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->doorbell_index &&
337 	    kfd_alloc_process_doorbells(dev->kfd, &pdd->doorbell_index) < 0) {
338 		err = -ENOMEM;
339 		goto err_alloc_doorbells;
340 	}
341 
342 	/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
343 	 * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
344 	 */
345 	if (dev->kfd->shared_resources.enable_mes &&
346 			((dev->adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK)
347 			>> AMDGPU_MES_API_VERSION_SHIFT) >= 2) {
348 		struct amdgpu_bo_va_mapping *wptr_mapping;
349 		struct amdgpu_vm *wptr_vm;
350 
351 		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
352 		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
353 		if (err)
354 			goto err_wptr_map_gart;
355 
356 		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
357 				wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
358 		amdgpu_bo_unreserve(wptr_vm->root.bo);
359 		if (!wptr_mapping) {
360 			pr_err("Failed to lookup wptr bo\n");
361 			err = -EINVAL;
362 			goto err_wptr_map_gart;
363 		}
364 
365 		wptr_bo = wptr_mapping->bo_va->base.bo;
366 		if (wptr_bo->tbo.base.size > PAGE_SIZE) {
367 			pr_err("Requested GART mapping for wptr bo larger than one page\n");
368 			err = -EINVAL;
369 			goto err_wptr_map_gart;
370 		}
371 
372 		err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
373 		if (err) {
374 			pr_err("Failed to map wptr bo to GART\n");
375 			goto err_wptr_map_gart;
376 		}
377 	}
378 
379 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
380 			p->pasid,
381 			dev->id);
382 
383 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
384 			NULL, NULL, NULL, &doorbell_offset_in_process);
385 	if (err != 0)
386 		goto err_create_queue;
387 
388 	args->queue_id = queue_id;
389 
390 
391 	/* Return gpu_id as doorbell offset for mmap usage */
392 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
393 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
394 	if (KFD_IS_SOC15(dev))
395 		/* On SOC15 ASICs, include the doorbell offset within the
396 		 * process doorbell frame, which is 2 pages.
397 		 */
398 		args->doorbell_offset |= doorbell_offset_in_process;
399 
400 	mutex_unlock(&p->mutex);
401 
402 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
403 
404 	pr_debug("Ring buffer address == 0x%016llX\n",
405 			args->ring_base_address);
406 
407 	pr_debug("Read ptr address    == 0x%016llX\n",
408 			args->read_pointer_address);
409 
410 	pr_debug("Write ptr address   == 0x%016llX\n",
411 			args->write_pointer_address);
412 
413 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
414 	return 0;
415 
416 err_create_queue:
417 	if (wptr_bo)
418 		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
419 err_wptr_map_gart:
420 err_alloc_doorbells:
421 err_bind_process:
422 err_pdd:
423 	mutex_unlock(&p->mutex);
424 	return err;
425 }
426 
427 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
428 					void *data)
429 {
430 	int retval;
431 	struct kfd_ioctl_destroy_queue_args *args = data;
432 
433 	pr_debug("Destroying queue id %d for pasid 0x%x\n",
434 				args->queue_id,
435 				p->pasid);
436 
437 	mutex_lock(&p->mutex);
438 
439 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
440 
441 	mutex_unlock(&p->mutex);
442 	return retval;
443 }
444 
445 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
446 					void *data)
447 {
448 	int retval;
449 	struct kfd_ioctl_update_queue_args *args = data;
450 	struct queue_properties properties;
451 
452 	/*
453 	 * Repurpose queue percentage to accommodate new features:
454 	 * bit 0-7: queue percentage
455 	 * bit 8-15: pm4_target_xcc
456 	 */
457 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
458 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
459 		return -EINVAL;
460 	}
461 
462 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
463 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
464 		return -EINVAL;
465 	}
466 
467 	if ((args->ring_base_address) &&
468 		(!access_ok((const void __user *) args->ring_base_address,
469 			sizeof(uint64_t)))) {
470 		pr_err("Can't access ring base address\n");
471 		return -EFAULT;
472 	}
473 
474 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
475 		pr_err("Ring size must be a power of 2 or 0\n");
476 		return -EINVAL;
477 	}
478 
479 	properties.queue_address = args->ring_base_address;
480 	properties.queue_size = args->ring_size;
481 	properties.queue_percent = args->queue_percentage & 0xFF;
482 	/* bit 8-15 are repurposed to be PM4 target XCC */
483 	properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
484 	properties.priority = args->queue_priority;
485 
486 	pr_debug("Updating queue id %d for pasid 0x%x\n",
487 			args->queue_id, p->pasid);
488 
489 	mutex_lock(&p->mutex);
490 
491 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
492 
493 	mutex_unlock(&p->mutex);
494 
495 	return retval;
496 }
497 
498 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
499 					void *data)
500 {
501 	int retval;
502 	const int max_num_cus = 1024;
503 	struct kfd_ioctl_set_cu_mask_args *args = data;
504 	struct mqd_update_info minfo = {0};
505 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
506 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
507 
508 	if ((args->num_cu_mask % 32) != 0) {
509 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
510 				args->num_cu_mask);
511 		return -EINVAL;
512 	}
513 
514 	minfo.cu_mask.count = args->num_cu_mask;
515 	if (minfo.cu_mask.count == 0) {
516 		pr_debug("CU mask cannot be 0");
517 		return -EINVAL;
518 	}
519 
520 	/* To prevent an unreasonably large CU mask size, set an arbitrary
521 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
522 	 * past max_num_cus bits and just use the first max_num_cus bits.
523 	 */
524 	if (minfo.cu_mask.count > max_num_cus) {
525 		pr_debug("CU mask cannot be greater than 1024 bits");
526 		minfo.cu_mask.count = max_num_cus;
527 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
528 	}
529 
530 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
531 	if (!minfo.cu_mask.ptr)
532 		return -ENOMEM;
533 
534 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
535 	if (retval) {
536 		pr_debug("Could not copy CU mask from userspace");
537 		retval = -EFAULT;
538 		goto out;
539 	}
540 
541 	mutex_lock(&p->mutex);
542 
543 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
544 
545 	mutex_unlock(&p->mutex);
546 
547 out:
548 	kfree(minfo.cu_mask.ptr);
549 	return retval;
550 }
551 
552 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
553 					  struct kfd_process *p, void *data)
554 {
555 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
556 	int r;
557 
558 	mutex_lock(&p->mutex);
559 
560 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
561 			       (void __user *)args->ctl_stack_address,
562 			       &args->ctl_stack_used_size,
563 			       &args->save_area_used_size);
564 
565 	mutex_unlock(&p->mutex);
566 
567 	return r;
568 }
569 
570 static int kfd_ioctl_set_memory_policy(struct file *filep,
571 					struct kfd_process *p, void *data)
572 {
573 	struct kfd_ioctl_set_memory_policy_args *args = data;
574 	int err = 0;
575 	struct kfd_process_device *pdd;
576 	enum cache_policy default_policy, alternate_policy;
577 
578 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
579 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
580 		return -EINVAL;
581 	}
582 
583 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
584 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
585 		return -EINVAL;
586 	}
587 
588 	mutex_lock(&p->mutex);
589 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
590 	if (!pdd) {
591 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
592 		err = -EINVAL;
593 		goto err_pdd;
594 	}
595 
596 	pdd = kfd_bind_process_to_device(pdd->dev, p);
597 	if (IS_ERR(pdd)) {
598 		err = -ESRCH;
599 		goto out;
600 	}
601 
602 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
603 			 ? cache_policy_coherent : cache_policy_noncoherent;
604 
605 	alternate_policy =
606 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
607 		   ? cache_policy_coherent : cache_policy_noncoherent;
608 
609 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
610 				&pdd->qpd,
611 				default_policy,
612 				alternate_policy,
613 				(void __user *)args->alternate_aperture_base,
614 				args->alternate_aperture_size))
615 		err = -EINVAL;
616 
617 out:
618 err_pdd:
619 	mutex_unlock(&p->mutex);
620 
621 	return err;
622 }
623 
624 static int kfd_ioctl_set_trap_handler(struct file *filep,
625 					struct kfd_process *p, void *data)
626 {
627 	struct kfd_ioctl_set_trap_handler_args *args = data;
628 	int err = 0;
629 	struct kfd_process_device *pdd;
630 
631 	mutex_lock(&p->mutex);
632 
633 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
634 	if (!pdd) {
635 		err = -EINVAL;
636 		goto err_pdd;
637 	}
638 
639 	pdd = kfd_bind_process_to_device(pdd->dev, p);
640 	if (IS_ERR(pdd)) {
641 		err = -ESRCH;
642 		goto out;
643 	}
644 
645 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
646 
647 out:
648 err_pdd:
649 	mutex_unlock(&p->mutex);
650 
651 	return err;
652 }
653 
654 static int kfd_ioctl_dbg_register(struct file *filep,
655 				struct kfd_process *p, void *data)
656 {
657 	return -EPERM;
658 }
659 
660 static int kfd_ioctl_dbg_unregister(struct file *filep,
661 				struct kfd_process *p, void *data)
662 {
663 	return -EPERM;
664 }
665 
666 static int kfd_ioctl_dbg_address_watch(struct file *filep,
667 					struct kfd_process *p, void *data)
668 {
669 	return -EPERM;
670 }
671 
672 /* Parse and generate fixed size data structure for wave control */
673 static int kfd_ioctl_dbg_wave_control(struct file *filep,
674 					struct kfd_process *p, void *data)
675 {
676 	return -EPERM;
677 }
678 
679 static int kfd_ioctl_get_clock_counters(struct file *filep,
680 				struct kfd_process *p, void *data)
681 {
682 	struct kfd_ioctl_get_clock_counters_args *args = data;
683 	struct kfd_process_device *pdd;
684 
685 	mutex_lock(&p->mutex);
686 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
687 	mutex_unlock(&p->mutex);
688 	if (pdd)
689 		/* Reading GPU clock counter from KGD */
690 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
691 	else
692 		/* Node without GPU resource */
693 		args->gpu_clock_counter = 0;
694 
695 	/* No access to rdtsc. Using raw monotonic time */
696 	args->cpu_clock_counter = ktime_get_raw_ns();
697 	args->system_clock_counter = ktime_get_boottime_ns();
698 
699 	/* Since the counter is in nano-seconds we use 1GHz frequency */
700 	args->system_clock_freq = 1000000000;
701 
702 	return 0;
703 }
704 
705 
706 static int kfd_ioctl_get_process_apertures(struct file *filp,
707 				struct kfd_process *p, void *data)
708 {
709 	struct kfd_ioctl_get_process_apertures_args *args = data;
710 	struct kfd_process_device_apertures *pAperture;
711 	int i;
712 
713 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
714 
