args->n_devices; i++) {
1301 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1302 		if (!peer_pdd) {
1303 			pr_debug("Getting device by id failed for 0x%x\n",
1304 				 devices_arr[i]);
1305 			err = -EINVAL;
1306 			goto get_mem_obj_from_handle_failed;
1307 		}
1308 
1309 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1310 		if (IS_ERR(peer_pdd)) {
1311 			err = PTR_ERR(peer_pdd);
1312 			goto get_mem_obj_from_handle_failed;
1313 		}
1314 
1315 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1316 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1317 			peer_pdd->drm_priv);
1318 		if (err) {
1319 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1320 
1321 			dev_err(dev->adev->dev,
1322 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1323 			       pci_domain_nr(pdev->bus),
1324 			       pdev->bus->number,
1325 			       PCI_SLOT(pdev->devfn),
1326 			       PCI_FUNC(pdev->devfn),
1327 			       ((struct kgd_mem *)mem)->domain);
1328 			goto map_memory_to_gpu_failed;
1329 		}
1330 		args->n_success = i+1;
1331 	}
1332 
1333 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1334 	if (err) {
1335 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1336 		goto sync_memory_failed;
1337 	}
1338 
1339 	mutex_unlock(&p->mutex);
1340 
1341 	/* Flush TLBs after waiting for the page table updates to complete */
1342 	for (i = 0; i < args->n_devices; i++) {
1343 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1344 		if (WARN_ON_ONCE(!peer_pdd))
1345 			continue;
1346 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1347 	}
1348 	kfree(devices_arr);
1349 
1350 	return err;
1351 
1352 get_process_device_data_failed:
1353 bind_process_to_device_failed:
1354 get_mem_obj_from_handle_failed:
1355 map_memory_to_gpu_failed:
1356 sync_memory_failed:
1357 	mutex_unlock(&p->mutex);
1358 copy_from_user_failed:
1359 	kfree(devices_arr);
1360 
1361 	return err;
1362 }
1363 
1364 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1365 					struct kfd_process *p, void *data)
1366 {
1367 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1368 	struct kfd_process_device *pdd, *peer_pdd;
1369 	void *mem;
1370 	long err = 0;
1371 	uint32_t *devices_arr = NULL, i;
1372 	bool flush_tlb;
1373 
1374 	if (!args->n_devices) {
1375 		pr_debug("Device IDs array empty\n");
1376 		return -EINVAL;
1377 	}
1378 	if (args->n_success > args->n_devices) {
1379 		pr_debug("n_success exceeds n_devices\n");
1380 		return -EINVAL;
1381 	}
1382 
1383 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1384 				    GFP_KERNEL);
1385 	if (!devices_arr)
1386 		return -ENOMEM;
1387 
1388 	err = copy_from_user(devices_arr,
1389 			     (void __user *)args->device_ids_array_ptr,
1390 			     args->n_devices * sizeof(*devices_arr));
1391 	if (err != 0) {
1392 		err = -EFAULT;
1393 		goto copy_from_user_failed;
1394 	}
1395 
1396 	mutex_lock(&p->mutex);
1397 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1398 	if (!pdd) {
1399 		err = -EINVAL;
1400 		goto bind_process_to_device_failed;
1401 	}
1402 
1403 	mem = kfd_process_device_translate_handle(pdd,
1404 						GET_IDR_HANDLE(args->handle));
1405 	if (!mem) {
1406 		err = -ENOMEM;
1407 		goto get_mem_obj_from_handle_failed;
1408 	}
1409 
1410 	for (i = args->n_success; i < args->n_devices; i++) {
1411 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1412 		if (!peer_pdd) {
1413 			err = -EINVAL;
1414 			goto get_mem_obj_from_handle_failed;
1415 		}
1416 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1417 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1418 		if (err) {
1419 			pr_err("Failed to unmap from gpu %d/%d\n",
1420 			       i, args->n_devices);
1421 			goto unmap_memory_from_gpu_failed;
1422 		}
1423 		args->n_success = i+1;
1424 	}
1425 
1426 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1427 	if (flush_tlb) {
1428 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1429 				(struct kgd_mem *) mem, true);
1430 		if (err) {
1431 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1432 			goto sync_memory_failed;
1433 		}
1434 	}
1435 
1436 	/* Flush TLBs after waiting for the page table updates to complete */
1437 	for (i = 0; i < args->n_devices; i++) {
1438 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1439 		if (WARN_ON_ONCE(!peer_pdd))
1440 			continue;
1441 		if (flush_tlb)
1442 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1443 
1444 		/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */
1445 		err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd->drm_priv);
1446 		if (err)
1447 			goto sync_memory_failed;
1448 	}
1449 
1450 	mutex_unlock(&p->mutex);
1451 
1452 	kfree(devices_arr);
1453 
1454 	return 0;
1455 
1456 bind_process_to_device_failed:
1457 get_mem_obj_from_handle_failed:
1458 unmap_memory_from_gpu_failed:
1459 sync_memory_failed:
1460 	mutex_unlock(&p->mutex);
1461 copy_from_user_failed:
1462 	kfree(devices_arr);
1463 	return err;
1464 }
1465 
1466 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1467 		struct kfd_process *p, void *data)
1468 {
1469 	int retval;
1470 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1471 	struct queue *q;
1472 	struct kfd_node *dev;
1473 
1474 	mutex_lock(&p->mutex);
1475 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1476 
1477 	if (q) {
1478 		dev = q->device;
1479 	} else {
1480 		retval = -EINVAL;
1481 		goto out_unlock;
1482 	}
1483 
1484 	if (!dev->gws) {
1485 		retval = -ENODEV;
1486 		goto out_unlock;
1487 	}
1488 
1489 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1490 		retval = -ENODEV;
1491 		goto out_unlock;
1492 	}
1493 
1494 	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1495 				      kfd_dbg_has_cwsr_workaround(dev))) {
1496 		retval = -EBUSY;
1497 		goto out_unlock;
1498 	}
1499 
1500 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1501 	mutex_unlock(&p->mutex);
1502 
1503 	args->first_gws = 0;
1504 	return retval;
1505 
1506 out_unlock:
1507 	mutex_unlock(&p->mutex);
1508 	return retval;
1509 }
1510 
1511 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1512 		struct kfd_process *p, void *data)
1513 {
1514 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1515 	struct kfd_node *dev = NULL;
1516 	struct amdgpu_device *dmabuf_adev;
1517 	void *metadata_buffer = NULL;
1518 	uint32_t flags;
1519 	int8_t xcp_id;
1520 	unsigned int i;
1521 	int r;
1522 
1523 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1524 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1525 		if (dev && !kfd_devcgroup_check_permission(dev))
1526 			break;
1527 	if (!dev)
1528 		return -EINVAL;
1529 
1530 	if (args->metadata_ptr) {
1531 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1532 		if (!metadata_buffer)
1533 			return -ENOMEM;
1534 	}
1535 
1536 	/* Get dmabuf info from KGD */
1537 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1538 					  &dmabuf_adev, &args->size,
1539 					  metadata_buffer, args->metadata_size,
1540 					  &args->metadata_size, &flags, &xcp_id);
1541 	if (r)
1542 		goto exit;
1543 
1544 	if (xcp_id >= 0)
1545 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1546 	else
1547 		args->gpu_id = dev->id;
1548 	args->flags = flags;
