xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c (revision e6a901a00822659181c93c86d8bbc2a17779fddc)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include <linux/device.h>
25 #include <linux/export.h>
26 #include <linux/err.h>
27 #include <linux/fs.h>
28 #include <linux/file.h>
29 #include <linux/sched.h>
30 #include <linux/slab.h>
31 #include <linux/uaccess.h>
32 #include <linux/compat.h>
33 #include <uapi/linux/kfd_ioctl.h>
34 #include <linux/time.h>
35 #include <linux/mm.h>
36 #include <linux/mman.h>
37 #include <linux/ptrace.h>
38 #include <linux/dma-buf.h>
39 #include <linux/fdtable.h>
40 #include <linux/processor.h>
41 #include "kfd_priv.h"
42 #include "kfd_device_queue_manager.h"
43 #include "kfd_svm.h"
44 #include "amdgpu_amdkfd.h"
45 #include "kfd_smi_events.h"
46 #include "amdgpu_dma_buf.h"
47 #include "kfd_debug.h"
48 
49 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
50 static int kfd_open(struct inode *, struct file *);
51 static int kfd_release(struct inode *, struct file *);
52 static int kfd_mmap(struct file *, struct vm_area_struct *);
53 
54 static const char kfd_dev_name[] = "kfd";
55 
56 static const struct file_operations kfd_fops = {
57 	.owner = THIS_MODULE,
58 	.unlocked_ioctl = kfd_ioctl,
59 	.compat_ioctl = compat_ptr_ioctl,
60 	.open = kfd_open,
61 	.release = kfd_release,
62 	.mmap = kfd_mmap,
63 };
64 
65 static int kfd_char_dev_major = -1;
66 struct device *kfd_device;
67 static const struct class kfd_class = {
68 	.name = kfd_dev_name,
69 };
70 
71 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
72 {
73 	struct kfd_process_device *pdd;
74 
75 	mutex_lock(&p->mutex);
76 	pdd = kfd_process_device_data_by_id(p, gpu_id);
77 
78 	if (pdd)
79 		return pdd;
80 
81 	mutex_unlock(&p->mutex);
82 	return NULL;
83 }
84 
85 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
86 {
87 	mutex_unlock(&pdd->process->mutex);
88 }
89 
90 int kfd_chardev_init(void)
91 {
92 	int err = 0;
93 
94 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
95 	err = kfd_char_dev_major;
96 	if (err < 0)
97 		goto err_register_chrdev;
98 
99 	err = class_register(&kfd_class);
100 	if (err)
101 		goto err_class_create;
102 
103 	kfd_device = device_create(&kfd_class, NULL,
104 				   MKDEV(kfd_char_dev_major, 0),
105 				   NULL, kfd_dev_name);
106 	err = PTR_ERR(kfd_device);
107 	if (IS_ERR(kfd_device))
108 		goto err_device_create;
109 
110 	return 0;
111 
112 err_device_create:
113 	class_unregister(&kfd_class);
114 err_class_create:
115 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
116 err_register_chrdev:
117 	return err;
118 }
119 
120 void kfd_chardev_exit(void)
121 {
122 	device_destroy(&kfd_class, MKDEV(kfd_char_dev_major, 0));
123 	class_unregister(&kfd_class);
124 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
125 	kfd_device = NULL;
126 }
127 
128 
129 static int kfd_open(struct inode *inode, struct file *filep)
130 {
131 	struct kfd_process *process;
132 	bool is_32bit_user_mode;
133 
134 	if (iminor(inode) != 0)
135 		return -ENODEV;
136 
137 	is_32bit_user_mode = in_compat_syscall();
138 
139 	if (is_32bit_user_mode) {
140 		dev_warn(kfd_device,
141 			"Process %d (32-bit) failed to open /dev/kfd\n"
142 			"32-bit processes are not supported by amdkfd\n",
143 			current->pid);
144 		return -EPERM;
145 	}
146 
147 	process = kfd_create_process(current);
148 	if (IS_ERR(process))
149 		return PTR_ERR(process);
150 
151 	if (kfd_process_init_cwsr_apu(process, filep)) {
152 		kfd_unref_process(process);
153 		return -EFAULT;
154 	}
155 
156 	/* filep now owns the reference returned by kfd_create_process */
157 	filep->private_data = process;
158 
159 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
160 		process->pasid, process->is_32bit_user_mode);
161 
162 	return 0;
163 }
164 
165 static int kfd_release(struct inode *inode, struct file *filep)
166 {
167 	struct kfd_process *process = filep->private_data;
168 
169 	if (process)
170 		kfd_unref_process(process);
171 
172 	return 0;
173 }
174 
175 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
176 					void *data)
177 {
178 	struct kfd_ioctl_get_version_args *args = data;
179 
180 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
181 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
182 
183 	return 0;
184 }
185 
186 static int set_queue_properties_from_user(struct queue_properties *q_properties,
187 				struct kfd_ioctl_create_queue_args *args)
188 {
189 	/*
190 	 * Repurpose queue percentage to accommodate new features:
191 	 * bit 0-7: queue percentage
192 	 * bit 8-15: pm4_target_xcc
193 	 */
194 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
195 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
196 		return -EINVAL;
197 	}
198 
199 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
200 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
201 		return -EINVAL;
202 	}
203 
204 	if ((args->ring_base_address) &&
205 		(!access_ok((const void __user *) args->ring_base_address,
206 			sizeof(uint64_t)))) {
207 		pr_err("Can't access ring base address\n");
208 		return -EFAULT;
209 	}
210 
211 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
212 		pr_err("Ring size must be a power of 2 or 0\n");
213 		return -EINVAL;
214 	}
215 
216 	if (!access_ok((const void __user *) args->read_pointer_address,
217 			sizeof(uint32_t))) {
218 		pr_err("Can't access read pointer\n");
219 		return -EFAULT;
220 	}
221 
222 	if (!access_ok((const void __user *) args->write_pointer_address,
223 			sizeof(uint32_t))) {
224 		pr_err("Can't access write pointer\n");
225 		return -EFAULT;
226 	}
227 
228 	if (args->eop_buffer_address &&
229 		!access_ok((const void __user *) args->eop_buffer_address,
230 			sizeof(uint32_t))) {
231 		pr_debug("Can't access eop buffer");
232 		return -EFAULT;
233 	}
234 
235 	if (args->ctx_save_restore_address &&
236 		!access_ok((const void __user *) args->ctx_save_restore_address,
237 			sizeof(uint32_t))) {
238 		pr_debug("Can't access ctx save restore buffer");
239 		return -EFAULT;
240 	}
241 
242 	q_properties->is_interop = false;
243 	q_properties->is_gws = false;
244 	q_properties->queue_percent = args->queue_percentage & 0xFF;
245 	/* bit 8-15 are repurposed to be PM4 target XCC */
246 	q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
247 	q_properties->priority = args->queue_priority;
248 	q_properties->queue_address = args->ring_base_address;
249 	q_properties->queue_size = args->ring_size;
250 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
251 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
252 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
253 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
254 	q_properties->ctx_save_restore_area_address =
255 			args->ctx_save_restore_address;
256 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
257 	q_properties->ctl_stack_size = args->ctl_stack_size;
258 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
259 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
260 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
261 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
262 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
263 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
264 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
265 	else
266 		return -ENOTSUPP;
267 
268 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
269 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
270 	else
271 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
272 
273 	pr_debug("Queue Percentage: %d, %d\n",
274 			q_properties->queue_percent, args->queue_percentage);
275 
276 	pr_debug("Queue Priority: %d, %d\n",
277 			q_properties->priority, args->queue_priority);
278 
279 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
280 			q_properties->queue_address, args->ring_base_address);
281 
282 	pr_debug("Queue Size: 0x%llX, %u\n",
283 			q_properties->queue_size, args->ring_size);
284 
285 	pr_debug("Queue r/w Pointers: %px, %px\n",
286 			q_properties->read_ptr,
287 			q_properties->write_ptr);
288 
289 	pr_debug("Queue Format: %d\n", q_properties->format);
290 
291 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
292 
293 	pr_debug("Queue CTX save area: 0x%llX\n",
294 			q_properties->ctx_save_restore_area_address);
295 
296 	return 0;
297 }
298 
299 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
300 					void *data)
301 {
302 	struct kfd_ioctl_create_queue_args *args = data;
303 	struct kfd_node *dev;
304 	int err = 0;
305 	unsigned int queue_id;
306 	struct kfd_process_device *pdd;
307 	struct queue_properties q_properties;
308 	uint32_t doorbell_offset_in_process = 0;
309 	struct amdgpu_bo *wptr_bo = NULL;
310 
311 	memset(&q_properties, 0, sizeof(struct queue_properties));
312 
313 	pr_debug("Creating queue ioctl\n");
314 
315 	err = set_queue_properties_from_user(&q_properties, args);
316 	if (err)
317 		return err;
318 
319 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
320 
321 	mutex_lock(&p->mutex);
322 
323 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
324 	if (!pdd) {
325 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
326 		err = -EINVAL;
327 		goto err_pdd;
328 	}
329 	dev = pdd->dev;
330 
331 	pdd = kfd_bind_process_to_device(dev, p);
332 	if (IS_ERR(pdd)) {
333 		err = -ESRCH;
334 		goto err_bind_process;
335 	}
336 
337 	if (!pdd->qpd.proc_doorbells) {
338 		err = kfd_alloc_process_doorbells(dev->kfd, pdd);
339 		if (err) {
340 			pr_debug("failed to allocate process doorbells\n");
341 			goto err_bind_process;
342 		}
343 	}
344 
345 	/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
346 	 * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
347 	 */
348 	if (dev->kfd->shared_resources.enable_mes &&
349 			((dev->adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK)
350 			>> AMDGPU_MES_API_VERSION_SHIFT) >= 2) {
351 		struct amdgpu_bo_va_mapping *wptr_mapping;
352 		struct amdgpu_vm *wptr_vm;
353 
354 		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
355 		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
356 		if (err)
357 			goto err_wptr_map_gart;
358 
359 		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
360 				wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
361 		amdgpu_bo_unreserve(wptr_vm->root.bo);
362 		if (!wptr_mapping) {
363 			pr_err("Failed to lookup wptr bo\n");
364 			err = -EINVAL;
365 			goto err_wptr_map_gart;
366 		}
367 
368 		wptr_bo = wptr_mapping->bo_va->base.bo;
369 		if (wptr_bo->tbo.base.size > PAGE_SIZE) {
370 			pr_err("Requested GART mapping for wptr bo larger than one page\n");
371 			err = -EINVAL;
372 			goto err_wptr_map_gart;
373 		}
374 		if (dev->adev != amdgpu_ttm_adev(wptr_bo->tbo.bdev)) {
375 			pr_err("Queue memory allocated to wrong device\n");
376 			err = -EINVAL;
377 			goto err_wptr_map_gart;
378 		}
379 
380 		err = amdgpu_amdkfd_map_gtt_bo_to_gart(wptr_bo);
381 		if (err) {
382 			pr_err("Failed to map wptr bo to GART\n");
383 			goto err_wptr_map_gart;
384 		}
385 	}
386 
387 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
388 			p->pasid,
389 			dev->id);
390 
391 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
392 			NULL, NULL, NULL, &doorbell_offset_in_process);
393 	if (err != 0)
394 		goto err_create_queue;
395 
396 	args->queue_id = queue_id;
397 
398 
399 	/* Return gpu_id as doorbell offset for mmap usage */
400 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
401 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
402 	if (KFD_IS_SOC15(dev))
403 		/* On SOC15 ASICs, include the doorbell offset within the
404 		 * process doorbell frame, which is 2 pages.