715 	args->num_of_nodes = 0;
716 
717 	mutex_lock(&p->mutex);
718 	/* Run over all pdd of the process */
719 	for (i = 0; i < p->n_pdds; i++) {
720 		struct kfd_process_device *pdd = p->pdds[i];
721 
722 		pAperture =
723 			&args->process_apertures[args->num_of_nodes];
724 		pAperture->gpu_id = pdd->dev->id;
725 		pAperture->lds_base = pdd->lds_base;
726 		pAperture->lds_limit = pdd->lds_limit;
727 		pAperture->gpuvm_base = pdd->gpuvm_base;
728 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
729 		pAperture->scratch_base = pdd->scratch_base;
730 		pAperture->scratch_limit = pdd->scratch_limit;
731 
732 		dev_dbg(kfd_device,
733 			"node id %u\n", args->num_of_nodes);
734 		dev_dbg(kfd_device,
735 			"gpu id %u\n", pdd->dev->id);
736 		dev_dbg(kfd_device,
737 			"lds_base %llX\n", pdd->lds_base);
738 		dev_dbg(kfd_device,
739 			"lds_limit %llX\n", pdd->lds_limit);
740 		dev_dbg(kfd_device,
741 			"gpuvm_base %llX\n", pdd->gpuvm_base);
742 		dev_dbg(kfd_device,
743 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
744 		dev_dbg(kfd_device,
745 			"scratch_base %llX\n", pdd->scratch_base);
746 		dev_dbg(kfd_device,
747 			"scratch_limit %llX\n", pdd->scratch_limit);
748 
749 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
750 			break;
751 	}
752 	mutex_unlock(&p->mutex);
753 
754 	return 0;
755 }
756 
757 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
758 				struct kfd_process *p, void *data)
759 {
760 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
761 	struct kfd_process_device_apertures *pa;
762 	int ret;
763 	int i;
764 
765 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
766 
767 	if (args->num_of_nodes == 0) {
768 		/* Return number of nodes, so that user space can alloacate
769 		 * sufficient memory
770 		 */
771 		mutex_lock(&p->mutex);
772 		args->num_of_nodes = p->n_pdds;
773 		goto out_unlock;
774 	}
775 
776 	/* Fill in process-aperture information for all available
777 	 * nodes, but not more than args->num_of_nodes as that is
778 	 * the amount of memory allocated by user
779 	 */
780 	pa = kzalloc((sizeof(struct kfd_process_device_apertures) *
781 				args->num_of_nodes), GFP_KERNEL);
782 	if (!pa)
783 		return -ENOMEM;
784 
785 	mutex_lock(&p->mutex);
786 
787 	if (!p->n_pdds) {
788 		args->num_of_nodes = 0;
789 		kfree(pa);
790 		goto out_unlock;
791 	}
792 
793 	/* Run over all pdd of the process */
794 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
795 		struct kfd_process_device *pdd = p->pdds[i];
796 
797 		pa[i].gpu_id = pdd->dev->id;
798 		pa[i].lds_base = pdd->lds_base;
799 		pa[i].lds_limit = pdd->lds_limit;
800 		pa[i].gpuvm_base = pdd->gpuvm_base;
801 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
802 		pa[i].scratch_base = pdd->scratch_base;
803 		pa[i].scratch_limit = pdd->scratch_limit;
804 
805 		dev_dbg(kfd_device,
806 			"gpu id %u\n", pdd->dev->id);
807 		dev_dbg(kfd_device,
808 			"lds_base %llX\n", pdd->lds_base);
809 		dev_dbg(kfd_device,
810 			"lds_limit %llX\n", pdd->lds_limit);
811 		dev_dbg(kfd_device,
812 			"gpuvm_base %llX\n", pdd->gpuvm_base);
813 		dev_dbg(kfd_device,
814 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
815 		dev_dbg(kfd_device,
816 			"scratch_base %llX\n", pdd->scratch_base);
817 		dev_dbg(kfd_device,
818 			"scratch_limit %llX\n", pdd->scratch_limit);
819 	}
820 	mutex_unlock(&p->mutex);
821 
822 	args->num_of_nodes = i;
823 	ret = copy_to_user(
824 			(void __user *)args->kfd_process_device_apertures_ptr,
825 			pa,
826 			(i * sizeof(struct kfd_process_device_apertures)));
827 	kfree(pa);
828 	return ret ? -EFAULT : 0;
829 
830 out_unlock:
831 	mutex_unlock(&p->mutex);
832 	return 0;
833 }
834 
835 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
836 					void *data)
837 {
838 	struct kfd_ioctl_create_event_args *args = data;
839 	int err;
840 
841 	/* For dGPUs the event page is allocated in user mode. The
842 	 * handle is passed to KFD with the first call to this IOCTL
843 	 * through the event_page_offset field.
844 	 */
845 	if (args->event_page_offset) {
846 		mutex_lock(&p->mutex);
847 		err = kfd_kmap_event_page(p, args->event_page_offset);
848 		mutex_unlock(&p->mutex);
849 		if (err)
850 			return err;
851 	}
852 
853 	err = kfd_event_create(filp, p, args->event_type,
854 				args->auto_reset != 0, args->node_id,
855 				&args->event_id, &args->event_trigger_data,
856 				&args->event_page_offset,
857 				&args->event_slot_index);
858 
859 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
860 	return err;
861 }
862 
863 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
864 					void *data)
865 {
866 	struct kfd_ioctl_destroy_event_args *args = data;
867 
868 	return kfd_event_destroy(p, args->event_id);
869 }
870 
871 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
872 				void *data)
873 {
874 	struct kfd_ioctl_set_event_args *args = data;
875 
876 	return kfd_set_event(p, args->event_id);
877 }
878 
879 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
880 				void *data)
881 {
882 	struct kfd_ioctl_reset_event_args *args = data;
883 
884 	return kfd_reset_event(p, args->event_id);
885 }
886 
887 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
888 				void *data)
889 {
890 	struct kfd_ioctl_wait_events_args *args = data;
891 
892 	return kfd_wait_on_events(p, args->num_events,
893 			(void __user *)args->events_ptr,
894 			(args->wait_for_all != 0),
895 			&args->timeout, &args->wait_result);
896 }
897 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
898 					struct kfd_process *p, void *data)
899 {
900 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
901 	struct kfd_process_device *pdd;
902 	struct kfd_node *dev;
903 	long err;
904 
905 	mutex_lock(&p->mutex);
906 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
907 	if (!pdd) {
908 		err = -EINVAL;
909 		goto err_pdd;
910 	}
911 	dev = pdd->dev;
912 
913 	pdd = kfd_bind_process_to_device(dev, p);
914 	if (IS_ERR(pdd)) {
915 		err = PTR_ERR(pdd);
916 		goto bind_process_to_device_fail;
917 	}
918 
919 	pdd->qpd.sh_hidden_private_base = args->va_addr;
920 
921 	mutex_unlock(&p->mutex);
922 
923 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
924 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
925 		dev->kfd2kgd->set_scratch_backing_va(
926 			dev->adev, args->va_addr, pdd->qpd.vmid);
927 
928 	return 0;
929 
930 bind_process_to_device_fail:
931 err_pdd:
932 	mutex_unlock(&p->mutex);
933 	return err;
934 }
935 
936 static int kfd_ioctl_get_tile_config(struct file *filep,
937 		struct kfd_process *p, void *data)
938 {
939 	struct kfd_ioctl_get_tile_config_args *args = data;
940 	struct kfd_process_device *pdd;
941 	struct tile_config config;
942 	int err = 0;
943 
944 	mutex_lock(&p->mutex);
945 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
946 	mutex_unlock(&p->mutex);
947 	if (!pdd)
948 		return -EINVAL;
949 
950 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
951 
952 	args->gb_addr_config = config.gb_addr_config;
953 	args->num_banks = config.num_banks;
954 	args->num_ranks = config.num_ranks;
955 
956 	if (args->num_tile_configs > config.num_tile_configs)
957 		args->num_tile_configs = config.num_tile_configs;
958 	err = copy_to_user((void __user *)args->tile_config_ptr,
959 			config.tile_config_ptr,
960 			args->num_tile_configs * sizeof(uint32_t));
961 	if (err) {
962 		args->num_tile_configs = 0;
963 		return -EFAULT;
964 	}
965 
966 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
967 		args->num_macro_tile_configs =
968 				config.num_macro_tile_configs;
969 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
970 			config.macro_tile_config_ptr,
971 			args->num_macro_tile_configs * sizeof(uint32_t));
972 	if (err) {
973 		args->num_macro_tile_configs = 0;
974 		return -EFAULT;
975 	}
976 
977 	return 0;
978 }
979 
980 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
981 				void *data)
982 {
983 	struct kfd_ioctl_acquire_vm_args *args = data;
984 	struct kfd_process_device *pdd;
985 	struct file *drm_file;
986 	int ret;
987 
988 	drm_file = fget(args->drm_fd);
989 	if (!drm_file)
990 		return -EINVAL;
991 
992 	mutex_lock(&p->mutex);
993 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
994 	if (!pdd) {
995 		ret = -EINVAL;
996 		goto err_pdd;
997 	}
998 
999 	if (pdd->drm_file) {
1000 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
1001 		goto err_drm_file;
1002 	}
1003 
1004 	ret = kfd_process_device_init_vm(pdd, drm_file);
1005 	if (ret)
1006 		goto err_unlock;
1007 
1008 	/* On success, the PDD keeps the drm_file reference */
1009 	mutex_unlock(&p->mutex);
1010 
1011 	return 0;
1012 
1013 err_unlock:
1014 err_pdd:
1015 err_drm_file:
1016 	mutex_unlock(&p->mutex);
1017 	fput(drm_file);
1018 	return ret;
1019 }
1020 
1021 bool kfd_dev_is_large_bar(struct kfd_node *dev)
1022 {
1023 	if (debug_largebar) {
1024 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1025 		return true;
1026 	}
1027 
1028 	if (dev->kfd->use_iommu_v2)
1029 		return false;
1030 
1031 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1032 	    dev->local_mem_info.local_mem_size_public > 0)
1033 		return true;
1034 
1035 	if (dev->local_mem_info.local_mem_size_public == 0 &&
1036 	    dev->kfd->adev->gmc.is_app_apu) {
1037 		pr_debug("APP APU, Consider like a large bar system\n");
1038 		return true;
1039 	}
1040 
1041 	return false;
1042 }
1043 
1044 static int kfd_ioctl_get_available_memory(struct file *filep,
1045 					  struct kfd_process *p, void *data)
1046 {
1047 	struct kfd_ioctl_get_available_memory_args *args = data;
1048 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1049 
1050 	if (!pdd)
1051 		return -EINVAL;
1052 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1053 							pdd->dev->node_id);
1054 	kfd_unlock_pdd(pdd);
1055 	return 0;
1056 }
1057 
1058 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1059 					struct kfd_process *p, void *data)
1060 {
1061 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1062 	struct kfd_process_device *pdd;
1063 	void *mem;
1064 	struct kfd_node *dev;
1065 	int idr_handle;
1066 	long err;
1067 	uint64_t offset = args->mmap_offset;
1068 	uint32_t flags = args->flags;
1069 
1070 	if (args->size == 0)
1071 		return -EINVAL;
1072 
1073 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1074 	/* Flush pending deferred work to avoid racing with deferred actions
1075 	 * from previous memory map changes (e.g. munmap).
1076 	 */
1077 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1078 	mutex_lock(&p->svms.lock);
1079 	mmap_write_unlock(current->mm);
1080 	if (interval_tree_iter_first(&p->svms.objects,
1081 				     args->va_addr >> PAGE_SHIFT,
1082 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1083 		pr_err("Address: 0x%llx already allocated by SVM\n",
1084 			args->va_addr);
1085 		mutex_unlock(&p->svms.lock);
1086 		return -EADDRINUSE;
1087 	}
1088 
1089 	/* When register user buffer check if it has been registered by svm by
1090 	 * buffer cpu virtual address.