1549 
1550 	/* Copy metadata buffer to user mode */
1551 	if (metadata_buffer) {
1552 		r = copy_to_user((void __user *)args->metadata_ptr,
1553 				 metadata_buffer, args->metadata_size);
1554 		if (r != 0)
1555 			r = -EFAULT;
1556 	}
1557 
1558 exit:
1559 	kfree(metadata_buffer);
1560 
1561 	return r;
1562 }
1563 
1564 static int kfd_ioctl_import_dmabuf(struct file *filep,
1565 				   struct kfd_process *p, void *data)
1566 {
1567 	struct kfd_ioctl_import_dmabuf_args *args = data;
1568 	struct kfd_process_device *pdd;
1569 	struct dma_buf *dmabuf;
1570 	int idr_handle;
1571 	uint64_t size;
1572 	void *mem;
1573 	int r;
1574 
1575 	dmabuf = dma_buf_get(args->dmabuf_fd);
1576 	if (IS_ERR(dmabuf))
1577 		return PTR_ERR(dmabuf);
1578 
1579 	mutex_lock(&p->mutex);
1580 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1581 	if (!pdd) {
1582 		r = -EINVAL;
1583 		goto err_unlock;
1584 	}
1585 
1586 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1587 	if (IS_ERR(pdd)) {
1588 		r = PTR_ERR(pdd);
1589 		goto err_unlock;
1590 	}
1591 
1592 	r = amdgpu_amdkfd_gpuvm_import_dmabuf(pdd->dev->adev, dmabuf,
1593 					      args->va_addr, pdd->drm_priv,
1594 					      (struct kgd_mem **)&mem, &size,
1595 					      NULL);
1596 	if (r)
1597 		goto err_unlock;
1598 
1599 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1600 	if (idr_handle < 0) {
1601 		r = -EFAULT;
1602 		goto err_free;
1603 	}
1604 
1605 	mutex_unlock(&p->mutex);
1606 	dma_buf_put(dmabuf);
1607 
1608 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1609 
1610 	return 0;
1611 
1612 err_free:
1613 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1614 					       pdd->drm_priv, NULL);
1615 err_unlock:
1616 	mutex_unlock(&p->mutex);
1617 	dma_buf_put(dmabuf);
1618 	return r;
1619 }
1620 
1621 static int kfd_ioctl_export_dmabuf(struct file *filep,
1622 				   struct kfd_process *p, void *data)
1623 {
1624 	struct kfd_ioctl_export_dmabuf_args *args = data;
1625 	struct kfd_process_device *pdd;
1626 	struct dma_buf *dmabuf;
1627 	struct kfd_node *dev;
1628 	void *mem;
1629 	int ret = 0;
1630 
1631 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1632 	if (!dev)
1633 		return -EINVAL;
1634 
1635 	mutex_lock(&p->mutex);
1636 
1637 	pdd = kfd_get_process_device_data(dev, p);
1638 	if (!pdd) {
1639 		ret = -EINVAL;
1640 		goto err_unlock;
1641 	}
1642 
1643 	mem = kfd_process_device_translate_handle(pdd,
1644 						GET_IDR_HANDLE(args->handle));
1645 	if (!mem) {
1646 		ret = -EINVAL;
1647 		goto err_unlock;
1648 	}
1649 
1650 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1651 	mutex_unlock(&p->mutex);
1652 	if (ret)
1653 		goto err_out;
1654 
1655 	ret = dma_buf_fd(dmabuf, args->flags);
1656 	if (ret < 0) {
1657 		dma_buf_put(dmabuf);
1658 		goto err_out;
1659 	}
1660 	/* dma_buf_fd assigns the reference count to the fd, no need to
1661 	 * put the reference here.
1662 	 */
1663 	args->dmabuf_fd = ret;
1664 
1665 	return 0;
1666 
1667 err_unlock:
1668 	mutex_unlock(&p->mutex);
1669 err_out:
1670 	return ret;
1671 }
1672 
1673 /* Handle requests for watching SMI events */
1674 static int kfd_ioctl_smi_events(struct file *filep,
1675 				struct kfd_process *p, void *data)
1676 {
1677 	struct kfd_ioctl_smi_events_args *args = data;
1678 	struct kfd_process_device *pdd;
1679 
1680 	mutex_lock(&p->mutex);
1681 
1682 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1683 	mutex_unlock(&p->mutex);
1684 	if (!pdd)
1685 		return -EINVAL;
1686 
1687 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1688 }
1689 
1690 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1691 
1692 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1693 				    struct kfd_process *p, void *data)
1694 {
1695 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1696 	int r = 0;
1697 
1698 	mutex_lock(&p->mutex);
1699 	if (args->xnack_enabled >= 0) {
1700 		if (!list_empty(&p->pqm.queues)) {
1701 			pr_debug("Process has user queues running\n");
1702 			r = -EBUSY;
1703 			goto out_unlock;
1704 		}
1705 
1706 		if (p->xnack_enabled == args->xnack_enabled)
1707 			goto out_unlock;
1708 
1709 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1710 			r = -EPERM;
1711 			goto out_unlock;
1712 		}
1713 
1714 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1715 	} else {
1716 		args->xnack_enabled = p->xnack_enabled;
1717 	}
1718 
1719 out_unlock:
1720 	mutex_unlock(&p->mutex);
1721 
1722 	return r;
1723 }
1724 
1725 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1726 {
1727 	struct kfd_ioctl_svm_args *args = data;
1728 	int r = 0;
1729 
1730 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1731 		 args->start_addr, args->size, args->op, args->nattr);
1732 
1733 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1734 		return -EINVAL;
1735 	if (!args->start_addr || !args->size)
1736 		return -EINVAL;
1737 
1738 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1739 		      args->attrs);
1740 
1741 	return r;
1742 }
1743 #else
1744 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1745 				    struct kfd_process *p, void *data)
1746 {
1747 	return -EPERM;
1748 }
1749 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1750 {
1751 	return -EPERM;
1752 }
1753 #endif
1754 
1755 static int criu_checkpoint_process(struct kfd_process *p,
1756 			     uint8_t __user *user_priv_data,
1757 			     uint64_t *priv_offset)
1758 {
1759 	struct kfd_criu_process_priv_data process_priv;
1760 	int ret;
1761 
1762 	memset(&process_priv, 0, sizeof(process_priv));
1763 
1764 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1765 	/* For CR, we don't consider negative xnack mode which is used for
1766 	 * querying without changing it, here 0 simply means disabled and 1
1767 	 * means enabled so retry for finding a valid PTE.
1768 	 */
1769 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1770 
1771 	ret = copy_to_user(user_priv_data + *priv_offset,
1772 				&process_priv, sizeof(process_priv));
1773 
1774 	if (ret) {
1775 		pr_err("Failed to copy process information to user\n");
1776 		ret = -EFAULT;
1777 	}
1778 
1779 	*priv_offset += sizeof(process_priv);
1780 	return ret;
1781 }
1782 
1783 static int criu_checkpoint_devices(struct kfd_process *p,
1784 			     uint32_t num_devices,
1785 			     uint8_t __user *user_addr,
1786 			     uint8_t __user *user_priv_data,
1787 			     uint64_t *priv_offset)
1788 {
1789 	struct kfd_criu_device_priv_data *device_priv = NULL;
1790 	struct kfd_criu_device_bucket *device_buckets = NULL;
1791 	int ret = 0, i;
1792 
1793 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1794 	if (!device_buckets) {
1795 		ret = -ENOMEM;
1796 		goto exit;
1797 	}
1798 
1799 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1800 	if (!device_priv) {
1801 		ret = -ENOMEM;
1802 		goto exit;
1803 	}
1804 
1805 	for (i = 0; i < num_devices; i++) {
1806 		struct kfd_process_device *pdd = p->pdds[i];
1807 
1808 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1809 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1810 
1811 		/*
1812 		 * priv_data does not contain useful information for now and is reserved for
1813 		 * future use, so we do not set its contents.