405 		 */
406 		args->doorbell_offset |= doorbell_offset_in_process;
407 
408 	mutex_unlock(&p->mutex);
409 
410 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
411 
412 	pr_debug("Ring buffer address == 0x%016llX\n",
413 			args->ring_base_address);
414 
415 	pr_debug("Read ptr address    == 0x%016llX\n",
416 			args->read_pointer_address);
417 
418 	pr_debug("Write ptr address   == 0x%016llX\n",
419 			args->write_pointer_address);
420 
421 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
422 	return 0;
423 
424 err_create_queue:
425 	if (wptr_bo)
426 		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
427 err_wptr_map_gart:
428 err_bind_process:
429 err_pdd:
430 	mutex_unlock(&p->mutex);
431 	return err;
432 }
433 
434 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
435 					void *data)
436 {
437 	int retval;
438 	struct kfd_ioctl_destroy_queue_args *args = data;
439 
440 	pr_debug("Destroying queue id %d for pasid 0x%x\n",
441 				args->queue_id,
442 				p->pasid);
443 
444 	mutex_lock(&p->mutex);
445 
446 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
447 
448 	mutex_unlock(&p->mutex);
449 	return retval;
450 }
451 
452 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
453 					void *data)
454 {
455 	int retval;
456 	struct kfd_ioctl_update_queue_args *args = data;
457 	struct queue_properties properties;
458 
459 	/*
460 	 * Repurpose queue percentage to accommodate new features:
461 	 * bit 0-7: queue percentage
462 	 * bit 8-15: pm4_target_xcc
463 	 */
464 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
465 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
466 		return -EINVAL;
467 	}
468 
469 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
470 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
471 		return -EINVAL;
472 	}
473 
474 	if ((args->ring_base_address) &&
475 		(!access_ok((const void __user *) args->ring_base_address,
476 			sizeof(uint64_t)))) {
477 		pr_err("Can't access ring base address\n");
478 		return -EFAULT;
479 	}
480 
481 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
482 		pr_err("Ring size must be a power of 2 or 0\n");
483 		return -EINVAL;
484 	}
485 
486 	properties.queue_address = args->ring_base_address;
487 	properties.queue_size = args->ring_size;
488 	properties.queue_percent = args->queue_percentage & 0xFF;
489 	/* bit 8-15 are repurposed to be PM4 target XCC */
490 	properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
491 	properties.priority = args->queue_priority;
492 
493 	pr_debug("Updating queue id %d for pasid 0x%x\n",
494 			args->queue_id, p->pasid);
495 
496 	mutex_lock(&p->mutex);
497 
498 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
499 
500 	mutex_unlock(&p->mutex);
501 
502 	return retval;
503 }
504 
505 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
506 					void *data)
507 {
508 	int retval;
509 	const int max_num_cus = 1024;
510 	struct kfd_ioctl_set_cu_mask_args *args = data;
511 	struct mqd_update_info minfo = {0};
512 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
513 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
514 
515 	if ((args->num_cu_mask % 32) != 0) {
516 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
517 				args->num_cu_mask);
518 		return -EINVAL;
519 	}
520 
521 	minfo.cu_mask.count = args->num_cu_mask;
522 	if (minfo.cu_mask.count == 0) {
523 		pr_debug("CU mask cannot be 0");
524 		return -EINVAL;
525 	}
526 
527 	/* To prevent an unreasonably large CU mask size, set an arbitrary
528 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
529 	 * past max_num_cus bits and just use the first max_num_cus bits.
530 	 */
531 	if (minfo.cu_mask.count > max_num_cus) {
532 		pr_debug("CU mask cannot be greater than 1024 bits");
533 		minfo.cu_mask.count = max_num_cus;
534 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
535 	}
536 
537 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
538 	if (!minfo.cu_mask.ptr)
539 		return -ENOMEM;
540 
541 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
542 	if (retval) {
543 		pr_debug("Could not copy CU mask from userspace");
544 		retval = -EFAULT;
545 		goto out;
546 	}
547 
548 	mutex_lock(&p->mutex);
549 
550 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
551 
552 	mutex_unlock(&p->mutex);
553 
554 out:
555 	kfree(minfo.cu_mask.ptr);
556 	return retval;
557 }
558 
559 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
560 					  struct kfd_process *p, void *data)
561 {
562 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
563 	int r;
564 
565 	mutex_lock(&p->mutex);
566 
567 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
568 			       (void __user *)args->ctl_stack_address,
569 			       &args->ctl_stack_used_size,
570 			       &args->save_area_used_size);
571 
572 	mutex_unlock(&p->mutex);
573 
574 	return r;
575 }
576 
577 static int kfd_ioctl_set_memory_policy(struct file *filep,
578 					struct kfd_process *p, void *data)
579 {
580 	struct kfd_ioctl_set_memory_policy_args *args = data;
581 	int err = 0;
582 	struct kfd_process_device *pdd;
583 	enum cache_policy default_policy, alternate_policy;
584 
585 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
586 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
587 		return -EINVAL;
588 	}
589 
590 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
591 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
592 		return -EINVAL;
593 	}
594 
595 	mutex_lock(&p->mutex);
596 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
597 	if (!pdd) {
598 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
599 		err = -EINVAL;
600 		goto err_pdd;
601 	}
602 
603 	pdd = kfd_bind_process_to_device(pdd->dev, p);
604 	if (IS_ERR(pdd)) {
605 		err = -ESRCH;
606 		goto out;
607 	}
608 
609 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
610 			 ? cache_policy_coherent : cache_policy_noncoherent;
611 
612 	alternate_policy =
613 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
614 		   ? cache_policy_coherent : cache_policy_noncoherent;
615 
616 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
617 				&pdd->qpd,
618 				default_policy,
619 				alternate_policy,
620 				(void __user *)args->alternate_aperture_base,
621 				args->alternate_aperture_size))
622 		err = -EINVAL;
623 
624 out:
625 err_pdd:
626 	mutex_unlock(&p->mutex);
627 
628 	return err;
629 }
630 
631 static int kfd_ioctl_set_trap_handler(struct file *filep,