1091 	 */
1092 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1093 	    interval_tree_iter_first(&p->svms.objects,
1094 				     args->mmap_offset >> PAGE_SHIFT,
1095 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1096 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1097 			args->mmap_offset);
1098 		mutex_unlock(&p->svms.lock);
1099 		return -EADDRINUSE;
1100 	}
1101 
1102 	mutex_unlock(&p->svms.lock);
1103 #endif
1104 	mutex_lock(&p->mutex);
1105 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1106 	if (!pdd) {
1107 		err = -EINVAL;
1108 		goto err_pdd;
1109 	}
1110 
1111 	dev = pdd->dev;
1112 
1113 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1114 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1115 		!kfd_dev_is_large_bar(dev)) {
1116 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1117 		err = -EINVAL;
1118 		goto err_large_bar;
1119 	}
1120 
1121 	pdd = kfd_bind_process_to_device(dev, p);
1122 	if (IS_ERR(pdd)) {
1123 		err = PTR_ERR(pdd);
1124 		goto err_unlock;
1125 	}
1126 
1127 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1128 		if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1129 			err = -EINVAL;
1130 			goto err_unlock;
1131 		}
1132 		offset = kfd_get_process_doorbells(pdd);
1133 		if (!offset) {
1134 			err = -ENOMEM;
1135 			goto err_unlock;
1136 		}
1137 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1138 		if (args->size != PAGE_SIZE) {
1139 			err = -EINVAL;
1140 			goto err_unlock;
1141 		}
1142 		offset = dev->adev->rmmio_remap.bus_addr;
1143 		if (!offset) {
1144 			err = -ENOMEM;
1145 			goto err_unlock;
1146 		}
1147 	}
1148 
1149 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1150 		dev->adev, args->va_addr, args->size,
1151 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1152 		flags, false);
1153 
1154 	if (err)
1155 		goto err_unlock;
1156 
1157 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1158 	if (idr_handle < 0) {
1159 		err = -EFAULT;
1160 		goto err_free;
1161 	}
1162 
1163 	/* Update the VRAM usage count */
1164 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1165 		uint64_t size = args->size;
1166 
1167 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1168 			size >>= 1;
1169 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1170 	}
1171 
1172 	mutex_unlock(&p->mutex);
1173 
1174 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1175 	args->mmap_offset = offset;
1176 
1177 	/* MMIO is mapped through kfd device
1178 	 * Generate a kfd mmap offset
1179 	 */
1180 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1181 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1182 					| KFD_MMAP_GPU_ID(args->gpu_id);
1183 
1184 	return 0;
1185 
1186 err_free:
1187 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1188 					       pdd->drm_priv, NULL);
1189 err_unlock:
1190 err_pdd:
1191 err_large_bar:
1192 	mutex_unlock(&p->mutex);
1193 	return err;
1194 }
1195 
1196 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1197 					struct kfd_process *p, void *data)
1198 {
1199 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1200 	struct kfd_process_device *pdd;
1201 	void *mem;
1202 	int ret;
1203 	uint64_t size = 0;
1204 
1205 	mutex_lock(&p->mutex);
1206 	/*
1207 	 * Safeguard to prevent user space from freeing signal BO.
1208 	 * It will be freed at process termination.
1209 	 */
1210 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1211 		pr_err("Free signal BO is not allowed\n");
1212 		ret = -EPERM;
1213 		goto err_unlock;
1214 	}
1215 
1216 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1217 	if (!pdd) {
1218 		pr_err("Process device data doesn't exist\n");
1219 		ret = -EINVAL;
1220 		goto err_pdd;
1221 	}
1222 
1223 	mem = kfd_process_device_translate_handle(
1224 		pdd, GET_IDR_HANDLE(args->handle));
1225 	if (!mem) {
1226 		ret = -EINVAL;
1227 		goto err_unlock;
1228 	}
1229 
1230 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1231 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1232 
1233 	/* If freeing the buffer failed, leave the handle in place for
1234 	 * clean-up during process tear-down.
1235 	 */
1236 	if (!ret)
1237 		kfd_process_device_remove_obj_handle(
1238 			pdd, GET_IDR_HANDLE(args->handle));
1239 
1240 	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1241 
1242 err_unlock:
1243 err_pdd:
1244 	mutex_unlock(&p->mutex);
1245 	return ret;
1246 }
1247 
1248 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1249 					struct kfd_process *p, void *data)
1250 {
1251 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1252 	struct kfd_process_device *pdd, *peer_pdd;
1253 	void *mem;
1254 	struct kfd_node *dev;
1255 	long err = 0;
1256 	int i;
1257 	uint32_t *devices_arr = NULL;
1258 
1259 	if (!args->n_devices) {
1260 		pr_debug("Device IDs array empty\n");
1261 		return -EINVAL;
1262 	}
1263 	if (args->n_success > args->n_devices) {
1264 		pr_debug("n_success exceeds n_devices\n");
1265 		return -EINVAL;
1266 	}
1267 
1268 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1269 				    GFP_KERNEL);
1270 	if (!devices_arr)
1271 		return -ENOMEM;
1272 
1273 	err = copy_from_user(devices_arr,
1274 			     (void __user *)args->device_ids_array_ptr,
1275 			     args->n_devices * sizeof(*devices_arr));
1276 	if (err != 0) {
1277 		err = -EFAULT;
1278 		goto copy_from_user_failed;
1279 	}
1280 
1281 	mutex_lock(&p->mutex);
1282 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1283 	if (!pdd) {
1284 		err = -EINVAL;
1285 		goto get_process_device_data_failed;
1286 	}
1287 	dev = pdd->dev;
1288 
1289 	pdd = kfd_bind_process_to_device(dev, p);
1290 	if (IS_ERR(pdd)) {
1291 		err = PTR_ERR(pdd);
1292 		goto bind_process_to_device_failed;
1293 	}
1294 
1295 	mem = kfd_process_device_translate_handle(pdd,
1296 						GET_IDR_HANDLE(args->handle));
1297 	if (!mem) {
1298 		err = -ENOMEM;
1299 		goto get_mem_obj_from_handle_failed;
1300 	}
1301 
1302 	for (i = args->n_success; i < args->n_devices; i++) {
1303 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1304 		if (!peer_pdd) {
1305 			pr_debug("Getting device by id failed for 0x%x\n",
1306 				 devices_arr[i]);
1307 			err = -EINVAL;
1308 			goto get_mem_obj_from_handle_failed;
1309 		}
1310 
1311 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1312 		if (IS_ERR(peer_pdd)) {
1313 			err = PTR_ERR(peer_pdd);
1314 			goto get_mem_obj_from_handle_failed;
1315 		}
1316 
1317 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1318 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1319 			peer_pdd->drm_priv);
1320 		if (err) {
1321 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1322 
1323 			dev_err(dev->adev->dev,
1324 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1325 			       pci_domain_nr(pdev->bus),
1326 			       pdev->bus->number,
1327 			       PCI_SLOT(pdev->devfn),
1328 			       PCI_FUNC(pdev->devfn),
1329 			       ((struct kgd_mem *)mem)->domain);
1330 			goto map_memory_to_gpu_failed;
1331 		}
1332 		args->n_success = i+1;
1333 	}
1334 
1335 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1336 	if (err) {
1337 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1338 		goto sync_memory_failed;
1339 	}
1340 
1341 	mutex_unlock(&p->mutex);
1342 
1343 	/* Flush TLBs after waiting for the page table updates to complete */
1344 	for (i = 0; i < args->n_devices; i++) {
1345 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1346 		if (WARN_ON_ONCE(!peer_pdd))
1347 			continue;
1348 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1349 	}
1350 	kfree(devices_arr);
1351 
1352 	return err;
1353 
1354 get_process_device_data_failed:
1355 bind_process_to_device_failed:
1356 get_mem_obj_from_handle_failed:
1357 map_memory_to_gpu_failed:
1358 sync_memory_failed:
1359 	mutex_unlock(&p->mutex);
1360 copy_from_user_failed:
1361 	kfree(devices_arr);
1362 
1363 	return err;
1364 }
1365 
1366 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1367 					struct kfd_process *p, void *data)
1368 {
1369 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1370 	struct kfd_process_device *pdd, *peer_pdd;
1371 	void *mem;
1372 	long err = 0;
1373 	uint32_t *devices_arr = NULL, i;
1374 	bool flush_tlb;
1375 
1376 	if (!args->n_devices) {
1377 		pr_debug("Device IDs array empty\n");
1378 		return -EINVAL;
1379 	}
1380 	if (args->n_success > args->n_devices) {
1381 		pr_debug("n_success exceeds n_devices\n");
1382 		return -EINVAL;
1383 	}
1384 
1385 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1386 				    GFP_KERNEL);
1387 	if (!devices_arr)
1388 		return -ENOMEM;
1389 
1390 	err = copy_from_user(devices_arr,
1391 			     (void __user *)args->device_ids_array_ptr,
1392 			     args->n_devices * sizeof(*devices_arr));
1393 	if (err != 0) {
1394 		err = -EFAULT;
1395 		goto copy_from_user_failed;
1396 	}
1397 
1398 	mutex_lock(&p->mutex);
1399 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1400 	if (!pdd) {
1401 		err = -EINVAL;
1402 		goto bind_process_to_device_failed;
1403 	}
1404 
1405 	mem = kfd_process_device_translate_handle(pdd,
1406 						GET_IDR_HANDLE(args->handle));
1407 	if (!mem) {
1408 		err = -ENOMEM;
1409 		goto get_mem_obj_from_handle_failed;
1410 	}
1411 
1412 	for (i = args->n_success; i < args->n_devices; i++) {
1413 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1414 		if (!peer_pdd) {
1415 			err = -EINVAL;
1416 			goto get_mem_obj_from_handle_failed;
1417 		}
1418 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1419 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1420 		if (err) {
1421 			pr_err("Failed to unmap from gpu %d/%d\n",
1422 			       i, args->n_devices);
1423 			goto unmap_memory_from_gpu_failed;
1424 		}
1425 		args->n_success = i+1;
1426 	}
1427 
1428 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1429 	if (flush_tlb) {
1430 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1431 				(struct kgd_mem *) mem, true);
1432 		if (err) {
1433 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1434 			goto sync_memory_failed;
1435 		}
1436 	}
1437 	mutex_unlock(&p->mutex);