1814 		 */
1815 	}
1816 
1817 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1818 	if (ret) {
1819 		pr_err("Failed to copy device information to user\n");
1820 		ret = -EFAULT;
1821 		goto exit;
1822 	}
1823 
1824 	ret = copy_to_user(user_priv_data + *priv_offset,
1825 			   device_priv,
1826 			   num_devices * sizeof(*device_priv));
1827 	if (ret) {
1828 		pr_err("Failed to copy device information to user\n");
1829 		ret = -EFAULT;
1830 	}
1831 	*priv_offset += num_devices * sizeof(*device_priv);
1832 
1833 exit:
1834 	kvfree(device_buckets);
1835 	kvfree(device_priv);
1836 	return ret;
1837 }
1838 
1839 static uint32_t get_process_num_bos(struct kfd_process *p)
1840 {
1841 	uint32_t num_of_bos = 0;
1842 	int i;
1843 
1844 	/* Run over all PDDs of the process */
1845 	for (i = 0; i < p->n_pdds; i++) {
1846 		struct kfd_process_device *pdd = p->pdds[i];
1847 		void *mem;
1848 		int id;
1849 
1850 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1851 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1852 
1853 			if (!kgd_mem->va || kgd_mem->va > pdd->gpuvm_base)
1854 				num_of_bos++;
1855 		}
1856 	}
1857 	return num_of_bos;
1858 }
1859 
1860 static int criu_get_prime_handle(struct kgd_mem *mem, int flags,
1861 				      u32 *shared_fd)
1862 {
1863 	struct dma_buf *dmabuf;
1864 	int ret;
1865 
1866 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1867 	if (ret) {
1868 		pr_err("dmabuf export failed for the BO\n");
1869 		return ret;
1870 	}
1871 
1872 	ret = dma_buf_fd(dmabuf, flags);
1873 	if (ret < 0) {
1874 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1875 		goto out_free_dmabuf;
1876 	}
1877 
1878 	*shared_fd = ret;
1879 	return 0;
1880 
1881 out_free_dmabuf:
1882 	dma_buf_put(dmabuf);
1883 	return ret;
1884 }
1885 
1886 static int criu_checkpoint_bos(struct kfd_process *p,
1887 			       uint32_t num_bos,
1888 			       uint8_t __user *user_bos,
1889 			       uint8_t __user *user_priv_data,
1890 			       uint64_t *priv_offset)
1891 {
1892 	struct kfd_criu_bo_bucket *bo_buckets;
1893 	struct kfd_criu_bo_priv_data *bo_privs;
1894 	int ret = 0, pdd_index, bo_index = 0, id;
1895 	void *mem;
1896 
1897 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1898 	if (!bo_buckets)
1899 		return -ENOMEM;
1900 
1901 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1902 	if (!bo_privs) {
1903 		ret = -ENOMEM;
1904 		goto exit;
1905 	}
1906 
1907 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1908 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1909 		struct amdgpu_bo *dumper_bo;
1910 		struct kgd_mem *kgd_mem;
1911 
1912 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1913 			struct kfd_criu_bo_bucket *bo_bucket;
1914 			struct kfd_criu_bo_priv_data *bo_priv;
1915 			int i, dev_idx = 0;
1916 
1917 			if (!mem) {
1918 				ret = -ENOMEM;
1919 				goto exit;
1920 			}
1921 
1922 			kgd_mem = (struct kgd_mem *)mem;
1923 			dumper_bo = kgd_mem->bo;
1924 
1925 			/* Skip checkpointing BOs that are used for Trap handler
1926 			 * code and state. Currently, these BOs have a VA that
1927 			 * is less GPUVM Base
1928 			 */
1929 			if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base)
1930 				continue;
1931 
1932 			bo_bucket = &bo_buckets[bo_index];
1933 			bo_priv = &bo_privs[bo_index];
1934 
1935 			bo_bucket->gpu_id = pdd->user_gpu_id;
1936 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1937 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1938 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1939 			bo_priv->idr_handle = id;
1940 
1941 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1942 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1943 								&bo_priv->user_addr);
1944 				if (ret) {
1945 					pr_err("Failed to obtain user address for user-pointer bo\n");
1946 					goto exit;
1947 				}
1948 			}
1949 			if (bo_bucket->alloc_flags
1950 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1951 				ret = criu_get_prime_handle(kgd_mem,
1952 						bo_bucket->alloc_flags &
1953 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1954 						&bo_bucket->dmabuf_fd);
1955 				if (ret)
1956 					goto exit;
1957 			} else {
1958 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1959 			}
1960 
1961 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1962 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1963 					KFD_MMAP_GPU_ID(pdd->dev->id);
1964 			else if (bo_bucket->alloc_flags &
1965 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1966 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1967 					KFD_MMAP_GPU_ID(pdd->dev->id);
1968 			else
1969 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1970 
1971 			for (i = 0; i < p->n_pdds; i++) {
1972 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->dev->adev, kgd_mem))
1973 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1974 			}
1975 
1976 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1977 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1978 					bo_bucket->size,
1979 					bo_bucket->addr,
1980 					bo_bucket->offset,
1981 					bo_bucket->gpu_id,
1982 					bo_bucket->alloc_flags,
1983 					bo_priv->idr_handle);
1984 			bo_index++;
1985 		}
1986 	}
1987 
1988 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1989 	if (ret) {
1990 		pr_err("Failed to copy BO information to user\n");
1991 		ret = -EFAULT;
1992 		goto exit;
1993 	}
1994 
1995 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1996 	if (ret) {
1997 		pr_err("Failed to copy BO priv information to user\n");
1998 		ret = -EFAULT;
1999 		goto exit;
2000 	}
2001 
2002 	*priv_offset += num_bos * sizeof(*bo_privs);
2003 
2004 exit:
2005 	while (ret && bo_index--) {
2006 		if (bo_buckets[bo_index].alloc_flags
2007 		    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2008 			close_fd(bo_buckets[bo_index].dmabuf_fd);
2009 	}
2010 
2011 	kvfree(bo_buckets);
2012 	kvfree(bo_privs);
2013 	return ret;
2014 }
2015 
2016 static int criu_get_process_object_info(struct kfd_process *p,
2017 					uint32_t *num_devices,
2018 					uint32_t *num_bos,
2019 					uint32_t *num_objects,
2020 					uint64_t *objs_priv_size)
2021 {
2022 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2023 	uint32_t num_queues, num_events, num_svm_ranges;
2024 	int ret;
2025 
2026 	*num_devices = p->n_pdds;
2027 	*num_bos = get_process_num_bos(p);
2028 
2029 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2030 	if (ret)
2031 		return ret;
2032 
2033 	num_events = kfd_get_num_events(p);
2034 
2035 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2036 	if (ret)
2037 		return ret;
2038 
2039 	*num_objects = num_queues + num_events + num_svm_ranges;
2040 
2041 	if (objs_priv_size) {
2042 		priv_size = sizeof(struct kfd_criu_process_priv_data);
2043 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2044 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2045 		priv_size += queues_priv_data_size;
2046 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2047 		priv_size += svm_priv_data_size;
2048 		*objs_priv_size = priv_size;
2049 	}
2050 	return 0;
2051 }
2052 
2053 static int criu_checkpoint(struct file *filep,
2054 			   struct kfd_process *p,
2055 			   struct kfd_ioctl_criu_args *args)
2056 {
2057 	int ret;
2058 	uint32_t num_devices, num_bos, num_objects;
2059 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2060 
2061 	if (!args->devices || !args->bos || !args->priv_data)
2062 		return -EINVAL;
2063 
2064 	mutex_lock(&p->mutex);
2065 
2066 	if (!p->n_pdds) {
2067 		pr_err("No pdd for given process\n");
2068 		ret = -ENODEV;
2069 		goto exit_unlock;
2070 	}
2071 
2072 	/* Confirm all process queues are evicted */
2073 	if (!p->queues_paused) {
2074 		pr_err("Cannot dump process when queues are not in evicted state\n");
2075 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2076 		ret = -EINVAL;
2077 		goto exit_unlock;
2078 	}
2079 
2080 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2081 	if (ret)
2082 		goto exit_unlock;
2083 
2084 	if (num_devices != args->num_devices ||
2085 	    num_bos != args->num_bos ||
2086 	    num_objects != args->num_objects ||
2087 	    priv_size != args->priv_data_size) {
2088 
2089 		ret = -EINVAL;
2090 		goto exit_unlock;
2091 	}
2092 
2093 	/* each function will store private data inside priv_data and adjust priv_offset */
2094 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2095 	if (ret)
2096 		goto exit_unlock;
2097 
2098 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2099 				(uint8_t __user *)args->priv_data, &priv_offset);
2100 	if (ret)
2101 		goto exit_unlock;
2102 
2103 	/* Leave room for BOs in the private data. They need to be restored
2104 	 * before events, but we checkpoint them last to simplify the error
2105 	 * handling.