632 					struct kfd_process *p, void *data)
633 {
634 	struct kfd_ioctl_set_trap_handler_args *args = data;
635 	int err = 0;
636 	struct kfd_process_device *pdd;
637 
638 	mutex_lock(&p->mutex);
639 
640 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
641 	if (!pdd) {
642 		err = -EINVAL;
643 		goto err_pdd;
644 	}
645 
646 	pdd = kfd_bind_process_to_device(pdd->dev, p);
647 	if (IS_ERR(pdd)) {
648 		err = -ESRCH;
649 		goto out;
650 	}
651 
652 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
653 
654 out:
655 err_pdd:
656 	mutex_unlock(&p->mutex);
657 
658 	return err;
659 }
660 
661 static int kfd_ioctl_dbg_register(struct file *filep,
662 				struct kfd_process *p, void *data)
663 {
664 	return -EPERM;
665 }
666 
667 static int kfd_ioctl_dbg_unregister(struct file *filep,
668 				struct kfd_process *p, void *data)
669 {
670 	return -EPERM;
671 }
672 
673 static int kfd_ioctl_dbg_address_watch(struct file *filep,
674 					struct kfd_process *p, void *data)
675 {
676 	return -EPERM;
677 }
678 
679 /* Parse and generate fixed size data structure for wave control */
680 static int kfd_ioctl_dbg_wave_control(struct file *filep,
681 					struct kfd_process *p, void *data)
682 {
683 	return -EPERM;
684 }
685 
686 static int kfd_ioctl_get_clock_counters(struct file *filep,
687 				struct kfd_process *p, void *data)
688 {
689 	struct kfd_ioctl_get_clock_counters_args *args = data;
690 	struct kfd_process_device *pdd;
691 
692 	mutex_lock(&p->mutex);
693 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
694 	mutex_unlock(&p->mutex);
695 	if (pdd)
696 		/* Reading GPU clock counter from KGD */
697 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
698 	else
699 		/* Node without GPU resource */
700 		args->gpu_clock_counter = 0;
701 
702 	/* No access to rdtsc. Using raw monotonic time */
703 	args->cpu_clock_counter = ktime_get_raw_ns();
704 	args->system_clock_counter = ktime_get_boottime_ns();
705 
706 	/* Since the counter is in nano-seconds we use 1GHz frequency */
707 	args->system_clock_freq = 1000000000;
708 
709 	return 0;
710 }
711 
712 
713 static int kfd_ioctl_get_process_apertures(struct file *filp,
714 				struct kfd_process *p, void *data)
715 {
716 	struct kfd_ioctl_get_process_apertures_args *args = data;
717 	struct kfd_process_device_apertures *pAperture;
718 	int i;
719 
720 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
721 
722 	args->num_of_nodes = 0;
723 
724 	mutex_lock(&p->mutex);
725 	/* Run over all pdd of the process */
726 	for (i = 0; i < p->n_pdds; i++) {
727 		struct kfd_process_device *pdd = p->pdds[i];
728 
729 		pAperture =
730 			&args->process_apertures[args->num_of_nodes];
731 		pAperture->gpu_id = pdd->dev->id;
732 		pAperture->lds_base = pdd->lds_base;
733 		pAperture->lds_limit = pdd->lds_limit;
734 		pAperture->gpuvm_base = pdd->gpuvm_base;
735 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
736 		pAperture->scratch_base = pdd->scratch_base;
737 		pAperture->scratch_limit = pdd->scratch_limit;
738 
739 		dev_dbg(kfd_device,
740 			"node id %u\n", args->num_of_nodes);
741 		dev_dbg(kfd_device,
742 			"gpu id %u\n", pdd->dev->id);
743 		dev_dbg(kfd_device,
744 			"lds_base %llX\n", pdd->lds_base);
745 		dev_dbg(kfd_device,
746 			"lds_limit %llX\n", pdd->lds_limit);
747 		dev_dbg(kfd_device,
748 			"gpuvm_base %llX\n", pdd->gpuvm_base);
749 		dev_dbg(kfd_device,
750 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
751 		dev_dbg(kfd_device,
752 			"scratch_base %llX\n", pdd->scratch_base);
753 		dev_dbg(kfd_device,
754 			"scratch_limit %llX\n", pdd->scratch_limit);
755 
756 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
757 			break;
758 	}
759 	mutex_unlock(&p->mutex);
760 
761 	return 0;
762 }
763 
764 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
765 				struct kfd_process *p, void *data)
766 {
767 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
768 	struct kfd_process_device_apertures *pa;
769 	int ret;
770 	int i;
771 
772 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
773 
774 	if (args->num_of_nodes == 0) {
775 		/* Return number of nodes, so that user space can alloacate
776 		 * sufficient memory
777 		 */
778 		mutex_lock(&p->mutex);
779 		args->num_of_nodes = p->n_pdds;
780 		goto out_unlock;
781 	}
782 
783 	/* Fill in process-aperture information for all available
784 	 * nodes, but not more than args->num_of_nodes as that is
785 	 * the amount of memory allocated by user
786 	 */
787 	pa = kcalloc(args->num_of_nodes, sizeof(struct kfd_process_device_apertures),
788 		     GFP_KERNEL);
789 	if (!pa)
790 		return -ENOMEM;
791 
792 	mutex_lock(&p->mutex);
793 
794 	if (!p->n_pdds) {
795 		args->num_of_nodes = 0;
796 		kfree(pa);
797 		goto out_unlock;
798 	}
799 
800 	/* Run over all pdd of the process */
801 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
802 		struct kfd_process_device *pdd = p->pdds[i];
803 
804 		pa[i].gpu_id = pdd->dev->id;
805 		pa[i].lds_base = pdd->lds_base;
806 		pa[i].lds_limit = pdd->lds_limit;
807 		pa[i].gpuvm_base = pdd->gpuvm_base;
808 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
809 		pa[i].scratch_base = pdd->scratch_base;
810 		pa[i].scratch_limit = pdd->scratch_limit;
811 
812 		dev_dbg(kfd_device,
813 			"gpu id %u\n", pdd->dev->id);
814 		dev_dbg(kfd_device,
815 			"lds_base %llX\n", pdd->lds_base);
816 		dev_dbg(kfd_device,
817 			"lds_limit %llX\n", pdd->lds_limit);
818 		dev_dbg(kfd_device,
819 			"gpuvm_base %llX\n", pdd->gpuvm_base);
820 		dev_dbg(kfd_device,
821 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
822 		dev_dbg(kfd_device,
823 			"scratch_base %llX\n", pdd->scratch_base);
824 		dev_dbg(kfd_device,
825 			"scratch_limit %llX\n", pdd->scratch_limit);
826 	}
827 	mutex_unlock(&p->mutex);
828 
829 	args->num_of_nodes = i;
830 	ret = copy_to_user(
831 			(void __user *)args->kfd_process_device_apertures_ptr,
832 			pa,
833 			(i * sizeof(struct kfd_process_device_apertures)));
834 	kfree(pa);
835 	return ret ? -EFAULT : 0;
836 
837 out_unlock:
838 	mutex_unlock(&p->mutex);
839 	return 0;
840 }
841 
842 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
843 					void *data)
844 {
845 	struct kfd_ioctl_create_event_args *args = data;
846 	int err;
847 
848 	/* For dGPUs the event page is allocated in user mode. The
849 	 * handle is passed to KFD with the first call to this IOCTL
850 	 * through the event_page_offset field.