1438 
1439 	if (flush_tlb) {
1440 		/* Flush TLBs after waiting for the page table updates to complete */
1441 		for (i = 0; i < args->n_devices; i++) {
1442 			peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1443 			if (WARN_ON_ONCE(!peer_pdd))
1444 				continue;
1445 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1446 		}
1447 	}
1448 	kfree(devices_arr);
1449 
1450 	return 0;
1451 
1452 bind_process_to_device_failed:
1453 get_mem_obj_from_handle_failed:
1454 unmap_memory_from_gpu_failed:
1455 sync_memory_failed:
1456 	mutex_unlock(&p->mutex);
1457 copy_from_user_failed:
1458 	kfree(devices_arr);
1459 	return err;
1460 }
1461 
1462 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1463 		struct kfd_process *p, void *data)
1464 {
1465 	int retval;
1466 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1467 	struct queue *q;
1468 	struct kfd_node *dev;
1469 
1470 	mutex_lock(&p->mutex);
1471 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1472 
1473 	if (q) {
1474 		dev = q->device;
1475 	} else {
1476 		retval = -EINVAL;
1477 		goto out_unlock;
1478 	}
1479 
1480 	if (!dev->gws) {
1481 		retval = -ENODEV;
1482 		goto out_unlock;
1483 	}
1484 
1485 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1486 		retval = -ENODEV;
1487 		goto out_unlock;
1488 	}
1489 
1490 	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1491 				      kfd_dbg_has_cwsr_workaround(dev))) {
1492 		retval = -EBUSY;
1493 		goto out_unlock;
1494 	}
1495 
1496 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1497 	mutex_unlock(&p->mutex);
1498 
1499 	args->first_gws = 0;
1500 	return retval;
1501 
1502 out_unlock:
1503 	mutex_unlock(&p->mutex);
1504 	return retval;
1505 }
1506 
1507 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1508 		struct kfd_process *p, void *data)
1509 {
1510 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1511 	struct kfd_node *dev = NULL;
1512 	struct amdgpu_device *dmabuf_adev;
1513 	void *metadata_buffer = NULL;
1514 	uint32_t flags;
1515 	int8_t xcp_id;
1516 	unsigned int i;
1517 	int r;
1518 
1519 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1520 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1521 		if (dev)
1522 			break;
1523 	if (!dev)
1524 		return -EINVAL;
1525 
1526 	if (args->metadata_ptr) {
1527 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1528 		if (!metadata_buffer)
1529 			return -ENOMEM;
1530 	}
1531 
1532 	/* Get dmabuf info from KGD */
1533 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1534 					  &dmabuf_adev, &args->size,
1535 					  metadata_buffer, args->metadata_size,
1536 					  &args->metadata_size, &flags, &xcp_id);
1537 	if (r)
1538 		goto exit;
1539 
1540 	if (xcp_id >= 0)
1541 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1542 	else
1543 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[0]->id;
1544 	args->flags = flags;
1545 
1546 	/* Copy metadata buffer to user mode */
1547 	if (metadata_buffer) {
1548 		r = copy_to_user((void __user *)args->metadata_ptr,
1549 				 metadata_buffer, args->metadata_size);
1550 		if (r != 0)
1551 			r = -EFAULT;
1552 	}
1553 
1554 exit:
1555 	kfree(metadata_buffer);
1556 
1557 	return r;
1558 }
1559 
1560 static int kfd_ioctl_import_dmabuf(struct file *filep,
1561 				   struct kfd_process *p, void *data)
1562 {
1563 	struct kfd_ioctl_import_dmabuf_args *args = data;
1564 	struct kfd_process_device *pdd;
1565 	struct dma_buf *dmabuf;
1566 	int idr_handle;
1567 	uint64_t size;
1568 	void *mem;
1569 	int r;
1570 
1571 	dmabuf = dma_buf_get(args->dmabuf_fd);
1572 	if (IS_ERR(dmabuf))
1573 		return PTR_ERR(dmabuf);
1574 
1575 	mutex_lock(&p->mutex);
1576 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1577 	if (!pdd) {
1578 		r = -EINVAL;
1579 		goto err_unlock;
1580 	}
1581 
1582 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1583 	if (IS_ERR(pdd)) {
1584 		r = PTR_ERR(pdd);
1585 		goto err_unlock;
1586 	}
1587 
1588 	r = amdgpu_amdkfd_gpuvm_import_dmabuf(pdd->dev->adev, dmabuf,
1589 					      args->va_addr, pdd->drm_priv,
1590 					      (struct kgd_mem **)&mem, &size,
1591 					      NULL);
1592 	if (r)
1593 		goto err_unlock;
1594 
1595 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1596 	if (idr_handle < 0) {
1597 		r = -EFAULT;
1598 		goto err_free;
1599 	}
1600 
1601 	mutex_unlock(&p->mutex);
1602 	dma_buf_put(dmabuf);
1603 
1604 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1605 
1606 	return 0;
1607 
1608 err_free:
1609 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1610 					       pdd->drm_priv, NULL);
1611 err_unlock:
1612 	mutex_unlock(&p->mutex);
1613 	dma_buf_put(dmabuf);
1614 	return r;
1615 }
1616 
1617 static int kfd_ioctl_export_dmabuf(struct file *filep,
1618 				   struct kfd_process *p, void *data)
1619 {
1620 	struct kfd_ioctl_export_dmabuf_args *args = data;
1621 	struct kfd_process_device *pdd;
1622 	struct dma_buf *dmabuf;
1623 	struct kfd_node *dev;
1624 	void *mem;
1625 	int ret = 0;
1626 
1627 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1628 	if (!dev)
1629 		return -EINVAL;
1630 
1631 	mutex_lock(&p->mutex);
1632 
1633 	pdd = kfd_get_process_device_data(dev, p);
1634 	if (!pdd) {
1635 		ret = -EINVAL;
1636 		goto err_unlock;
1637 	}
1638 
1639 	mem = kfd_process_device_translate_handle(pdd,
1640 						GET_IDR_HANDLE(args->handle));
1641 	if (!mem) {
1642 		ret = -EINVAL;
1643 		goto err_unlock;
1644 	}
1645 
1646 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1647 	mutex_unlock(&p->mutex);
1648 	if (ret)
1649 		goto err_out;
1650 
1651 	ret = dma_buf_fd(dmabuf, args->flags);
1652 	if (ret < 0) {
1653 		dma_buf_put(dmabuf);
1654 		goto err_out;
1655 	}
1656 	/* dma_buf_fd assigns the reference count to the fd, no need to
1657 	 * put the reference here.
1658 	 */
1659 	args->dmabuf_fd = ret;
1660 
1661 	return 0;
1662 
1663 err_unlock:
1664 	mutex_unlock(&p->mutex);
1665 err_out:
1666 	return ret;
1667 }
1668 
1669 /* Handle requests for watching SMI events */
1670 static int kfd_ioctl_smi_events(struct file *filep,
1671 				struct kfd_process *p, void *data)
1672 {
1673 	struct kfd_ioctl_smi_events_args *args = data;
1674 	struct kfd_process_device *pdd;
1675 
1676 	mutex_lock(&p->mutex);
1677 
1678 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1679 	mutex_unlock(&p->mutex);
1680 	if (!pdd)
1681 		return -EINVAL;
1682 
1683 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1684 }
1685 
1686 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1687 
1688 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1689 				    struct kfd_process *p, void *data)
1690 {
1691 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1692 	int r = 0;
1693 
1694 	mutex_lock(&p->mutex);
1695 	if (args->xnack_enabled >= 0) {
1696 		if (!list_empty(&p->pqm.queues)) {
1697 			pr_debug("Process has user queues running\n");
1698 			r = -EBUSY;
1699 			goto out_unlock;
1700 		}
1701 
1702 		if (p->xnack_enabled == args->xnack_enabled)
1703 			goto out_unlock;
1704 
1705 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1706 			r = -EPERM;
1707 			goto out_unlock;
1708 		}
1709 
1710 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1711 	} else {
1712 		args->xnack_enabled = p->xnack_enabled;
1713 	}
1714 
1715 out_unlock:
1716 	mutex_unlock(&p->mutex);
1717 
1718 	return r;
1719 }
1720 
1721 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1722 {
1723 	struct kfd_ioctl_svm_args *args = data;
1724 	int r = 0;
1725 
1726 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1727 		 args->start_addr, args->size, args->op, args->nattr);
1728 
1729 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1730 		return -EINVAL;
1731 	if (!args->start_addr || !args->size)
1732 		return -EINVAL;
1733 
1734 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1735 		      args->attrs);
1736 
1737 	return r;
1738 }
1739 #else
1740 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1741 				    struct kfd_process *p, void *data)
1742 {
1743 	return -EPERM;
1744 }
1745 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1746 {
1747 	return -EPERM;
1748 }
1749 #endif
1750 
1751 static int criu_checkpoint_process(struct kfd_process *p,
1752 			     uint8_t __user *user_priv_data,
1753 			     uint64_t *priv_offset)
1754 {
1755 	struct kfd_criu_process_priv_data process_priv;
1756 	int ret;
1757 
1758 	memset(&process_priv, 0, sizeof(process_priv));
1759 
1760 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1761 	/* For CR, we don't consider negative xnack mode which is used for
1762 	 * querying without changing it, here 0 simply means disabled and 1
1763 	 * means enabled so retry for finding a valid PTE.
1764 	 */
1765 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1766 
1767 	ret = copy_to_user(user_priv_data + *priv_offset,
1768 				&process_priv, sizeof(process_priv));
1769 
1770 	if (ret) {
1771 		pr_err("Failed to copy process information to user\n");
1772 		ret = -EFAULT;
1773 	}
1774 
1775 	*priv_offset += sizeof(process_priv);
1776 	return ret;
1777 }
1778 
1779 static int criu_checkpoint_devices(struct kfd_process *p,
1780 			     uint32_t num_devices,
1781 			     uint8_t __user *user_addr,
1782 			     uint8_t __user *user_priv_data,
1783 			     uint64_t *priv_offset)
1784 {
1785 	struct kfd_criu_device_priv_data *device_priv = NULL;
1786 	struct kfd_criu_device_bucket *device_buckets = NULL;
1787 	int ret = 0, i;
1788 
1789 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1790 	if (!device_buckets) {
1791 		ret = -ENOMEM;
1792 		goto exit;
1793 	}
1794 
1795 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1796 	if (!device_priv) {
1797 		ret = -ENOMEM;
1798 		goto exit;
1799 	}
1800 
1801 	for (i = 0; i < num_devices; i++) {
1802 		struct kfd_process_device *pdd = p->pdds[i];
1803 
1804 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1805 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1806 
1807 		/*
1808 		 * priv_data does not contain useful information for now and is reserved for
1809 		 * future use, so we do not set its contents.