2106 	 */
2107 	bo_priv_offset = priv_offset;
2108 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2109 
2110 	if (num_objects) {
2111 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2112 						 &priv_offset);
2113 		if (ret)
2114 			goto exit_unlock;
2115 
2116 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2117 						 &priv_offset);
2118 		if (ret)
2119 			goto exit_unlock;
2120 
2121 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2122 		if (ret)
2123 			goto exit_unlock;
2124 	}
2125 
2126 	/* This must be the last thing in this function that can fail.
2127 	 * Otherwise we leak dmabuf file descriptors.
2128 	 */
2129 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2130 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2131 
2132 exit_unlock:
2133 	mutex_unlock(&p->mutex);
2134 	if (ret)
2135 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2136 	else
2137 		pr_debug("CRIU dump ret:%d\n", ret);
2138 
2139 	return ret;
2140 }
2141 
2142 static int criu_restore_process(struct kfd_process *p,
2143 				struct kfd_ioctl_criu_args *args,
2144 				uint64_t *priv_offset,
2145 				uint64_t max_priv_data_size)
2146 {
2147 	int ret = 0;
2148 	struct kfd_criu_process_priv_data process_priv;
2149 
2150 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2151 		return -EINVAL;
2152 
2153 	ret = copy_from_user(&process_priv,
2154 				(void __user *)(args->priv_data + *priv_offset),
2155 				sizeof(process_priv));
2156 	if (ret) {
2157 		pr_err("Failed to copy process private information from user\n");
2158 		ret = -EFAULT;
2159 		goto exit;
2160 	}
2161 	*priv_offset += sizeof(process_priv);
2162 
2163 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2164 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2165 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2166 		return -EINVAL;
2167 	}
2168 
2169 	pr_debug("Setting XNACK mode\n");
2170 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2171 		pr_err("xnack mode cannot be set\n");
2172 		ret = -EPERM;
2173 		goto exit;
2174 	} else {
2175 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2176 		p->xnack_enabled = process_priv.xnack_mode;
2177 	}
2178 
2179 exit:
2180 	return ret;
2181 }
2182 
2183 static int criu_restore_devices(struct kfd_process *p,
2184 				struct kfd_ioctl_criu_args *args,
2185 				uint64_t *priv_offset,
2186 				uint64_t max_priv_data_size)
2187 {
2188 	struct kfd_criu_device_bucket *device_buckets;
2189 	struct kfd_criu_device_priv_data *device_privs;
2190 	int ret = 0;
2191 	uint32_t i;
2192 
2193 	if (args->num_devices != p->n_pdds)
2194 		return -EINVAL;
2195 
2196 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2197 		return -EINVAL;
2198 
2199 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2200 	if (!device_buckets)
2201 		return -ENOMEM;
2202 
2203 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2204 				args->num_devices * sizeof(*device_buckets));
2205 	if (ret) {
2206 		pr_err("Failed to copy devices buckets from user\n");
2207 		ret = -EFAULT;
2208 		goto exit;
2209 	}
2210 
2211 	for (i = 0; i < args->num_devices; i++) {
2212 		struct kfd_node *dev;
2213 		struct kfd_process_device *pdd;
2214 		struct file *drm_file;
2215 
2216 		/* device private data is not currently used */
2217 
2218 		if (!device_buckets[i].user_gpu_id) {
2219 			pr_err("Invalid user gpu_id\n");
2220 			ret = -EINVAL;
2221 			goto exit;
2222 		}
2223 
2224 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2225 		if (!dev) {
2226 			pr_err("Failed to find device with gpu_id = %x\n",
2227 				device_buckets[i].actual_gpu_id);
2228 			ret = -EINVAL;
2229 			goto exit;
2230 		}
2231 
2232 		pdd = kfd_get_process_device_data(dev, p);
2233 		if (!pdd) {
2234 			pr_err("Failed to get pdd for gpu_id = %x\n",
2235 					device_buckets[i].actual_gpu_id);
2236 			ret = -EINVAL;
2237 			goto exit;
2238 		}
2239 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2240 
2241 		drm_file = fget(device_buckets[i].drm_fd);
2242 		if (!drm_file) {
2243 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2244 				device_buckets[i].drm_fd);
2245 			ret = -EINVAL;
2246 			goto exit;
2247 		}
2248 
2249 		if (pdd->drm_file) {
2250 			ret = -EINVAL;
2251 			goto exit;
2252 		}
2253 
2254 		/* create the vm using render nodes for kfd pdd */
2255 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2256 			pr_err("could not init vm for given pdd\n");
2257 			/* On success, the PDD keeps the drm_file reference */
2258 			fput(drm_file);
2259 			ret = -EINVAL;
2260 			goto exit;
2261 		}
2262 		/*
2263 		 * pdd now already has the vm bound to render node so below api won't create a new
2264 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2265 		 * for iommu v2 binding  and runtime pm.
2266 		 */
2267 		pdd = kfd_bind_process_to_device(dev, p);
2268 		if (IS_ERR(pdd)) {
2269 			ret = PTR_ERR(pdd);
2270 			goto exit;
2271 		}
2272 
2273 		if (!pdd->qpd.proc_doorbells) {
2274 			ret = kfd_alloc_process_doorbells(dev->kfd, pdd);
2275 			if (ret)
2276 				goto exit;
2277 		}
2278 	}
2279 
2280 	/*
2281 	 * We are not copying device private data from user as we are not using the data for now,
2282 	 * but we still adjust for its private data.