851 	 */
852 	if (args->event_page_offset) {
853 		mutex_lock(&p->mutex);
854 		err = kfd_kmap_event_page(p, args->event_page_offset);
855 		mutex_unlock(&p->mutex);
856 		if (err)
857 			return err;
858 	}
859 
860 	err = kfd_event_create(filp, p, args->event_type,
861 				args->auto_reset != 0, args->node_id,
862 				&args->event_id, &args->event_trigger_data,
863 				&args->event_page_offset,
864 				&args->event_slot_index);
865 
866 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
867 	return err;
868 }
869 
870 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
871 					void *data)
872 {
873 	struct kfd_ioctl_destroy_event_args *args = data;
874 
875 	return kfd_event_destroy(p, args->event_id);
876 }
877 
878 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
879 				void *data)
880 {
881 	struct kfd_ioctl_set_event_args *args = data;
882 
883 	return kfd_set_event(p, args->event_id);
884 }
885 
886 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
887 				void *data)
888 {
889 	struct kfd_ioctl_reset_event_args *args = data;
890 
891 	return kfd_reset_event(p, args->event_id);
892 }
893 
894 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
895 				void *data)
896 {
897 	struct kfd_ioctl_wait_events_args *args = data;
898 
899 	return kfd_wait_on_events(p, args->num_events,
900 			(void __user *)args->events_ptr,
901 			(args->wait_for_all != 0),
902 			&args->timeout, &args->wait_result);
903 }
904 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
905 					struct kfd_process *p, void *data)
906 {
907 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
908 	struct kfd_process_device *pdd;
909 	struct kfd_node *dev;
910 	long err;
911 
912 	mutex_lock(&p->mutex);
913 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
914 	if (!pdd) {
915 		err = -EINVAL;
916 		goto err_pdd;
917 	}
918 	dev = pdd->dev;
919 
920 	pdd = kfd_bind_process_to_device(dev, p);
921 	if (IS_ERR(pdd)) {
922 		err = PTR_ERR(pdd);
923 		goto bind_process_to_device_fail;
924 	}
925 
926 	pdd->qpd.sh_hidden_private_base = args->va_addr;
927 
928 	mutex_unlock(&p->mutex);
929 
930 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
931 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
932 		dev->kfd2kgd->set_scratch_backing_va(
933 			dev->adev, args->va_addr, pdd->qpd.vmid);
934 
935 	return 0;
936 
937 bind_process_to_device_fail:
938 err_pdd:
939 	mutex_unlock(&p->mutex);
940 	return err;
941 }
942 
943 static int kfd_ioctl_get_tile_config(struct file *filep,
944 		struct kfd_process *p, void *data)
945 {
946 	struct kfd_ioctl_get_tile_config_args *args = data;
947 	struct kfd_process_device *pdd;
948 	struct tile_config config;
949 	int err = 0;
950 
951 	mutex_lock(&p->mutex);
952 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
953 	mutex_unlock(&p->mutex);
954 	if (!pdd)
955 		return -EINVAL;
956 
957 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
958 
959 	args->gb_addr_config = config.gb_addr_config;
960 	args->num_banks = config.num_banks;
961 	args->num_ranks = config.num_ranks;
962 
963 	if (args->num_tile_configs > config.num_tile_configs)
964 		args->num_tile_configs = config.num_tile_configs;
965 	err = copy_to_user((void __user *)args->tile_config_ptr,
966 			config.tile_config_ptr,
967 			args->num_tile_configs * sizeof(uint32_t));
968 	if (err) {
969 		args->num_tile_configs = 0;
970 		return -EFAULT;
971 	}
972 
973 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
974 		args->num_macro_tile_configs =
975 				config.num_macro_tile_configs;
976 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
977 			config.macro_tile_config_ptr,
978 			args->num_macro_tile_configs * sizeof(uint32_t));
979 	if (err) {
980 		args->num_macro_tile_configs = 0;
981 		return -EFAULT;
982 	}
983 
984 	return 0;
985 }
986 
987 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
988 				void *data)
989 {
990 	struct kfd_ioctl_acquire_vm_args *args = data;
991 	struct kfd_process_device *pdd;
992 	struct file *drm_file;
993 	int ret;
994 
995 	drm_file = fget(args->drm_fd);
996 	if (!drm_file)
997 		return -EINVAL;
998 
999 	mutex_lock(&p->mutex);
1000 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1001 	if (!pdd) {
1002 		ret = -EINVAL;
1003 		goto err_pdd;
1004 	}
1005 
1006 	if (pdd->drm_file) {
1007 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
1008 		goto err_drm_file;
1009 	}
1010 
1011 	ret = kfd_process_device_init_vm(pdd, drm_file);
1012 	if (ret)
1013 		goto err_unlock;
1014 
1015 	/* On success, the PDD keeps the drm_file reference */
1016 	mutex_unlock(&p->mutex);
1017 
1018 	return 0;
1019 
1020 err_unlock:
1021 err_pdd:
1022 err_drm_file:
1023 	mutex_unlock(&p->mutex);
1024 	fput(drm_file);
1025 	return ret;
1026 }
1027 
1028 bool kfd_dev_is_large_bar(struct kfd_node *dev)
1029 {
1030 	if (dev->kfd->adev->debug_largebar) {
1031 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1032 		return true;
1033 	}
1034 
1035 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1036 	    dev->local_mem_info.local_mem_size_public > 0)
1037 		return true;
1038 
1039 	if (dev->local_mem_info.local_mem_size_public == 0 &&
1040 	    dev->kfd->adev->gmc.is_app_apu) {
1041 		pr_debug("APP APU, Consider like a large bar system\n");
1042 		return true;
1043 	}
1044 
1045 	return false;
1046 }
1047 
1048 static int kfd_ioctl_get_available_memory(struct file *filep,
1049 					  struct kfd_process *p, void *data)
1050 {
1051 	struct kfd_ioctl_get_available_memory_args *args = data;
1052 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1053 
1054 	if (!pdd)
1055 		return -EINVAL;
1056 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1057 							pdd->dev->node_id);
1058 	kfd_unlock_pdd(pdd);
1059 	return 0;
1060 }
1061 
1062 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1063 					struct kfd_process *p, void *data)
1064 {
1065 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1066 	struct kfd_process_device *pdd;
1067 	void *mem;
1068 	struct kfd_node *dev;
1069 	int idr_handle;
1070 	long err;
1071 	uint64_t offset = args->mmap_offset;
1072 	uint32_t flags = args->flags;
1073 
1074 	if (args->size == 0)
1075 		return -EINVAL;
1076 
1077 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1078 	/* Flush pending deferred work to avoid racing with deferred actions
1079 	 * from previous memory map changes (e.g. munmap).
1080 	 */
1081 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1082 	mutex_lock(&p->svms.lock);
1083 	mmap_write_unlock(current->mm);
1084 	if (interval_tree_iter_first(&p->svms.objects,
1085 				     args->va_addr >> PAGE_SHIFT,
1086 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1087 		pr_err("Address: 0x%llx already allocated by SVM\n",
1088 			args->va_addr);
1089 		mutex_unlock(&p->svms.lock);
1090 		return -EADDRINUSE;
1091 	}
1092 
1093 	/* When register user buffer check if it has been registered by svm by
1094 	 * buffer cpu virtual address.
1095 	 */
1096 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1097 	    interval_tree_iter_first(&p->svms.objects,
1098 				     args->mmap_offset >> PAGE_SHIFT,
1099 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1100 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1101 			args->mmap_offset);
1102 		mutex_unlock(&p->svms.lock);
1103 		return -EADDRINUSE;
1104 	}
1105 
1106 	mutex_unlock(&p->svms.lock);
1107 #endif
1108 	mutex_lock(&p->mutex);
1109 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1110 	if (!pdd) {
1111 		err = -EINVAL;
1112 		goto err_pdd;
1113 	}
1114 
1115 	dev = pdd->dev;
1116 
1117 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1118 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1119 		!kfd_dev_is_large_bar(dev)) {
1120 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1121 		err = -EINVAL;
1122 		goto err_large_bar;
1123 	}
1124 
1125 	pdd = kfd_bind_process_to_device(dev, p);
1126 	if (IS_ERR(pdd)) {
1127 		err = PTR_ERR(pdd);
1128 		goto err_unlock;
1129 	}
1130 
1131 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1132 		if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1133 			err = -EINVAL;
1134 			goto err_unlock;
1135 		}
1136 		offset = kfd_get_process_doorbells(pdd);
1137 		if (!offset) {
1138 			err = -ENOMEM;
1139 			goto err_unlock;
1140 		}
1141 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1142 		if (args->size != PAGE_SIZE) {
1143 			err = -EINVAL;
1144 			goto err_unlock;
1145 		}
1146 		offset = dev->adev->rmmio_remap.bus_addr;
1147 		if (!offset) {
1148 			err = -ENOMEM;
1149 			goto err_unlock;
1150 		}
1151 	}
1152 
1153 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1154 		dev->adev, args->va_addr, args->size,
1155 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1156 		flags, false);
1157 
1158 	if (err)
1159 		goto err_unlock;
1160 
1161 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1162 	if (idr_handle < 0) {
1163 		err = -EFAULT;
1164 		goto err_free;
1165 	}
1166 
1167 	/* Update the VRAM usage count */
1168 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1169 		uint64_t size = args->size;
1170 
1171 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1172 			size >>= 1;
1173 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1174 	}
1175 
1176 	mutex_unlock(&p->mutex);
1177 
1178 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1179 	args->mmap_offset = offset;
1180 
1181 	/* MMIO is mapped through kfd device
1182 	 * Generate a kfd mmap offset
1183 	 */
1184 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1185 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1186 					| KFD_MMAP_GPU_ID(args->gpu_id);
1187 
1188 	return 0;
1189 
1190 err_free:
1191 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1192 					       pdd->drm_priv, NULL);
1193 err_unlock:
1194 err_pdd:
1195 err_large_bar:
1196 	mutex_unlock(&p->mutex);
1197 	return err;
1198 }
1199 
1200 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1201 					struct kfd_process *p, void *data)
1202 {
1203 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1204 	struct kfd_process_device *pdd;
1205 	void *mem;
1206 	int ret;
1207 	uint64_t size = 0;
1208 
1209 	mutex_lock(&p->mutex);
1210 	/*
1211 	 * Safeguard to prevent user space from freeing signal BO.
1212 	 * It will be freed at process termination.
1213 	 */
1214 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1215 		pr_err("Free signal BO is not allowed\n");
1216 		ret = -EPERM;
1217 		goto err_unlock;
1218 	}
1219 
1220 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1221 	if (!pdd) {
1222 		pr_err("Process device data doesn't exist\n");
1223 		ret = -EINVAL;
1224 		goto err_pdd;
1225 	}
1226 
1227 	mem = kfd_process_device_translate_handle(
1228 		pdd, GET_IDR_HANDLE(args->handle));
1229 	if (!mem) {
1230 		ret = -EINVAL;
1231 		goto err_unlock;
1232 	}
1233 
1234 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1235 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1236 
1237 	/* If freeing the buffer failed, leave the handle in place for
1238 	 * clean-up during process tear-down.