1810 		 */
1811 	}
1812 
1813 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1814 	if (ret) {
1815 		pr_err("Failed to copy device information to user\n");
1816 		ret = -EFAULT;
1817 		goto exit;
1818 	}
1819 
1820 	ret = copy_to_user(user_priv_data + *priv_offset,
1821 			   device_priv,
1822 			   num_devices * sizeof(*device_priv));
1823 	if (ret) {
1824 		pr_err("Failed to copy device information to user\n");
1825 		ret = -EFAULT;
1826 	}
1827 	*priv_offset += num_devices * sizeof(*device_priv);
1828 
1829 exit:
1830 	kvfree(device_buckets);
1831 	kvfree(device_priv);
1832 	return ret;
1833 }
1834 
1835 static uint32_t get_process_num_bos(struct kfd_process *p)
1836 {
1837 	uint32_t num_of_bos = 0;
1838 	int i;
1839 
1840 	/* Run over all PDDs of the process */
1841 	for (i = 0; i < p->n_pdds; i++) {
1842 		struct kfd_process_device *pdd = p->pdds[i];
1843 		void *mem;
1844 		int id;
1845 
1846 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1847 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1848 
1849 			if ((uint64_t)kgd_mem->va > pdd->gpuvm_base)
1850 				num_of_bos++;
1851 		}
1852 	}
1853 	return num_of_bos;
1854 }
1855 
1856 static int criu_get_prime_handle(struct kgd_mem *mem, int flags,
1857 				      u32 *shared_fd)
1858 {
1859 	struct dma_buf *dmabuf;
1860 	int ret;
1861 
1862 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1863 	if (ret) {
1864 		pr_err("dmabuf export failed for the BO\n");
1865 		return ret;
1866 	}
1867 
1868 	ret = dma_buf_fd(dmabuf, flags);
1869 	if (ret < 0) {
1870 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1871 		goto out_free_dmabuf;
1872 	}
1873 
1874 	*shared_fd = ret;
1875 	return 0;
1876 
1877 out_free_dmabuf:
1878 	dma_buf_put(dmabuf);
1879 	return ret;
1880 }
1881 
1882 static int criu_checkpoint_bos(struct kfd_process *p,
1883 			       uint32_t num_bos,
1884 			       uint8_t __user *user_bos,
1885 			       uint8_t __user *user_priv_data,
1886 			       uint64_t *priv_offset)
1887 {
1888 	struct kfd_criu_bo_bucket *bo_buckets;
1889 	struct kfd_criu_bo_priv_data *bo_privs;
1890 	int ret = 0, pdd_index, bo_index = 0, id;
1891 	void *mem;
1892 
1893 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1894 	if (!bo_buckets)
1895 		return -ENOMEM;
1896 
1897 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1898 	if (!bo_privs) {
1899 		ret = -ENOMEM;
1900 		goto exit;
1901 	}
1902 
1903 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1904 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1905 		struct amdgpu_bo *dumper_bo;
1906 		struct kgd_mem *kgd_mem;
1907 
1908 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1909 			struct kfd_criu_bo_bucket *bo_bucket;
1910 			struct kfd_criu_bo_priv_data *bo_priv;
1911 			int i, dev_idx = 0;
1912 
1913 			if (!mem) {
1914 				ret = -ENOMEM;
1915 				goto exit;
1916 			}
1917 
1918 			kgd_mem = (struct kgd_mem *)mem;
1919 			dumper_bo = kgd_mem->bo;
1920 
1921 			if ((uint64_t)kgd_mem->va <= pdd->gpuvm_base)
1922 				continue;
1923 
1924 			bo_bucket = &bo_buckets[bo_index];
1925 			bo_priv = &bo_privs[bo_index];
1926 
1927 			bo_bucket->gpu_id = pdd->user_gpu_id;
1928 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1929 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1930 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1931 			bo_priv->idr_handle = id;
1932 
1933 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1934 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1935 								&bo_priv->user_addr);
1936 				if (ret) {
1937 					pr_err("Failed to obtain user address for user-pointer bo\n");
1938 					goto exit;
1939 				}
1940 			}
1941 			if (bo_bucket->alloc_flags
1942 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1943 				ret = criu_get_prime_handle(kgd_mem,
1944 						bo_bucket->alloc_flags &
1945 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1946 						&bo_bucket->dmabuf_fd);
1947 				if (ret)
1948 					goto exit;
1949 			} else {
1950 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1951 			}
1952 
1953 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1954 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1955 					KFD_MMAP_GPU_ID(pdd->dev->id);
1956 			else if (bo_bucket->alloc_flags &
1957 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1958 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1959 					KFD_MMAP_GPU_ID(pdd->dev->id);
1960 			else
1961 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1962 
1963 			for (i = 0; i < p->n_pdds; i++) {
1964 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->dev->adev, kgd_mem))
1965 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1966 			}
1967 
1968 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1969 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1970 					bo_bucket->size,
1971 					bo_bucket->addr,
1972 					bo_bucket->offset,
1973 					bo_bucket->gpu_id,
1974 					bo_bucket->alloc_flags,
1975 					bo_priv->idr_handle);
1976 			bo_index++;
1977 		}
1978 	}
1979 
1980 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1981 	if (ret) {
1982 		pr_err("Failed to copy BO information to user\n");
1983 		ret = -EFAULT;
1984 		goto exit;
1985 	}
1986 
1987 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1988 	if (ret) {
1989 		pr_err("Failed to copy BO priv information to user\n");
1990 		ret = -EFAULT;
1991 		goto exit;
1992 	}
1993 
1994 	*priv_offset += num_bos * sizeof(*bo_privs);
1995 
1996 exit:
1997 	while (ret && bo_index--) {
1998 		if (bo_buckets[bo_index].alloc_flags
1999 		    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2000 			close_fd(bo_buckets[bo_index].dmabuf_fd);
2001 	}
2002 
2003 	kvfree(bo_buckets);
2004 	kvfree(bo_privs);
2005 	return ret;
2006 }
2007 
2008 static int criu_get_process_object_info(struct kfd_process *p,
2009 					uint32_t *num_devices,
2010 					uint32_t *num_bos,
2011 					uint32_t *num_objects,
2012 					uint64_t *objs_priv_size)
2013 {
2014 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2015 	uint32_t num_queues, num_events, num_svm_ranges;
2016 	int ret;
2017 
2018 	*num_devices = p->n_pdds;
2019 	*num_bos = get_process_num_bos(p);
2020 
2021 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2022 	if (ret)
2023 		return ret;
2024 
2025 	num_events = kfd_get_num_events(p);
2026 
2027 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2028 	if (ret)
2029 		return ret;
2030 
2031 	*num_objects = num_queues + num_events + num_svm_ranges;
2032 
2033 	if (objs_priv_size) {
2034 		priv_size = sizeof(struct kfd_criu_process_priv_data);
2035 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2036 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2037 		priv_size += queues_priv_data_size;
2038 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2039 		priv_size += svm_priv_data_size;
2040 		*objs_priv_size = priv_size;
2041 	}
2042 	return 0;
2043 }
2044 
2045 static int criu_checkpoint(struct file *filep,
2046 			   struct kfd_process *p,
2047 			   struct kfd_ioctl_criu_args *args)
2048 {
2049 	int ret;
2050 	uint32_t num_devices, num_bos, num_objects;
2051 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2052 
2053 	if (!args->devices || !args->bos || !args->priv_data)
2054 		return -EINVAL;
2055 
2056 	mutex_lock(&p->mutex);
2057 
2058 	if (!p->n_pdds) {
2059 		pr_err("No pdd for given process\n");
2060 		ret = -ENODEV;
2061 		goto exit_unlock;
2062 	}
2063 
2064 	/* Confirm all process queues are evicted */
2065 	if (!p->queues_paused) {
2066 		pr_err("Cannot dump process when queues are not in evicted state\n");
2067 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2068 		ret = -EINVAL;
2069 		goto exit_unlock;
2070 	}
2071 
2072 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2073 	if (ret)
2074 		goto exit_unlock;
2075 
2076 	if (num_devices != args->num_devices ||
2077 	    num_bos != args->num_bos ||
2078 	    num_objects != args->num_objects ||
2079 	    priv_size != args->priv_data_size) {
2080 
2081 		ret = -EINVAL;
2082 		goto exit_unlock;
2083 	}
2084 
2085 	/* each function will store private data inside priv_data and adjust priv_offset */
2086 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2087 	if (ret)
2088 		goto exit_unlock;
2089 
2090 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2091 				(uint8_t __user *)args->priv_data, &priv_offset);
2092 	if (ret)
2093 		goto exit_unlock;
2094 
2095 	/* Leave room for BOs in the private data. They need to be restored
2096 	 * before events, but we checkpoint them last to simplify the error
2097 	 * handling.
2098 	 */
2099 	bo_priv_offset = priv_offset;
2100 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2101 
2102 	if (num_objects) {
2103 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2104 						 &priv_offset);
2105 		if (ret)
2106 			goto exit_unlock;
2107 
2108 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2109 						 &priv_offset);
2110 		if (ret)
2111 			goto exit_unlock;
2112 
2113 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2114 		if (ret)
2115 			goto exit_unlock;
2116 	}
2117 
2118 	/* This must be the last thing in this function that can fail.
2119 	 * Otherwise we leak dmabuf file descriptors.
2120 	 */
2121 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2122 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2123 
2124 exit_unlock:
2125 	mutex_unlock(&p->mutex);
2126 	if (ret)
2127 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2128 	else
2129 		pr_debug("CRIU dump ret:%d\n", ret);
2130 
2131 	return ret;
2132 }
2133 
2134 static int criu_restore_process(struct kfd_process *p,
2135 				struct kfd_ioctl_criu_args *args,
2136 				uint64_t *priv_offset,
2137 				uint64_t max_priv_data_size)
2138 {
2139 	int ret = 0;
2140 	struct kfd_criu_process_priv_data process_priv;
2141 
2142 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2143 		return -EINVAL;
2144 
2145 	ret = copy_from_user(&process_priv,
2146 				(void __user *)(args->priv_data + *priv_offset),
2147 				sizeof(process_priv));
2148 	if (ret) {
2149 		pr_err("Failed to copy process private information from user\n");
2150 		ret = -EFAULT;
2151 		goto exit;
2152 	}
2153 	*priv_offset += sizeof(process_priv);
2154 
2155 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2156 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2157 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2158 		return -EINVAL;
2159 	}
2160 
2161 	pr_debug("Setting XNACK mode\n");
2162 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2163 		pr_err("xnack mode cannot be set\n");
2164 		ret = -EPERM;
2165 		goto exit;
2166 	} else {
2167 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2168 		p->xnack_enabled = process_priv.xnack_mode;
2169 	}
2170 
2171 exit:
2172 	return ret;
2173 }
2174 
2175 static int criu_restore_devices(struct kfd_process *p,
2176 				struct kfd_ioctl_criu_args *args,
2177 				uint64_t *priv_offset,
2178 				uint64_t max_priv_data_size)
2179 {
2180 	struct kfd_criu_device_bucket *device_buckets;
2181 	struct kfd_criu_device_priv_data *device_privs;
2182 	int ret = 0;
2183 	uint32_t i;
2184 
2185 	if (args->num_devices != p->n_pdds)
2186 		return -EINVAL;
2187 
2188 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2189 		return -EINVAL;
2190 
2191 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2192 	if (!device_buckets)
2193 		return -ENOMEM;
2194 
2195 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2196 				args->num_devices * sizeof(*device_buckets));
2197 	if (ret) {
2198 		pr_err("Failed to copy devices buckets from user\n");
2199 		ret = -EFAULT;
2200 		goto exit;
2201 	}
2202 
2203 	for (i = 0; i < args->num_devices; i++) {
2204 		struct kfd_node *dev;
2205 		struct kfd_process_device *pdd;
2206 		struct file *drm_file;
2207 
2208 		/* device private data is not currently used */
2209 
2210 		if (!device_buckets[i].user_gpu_id) {
2211 			pr_err("Invalid user gpu_id\n");
2212 			ret = -EINVAL;
2213 			goto exit;
2214 		}
2215 
2216 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2217 		if (!dev) {
2218 			pr_err("Failed to find device with gpu_id = %x\n",
2219 				device_buckets[i].actual_gpu_id);
2220 			ret = -EINVAL;
2221 			goto exit;
2222 		}
2223 
2224 		pdd = kfd_get_process_device_data(dev, p);
2225 		if (!pdd) {
2226 			pr_err("Failed to get pdd for gpu_id = %x\n",
2227 					device_buckets[i].actual_gpu_id);
2228 			ret = -EINVAL;
2229 			goto exit;
2230 		}
2231 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2232 
2233 		drm_file = fget(device_buckets[i].drm_fd);
2234 		if (!drm_file) {
2235 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2236 				device_buckets[i].drm_fd);
2237 			ret = -EINVAL;
2238 			goto exit;
2239 		}
2240 
2241 		if (pdd->drm_file) {
2242 			ret = -EINVAL;
2243 			goto exit;
2244 		}
2245 
2246 		/* create the vm using render nodes for kfd pdd */
2247 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2248 			pr_err("could not init vm for given pdd\n");
2249 			/* On success, the PDD keeps the drm_file reference */
2250 			fput(drm_file);
2251 			ret = -EINVAL;
2252 			goto exit;
2253 		}
2254 		/*
2255 		 * pdd now already has the vm bound to render node so below api won't create a new
2256 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2257 		 * for iommu v2 binding  and runtime pm.