2283 	 */
2284 	*priv_offset += args->num_devices * sizeof(*device_privs);
2285 
2286 exit:
2287 	kfree(device_buckets);
2288 	return ret;
2289 }
2290 
2291 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2292 				      struct kfd_criu_bo_bucket *bo_bucket,
2293 				      struct kfd_criu_bo_priv_data *bo_priv,
2294 				      struct kgd_mem **kgd_mem)
2295 {
2296 	int idr_handle;
2297 	int ret;
2298 	const bool criu_resume = true;
2299 	u64 offset;
2300 
2301 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2302 		if (bo_bucket->size !=
2303 				kfd_doorbell_process_slice(pdd->dev->kfd))
2304 			return -EINVAL;
2305 
2306 		offset = kfd_get_process_doorbells(pdd);
2307 		if (!offset)
2308 			return -ENOMEM;
2309 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2310 		/* MMIO BOs need remapped bus address */
2311 		if (bo_bucket->size != PAGE_SIZE) {
2312 			pr_err("Invalid page size\n");
2313 			return -EINVAL;
2314 		}
2315 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2316 		if (!offset || (PAGE_SIZE > 4096)) {
2317 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2318 			return -ENOMEM;
2319 		}
2320 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2321 		offset = bo_priv->user_addr;
2322 	}
2323 	/* Create the BO */
2324 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2325 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2326 						      &offset, bo_bucket->alloc_flags, criu_resume);
2327 	if (ret) {
2328 		pr_err("Could not create the BO\n");
2329 		return ret;
2330 	}
2331 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2332 		 bo_bucket->size, bo_bucket->addr, offset);
2333 
2334 	/* Restore previous IDR handle */
2335 	pr_debug("Restoring old IDR handle for the BO");
2336 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2337 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2338 
2339 	if (idr_handle < 0) {
2340 		pr_err("Could not allocate idr\n");
2341 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2342 						       NULL);
2343 		return -ENOMEM;
2344 	}
2345 
2346 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2347 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2348 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2349 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2350 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2351 		bo_bucket->restored_offset = offset;
2352 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2353 		bo_bucket->restored_offset = offset;
2354 		/* Update the VRAM usage count */
2355 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2356 	}
2357 	return 0;
2358 }
2359 
2360 static int criu_restore_bo(struct kfd_process *p,
2361 			   struct kfd_criu_bo_bucket *bo_bucket,
2362 			   struct kfd_criu_bo_priv_data *bo_priv)
2363 {
2364 	struct kfd_process_device *pdd;
2365 	struct kgd_mem *kgd_mem;
2366 	int ret;
2367 	int j;
2368 
2369 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2370 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2371 		 bo_priv->idr_handle);
2372 
2373 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2374 	if (!pdd) {
2375 		pr_err("Failed to get pdd\n");
2376 		return -ENODEV;
2377 	}
2378 
2379 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2380 	if (ret)
2381 		return ret;
2382 
2383 	/* now map these BOs to GPU/s */
2384 	for (j = 0; j < p->n_pdds; j++) {
2385 		struct kfd_node *peer;
2386 		struct kfd_process_device *peer_pdd;
2387 
2388 		if (!bo_priv->mapped_gpuids[j])
2389 			break;
2390 
2391 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2392 		if (!peer_pdd)
2393 			return -EINVAL;
2394 
2395 		peer = peer_pdd->dev;
2396 
2397 		peer_pdd = kfd_bind_process_to_device(peer, p);
2398 		if (IS_ERR(peer_pdd))
2399 			return PTR_ERR(peer_pdd);
2400 
2401 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2402 							    peer_pdd->drm_priv);
2403 		if (ret) {
2404 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2405 			return ret;
2406 		}
2407 	}
2408 
2409 	pr_debug("map memory was successful for the BO\n");
2410 	/* create the dmabuf object and export the bo */
2411 	if (bo_bucket->alloc_flags
2412 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2413 		ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2414 					    &bo_bucket->dmabuf_fd);
2415 		if (ret)
2416 			return ret;
2417 	} else {
2418 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2419 	}
2420 
2421 	return 0;
2422 }
2423 
2424 static int criu_restore_bos(struct kfd_process *p,
2425 			    struct kfd_ioctl_criu_args *args,
2426 			    uint64_t *priv_offset,
2427 			    uint64_t max_priv_data_size)
2428 {
2429 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2430 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2431 	int ret = 0;
2432 	uint32_t i = 0;
2433 
2434 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2435 		return -EINVAL;
2436 
2437 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2438 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2439 
2440 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2441 	if (!bo_buckets)
2442 		return -ENOMEM;
2443 
2444 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2445 			     args->num_bos * sizeof(*bo_buckets));
2446 	if (ret) {
2447 		pr_err("Failed to copy BOs information from user\n");
2448 		ret = -EFAULT;
2449 		goto exit;
2450 	}
2451 
2452 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2453 	if (!bo_privs) {
2454 		ret = -ENOMEM;
2455 		goto exit;
2456 	}
2457 
2458 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2459 			     args->num_bos * sizeof(*bo_privs));
2460 	if (ret) {
2461 		pr_err("Failed to copy BOs information from user\n");
2462 		ret = -EFAULT;
2463 		goto exit;
2464 	}
2465 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2466 
2467 	/* Create and map new BOs */
2468 	for (; i < args->num_bos; i++) {
2469 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2470 		if (ret) {
2471 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2472 			goto exit;
2473 		}
2474 	} /* done */
2475 
2476 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2477 	ret = copy_to_user((void __user *)args->bos,
2478 				bo_buckets,
2479 				(args->num_bos * sizeof(*bo_buckets)));
2480 	if (ret)
2481 		ret = -EFAULT;
2482 
2483 exit:
2484 	while (ret && i--) {
2485 		if (bo_buckets[i].alloc_flags
2486 		   & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2487 			close_fd(bo_buckets[i].dmabuf_fd);
2488 	}
2489 	kvfree(bo_buckets);
2490 	kvfree(bo_privs);
2491 	return ret;
2492 }
2493 
2494 static int criu_restore_objects(struct file *filep,
2495 				struct kfd_process *p,
2496 				struct kfd_ioctl_criu_args *args,
2497 				uint64_t *priv_offset,
2498 				uint64_t max_priv_data_size)
2499 {
2500 	int ret = 0;
2501 	uint32_t i;
2502 
2503 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2504 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2505 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2506 
2507 	for (i = 0; i < args->num_objects; i++) {
2508 		uint32_t object_type;
2509 
2510 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2511 			pr_err("Invalid private data size\n");
2512 			return -EINVAL;
2513 		}
2514 
2515 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2516 		if (ret) {
2517 			pr_err("Failed to copy private information from user\n");
2518 			goto exit;
2519 		}
2520 
2521 		switch (object_type) {
2522 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2523 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2524 						     priv_offset, max_priv_data_size);
2525 			if (ret)
2526 				goto exit;
2527 			break;
2528 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2529 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2530 						     priv_offset, max_priv_data_size);
2531 			if (ret)
2532 				goto exit;
2533 			break;
2534 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2535 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2536 						     priv_offset, max_priv_data_size);
2537 			if (ret)
2538 				goto exit;
2539 			break;
2540 		default:
2541 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2542 			ret = -EINVAL;
2543 			goto exit;
2544 		}
2545 	}
2546 exit:
2547 	return ret;
2548 }
2549 
2550 static int criu_restore(struct file *filep,
2551 			struct kfd_process *p,
2552 			struct kfd_ioctl_criu_args *args)
2553 {
2554 	uint64_t priv_offset = 0;
2555 	int ret = 0;
2556 
2557 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2558 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2559 
2560 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2561 	    !args->num_devices || !args->num_bos)
2562 		return -EINVAL;
2563 
2564 	mutex_lock(&p->mutex);
2565 
2566 	/*
2567 	 * Set the process to evicted state to avoid running any new queues before all the memory
2568 	 * mappings are ready.