1239 	 */
1240 	if (!ret)
1241 		kfd_process_device_remove_obj_handle(
1242 			pdd, GET_IDR_HANDLE(args->handle));
1243 
1244 	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1245 
1246 err_unlock:
1247 err_pdd:
1248 	mutex_unlock(&p->mutex);
1249 	return ret;
1250 }
1251 
1252 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1253 					struct kfd_process *p, void *data)
1254 {
1255 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1256 	struct kfd_process_device *pdd, *peer_pdd;
1257 	void *mem;
1258 	struct kfd_node *dev;
1259 	long err = 0;
1260 	int i;
1261 	uint32_t *devices_arr = NULL;
1262 
1263 	if (!args->n_devices) {
1264 		pr_debug("Device IDs array empty\n");
1265 		return -EINVAL;
1266 	}
1267 	if (args->n_success > args->n_devices) {
1268 		pr_debug("n_success exceeds n_devices\n");
1269 		return -EINVAL;
1270 	}
1271 
1272 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1273 				    GFP_KERNEL);
1274 	if (!devices_arr)
1275 		return -ENOMEM;
1276 
1277 	err = copy_from_user(devices_arr,
1278 			     (void __user *)args->device_ids_array_ptr,
1279 			     args->n_devices * sizeof(*devices_arr));
1280 	if (err != 0) {
1281 		err = -EFAULT;
1282 		goto copy_from_user_failed;
1283 	}
1284 
1285 	mutex_lock(&p->mutex);
1286 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1287 	if (!pdd) {
1288 		err = -EINVAL;
1289 		goto get_process_device_data_failed;
1290 	}
1291 	dev = pdd->dev;
1292 
1293 	pdd = kfd_bind_process_to_device(dev, p);
1294 	if (IS_ERR(pdd)) {
1295 		err = PTR_ERR(pdd);
1296 		goto bind_process_to_device_failed;
1297 	}
1298 
1299 	mem = kfd_process_device_translate_handle(pdd,
1300 						GET_IDR_HANDLE(args->handle));
1301 	if (!mem) {
1302 		err = -ENOMEM;
1303 		goto get_mem_obj_from_handle_failed;
1304 	}
1305 
1306 	for (i = args->n_success; i < args->n_devices; i++) {
1307 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1308 		if (!peer_pdd) {
1309 			pr_debug("Getting device by id failed for 0x%x\n",
1310 				 devices_arr[i]);
1311 			err = -EINVAL;
1312 			goto get_mem_obj_from_handle_failed;
1313 		}
1314 
1315 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1316 		if (IS_ERR(peer_pdd)) {
1317 			err = PTR_ERR(peer_pdd);
1318 			goto get_mem_obj_from_handle_failed;
1319 		}
1320 
1321 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1322 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1323 			peer_pdd->drm_priv);
1324 		if (err) {
1325 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1326 
1327 			dev_err(dev->adev->dev,
1328 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1329 			       pci_domain_nr(pdev->bus),
1330 			       pdev->bus->number,
1331 			       PCI_SLOT(pdev->devfn),
1332 			       PCI_FUNC(pdev->devfn),
1333 			       ((struct kgd_mem *)mem)->domain);
1334 			goto map_memory_to_gpu_failed;
1335 		}
1336 		args->n_success = i+1;
1337 	}
1338 
1339 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1340 	if (err) {
1341 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1342 		goto sync_memory_failed;
1343 	}
1344 
1345 	mutex_unlock(&p->mutex);
1346 
1347 	/* Flush TLBs after waiting for the page table updates to complete */
1348 	for (i = 0; i < args->n_devices; i++) {
1349 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1350 		if (WARN_ON_ONCE(!peer_pdd))
1351 			continue;
1352 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1353 	}
1354 	kfree(devices_arr);
1355 
1356 	return err;
1357 
1358 get_process_device_data_failed:
1359 bind_process_to_device_failed:
1360 get_mem_obj_from_handle_failed:
1361 map_memory_to_gpu_failed:
1362 sync_memory_failed:
1363 	mutex_unlock(&p->mutex);
1364 copy_from_user_failed:
1365 	kfree(devices_arr);
1366 
1367 	return err;
1368 }
1369 
1370 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1371 					struct kfd_process *p, void *data)
1372 {
1373 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1374 	struct kfd_process_device *pdd, *peer_pdd;
1375 	void *mem;
1376 	long err = 0;
1377 	uint32_t *devices_arr = NULL, i;
1378 	bool flush_tlb;
1379 
1380 	if (!args->n_devices) {
1381 		pr_debug("Device IDs array empty\n");
1382 		return -EINVAL;
1383 	}
1384 	if (args->n_success > args->n_devices) {
1385 		pr_debug("n_success exceeds n_devices\n");
1386 		return -EINVAL;
1387 	}
1388 
1389 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1390 				    GFP_KERNEL);
1391 	if (!devices_arr)
1392 		return -ENOMEM;
1393 
1394 	err = copy_from_user(devices_arr,
1395 			     (void __user *)args->device_ids_array_ptr,
1396 			     args->n_devices * sizeof(*devices_arr));
1397 	if (err != 0) {
1398 		err = -EFAULT;
1399 		goto copy_from_user_failed;
1400 	}
1401 
1402 	mutex_lock(&p->mutex);
1403 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1404 	if (!pdd) {
1405 		err = -EINVAL;
1406 		goto bind_process_to_device_failed;
1407 	}
1408 
1409 	mem = kfd_process_device_translate_handle(pdd,
1410 						GET_IDR_HANDLE(args->handle));
1411 	if (!mem) {
1412 		err = -ENOMEM;
1413 		goto get_mem_obj_from_handle_failed;
1414 	}
1415 
1416 	for (i = args->n_success; i < args->n_devices; i++) {
1417 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1418 		if (!peer_pdd) {
1419 			err = -EINVAL;
1420 			goto get_mem_obj_from_handle_failed;
1421 		}
1422 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1423 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1424 		if (err) {
1425 			pr_err("Failed to unmap from gpu %d/%d\n",
1426 			       i, args->n_devices);
1427 			goto unmap_memory_from_gpu_failed;
1428 		}
1429 		args->n_success = i+1;
1430 	}
1431 
1432 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1433 	if (flush_tlb) {
1434 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1435 				(struct kgd_mem *) mem, true);
1436 		if (err) {
1437 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1438 			goto sync_memory_failed;
1439 		}
1440 	}
1441 
1442 	/* Flush TLBs after waiting for the page table updates to complete */
1443 	for (i = 0; i < args->n_devices; i++) {
1444 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1445 		if (WARN_ON_ONCE(!peer_pdd))
1446 			continue;
1447 		if (flush_tlb)
1448 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1449 
1450 		/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */
1451 		err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd->drm_priv);
1452 		if (err)
1453 			goto sync_memory_failed;
1454 	}
1455 
1456 	mutex_unlock(&p->mutex);
1457 
1458 	kfree(devices_arr);
1459 
1460 	return 0;
1461 
1462 bind_process_to_device_failed:
1463 get_mem_obj_from_handle_failed:
1464 unmap_memory_from_gpu_failed:
1465 sync_memory_failed:
1466 	mutex_unlock(&p->mutex);
1467 copy_from_user_failed:
1468 	kfree(devices_arr);
1469 	return err;
1470 }
1471 
1472 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1473 		struct kfd_process *p, void *data)
1474 {
1475 	int retval;
1476 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1477 	struct queue *q;
1478 	struct kfd_node *dev;
1479 
1480 	mutex_lock(&p->mutex);
1481 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1482 
1483 	if (q) {
1484 		dev = q->device;
1485 	} else {
1486 		retval = -EINVAL;
1487 		goto out_unlock;
1488 	}
1489 
1490 	if (!dev->gws) {
1491 		retval = -ENODEV;
1492 		goto out_unlock;
1493 	}
1494 
1495 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1496 		retval = -ENODEV;
1497 		goto out_unlock;
1498 	}
1499 
1500 	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1501 				      kfd_dbg_has_cwsr_workaround(dev))) {
1502 		retval = -EBUSY;
1503 		goto out_unlock;
1504 	}
1505 
1506 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1507 	mutex_unlock(&p->mutex);
1508 
1509 	args->first_gws = 0;
1510 	return retval;
1511 
1512 out_unlock:
1513 	mutex_unlock(&p->mutex);
1514 	return retval;
1515 }
1516 
1517 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1518 		struct kfd_process *p, void *data)
1519 {
1520 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1521 	struct kfd_node *dev = NULL;