2258 		 */
2259 		pdd = kfd_bind_process_to_device(dev, p);
2260 		if (IS_ERR(pdd)) {
2261 			ret = PTR_ERR(pdd);
2262 			goto exit;
2263 		}
2264 
2265 		if (!pdd->doorbell_index &&
2266 		    kfd_alloc_process_doorbells(pdd->dev->kfd, &pdd->doorbell_index) < 0) {
2267 			ret = -ENOMEM;
2268 			goto exit;
2269 		}
2270 	}
2271 
2272 	/*
2273 	 * We are not copying device private data from user as we are not using the data for now,
2274 	 * but we still adjust for its private data.
2275 	 */
2276 	*priv_offset += args->num_devices * sizeof(*device_privs);
2277 
2278 exit:
2279 	kfree(device_buckets);
2280 	return ret;
2281 }
2282 
2283 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2284 				      struct kfd_criu_bo_bucket *bo_bucket,
2285 				      struct kfd_criu_bo_priv_data *bo_priv,
2286 				      struct kgd_mem **kgd_mem)
2287 {
2288 	int idr_handle;
2289 	int ret;
2290 	const bool criu_resume = true;
2291 	u64 offset;
2292 
2293 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2294 		if (bo_bucket->size !=
2295 				kfd_doorbell_process_slice(pdd->dev->kfd))
2296 			return -EINVAL;
2297 
2298 		offset = kfd_get_process_doorbells(pdd);
2299 		if (!offset)
2300 			return -ENOMEM;
2301 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2302 		/* MMIO BOs need remapped bus address */
2303 		if (bo_bucket->size != PAGE_SIZE) {
2304 			pr_err("Invalid page size\n");
2305 			return -EINVAL;
2306 		}
2307 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2308 		if (!offset) {
2309 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2310 			return -ENOMEM;
2311 		}
2312 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2313 		offset = bo_priv->user_addr;
2314 	}
2315 	/* Create the BO */
2316 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2317 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2318 						      &offset, bo_bucket->alloc_flags, criu_resume);
2319 	if (ret) {
2320 		pr_err("Could not create the BO\n");
2321 		return ret;
2322 	}
2323 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2324 		 bo_bucket->size, bo_bucket->addr, offset);
2325 
2326 	/* Restore previous IDR handle */
2327 	pr_debug("Restoring old IDR handle for the BO");
2328 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2329 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2330 
2331 	if (idr_handle < 0) {
2332 		pr_err("Could not allocate idr\n");
2333 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2334 						       NULL);
2335 		return -ENOMEM;
2336 	}
2337 
2338 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2339 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2340 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2341 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2342 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2343 		bo_bucket->restored_offset = offset;
2344 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2345 		bo_bucket->restored_offset = offset;
2346 		/* Update the VRAM usage count */
2347 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2348 	}
2349 	return 0;
2350 }
2351 
2352 static int criu_restore_bo(struct kfd_process *p,
2353 			   struct kfd_criu_bo_bucket *bo_bucket,
2354 			   struct kfd_criu_bo_priv_data *bo_priv)
2355 {
2356 	struct kfd_process_device *pdd;
2357 	struct kgd_mem *kgd_mem;
2358 	int ret;
2359 	int j;
2360 
2361 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2362 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2363 		 bo_priv->idr_handle);
2364 
2365 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2366 	if (!pdd) {
2367 		pr_err("Failed to get pdd\n");
2368 		return -ENODEV;
2369 	}
2370 
2371 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2372 	if (ret)
2373 		return ret;
2374 
2375 	/* now map these BOs to GPU/s */
2376 	for (j = 0; j < p->n_pdds; j++) {
2377 		struct kfd_node *peer;
2378 		struct kfd_process_device *peer_pdd;
2379 
2380 		if (!bo_priv->mapped_gpuids[j])
2381 			break;
2382 
2383 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2384 		if (!peer_pdd)
2385 			return -EINVAL;
2386 
2387 		peer = peer_pdd->dev;
2388 
2389 		peer_pdd = kfd_bind_process_to_device(peer, p);
2390 		if (IS_ERR(peer_pdd))
2391 			return PTR_ERR(peer_pdd);
2392 
2393 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2394 							    peer_pdd->drm_priv);
2395 		if (ret) {
2396 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2397 			return ret;
2398 		}
2399 	}
2400 
2401 	pr_debug("map memory was successful for the BO\n");
2402 	/* create the dmabuf object and export the bo */
2403 	if (bo_bucket->alloc_flags
2404 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2405 		ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2406 					    &bo_bucket->dmabuf_fd);
2407 		if (ret)
2408 			return ret;
2409 	} else {
2410 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2411 	}
2412 
2413 	return 0;
2414 }
2415 
2416 static int criu_restore_bos(struct kfd_process *p,
2417 			    struct kfd_ioctl_criu_args *args,
2418 			    uint64_t *priv_offset,
2419 			    uint64_t max_priv_data_size)
2420 {
2421 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2422 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2423 	int ret = 0;
2424 	uint32_t i = 0;
2425 
2426 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2427 		return -EINVAL;
2428 
2429 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2430 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2431 
2432 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2433 	if (!bo_buckets)
2434 		return -ENOMEM;
2435 
2436 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2437 			     args->num_bos * sizeof(*bo_buckets));
2438 	if (ret) {
2439 		pr_err("Failed to copy BOs information from user\n");
2440 		ret = -EFAULT;
2441 		goto exit;
2442 	}
2443 
2444 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2445 	if (!bo_privs) {
2446 		ret = -ENOMEM;
2447 		goto exit;
2448 	}
2449 
2450 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2451 			     args->num_bos * sizeof(*bo_privs));
2452 	if (ret) {
2453 		pr_err("Failed to copy BOs information from user\n");
2454 		ret = -EFAULT;
2455 		goto exit;
2456 	}
2457 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2458 
2459 	/* Create and map new BOs */
2460 	for (; i < args->num_bos; i++) {
2461 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2462 		if (ret) {
2463 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2464 			goto exit;
2465 		}
2466 	} /* done */
2467 
2468 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2469 	ret = copy_to_user((void __user *)args->bos,
2470 				bo_buckets,
2471 				(args->num_bos * sizeof(*bo_buckets)));
2472 	if (ret)
2473 		ret = -EFAULT;
2474 
2475 exit:
2476 	while (ret && i--) {
2477 		if (bo_buckets[i].alloc_flags
2478 		   & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2479 			close_fd(bo_buckets[i].dmabuf_fd);
2480 	}
2481 	kvfree(bo_buckets);
2482 	kvfree(bo_privs);
2483 	return ret;
2484 }
2485 
2486 static int criu_restore_objects(struct file *filep,
2487 				struct kfd_process *p,
2488 				struct kfd_ioctl_criu_args *args,
2489 				uint64_t *priv_offset,
2490 				uint64_t max_priv_data_size)
2491 {
2492 	int ret = 0;
2493 	uint32_t i;
2494 
2495 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2496 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2497 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2498 
2499 	for (i = 0; i < args->num_objects; i++) {
2500 		uint32_t object_type;
2501 
2502 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2503 			pr_err("Invalid private data size\n");
2504 			return -EINVAL;
2505 		}
2506 
2507 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2508 		if (ret) {
2509 			pr_err("Failed to copy private information from user\n");
2510 			goto exit;
2511 		}
2512 
2513 		switch (object_type) {
2514 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2515 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2516 						     priv_offset, max_priv_data_size);
2517 			if (ret)
2518 				goto exit;
2519 			break;
2520 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2521 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2522 						     priv_offset, max_priv_data_size);
2523 			if (ret)
2524 				goto exit;
2525 			break;
2526 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2527 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2528 						     priv_offset, max_priv_data_size);
2529 			if (ret)
2530 				goto exit;
2531 			break;
2532 		default:
2533 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2534 			ret = -EINVAL;
2535 			goto exit;
2536 		}
2537 	}
2538 exit:
2539 	return ret;
2540 }
2541 
2542 static int criu_restore(struct file *filep,
2543 			struct kfd_process *p,
2544 			struct kfd_ioctl_criu_args *args)
2545 {
2546 	uint64_t priv_offset = 0;
2547 	int ret = 0;
2548 
2549 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2550 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2551 
2552 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2553 	    !args->num_devices || !args->num_bos)
2554 		return -EINVAL;
2555 
2556 	mutex_lock(&p->mutex);
2557 
2558 	/*
2559 	 * Set the process to evicted state to avoid running any new queues before all the memory
2560 	 * mappings are ready.