2569 	 */
2570 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2571 	if (ret)
2572 		goto exit_unlock;
2573 
2574 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2575 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2576 	if (ret)
2577 		goto exit_unlock;
2578 
2579 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2580 	if (ret)
2581 		goto exit_unlock;
2582 
2583 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2584 	if (ret)
2585 		goto exit_unlock;
2586 
2587 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2588 	if (ret)
2589 		goto exit_unlock;
2590 
2591 	if (priv_offset != args->priv_data_size) {
2592 		pr_err("Invalid private data size\n");
2593 		ret = -EINVAL;
2594 	}
2595 
2596 exit_unlock:
2597 	mutex_unlock(&p->mutex);
2598 	if (ret)
2599 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2600 	else
2601 		pr_debug("CRIU restore successful\n");
2602 
2603 	return ret;
2604 }
2605 
2606 static int criu_unpause(struct file *filep,
2607 			struct kfd_process *p,
2608 			struct kfd_ioctl_criu_args *args)
2609 {
2610 	int ret;
2611 
2612 	mutex_lock(&p->mutex);
2613 
2614 	if (!p->queues_paused) {
2615 		mutex_unlock(&p->mutex);
2616 		return -EINVAL;
2617 	}
2618 
2619 	ret = kfd_process_restore_queues(p);
2620 	if (ret)
2621 		pr_err("Failed to unpause queues ret:%d\n", ret);
2622 	else
2623 		p->queues_paused = false;
2624 
2625 	mutex_unlock(&p->mutex);
2626 
2627 	return ret;
2628 }
2629 
2630 static int criu_resume(struct file *filep,
2631 			struct kfd_process *p,
2632 			struct kfd_ioctl_criu_args *args)
2633 {
2634 	struct kfd_process *target = NULL;
2635 	struct pid *pid = NULL;
2636 	int ret = 0;
2637 
2638 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2639 		 args->pid);
2640 
2641 	pid = find_get_pid(args->pid);
2642 	if (!pid) {
2643 		pr_err("Cannot find pid info for %i\n", args->pid);
2644 		return -ESRCH;
2645 	}
2646 
2647 	pr_debug("calling kfd_lookup_process_by_pid\n");
2648 	target = kfd_lookup_process_by_pid(pid);
2649 
2650 	put_pid(pid);
2651 
2652 	if (!target) {
2653 		pr_debug("Cannot find process info for %i\n", args->pid);
2654 		return -ESRCH;
2655 	}
2656 
2657 	mutex_lock(&target->mutex);
2658 	ret = kfd_criu_resume_svm(target);
2659 	if (ret) {
2660 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2661 		goto exit;
2662 	}
2663 
2664 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2665 	if (ret)
2666 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2667 
2668 exit:
2669 	mutex_unlock(&target->mutex);
2670 
2671 	kfd_unref_process(target);
2672 	return ret;
2673 }
2674 
2675 static int criu_process_info(struct file *filep,
2676 				struct kfd_process *p,
2677 				struct kfd_ioctl_criu_args *args)
2678 {
2679 	int ret = 0;
2680 
2681 	mutex_lock(&p->mutex);
2682 
2683 	if (!p->n_pdds) {
2684 		pr_err("No pdd for given process\n");
2685 		ret = -ENODEV;
2686 		goto err_unlock;
2687 	}
2688 
2689 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2690 	if (ret)
2691 		goto err_unlock;
2692 
2693 	p->queues_paused = true;
2694 
2695 	args->pid = task_pid_nr_ns(p->lead_thread,
2696 					task_active_pid_ns(p->lead_thread));
2697 
2698 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2699 					   &args->num_objects, &args->priv_data_size);
2700 	if (ret)
2701 		goto err_unlock;
2702 
2703 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2704 				args->num_devices, args->num_bos, args->num_objects,
2705 				args->priv_data_size);
2706 
2707 err_unlock:
2708 	if (ret) {
2709 		kfd_process_restore_queues(p);
2710 		p->queues_paused = false;
2711 	}
2712 	mutex_unlock(&p->mutex);
2713 	return ret;
2714 }
2715 
2716 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2717 {
2718 	struct kfd_ioctl_criu_args *args = data;
2719 	int ret;
2720 
2721 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2722 	switch (args->op) {
2723 	case KFD_CRIU_OP_PROCESS_INFO:
2724 		ret = criu_process_info(filep, p, args);
2725 		break;
2726 	case KFD_CRIU_OP_CHECKPOINT:
2727 		ret = criu_checkpoint(filep, p, args);
2728 		break;
2729 	case KFD_CRIU_OP_UNPAUSE:
2730 		ret = criu_unpause(filep, p, args);
2731 		break;
2732 	case KFD_CRIU_OP_RESTORE:
2733 		ret = criu_restore(filep, p, args);
2734 		break;
2735 	case KFD_CRIU_OP_RESUME:
2736 		ret = criu_resume(filep, p, args);
2737 		break;
2738 	default:
2739 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2740 		ret = -EINVAL;
2741 		break;
2742 	}
2743 
2744 	if (ret)
2745 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2746 
2747 	return ret;
2748 }
2749 
2750 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2751 			bool enable_ttmp_setup)
2752 {
2753 	int i = 0, ret = 0;
2754 
2755 	if (p->is_runtime_retry)
2756 		goto retry;
2757 
2758 	if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2759 		return -EBUSY;
2760 
2761 	for (i = 0; i < p->n_pdds; i++) {
2762 		struct kfd_process_device *pdd = p->pdds[i];
2763 
2764 		if (pdd->qpd.queue_count)
2765 			return -EEXIST;
2766 
2767 		/*
2768 		 * Setup TTMPs by default.
2769 		 * Note that this call must remain here for MES ADD QUEUE to
2770 		 * skip_process_ctx_clear unconditionally as the first call to
2771 		 * SET_SHADER_DEBUGGER clears any stale process context data
2772 		 * saved in MES.
2773 		 */
2774 		if (pdd->dev->kfd->shared_resources.enable_mes)
2775 			kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
2776 	}
2777 
2778 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2779 	p->runtime_info.r_debug = r_debug;
2780 	p->runtime_info.ttmp_setup = enable_ttmp_setup;
2781 
2782 	if (p->runtime_info.ttmp_setup) {
2783 		for (i = 0; i < p->n_pdds; i++) {
2784 			struct kfd_process_device *pdd = p->pdds[i];
2785 
2786 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2787 				amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2788 				pdd->dev->kfd2kgd->enable_debug_trap(
2789 						pdd->dev->adev,
2790 						true,
2791 						pdd->dev->vm_info.last_vmid_kfd);
2792 			} else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2793 				pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2794 						pdd->dev->adev,
2795 						false,
2796 						0);
2797 			}
2798 		}
2799 	}
2800 
2801 retry:
2802 	if (p->debug_trap_enabled) {
2803 		if (!p->is_runtime_retry) {
2804 			kfd_dbg_trap_activate(p);
2805 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2806 					p, NULL, 0, false, NULL, 0);
2807 		}
2808 
2809 		mutex_unlock(&p->mutex);
2810 		ret = down_interruptible(&p->runtime_enable_sema);
2811 		mutex_lock(&p->mutex);
2812 
2813 		p->is_runtime_retry = !!ret;
2814 	}
2815 
2816 	return ret;
2817 }
2818 
2819 static int runtime_disable(struct kfd_process *p)
2820 {
2821 	int i = 0, ret;
2822 	bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2823 
2824 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2825 	p->runtime_info.r_debug = 0;
2826 
2827 	if (p->debug_trap_enabled) {
2828 		if (was_enabled)
2829 			kfd_dbg_trap_deactivate(p, false, 0);
2830 
2831 		if (!p->is_runtime_retry)
2832 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2833 					p, NULL, 0, false, NULL, 0);
2834 
2835 		mutex_unlock(&p->mutex);
2836 		ret = down_interruptible(&p->runtime_enable_sema);
2837 		mutex_lock(&p->mutex);
2838 
2839 		p->is_runtime_retry = !!ret;
2840 		if (ret)
2841 			return ret;
2842 	}
2843 
2844 	if (was_enabled && p->runtime_info.ttmp_setup) {
2845 		for (i = 0; i < p->n_pdds; i++) {
2846 			struct kfd_process_device *pdd = p->pdds[i];
2847 
2848 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2849 				amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2850 		}
2851 	}
2852 
2853 	p->runtime_info.ttmp_setup = false;
2854 
2855 	/* disable ttmp setup */
2856 	for (i = 0; i < p->n_pdds; i++) {
2857 		struct kfd_process_device *pdd = p->pdds[i];
2858 
2859 		if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2860 			pdd->spi_dbg_override =
2861 					pdd->dev->kfd2kgd->disable_debug_trap(
2862 					pdd->dev->adev,
2863 					false,
2864 					pdd->dev->vm_info.last_vmid_kfd);
2865 
2866 			if (!pdd->dev->kfd->shared_resources.enable_mes)
2867 				debug_refresh_runlist(pdd->dev->dqm);
2868 			else
2869 				kfd_dbg_set_mes_debug_mode(pdd,
2870 							   !kfd_dbg_has_cwsr_workaround(pdd->dev));
2871 		}
2872 	}
2873 
2874 	return 0;
2875 }
2876 
2877 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2878 {
2879 	struct kfd_ioctl_runtime_enable_args *args = data;
2880 	int r;
2881 
2882 	mutex_lock(&p->mutex);
2883 
2884 	if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2885 		r = runtime_enable(p, args->r_debug,