1522 	struct amdgpu_device *dmabuf_adev;
1523 	void *metadata_buffer = NULL;
1524 	uint32_t flags;
1525 	int8_t xcp_id;
1526 	unsigned int i;
1527 	int r;
1528 
1529 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1530 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1531 		if (dev && !kfd_devcgroup_check_permission(dev))
1532 			break;
1533 	if (!dev)
1534 		return -EINVAL;
1535 
1536 	if (args->metadata_ptr) {
1537 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1538 		if (!metadata_buffer)
1539 			return -ENOMEM;
1540 	}
1541 
1542 	/* Get dmabuf info from KGD */
1543 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1544 					  &dmabuf_adev, &args->size,
1545 					  metadata_buffer, args->metadata_size,
1546 					  &args->metadata_size, &flags, &xcp_id);
1547 	if (r)
1548 		goto exit;
1549 
1550 	if (xcp_id >= 0)
1551 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1552 	else
1553 		args->gpu_id = dev->id;
1554 	args->flags = flags;
1555 
1556 	/* Copy metadata buffer to user mode */
1557 	if (metadata_buffer) {
1558 		r = copy_to_user((void __user *)args->metadata_ptr,
1559 				 metadata_buffer, args->metadata_size);
1560 		if (r != 0)
1561 			r = -EFAULT;
1562 	}
1563 
1564 exit:
1565 	kfree(metadata_buffer);
1566 
1567 	return r;
1568 }
1569 
1570 static int kfd_ioctl_import_dmabuf(struct file *filep,
1571 				   struct kfd_process *p, void *data)
1572 {
1573 	struct kfd_ioctl_import_dmabuf_args *args = data;
1574 	struct kfd_process_device *pdd;
1575 	int idr_handle;
1576 	uint64_t size;
1577 	void *mem;
1578 	int r;
1579 
1580 	mutex_lock(&p->mutex);
1581 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1582 	if (!pdd) {
1583 		r = -EINVAL;
1584 		goto err_unlock;
1585 	}
1586 
1587 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1588 	if (IS_ERR(pdd)) {
1589 		r = PTR_ERR(pdd);
1590 		goto err_unlock;
1591 	}
1592 
1593 	r = amdgpu_amdkfd_gpuvm_import_dmabuf_fd(pdd->dev->adev, args->dmabuf_fd,
1594 						 args->va_addr, pdd->drm_priv,
1595 						 (struct kgd_mem **)&mem, &size,
1596 						 NULL);
1597 	if (r)
1598 		goto err_unlock;
1599 
1600 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1601 	if (idr_handle < 0) {
1602 		r = -EFAULT;
1603 		goto err_free;
1604 	}
1605 
1606 	mutex_unlock(&p->mutex);
1607 
1608 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1609 
1610 	return 0;
1611 
1612 err_free:
1613 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1614 					       pdd->drm_priv, NULL);
1615 err_unlock:
1616 	mutex_unlock(&p->mutex);
1617 	return r;
1618 }
1619 
1620 static int kfd_ioctl_export_dmabuf(struct file *filep,
1621 				   struct kfd_process *p, void *data)
1622 {
1623 	struct kfd_ioctl_export_dmabuf_args *args = data;
1624 	struct kfd_process_device *pdd;
1625 	struct dma_buf *dmabuf;
1626 	struct kfd_node *dev;
1627 	void *mem;
1628 	int ret = 0;
1629 
1630 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1631 	if (!dev)
1632 		return -EINVAL;
1633 
1634 	mutex_lock(&p->mutex);
1635 
1636 	pdd = kfd_get_process_device_data(dev, p);
1637 	if (!pdd) {
1638 		ret = -EINVAL;
1639 		goto err_unlock;
1640 	}
1641 
1642 	mem = kfd_process_device_translate_handle(pdd,
1643 						GET_IDR_HANDLE(args->handle));
1644 	if (!mem) {
1645 		ret = -EINVAL;
1646 		goto err_unlock;
1647 	}
1648 
1649 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1650 	mutex_unlock(&p->mutex);
1651 	if (ret)
1652 		goto err_out;
1653 
1654 	ret = dma_buf_fd(dmabuf, args->flags);
1655 	if (ret < 0) {
1656 		dma_buf_put(dmabuf);
1657 		goto err_out;
1658 	}
1659 	/* dma_buf_fd assigns the reference count to the fd, no need to
1660 	 * put the reference here.
1661 	 */
1662 	args->dmabuf_fd = ret;
1663 
1664 	return 0;
1665 
1666 err_unlock:
1667 	mutex_unlock(&p->mutex);
1668 err_out:
1669 	return ret;
1670 }
1671 
1672 /* Handle requests for watching SMI events */
1673 static int kfd_ioctl_smi_events(struct file *filep,
1674 				struct kfd_process *p, void *data)
1675 {
1676 	struct kfd_ioctl_smi_events_args *args = data;
1677 	struct kfd_process_device *pdd;
1678 
1679 	mutex_lock(&p->mutex);
1680 
1681 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1682 	mutex_unlock(&p->mutex);
1683 	if (!pdd)
1684 		return -EINVAL;
1685 
1686 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1687 }
1688 
1689 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1690 
1691 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1692 				    struct kfd_process *p, void *data)
1693 {
1694 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1695 	int r = 0;
1696 
1697 	mutex_lock(&p->mutex);
1698 	if (args->xnack_enabled >= 0) {
1699 		if (!list_empty(&p->pqm.queues)) {
1700 			pr_debug("Process has user queues running\n");
1701 			r = -EBUSY;
1702 			goto out_unlock;
1703 		}
1704 
1705 		if (p->xnack_enabled == args->xnack_enabled)
1706 			goto out_unlock;
1707 
1708 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1709 			r = -EPERM;
1710 			goto out_unlock;
1711 		}
1712 
1713 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1714 	} else {
1715 		args->xnack_enabled = p->xnack_enabled;
1716 	}
1717 
1718 out_unlock:
1719 	mutex_unlock(&p->mutex);
1720 
1721 	return r;
1722 }
1723 
1724 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1725 {
1726 	struct kfd_ioctl_svm_args *args = data;
1727 	int r = 0;
1728 
1729 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1730 		 args->start_addr, args->size, args->op, args->nattr);
1731 
1732 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1733 		return -EINVAL;
1734 	if (!args->start_addr || !args->size)
1735 		return -EINVAL;
1736 
1737 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1738 		      args->attrs);
1739 
1740 	return r;
1741 }
1742 #else
1743 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1744 				    struct kfd_process *p, void *data)
1745 {
1746 	return -EPERM;
1747 }
1748 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1749 {
1750 	return -EPERM;
1751 }
1752 #endif
1753 
1754 static int criu_checkpoint_process(struct kfd_process *p,
1755 			     uint8_t __user *user_priv_data,
1756 			     uint64_t *priv_offset)
1757 {
1758 	struct kfd_criu_process_priv_data process_priv;
1759 	int ret;
1760 
1761 	memset(&process_priv, 0, sizeof(process_priv));
1762 
1763 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1764 	/* For CR, we don't consider negative xnack mode which is used for
1765 	 * querying without changing it, here 0 simply means disabled and 1
1766 	 * means enabled so retry for finding a valid PTE.
1767 	 */
1768 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1769 
1770 	ret = copy_to_user(user_priv_data + *priv_offset,
1771 				&process_priv, sizeof(process_priv));
1772 
1773 	if (ret) {
1774 		pr_err("Failed to copy process information to user\n");
1775 		ret = -EFAULT;
1776 	}
1777 
1778 	*priv_offset += sizeof(process_priv);
1779 	return ret;
1780 }
1781 
1782 static int criu_checkpoint_devices(struct kfd_process *p,
1783 			     uint32_t num_devices,
1784 			     uint8_t __user *user_addr,
1785 			     uint8_t __user *user_priv_data,
1786 			     uint64_t *priv_offset)
1787 {
1788 	struct kfd_criu_device_priv_data *device_priv = NULL;
1789 	struct kfd_criu_device_bucket *device_buckets = NULL;
1790 	int ret = 0, i;
1791 
1792 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1793 	if (!device_buckets) {
1794 		ret = -ENOMEM;
1795 		goto exit;
1796 	}
1797 
1798 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1799 	if (!device_priv) {
1800 		ret = -ENOMEM;
1801 		goto exit;
1802 	}
1803 
1804 	for (i = 0; i < num_devices; i++) {
1805 		struct kfd_process_device *pdd = p->pdds[i];
1806 
1807 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1808 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1809 
1810 		/*
1811 		 * priv_data does not contain useful information for now and is reserved for
1812 		 * future use, so we do not set its contents.