2561 	 */
2562 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2563 	if (ret)
2564 		goto exit_unlock;
2565 
2566 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2567 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2568 	if (ret)
2569 		goto exit_unlock;
2570 
2571 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2572 	if (ret)
2573 		goto exit_unlock;
2574 
2575 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2576 	if (ret)
2577 		goto exit_unlock;
2578 
2579 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2580 	if (ret)
2581 		goto exit_unlock;
2582 
2583 	if (priv_offset != args->priv_data_size) {
2584 		pr_err("Invalid private data size\n");
2585 		ret = -EINVAL;
2586 	}
2587 
2588 exit_unlock:
2589 	mutex_unlock(&p->mutex);
2590 	if (ret)
2591 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2592 	else
2593 		pr_debug("CRIU restore successful\n");
2594 
2595 	return ret;
2596 }
2597 
2598 static int criu_unpause(struct file *filep,
2599 			struct kfd_process *p,
2600 			struct kfd_ioctl_criu_args *args)
2601 {
2602 	int ret;
2603 
2604 	mutex_lock(&p->mutex);
2605 
2606 	if (!p->queues_paused) {
2607 		mutex_unlock(&p->mutex);
2608 		return -EINVAL;
2609 	}
2610 
2611 	ret = kfd_process_restore_queues(p);
2612 	if (ret)
2613 		pr_err("Failed to unpause queues ret:%d\n", ret);
2614 	else
2615 		p->queues_paused = false;
2616 
2617 	mutex_unlock(&p->mutex);
2618 
2619 	return ret;
2620 }
2621 
2622 static int criu_resume(struct file *filep,
2623 			struct kfd_process *p,
2624 			struct kfd_ioctl_criu_args *args)
2625 {
2626 	struct kfd_process *target = NULL;
2627 	struct pid *pid = NULL;
2628 	int ret = 0;
2629 
2630 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2631 		 args->pid);
2632 
2633 	pid = find_get_pid(args->pid);
2634 	if (!pid) {
2635 		pr_err("Cannot find pid info for %i\n", args->pid);
2636 		return -ESRCH;
2637 	}
2638 
2639 	pr_debug("calling kfd_lookup_process_by_pid\n");
2640 	target = kfd_lookup_process_by_pid(pid);
2641 
2642 	put_pid(pid);
2643 
2644 	if (!target) {
2645 		pr_debug("Cannot find process info for %i\n", args->pid);
2646 		return -ESRCH;
2647 	}
2648 
2649 	mutex_lock(&target->mutex);
2650 	ret = kfd_criu_resume_svm(target);
2651 	if (ret) {
2652 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2653 		goto exit;
2654 	}
2655 
2656 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2657 	if (ret)
2658 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2659 
2660 exit:
2661 	mutex_unlock(&target->mutex);
2662 
2663 	kfd_unref_process(target);
2664 	return ret;
2665 }
2666 
2667 static int criu_process_info(struct file *filep,
2668 				struct kfd_process *p,
2669 				struct kfd_ioctl_criu_args *args)
2670 {
2671 	int ret = 0;
2672 
2673 	mutex_lock(&p->mutex);
2674 
2675 	if (!p->n_pdds) {
2676 		pr_err("No pdd for given process\n");
2677 		ret = -ENODEV;
2678 		goto err_unlock;
2679 	}
2680 
2681 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2682 	if (ret)
2683 		goto err_unlock;
2684 
2685 	p->queues_paused = true;
2686 
2687 	args->pid = task_pid_nr_ns(p->lead_thread,
2688 					task_active_pid_ns(p->lead_thread));
2689 
2690 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2691 					   &args->num_objects, &args->priv_data_size);
2692 	if (ret)
2693 		goto err_unlock;
2694 
2695 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2696 				args->num_devices, args->num_bos, args->num_objects,
2697 				args->priv_data_size);
2698 
2699 err_unlock:
2700 	if (ret) {
2701 		kfd_process_restore_queues(p);
2702 		p->queues_paused = false;
2703 	}
2704 	mutex_unlock(&p->mutex);
2705 	return ret;
2706 }
2707 
2708 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2709 {
2710 	struct kfd_ioctl_criu_args *args = data;
2711 	int ret;
2712 
2713 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2714 	switch (args->op) {
2715 	case KFD_CRIU_OP_PROCESS_INFO:
2716 		ret = criu_process_info(filep, p, args);
2717 		break;
2718 	case KFD_CRIU_OP_CHECKPOINT:
2719 		ret = criu_checkpoint(filep, p, args);
2720 		break;
2721 	case KFD_CRIU_OP_UNPAUSE:
2722 		ret = criu_unpause(filep, p, args);
2723 		break;
2724 	case KFD_CRIU_OP_RESTORE:
2725 		ret = criu_restore(filep, p, args);
2726 		break;
2727 	case KFD_CRIU_OP_RESUME:
2728 		ret = criu_resume(filep, p, args);
2729 		break;
2730 	default:
2731 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2732 		ret = -EINVAL;
2733 		break;
2734 	}
2735 
2736 	if (ret)
2737 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2738 
2739 	return ret;
2740 }
2741 
2742 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2743 			bool enable_ttmp_setup)
2744 {
2745 	int i = 0, ret = 0;
2746 
2747 	if (p->is_runtime_retry)
2748 		goto retry;
2749 
2750 	if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2751 		return -EBUSY;
2752 
2753 	for (i = 0; i < p->n_pdds; i++) {
2754 		struct kfd_process_device *pdd = p->pdds[i];
2755 
2756 		if (pdd->qpd.queue_count)
2757 			return -EEXIST;
2758 	}
2759 
2760 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2761 	p->runtime_info.r_debug = r_debug;
2762 	p->runtime_info.ttmp_setup = enable_ttmp_setup;
2763 
2764 	if (p->runtime_info.ttmp_setup) {
2765 		for (i = 0; i < p->n_pdds; i++) {
2766 			struct kfd_process_device *pdd = p->pdds[i];
2767 
2768 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2769 				amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2770 				pdd->dev->kfd2kgd->enable_debug_trap(
2771 						pdd->dev->adev,
2772 						true,
2773 						pdd->dev->vm_info.last_vmid_kfd);
2774 			} else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2775 				pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2776 						pdd->dev->adev,
2777 						false,
2778 						0);
2779 			}
2780 		}
2781 	}
2782 
2783 retry:
2784 	if (p->debug_trap_enabled) {
2785 		if (!p->is_runtime_retry) {
2786 			kfd_dbg_trap_activate(p);
2787 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2788 					p, NULL, 0, false, NULL, 0);
2789 		}
2790 
2791 		mutex_unlock(&p->mutex);
2792 		ret = down_interruptible(&p->runtime_enable_sema);
2793 		mutex_lock(&p->mutex);
2794 
2795 		p->is_runtime_retry = !!ret;
2796 	}
2797 
2798 	return ret;
2799 }
2800 
2801 static int runtime_disable(struct kfd_process *p)
2802 {
2803 	int i = 0, ret;
2804 	bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2805 
2806 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2807 	p->runtime_info.r_debug = 0;
2808 
2809 	if (p->debug_trap_enabled) {
2810 		if (was_enabled)
2811 			kfd_dbg_trap_deactivate(p, false, 0);
2812 
2813 		if (!p->is_runtime_retry)
2814 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2815 					p, NULL, 0, false, NULL, 0);
2816 
2817 		mutex_unlock(&p->mutex);
2818 		ret = down_interruptible(&p->runtime_enable_sema);
2819 		mutex_lock(&p->mutex);
2820 
2821 		p->is_runtime_retry = !!ret;
2822 		if (ret)
2823 			return ret;
2824 	}
2825 
2826 	if (was_enabled && p->runtime_info.ttmp_setup) {
2827 		for (i = 0; i < p->n_pdds; i++) {
2828 			struct kfd_process_device *pdd = p->pdds[i];
2829 
2830 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2831 				amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2832 		}
2833 	}
2834 
2835 	p->runtime_info.ttmp_setup = false;
2836 
2837 	/* disable ttmp setup */
2838 	for (i = 0; i < p->n_pdds; i++) {
2839 		struct kfd_process_device *pdd = p->pdds[i];
2840 
2841 		if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2842 			pdd->spi_dbg_override =
2843 					pdd->dev->kfd2kgd->disable_debug_trap(
2844 					pdd->dev->adev,
2845 					false,
2846 					pdd->dev->vm_info.last_vmid_kfd);
2847 
2848 			if (!pdd->dev->kfd->shared_resources.enable_mes)
2849 				debug_refresh_runlist(pdd->dev->dqm);
2850 			else
2851 				kfd_dbg_set_mes_debug_mode(pdd);
2852 		}
2853 	}
2854 
2855 	return 0;
2856 }
2857 
2858 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2859 {
2860 	struct kfd_ioctl_runtime_enable_args *args = data;
2861 	int r;
2862 
2863 	mutex_lock(&p->mutex);
2864 
2865 	if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2866 		r = runtime_enable(p, args->r_debug,
2867 				!!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2868 	else
2869 		r = runtime_disable(p);
2870 
2871 	mutex_unlock(&p->mutex);
2872 
2873 	return r;
2874 }
2875 
2876 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2877 {
2878 	struct kfd_ioctl_dbg_trap_args *args = data;
2879 	struct task_struct *thread = NULL;
2880 	struct mm_struct *mm = NULL;
2881 	struct pid *pid = NULL;
2882 	struct kfd_process *target = NULL;
2883 	struct kfd_process_device *pdd = NULL;
2884 	int r = 0;
2885 
2886 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2887 		pr_err("Debugging does not support sched_policy %i", sched_policy);
2888 		return -EINVAL;
2889 	}
2890 
2891 	pid = find_get_pid(args->pid);
2892 	if (!pid) {
2893 		pr_debug("Cannot find pid info for %i\n", args->pid);
2894 		r = -ESRCH;
2895 		goto out;
2896 	}
2897 
2898 	thread = get_pid_task(pid, PIDTYPE_PID);
2899 	if (!thread) {
2900 		r = -ESRCH;
2901 		goto out;
2902 	}
2903 
2904 	mm = get_task_mm(thread);
2905 	if (!mm) {
2906 		r = -ESRCH;
2907 		goto out;
2908 	}
2909 
2910 	if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2911 		bool create_process;
2912 
2913 		rcu_read_lock();
2914 		create_process = thread && thread != current && ptrace_parent(thread) == current;
2915 		rcu_read_unlock();
2916 
2917 		target = create_process ? kfd_create_process(thread) :
2918 					kfd_lookup_process_by_pid(pid);
2919 	} else {
2920 		target = kfd_lookup_process_by_pid(pid);
2921 	}
2922 
2923 	if (IS_ERR_OR_NULL(target)) {
2924 		pr_debug("Cannot find process PID %i to debug\n", args->pid);
2925 		r = target ? PTR_ERR(target) : -ESRCH;
2926 		goto out;
2927 	}
2928 
2929 	/* Check if target is still PTRACED. */
2930 	rcu_read_lock();
2931 	if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2932 				&& ptrace_parent(target->lead_thread) != current) {
2933 		pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2934 		r = -EPERM;
2935 	}
2936 	rcu_read_unlock();
2937 
2938 	if (r)
2939 		goto out;
2940 
2941 	mutex_lock(&target->mutex);
2942 
2943 	if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2944 		pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2945 		r = -EINVAL;
2946 		goto unlock_out;
2947 	}
2948 
2949 	if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2950 			(args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2951 			 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2952 			 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2953 			 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2954 			 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2955 			 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2956 			 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2957 		r = -EPERM;
2958 		goto unlock_out;
2959 	}
2960 
2961 	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2962 	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2963 		int user_gpu_id = kfd_process_get_user_gpu_id(target,
2964 				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2965 					args->set_node_address_watch.gpu_id :
2966 					args->clear_node_address_watch.gpu_id);
2967 
2968 		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
2969 		if (user_gpu_id == -EINVAL || !pdd) {
2970 			r = -ENODEV;
2971 			goto unlock_out;
2972 		}
2973 	}
2974 
2975 	switch (args->op) {
2976 	case KFD_IOC_DBG_TRAP_ENABLE:
2977 		if (target != p)
2978 			target->debugger_process = p;
2979 
2980 		r = kfd_dbg_trap_enable(target,
2981 					args->enable.dbg_fd,
2982 					(void __user *)args->enable.rinfo_ptr,
2983 					&args->enable.rinfo_size);
2984 		if (!r)
2985 			target->exception_enable_mask = args->enable.exception_mask;
2986 
2987 		break;
2988 	case KFD_IOC_DBG_TRAP_DISABLE:
2989 		r = kfd_dbg_trap_disable(target);
2990 		break;
2991 	case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
2992 		r = kfd_dbg_send_exception_to_runtime(target,
2993 				args->send_runtime_event.gpu_id,
2994 				args->send_runtime_event.queue_id,
2995 				args->send_runtime_event.exception_mask);
2996 		break;
2997 	case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
2998 		kfd_dbg_set_enabled_debug_exception_mask(target,
2999 				args->set_exceptions_enabled.exception_mask);
3000 		break;
3001 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
3002 		r = kfd_dbg_trap_set_wave_launch_override(target,
3003 				args->launch_override.override_mode,
3004 				args->launch_override.enable_mask,
3005 				args->launch_override.support_request_mask,