2886 				!!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2887 	else
2888 		r = runtime_disable(p);
2889 
2890 	mutex_unlock(&p->mutex);
2891 
2892 	return r;
2893 }
2894 
2895 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2896 {
2897 	struct kfd_ioctl_dbg_trap_args *args = data;
2898 	struct task_struct *thread = NULL;
2899 	struct mm_struct *mm = NULL;
2900 	struct pid *pid = NULL;
2901 	struct kfd_process *target = NULL;
2902 	struct kfd_process_device *pdd = NULL;
2903 	int r = 0;
2904 
2905 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2906 		pr_err("Debugging does not support sched_policy %i", sched_policy);
2907 		return -EINVAL;
2908 	}
2909 
2910 	pid = find_get_pid(args->pid);
2911 	if (!pid) {
2912 		pr_debug("Cannot find pid info for %i\n", args->pid);
2913 		r = -ESRCH;
2914 		goto out;
2915 	}
2916 
2917 	thread = get_pid_task(pid, PIDTYPE_PID);
2918 	if (!thread) {
2919 		r = -ESRCH;
2920 		goto out;
2921 	}
2922 
2923 	mm = get_task_mm(thread);
2924 	if (!mm) {
2925 		r = -ESRCH;
2926 		goto out;
2927 	}
2928 
2929 	if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2930 		bool create_process;
2931 
2932 		rcu_read_lock();
2933 		create_process = thread && thread != current && ptrace_parent(thread) == current;
2934 		rcu_read_unlock();
2935 
2936 		target = create_process ? kfd_create_process(thread) :
2937 					kfd_lookup_process_by_pid(pid);
2938 	} else {
2939 		target = kfd_lookup_process_by_pid(pid);
2940 	}
2941 
2942 	if (IS_ERR_OR_NULL(target)) {
2943 		pr_debug("Cannot find process PID %i to debug\n", args->pid);
2944 		r = target ? PTR_ERR(target) : -ESRCH;
2945 		goto out;
2946 	}
2947 
2948 	/* Check if target is still PTRACED. */
2949 	rcu_read_lock();
2950 	if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2951 				&& ptrace_parent(target->lead_thread) != current) {
2952 		pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2953 		r = -EPERM;
2954 	}
2955 	rcu_read_unlock();
2956 
2957 	if (r)
2958 		goto out;
2959 
2960 	mutex_lock(&target->mutex);
2961 
2962 	if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2963 		pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2964 		r = -EINVAL;
2965 		goto unlock_out;
2966 	}
2967 
2968 	if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2969 			(args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2970 			 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2971 			 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2972 			 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2973 			 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2974 			 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2975 			 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2976 		r = -EPERM;
2977 		goto unlock_out;
2978 	}
2979 
2980 	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2981 	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2982 		int user_gpu_id = kfd_process_get_user_gpu_id(target,
2983 				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2984 					args->set_node_address_watch.gpu_id :
2985 					args->clear_node_address_watch.gpu_id);
2986 
2987 		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
2988 		if (user_gpu_id == -EINVAL || !pdd) {
2989 			r = -ENODEV;
2990 			goto unlock_out;
2991 		}
2992 	}
2993 
2994 	switch (args->op) {
2995 	case KFD_IOC_DBG_TRAP_ENABLE:
2996 		if (target != p)
2997 			target->debugger_process = p;
2998 
2999 		r = kfd_dbg_trap_enable(target,
3000 					args->enable.dbg_fd,
3001 					(void __user *)args->enable.rinfo_ptr,
3002 					&args->enable.rinfo_size);
3003 		if (!r)
3004 			target->exception_enable_mask = args->enable.exception_mask;
3005 
3006 		break;
3007 	case KFD_IOC_DBG_TRAP_DISABLE:
3008 		r = kfd_dbg_trap_disable(target);
3009 		break;
3010 	case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
3011 		r = kfd_dbg_send_exception_to_runtime(target,
3012 				args->send_runtime_event.gpu_id,
3013 				args->send_runtime_event.queue_id,
3014 				args->send_runtime_event.exception_mask);
3015 		break;
3016 	case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
3017 		kfd_dbg_set_enabled_debug_exception_mask(target,
3018 				args->set_exceptions_enabled.exception_mask);
3019 		break;
3020 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
3021 		r = kfd_dbg_trap_set_wave_launch_override(target,
3022 				args->launch_override.override_mode,
3023 				args->launch_override.enable_mask,
3024 				args->launch_override.support_request_mask,
3025 				&args->launch_override.enable_mask,
3026 				&args->launch_override.support_request_mask);
3027 		break;
3028 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
3029 		r = kfd_dbg_trap_set_wave_launch_mode(target,
3030 				args->launch_mode.launch_mode);
3031 		break;
3032 	case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
3033 		r = suspend_queues(target,
3034 				args->suspend_queues.num_queues,
3035 				args->suspend_queues.grace_period,
3036 				args->suspend_queues.exception_mask,
3037 				(uint32_t *)args->suspend_queues.queue_array_ptr);
3038 
3039 		break;
3040 	case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
3041 		r = resume_queues(target, args->resume_queues.num_queues,
3042 				(uint32_t *)args->resume_queues.queue_array_ptr);
3043 		break;
3044 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3045 		r = kfd_dbg_trap_set_dev_address_watch(pdd,
3046 				args->set_node_address_watch.address,
3047 				args->set_node_address_watch.mask,
3048 				&args->set_node_address_watch.id,
3049 				args->set_node_address_watch.mode);
3050 		break;
3051 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3052 		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3053 				args->clear_node_address_watch.id);
3054 		break;
3055 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
3056 		r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3057 		break;
3058 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3059 		r = kfd_dbg_ev_query_debug_event(target,
3060 				&args->query_debug_event.queue_id,
3061 				&args->query_debug_event.gpu_id,
3062 				args->query_debug_event.exception_mask,
3063 				&args->query_debug_event.exception_mask);
3064 		break;
3065 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3066 		r = kfd_dbg_trap_query_exception_info(target,
3067 				args->query_exception_info.source_id,
3068 				args->query_exception_info.exception_code,
3069 				args->query_exception_info.clear_exception,
3070 				(void __user *)args->query_exception_info.info_ptr,
3071 				&args->query_exception_info.info_size);
3072 		break;
3073 	case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3074 		r = pqm_get_queue_snapshot(&target->pqm,
3075 				args->queue_snapshot.exception_mask,
3076 				(void __user *)args->queue_snapshot.snapshot_buf_ptr,
3077 				&args->queue_snapshot.num_queues,
3078 				&args->queue_snapshot.entry_size);
3079 		break;
3080 	case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3081 		r = kfd_dbg_trap_device_snapshot(target,
3082 				args->device_snapshot.exception_mask,
3083 				(void __user *)args->device_snapshot.snapshot_buf_ptr,
3084 				&args->device_snapshot.num_devices,
3085 				&args->device_snapshot.entry_size);
3086 		break;
3087 	default:
3088 		pr_err("Invalid option: %i\n", args->op);
3089 		r = -EINVAL;
3090 	}
3091 
3092 unlock_out:
3093 	mutex_unlock(&target->mutex);
3094 
3095 out:
3096 	if (thread)
3097 		put_task_struct(thread);
3098 
3099 	if (mm)
3100 		mmput(mm);
3101 
3102 	if (pid)
3103 		put_pid(pid);
3104 
3105 	if (target)
3106 		kfd_unref_process(target);
3107 
3108 	return r;
3109 }
3110 
3111 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3112 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3113 			    .cmd_drv = 0, .name = #ioctl}
3114 
3115 /** Ioctl table */
3116 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3117 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3118 			kfd_ioctl_get_version, 0),
3119 
3120 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3121 			kfd_ioctl_create_queue, 0),
3122 
3123 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3124 			kfd_ioctl_destroy_queue, 0),
3125 
3126 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3127 			kfd_ioctl_set_memory_policy, 0),
3128 
3129 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3130 			kfd_ioctl_get_clock_counters, 0),
3131 
3132 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3133 			kfd_ioctl_get_process_apertures, 0),
3134 
3135 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3136 			kfd_ioctl_update_queue, 0),