1813 		 */
1814 	}
1815 
1816 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1817 	if (ret) {
1818 		pr_err("Failed to copy device information to user\n");
1819 		ret = -EFAULT;
1820 		goto exit;
1821 	}
1822 
1823 	ret = copy_to_user(user_priv_data + *priv_offset,
1824 			   device_priv,
1825 			   num_devices * sizeof(*device_priv));
1826 	if (ret) {
1827 		pr_err("Failed to copy device information to user\n");
1828 		ret = -EFAULT;
1829 	}
1830 	*priv_offset += num_devices * sizeof(*device_priv);
1831 
1832 exit:
1833 	kvfree(device_buckets);
1834 	kvfree(device_priv);
1835 	return ret;
1836 }
1837 
1838 static uint32_t get_process_num_bos(struct kfd_process *p)
1839 {
1840 	uint32_t num_of_bos = 0;
1841 	int i;
1842 
1843 	/* Run over all PDDs of the process */
1844 	for (i = 0; i < p->n_pdds; i++) {
1845 		struct kfd_process_device *pdd = p->pdds[i];
1846 		void *mem;
1847 		int id;
1848 
1849 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1850 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1851 
1852 			if (!kgd_mem->va || kgd_mem->va > pdd->gpuvm_base)
1853 				num_of_bos++;
1854 		}
1855 	}
1856 	return num_of_bos;
1857 }
1858 
1859 static int criu_get_prime_handle(struct kgd_mem *mem,
1860 				 int flags, u32 *shared_fd)
1861 {
1862 	struct dma_buf *dmabuf;
1863 	int ret;
1864 
1865 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1866 	if (ret) {
1867 		pr_err("dmabuf export failed for the BO\n");
1868 		return ret;
1869 	}
1870 
1871 	ret = dma_buf_fd(dmabuf, flags);
1872 	if (ret < 0) {
1873 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1874 		goto out_free_dmabuf;
1875 	}
1876 
1877 	*shared_fd = ret;
1878 	return 0;
1879 
1880 out_free_dmabuf:
1881 	dma_buf_put(dmabuf);
1882 	return ret;
1883 }
1884 
1885 static int criu_checkpoint_bos(struct kfd_process *p,
1886 			       uint32_t num_bos,
1887 			       uint8_t __user *user_bos,
1888 			       uint8_t __user *user_priv_data,
1889 			       uint64_t *priv_offset)
1890 {
1891 	struct kfd_criu_bo_bucket *bo_buckets;
1892 	struct kfd_criu_bo_priv_data *bo_privs;
1893 	int ret = 0, pdd_index, bo_index = 0, id;
1894 	void *mem;
1895 
1896 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1897 	if (!bo_buckets)
1898 		return -ENOMEM;
1899 
1900 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1901 	if (!bo_privs) {
1902 		ret = -ENOMEM;
1903 		goto exit;
1904 	}
1905 
1906 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1907 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1908 		struct amdgpu_bo *dumper_bo;
1909 		struct kgd_mem *kgd_mem;
1910 
1911 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1912 			struct kfd_criu_bo_bucket *bo_bucket;
1913 			struct kfd_criu_bo_priv_data *bo_priv;
1914 			int i, dev_idx = 0;
1915 
1916 			if (!mem) {
1917 				ret = -ENOMEM;
1918 				goto exit;
1919 			}
1920 
1921 			kgd_mem = (struct kgd_mem *)mem;
1922 			dumper_bo = kgd_mem->bo;
1923 
1924 			/* Skip checkpointing BOs that are used for Trap handler
1925 			 * code and state. Currently, these BOs have a VA that
1926 			 * is less GPUVM Base
1927 			 */
1928 			if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base)
1929 				continue;
1930 
1931 			bo_bucket = &bo_buckets[bo_index];
1932 			bo_priv = &bo_privs[bo_index];
1933 
1934 			bo_bucket->gpu_id = pdd->user_gpu_id;
1935 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1936 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1937 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1938 			bo_priv->idr_handle = id;
1939 
1940 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1941 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1942 								&bo_priv->user_addr);
1943 				if (ret) {
1944 					pr_err("Failed to obtain user address for user-pointer bo\n");
1945 					goto exit;
1946 				}
1947 			}
1948 			if (bo_bucket->alloc_flags
1949 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1950 				ret = criu_get_prime_handle(kgd_mem,
1951 						bo_bucket->alloc_flags &
1952 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1953 						&bo_bucket->dmabuf_fd);
1954 				if (ret)
1955 					goto exit;
1956 			} else {
1957 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1958 			}
1959 
1960 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1961 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1962 					KFD_MMAP_GPU_ID(pdd->dev->id);
1963 			else if (bo_bucket->alloc_flags &
1964 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1965 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1966 					KFD_MMAP_GPU_ID(pdd->dev->id);
1967 			else
1968 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1969 
1970 			for (i = 0; i < p->n_pdds; i++) {
1971 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->dev->adev, kgd_mem))
1972 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1973 			}
1974 
1975 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1976 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1977 					bo_bucket->size,
1978 					bo_bucket->addr,
1979 					bo_bucket->offset,
1980 					bo_bucket->gpu_id,
1981 					bo_bucket->alloc_flags,
1982 					bo_priv->idr_handle);
1983 			bo_index++;
1984 		}
1985 	}
1986 
1987 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1988 	if (ret) {
1989 		pr_err("Failed to copy BO information to user\n");
1990 		ret = -EFAULT;
1991 		goto exit;
1992 	}
1993 
1994 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1995 	if (ret) {
1996 		pr_err("Failed to copy BO priv information to user\n");
1997 		ret = -EFAULT;
1998 		goto exit;
1999 	}
2000 
2001 	*priv_offset += num_bos * sizeof(*bo_privs);
2002 
2003 exit:
2004 	while (ret && bo_index--) {
2005 		if (bo_buckets[bo_index].alloc_flags
2006 		    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2007 			close_fd(bo_buckets[bo_index].dmabuf_fd);
2008 	}
2009 
2010 	kvfree(bo_buckets);
2011 	kvfree(bo_privs);
2012 	return ret;
2013 }
2014 
2015 static int criu_get_process_object_info(struct kfd_process *p,
2016 					uint32_t *num_devices,
2017 					uint32_t *num_bos,
2018 					uint32_t *num_objects,
2019 					uint64_t *objs_priv_size)
2020 {
2021 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2022 	uint32_t num_queues, num_events, num_svm_ranges;
2023 	int ret;
2024 
2025 	*num_devices = p->n_pdds;
2026 	*num_bos = get_process_num_bos(p);
2027 
2028 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2029 	if (ret)
2030 		return ret;
2031 
2032 	num_events = kfd_get_num_events(p);
2033 
2034 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2035 	if (ret)
2036 		return ret;
2037 
2038 	*num_objects = num_queues + num_events + num_svm_ranges;
2039 
2040 	if (objs_priv_size) {
2041 		priv_size = sizeof(struct kfd_criu_process_priv_data);
2042 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2043 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2044 		priv_size += queues_priv_data_size;
2045 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2046 		priv_size += svm_priv_data_size;
2047 		*objs_priv_size = priv_size;
2048 	}
2049 	return 0;
2050 }
2051 
2052 static int criu_checkpoint(struct file *filep,
2053 			   struct kfd_process *p,
2054 			   struct kfd_ioctl_criu_args *args)
2055 {
2056 	int ret;
2057 	uint32_t num_devices, num_bos, num_objects;
2058 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2059 
2060 	if (!args->devices || !args->bos || !args->priv_data)
2061 		return -EINVAL;
2062 
2063 	mutex_lock(&p->mutex);
2064 
2065 	if (!p->n_pdds) {
2066 		pr_err("No pdd for given process\n");
2067 		ret = -ENODEV;
2068 		goto exit_unlock;
2069 	}
2070 
2071 	/* Confirm all process queues are evicted */
2072 	if (!p->queues_paused) {
2073 		pr_err("Cannot dump process when queues are not in evicted state\n");
2074 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2075 		ret = -EINVAL;
2076 		goto exit_unlock;
2077 	}
2078 
2079 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2080 	if (ret)
2081 		goto exit_unlock;
2082 
2083 	if (num_devices != args->num_devices ||
2084 	    num_bos != args->num_bos ||
2085 	    num_objects != args->num_objects ||
2086 	    priv_size != args->priv_data_size) {
2087 
2088 		ret = -EINVAL;
2089 		goto exit_unlock;
2090 	}
2091 
2092 	/* each function will store private data inside priv_data and adjust priv_offset */
2093 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2094 	if (ret)
2095 		goto exit_unlock;
2096 
2097 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2098 				(uint8_t __user *)args->priv_data, &priv_offset);
2099 	if (ret)
2100 		goto exit_unlock;
2101 
2102 	/* Leave room for BOs in the private data. They need to be restored
2103 	 * before events, but we checkpoint them last to simplify the error
2104 	 * handling.
2105 	 */
2106 	bo_priv_offset = priv_offset;
2107 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2108 
2109 	if (num_objects) {
2110 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2111 						 &priv_offset);
2112 		if (ret)
2113 			goto exit_unlock;
2114 
2115 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2116 						 &priv_offset);
2117 		if (ret)
2118 			goto exit_unlock;
2119 
2120 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2121 		if (ret)
2122 			goto exit_unlock;
2123 	}
2124 
2125 	/* This must be the last thing in this function that can fail.
2126 	 * Otherwise we leak dmabuf file descriptors.