3006 				&args->launch_override.enable_mask,
3007 				&args->launch_override.support_request_mask);
3008 		break;
3009 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
3010 		r = kfd_dbg_trap_set_wave_launch_mode(target,
3011 				args->launch_mode.launch_mode);
3012 		break;
3013 	case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
3014 		r = suspend_queues(target,
3015 				args->suspend_queues.num_queues,
3016 				args->suspend_queues.grace_period,
3017 				args->suspend_queues.exception_mask,
3018 				(uint32_t *)args->suspend_queues.queue_array_ptr);
3019 
3020 		break;
3021 	case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
3022 		r = resume_queues(target, args->resume_queues.num_queues,
3023 				(uint32_t *)args->resume_queues.queue_array_ptr);
3024 		break;
3025 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3026 		r = kfd_dbg_trap_set_dev_address_watch(pdd,
3027 				args->set_node_address_watch.address,
3028 				args->set_node_address_watch.mask,
3029 				&args->set_node_address_watch.id,
3030 				args->set_node_address_watch.mode);
3031 		break;
3032 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3033 		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3034 				args->clear_node_address_watch.id);
3035 		break;
3036 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
3037 		r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3038 		break;
3039 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3040 		r = kfd_dbg_ev_query_debug_event(target,
3041 				&args->query_debug_event.queue_id,
3042 				&args->query_debug_event.gpu_id,
3043 				args->query_debug_event.exception_mask,
3044 				&args->query_debug_event.exception_mask);
3045 		break;
3046 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3047 		r = kfd_dbg_trap_query_exception_info(target,
3048 				args->query_exception_info.source_id,
3049 				args->query_exception_info.exception_code,
3050 				args->query_exception_info.clear_exception,
3051 				(void __user *)args->query_exception_info.info_ptr,
3052 				&args->query_exception_info.info_size);
3053 		break;
3054 	case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3055 		r = pqm_get_queue_snapshot(&target->pqm,
3056 				args->queue_snapshot.exception_mask,
3057 				(void __user *)args->queue_snapshot.snapshot_buf_ptr,
3058 				&args->queue_snapshot.num_queues,
3059 				&args->queue_snapshot.entry_size);
3060 		break;
3061 	case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3062 		r = kfd_dbg_trap_device_snapshot(target,
3063 				args->device_snapshot.exception_mask,
3064 				(void __user *)args->device_snapshot.snapshot_buf_ptr,
3065 				&args->device_snapshot.num_devices,
3066 				&args->device_snapshot.entry_size);
3067 		break;
3068 	default:
3069 		pr_err("Invalid option: %i\n", args->op);
3070 		r = -EINVAL;
3071 	}
3072 
3073 unlock_out:
3074 	mutex_unlock(&target->mutex);
3075 
3076 out:
3077 	if (thread)
3078 		put_task_struct(thread);
3079 
3080 	if (mm)
3081 		mmput(mm);
3082 
3083 	if (pid)
3084 		put_pid(pid);
3085 
3086 	if (target)
3087 		kfd_unref_process(target);
3088 
3089 	return r;
3090 }
3091 
3092 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3093 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3094 			    .cmd_drv = 0, .name = #ioctl}
3095 
3096 /** Ioctl table */
3097 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3098 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3099 			kfd_ioctl_get_version, 0),
3100 
3101 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3102 			kfd_ioctl_create_queue, 0),
3103 
3104 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3105 			kfd_ioctl_destroy_queue, 0),
3106 
3107 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3108 			kfd_ioctl_set_memory_policy, 0),
3109 
3110 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3111 			kfd_ioctl_get_clock_counters, 0),
3112 
3113 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3114 			kfd_ioctl_get_process_apertures, 0),
3115 
3116 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3117 			kfd_ioctl_update_queue, 0),
3118 
3119 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3120 			kfd_ioctl_create_event, 0),
3121 
3122 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3123 			kfd_ioctl_destroy_event, 0),
3124 
3125 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3126 			kfd_ioctl_set_event, 0),
3127 
3128 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3129 			kfd_ioctl_reset_event, 0),
3130 
3131 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3132 			kfd_ioctl_wait_events, 0),
3133 
3134 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3135 			kfd_ioctl_dbg_register, 0),
3136 
3137 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3138 			kfd_ioctl_dbg_unregister, 0),
3139 
3140 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3141 			kfd_ioctl_dbg_address_watch, 0),
3142 
3143 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3144 			kfd_ioctl_dbg_wave_control, 0),
3145 
3146 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3147 			kfd_ioctl_set_scratch_backing_va, 0),
3148 
3149 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3150 			kfd_ioctl_get_tile_config, 0),
3151 
3152 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3153 			kfd_ioctl_set_trap_handler, 0),
3154 
3155 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3156 			kfd_ioctl_get_process_apertures_new, 0),
3157 
3158 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3159 			kfd_ioctl_acquire_vm, 0),
3160 
3161 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3162 			kfd_ioctl_alloc_memory_of_gpu, 0),
3163 
3164 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3165 			kfd_ioctl_free_memory_of_gpu, 0),
3166 
3167 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3168 			kfd_ioctl_map_memory_to_gpu, 0),
3169 
3170 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3171 			kfd_ioctl_unmap_memory_from_gpu, 0),
3172 
3173 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3174 			kfd_ioctl_set_cu_mask, 0),
3175 
3176 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3177 			kfd_ioctl_get_queue_wave_state, 0),
3178 
3179 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3180 				kfd_ioctl_get_dmabuf_info, 0),
3181 
3182 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3183 				kfd_ioctl_import_dmabuf, 0),
3184 
3185 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3186 			kfd_ioctl_alloc_queue_gws, 0),
3187 
3188 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3189 			kfd_ioctl_smi_events, 0),
3190 
3191 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3192 
3193 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3194 			kfd_ioctl_set_xnack_mode, 0),
3195 
3196 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3197 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3198 
3199 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3200 			kfd_ioctl_get_available_memory, 0),
3201 
3202 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3203 				kfd_ioctl_export_dmabuf, 0),
3204 
3205 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3206 			kfd_ioctl_runtime_enable, 0),
3207 
3208 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3209 			kfd_ioctl_set_debug_trap, 0),
3210 };
3211 
3212 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
3213 
3214 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3215 {
3216 	struct kfd_process *process;
3217 	amdkfd_ioctl_t *func;
3218 	const struct amdkfd_ioctl_desc *ioctl = NULL;
3219 	unsigned int nr = _IOC_NR(cmd);
3220 	char stack_kdata[128];
3221 	char *kdata = NULL;
3222 	unsigned int usize, asize;
3223 	int retcode = -EINVAL;
3224 	bool ptrace_attached = false;
3225 
3226 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3227 		goto err_i1;
3228 
3229 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3230 		u32 amdkfd_size;
3231 
3232 		ioctl = &amdkfd_ioctls[nr];
3233 
3234 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
3235 		usize = asize = _IOC_SIZE(cmd);
3236 		if (amdkfd_size > asize)
3237 			asize = amdkfd_size;
3238 
3239 		cmd = ioctl->cmd;
3240 	} else
3241 		goto err_i1;
3242 
3243 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3244 
3245 	/* Get the process struct from the filep. Only the process
3246 	 * that opened /dev/kfd can use the file descriptor. Child
3247 	 * processes need to create their own KFD device context.
3248 	 */
3249 	process = filep->private_data;
3250 
3251 	rcu_read_lock();
3252 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3253 	    ptrace_parent(process->lead_thread) == current)
3254 		ptrace_attached = true;
3255 	rcu_read_unlock();
3256 
3257 	if (process->lead_thread != current->group_leader
3258 	    && !ptrace_attached) {
3259 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3260 		retcode = -EBADF;
3261 		goto err_i1;
3262 	}
3263 
3264 	/* Do not trust userspace, use our own definition */
3265 	func = ioctl->func;
3266 
3267 	if (unlikely(!func)) {
3268 		dev_dbg(kfd_device, "no function\n");
3269 		retcode = -EINVAL;
3270 		goto err_i1;
3271 	}
3272 
3273 	/*
3274 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3275 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3276 	 * more priviledged access.
3277 	 */
3278 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3279 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
3280 						!capable(CAP_SYS_ADMIN)) {
3281 			retcode = -EACCES;
3282 			goto err_i1;
3283 		}
3284 	}
3285 
3286 	if (cmd & (IOC_IN | IOC_OUT)) {
3287 		if (asize <= sizeof(stack_kdata)) {
3288 			kdata = stack_kdata;
3289 		} else {
3290 			kdata = kmalloc(asize, GFP_KERNEL);
3291 			if (!kdata) {
3292 				retcode = -ENOMEM;
3293 				goto err_i1;
3294 			}
3295 		}
3296 		if (asize > usize)
3297 			memset(kdata + usize, 0, asize - usize);
3298 	}
3299 
3300 	if (cmd & IOC_IN) {
3301 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3302 			retcode = -EFAULT;
3303 			goto err_i1;
3304 		}
3305 	} else if (cmd & IOC_OUT) {
3306 		memset(kdata, 0, usize);
3307 	}
3308 
3309 	retcode = func(filep, process, kdata);
3310 
3311 	if (cmd & IOC_OUT)
3312 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3313 			retcode = -EFAULT;
3314 
3315 err_i1:
3316 	if (!ioctl)
3317 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3318 			  task_pid_nr(current), cmd, nr);
3319 
3320 	if (kdata != stack_kdata)
3321 		kfree(kdata);
3322 
3323 	if (retcode)
3324 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3325 				nr, arg, retcode);
3326 
3327 	return retcode;
3328 }
3329 
3330 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3331 		      struct vm_area_struct *vma)
3332 {
3333 	phys_addr_t address;
3334 
3335 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3336 		return -EINVAL;
3337 
3338 	address = dev->adev->rmmio_remap.bus_addr;
3339 
3340 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3341 				VM_DONTDUMP | VM_PFNMAP);
3342 
3343 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3344 
3345 	pr_debug("pasid 0x%x mapping mmio page\n"
3346 		 "     target user address == 0x%08llX\n"
3347 		 "     physical address    == 0x%08llX\n"
3348 		 "     vm_flags            == 0x%04lX\n"
3349 		 "     size                == 0x%04lX\n",
3350 		 process->pasid, (unsigned long long) vma->vm_start,
3351 		 address, vma->vm_flags, PAGE_SIZE);
3352 
3353 	return io_remap_pfn_range(vma,
3354 				vma->vm_start,
3355 				address >> PAGE_SHIFT,
3356 				PAGE_SIZE,
3357 				vma->vm_page_prot);
3358 }
3359 
3360 
3361 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3362 {
3363 	struct kfd_process *process;
3364 	struct kfd_node *dev = NULL;
3365 	unsigned long mmap_offset;
3366 	unsigned int gpu_id;
3367 
3368 	process = kfd_get_process(current);
3369 	if (IS_ERR(process))
3370 		return PTR_ERR(process);
3371 
3372 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3373 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3374 	if (gpu_id)
3375 		dev = kfd_device_by_id(gpu_id);
3376 
3377 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3378 	case KFD_MMAP_TYPE_DOORBELL:
3379 		if (!dev)
3380 			return -ENODEV;
3381 		return kfd_doorbell_mmap(dev, process, vma);
3382 
3383 	case KFD_MMAP_TYPE_EVENTS:
3384 		return kfd_event_mmap(process, vma);
3385 
3386 	case KFD_MMAP_TYPE_RESERVED_MEM:
3387 		if (!dev)
3388 			return -ENODEV;
3389 		return kfd_reserved_mem_mmap(dev, process, vma);
3390 	case KFD_MMAP_TYPE_MMIO:
3391 		if (!dev)
3392 			return -ENODEV;
3393 		return kfd_mmio_mmap(dev, process, vma);
3394 	}
3395 
3396 	return -EFAULT;
3397 }
3398