3137 
3138 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3139 			kfd_ioctl_create_event, 0),
3140 
3141 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3142 			kfd_ioctl_destroy_event, 0),
3143 
3144 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3145 			kfd_ioctl_set_event, 0),
3146 
3147 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3148 			kfd_ioctl_reset_event, 0),
3149 
3150 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3151 			kfd_ioctl_wait_events, 0),
3152 
3153 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3154 			kfd_ioctl_dbg_register, 0),
3155 
3156 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3157 			kfd_ioctl_dbg_unregister, 0),
3158 
3159 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3160 			kfd_ioctl_dbg_address_watch, 0),
3161 
3162 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3163 			kfd_ioctl_dbg_wave_control, 0),
3164 
3165 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3166 			kfd_ioctl_set_scratch_backing_va, 0),
3167 
3168 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3169 			kfd_ioctl_get_tile_config, 0),
3170 
3171 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3172 			kfd_ioctl_set_trap_handler, 0),
3173 
3174 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3175 			kfd_ioctl_get_process_apertures_new, 0),
3176 
3177 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3178 			kfd_ioctl_acquire_vm, 0),
3179 
3180 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3181 			kfd_ioctl_alloc_memory_of_gpu, 0),
3182 
3183 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3184 			kfd_ioctl_free_memory_of_gpu, 0),
3185 
3186 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3187 			kfd_ioctl_map_memory_to_gpu, 0),
3188 
3189 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3190 			kfd_ioctl_unmap_memory_from_gpu, 0),
3191 
3192 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3193 			kfd_ioctl_set_cu_mask, 0),
3194 
3195 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3196 			kfd_ioctl_get_queue_wave_state, 0),
3197 
3198 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3199 				kfd_ioctl_get_dmabuf_info, 0),
3200 
3201 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3202 				kfd_ioctl_import_dmabuf, 0),
3203 
3204 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3205 			kfd_ioctl_alloc_queue_gws, 0),
3206 
3207 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3208 			kfd_ioctl_smi_events, 0),
3209 
3210 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3211 
3212 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3213 			kfd_ioctl_set_xnack_mode, 0),
3214 
3215 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3216 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3217 
3218 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3219 			kfd_ioctl_get_available_memory, 0),
3220 
3221 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3222 				kfd_ioctl_export_dmabuf, 0),
3223 
3224 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3225 			kfd_ioctl_runtime_enable, 0),
3226 
3227 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3228 			kfd_ioctl_set_debug_trap, 0),
3229 };
3230 
3231 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
3232 
3233 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3234 {
3235 	struct kfd_process *process;
3236 	amdkfd_ioctl_t *func;
3237 	const struct amdkfd_ioctl_desc *ioctl = NULL;
3238 	unsigned int nr = _IOC_NR(cmd);
3239 	char stack_kdata[128];
3240 	char *kdata = NULL;
3241 	unsigned int usize, asize;
3242 	int retcode = -EINVAL;
3243 	bool ptrace_attached = false;
3244 
3245 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3246 		goto err_i1;
3247 
3248 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3249 		u32 amdkfd_size;
3250 
3251 		ioctl = &amdkfd_ioctls[nr];
3252 
3253 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
3254 		usize = asize = _IOC_SIZE(cmd);
3255 		if (amdkfd_size > asize)
3256 			asize = amdkfd_size;
3257 
3258 		cmd = ioctl->cmd;
3259 	} else
3260 		goto err_i1;
3261 
3262 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3263 
3264 	/* Get the process struct from the filep. Only the process
3265 	 * that opened /dev/kfd can use the file descriptor. Child
3266 	 * processes need to create their own KFD device context.
3267 	 */
3268 	process = filep->private_data;
3269 
3270 	rcu_read_lock();
3271 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3272 	    ptrace_parent(process->lead_thread) == current)
3273 		ptrace_attached = true;
3274 	rcu_read_unlock();
3275 
3276 	if (process->lead_thread != current->group_leader
3277 	    && !ptrace_attached) {
3278 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3279 		retcode = -EBADF;
3280 		goto err_i1;
3281 	}
3282 
3283 	/* Do not trust userspace, use our own definition */
3284 	func = ioctl->func;
3285 
3286 	if (unlikely(!func)) {
3287 		dev_dbg(kfd_device, "no function\n");
3288 		retcode = -EINVAL;
3289 		goto err_i1;
3290 	}
3291 
3292 	/*
3293 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3294 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3295 	 * more priviledged access.
3296 	 */
3297 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3298 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
3299 						!capable(CAP_SYS_ADMIN)) {
3300 			retcode = -EACCES;
3301 			goto err_i1;
3302 		}
3303 	}
3304 
3305 	if (cmd & (IOC_IN | IOC_OUT)) {
3306 		if (asize <= sizeof(stack_kdata)) {
3307 			kdata = stack_kdata;
3308 		} else {
3309 			kdata = kmalloc(asize, GFP_KERNEL);
3310 			if (!kdata) {
3311 				retcode = -ENOMEM;
3312 				goto err_i1;
3313 			}
3314 		}
3315 		if (asize > usize)
3316 			memset(kdata + usize, 0, asize - usize);
3317 	}
3318 
3319 	if (cmd & IOC_IN) {
3320 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3321 			retcode = -EFAULT;
3322 			goto err_i1;
3323 		}
3324 	} else if (cmd & IOC_OUT) {
3325 		memset(kdata, 0, usize);
3326 	}
3327 
3328 	retcode = func(filep, process, kdata);
3329 
3330 	if (cmd & IOC_OUT)
3331 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3332 			retcode = -EFAULT;
3333 
3334 err_i1:
3335 	if (!ioctl)
3336 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3337 			  task_pid_nr(current), cmd, nr);
3338 
3339 	if (kdata != stack_kdata)
3340 		kfree(kdata);
3341 
3342 	if (retcode)
3343 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3344 				nr, arg, retcode);
3345 
3346 	return retcode;
3347 }
3348 
3349 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3350 		      struct vm_area_struct *vma)
3351 {
3352 	phys_addr_t address;
3353 
3354 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3355 		return -EINVAL;
3356 
3357 	if (PAGE_SIZE > 4096)
3358 		return -EINVAL;
3359 
3360 	address = dev->adev->rmmio_remap.bus_addr;
3361 
3362 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3363 				VM_DONTDUMP | VM_PFNMAP);
3364 
3365 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3366 
3367 	pr_debug("pasid 0x%x mapping mmio page\n"
3368 		 "     target user address == 0x%08llX\n"
3369 		 "     physical address    == 0x%08llX\n"
3370 		 "     vm_flags            == 0x%04lX\n"
3371 		 "     size                == 0x%04lX\n",
3372 		 process->pasid, (unsigned long long) vma->vm_start,
3373 		 address, vma->vm_flags, PAGE_SIZE);
3374 
3375 	return io_remap_pfn_range(vma,
3376 				vma->vm_start,
3377 				address >> PAGE_SHIFT,
3378 				PAGE_SIZE,
3379 				vma->vm_page_prot);
3380 }
3381 
3382 
3383 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3384 {
3385 	struct kfd_process *process;
3386 	struct kfd_node *dev = NULL;
3387 	unsigned long mmap_offset;
3388 	unsigned int gpu_id;
3389 
3390 	process = kfd_get_process(current);
3391 	if (IS_ERR(process))
3392 		return PTR_ERR(process);
3393 
3394 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3395 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3396 	if (gpu_id)
3397 		dev = kfd_device_by_id(gpu_id);
3398 
3399 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3400 	case KFD_MMAP_TYPE_DOORBELL:
3401 		if (!dev)
3402 			return -ENODEV;
3403 		return kfd_doorbell_mmap(dev, process, vma);
3404 
3405 	case KFD_MMAP_TYPE_EVENTS:
3406 		return kfd_event_mmap(process, vma);
3407 
3408 	case KFD_MMAP_TYPE_RESERVED_MEM:
3409 		if (!dev)
3410 			return -ENODEV;
3411 		return kfd_reserved_mem_mmap(dev, process, vma);
3412 	case KFD_MMAP_TYPE_MMIO:
3413 		if (!dev)
3414 			return -ENODEV;
3415 		return kfd_mmio_mmap(dev, process, vma);
3416 	}
3417 
3418 	return -EFAULT;
3419 }
3420