2127 	 */
2128 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2129 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2130 
2131 exit_unlock:
2132 	mutex_unlock(&p->mutex);
2133 	if (ret)
2134 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2135 	else
2136 		pr_debug("CRIU dump ret:%d\n", ret);
2137 
2138 	return ret;
2139 }
2140 
2141 static int criu_restore_process(struct kfd_process *p,
2142 				struct kfd_ioctl_criu_args *args,
2143 				uint64_t *priv_offset,
2144 				uint64_t max_priv_data_size)
2145 {
2146 	int ret = 0;
2147 	struct kfd_criu_process_priv_data process_priv;
2148 
2149 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2150 		return -EINVAL;
2151 
2152 	ret = copy_from_user(&process_priv,
2153 				(void __user *)(args->priv_data + *priv_offset),
2154 				sizeof(process_priv));
2155 	if (ret) {
2156 		pr_err("Failed to copy process private information from user\n");
2157 		ret = -EFAULT;
2158 		goto exit;
2159 	}
2160 	*priv_offset += sizeof(process_priv);
2161 
2162 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2163 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2164 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2165 		return -EINVAL;
2166 	}
2167 
2168 	pr_debug("Setting XNACK mode\n");
2169 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2170 		pr_err("xnack mode cannot be set\n");
2171 		ret = -EPERM;
2172 		goto exit;
2173 	} else {
2174 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2175 		p->xnack_enabled = process_priv.xnack_mode;
2176 	}
2177 
2178 exit:
2179 	return ret;
2180 }
2181 
2182 static int criu_restore_devices(struct kfd_process *p,
2183 				struct kfd_ioctl_criu_args *args,
2184 				uint64_t *priv_offset,
2185 				uint64_t max_priv_data_size)
2186 {
2187 	struct kfd_criu_device_bucket *device_buckets;
2188 	struct kfd_criu_device_priv_data *device_privs;
2189 	int ret = 0;
2190 	uint32_t i;
2191 
2192 	if (args->num_devices != p->n_pdds)
2193 		return -EINVAL;
2194 
2195 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2196 		return -EINVAL;
2197 
2198 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2199 	if (!device_buckets)
2200 		return -ENOMEM;
2201 
2202 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2203 				args->num_devices * sizeof(*device_buckets));
2204 	if (ret) {
2205 		pr_err("Failed to copy devices buckets from user\n");
2206 		ret = -EFAULT;
2207 		goto exit;
2208 	}
2209 
2210 	for (i = 0; i < args->num_devices; i++) {
2211 		struct kfd_node *dev;
2212 		struct kfd_process_device *pdd;
2213 		struct file *drm_file;
2214 
2215 		/* device private data is not currently used */
2216 
2217 		if (!device_buckets[i].user_gpu_id) {
2218 			pr_err("Invalid user gpu_id\n");
2219 			ret = -EINVAL;
2220 			goto exit;
2221 		}
2222 
2223 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2224 		if (!dev) {
2225 			pr_err("Failed to find device with gpu_id = %x\n",
2226 				device_buckets[i].actual_gpu_id);
2227 			ret = -EINVAL;
2228 			goto exit;
2229 		}
2230 
2231 		pdd = kfd_get_process_device_data(dev, p);
2232 		if (!pdd) {
2233 			pr_err("Failed to get pdd for gpu_id = %x\n",
2234 					device_buckets[i].actual_gpu_id);
2235 			ret = -EINVAL;
2236 			goto exit;
2237 		}
2238 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2239 
2240 		drm_file = fget(device_buckets[i].drm_fd);
2241 		if (!drm_file) {
2242 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2243 				device_buckets[i].drm_fd);
2244 			ret = -EINVAL;
2245 			goto exit;
2246 		}
2247 
2248 		if (pdd->drm_file) {
2249 			ret = -EINVAL;
2250 			goto exit;
2251 		}
2252 
2253 		/* create the vm using render nodes for kfd pdd */
2254 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2255 			pr_err("could not init vm for given pdd\n");
2256 			/* On success, the PDD keeps the drm_file reference */
2257 			fput(drm_file);
2258 			ret = -EINVAL;
2259 			goto exit;
2260 		}
2261 		/*
2262 		 * pdd now already has the vm bound to render node so below api won't create a new
2263 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2264 		 * for iommu v2 binding  and runtime pm.
2265 		 */
2266 		pdd = kfd_bind_process_to_device(dev, p);
2267 		if (IS_ERR(pdd)) {
2268 			ret = PTR_ERR(pdd);
2269 			goto exit;
2270 		}
2271 
2272 		if (!pdd->qpd.proc_doorbells) {
2273 			ret = kfd_alloc_process_doorbells(dev->kfd, pdd);
2274 			if (ret)
2275 				goto exit;
2276 		}
2277 	}
2278 
2279 	/*
2280 	 * We are not copying device private data from user as we are not using the data for now,
2281 	 * but we still adjust for its private data.
2282 	 */
2283 	*priv_offset += args->num_devices * sizeof(*device_privs);
2284 
2285 exit:
2286 	kfree(device_buckets);
2287 	return ret;
2288 }
2289 
2290 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2291 				      struct kfd_criu_bo_bucket *bo_bucket,
2292 				      struct kfd_criu_bo_priv_data *bo_priv,
2293 				      struct kgd_mem **kgd_mem)
2294 {
2295 	int idr_handle;
2296 	int ret;
2297 	const bool criu_resume = true;
2298 	u64 offset;
2299 
2300 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2301 		if (bo_bucket->size !=
2302 				kfd_doorbell_process_slice(pdd->dev->kfd))
2303 			return -EINVAL;
2304 
2305 		offset = kfd_get_process_doorbells(pdd);
2306 		if (!offset)
2307 			return -ENOMEM;
2308 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2309 		/* MMIO BOs need remapped bus address */
2310 		if (bo_bucket->size != PAGE_SIZE) {
2311 			pr_err("Invalid page size\n");
2312 			return -EINVAL;
2313 		}
2314 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2315 		if (!offset) {
2316 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2317 			return -ENOMEM;
2318 		}
2319 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2320 		offset = bo_priv->user_addr;
2321 	}
2322 	/* Create the BO */
2323 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2324 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2325 						      &offset, bo_bucket->alloc_flags, criu_resume);
2326 	if (ret) {
2327 		pr_err("Could not create the BO\n");
2328 		return ret;
2329 	}
2330 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2331 		 bo_bucket->size, bo_bucket->addr, offset);
2332 
2333 	/* Restore previous IDR handle */
2334 	pr_debug("Restoring old IDR handle for the BO");
2335 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2336 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2337 
2338 	if (idr_handle < 0) {
2339 		pr_err("Could not allocate idr\n");
2340 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2341 						       NULL);
2342 		return -ENOMEM;
2343 	}
2344 
2345 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2346 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2347 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2348 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2349 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2350 		bo_bucket->restored_offset = offset;
2351 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2352 		bo_bucket->restored_offset = offset;
2353 		/* Update the VRAM usage count */
2354 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2355 	}
2356 	return 0;
2357 }
2358 
2359 static int criu_restore_bo(struct kfd_process *p,
2360 			   struct kfd_criu_bo_bucket *bo_bucket,
2361 			   struct kfd_criu_bo_priv_data *bo_priv)
2362 {
2363 	struct kfd_process_device *pdd;
2364 	struct kgd_mem *kgd_mem;
2365 	int ret;
2366 	int j;
2367 
2368 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2369 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2370 		 bo_priv->idr_handle);
2371 
2372 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2373 	if (!pdd) {
2374 		pr_err("Failed to get pdd\n");
2375 		return -ENODEV;
2376 	}
2377 
2378 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2379 	if (ret)
2380 		return ret;
2381 
2382 	/* now map these BOs to GPU/s */
2383 	for (j = 0; j < p->n_pdds; j++) {
2384 		struct kfd_node *peer;
2385 		struct kfd_process_device *peer_pdd;
2386 
2387 		if (!bo_priv->mapped_gpuids[j])
2388 			break;
2389 
2390 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2391 		if (!peer_pdd)
2392 			return -EINVAL;
2393 
2394 		peer = peer_pdd->dev;
2395 
2396 		peer_pdd = kfd_bind_process_to_device(peer, p);
2397 		if (IS_ERR(peer_pdd))
2398 			return PTR_ERR(peer_pdd);
2399 
2400 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2401 							    peer_pdd->drm_priv);
2402 		if (ret) {
2403 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2404 			return ret;
2405 		}
2406 	}
2407 
2408 	pr_debug("map memory was successful for the BO\n");
2409 	/* create the dmabuf object and export the bo */
2410 	if (bo_bucket->alloc_flags
2411 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2412 		ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2413 					    &bo_bucket->dmabuf_fd);
2414 		if (ret)
2415 			return ret;
2416 	} else {
2417 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2418 	}
2419 
2420 	return 0;
2421 }
2422 
2423 static int criu_restore_bos(struct kfd_process *p,
2424 			    struct kfd_ioctl_criu_args *args,
2425 			    uint64_t *priv_offset,
2426 			    uint64_t max_priv_data_size)
2427 {
2428 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2429 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2430 	int ret = 0;
2431 	uint32_t i = 0;
2432 
2433 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2434 		return -EINVAL;
2435 
2436 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2437 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2438 
2439 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2440 	if (!bo_buckets)
2441 		return -ENOMEM;
2442 
2443 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2444 			     args->num_bos * sizeof(*bo_buckets));
2445 	if (ret) {
2446 		pr_err("Failed to copy BOs information from user\n");
2447 		ret = -EFAULT;
2448 		goto exit;
2449 	}
2450 
2451 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2452 	if (!bo_privs) {
2453 		ret = -ENOMEM;
2454 		goto exit;
2455 	}
2456 
2457 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2458 			     args->num_bos * sizeof(*bo_privs));
2459 	if (ret) {
2460 		pr_err("Failed to copy BOs information from user\n");
2461 		ret = -EFAULT;
2462 		goto exit;
2463 	}
2464 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2465 
2466 	/* Create and map new BOs */
2467 	for (; i < args->num_bos; i++) {
2468 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2469 		if (ret) {
2470 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2471 			goto exit;
2472 		}
2473 	} /* done */
2474 
2475 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2476 	ret = copy_to_user((void __user *)args->bos,
2477 				bo_buckets,
2478 				(args->num_bos * sizeof(*bo_buckets)));
2479 	if (ret)
2480 		ret = -EFAULT;
2481 
2482 exit:
2483 	while (ret && i--) {
2484 		if (bo_buckets[i].alloc_flags
2485 		   & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2486 			close_fd(bo_buckets[i].dmabuf_fd);
2487 	}
2488 	kvfree(bo_buckets);
2489 	kvfree(bo_privs);
2490 	return ret;
2491 }
2492 
2493 static int criu_restore_objects(struct file *filep,
2494 				struct kfd_process *p,
2495 				struct kfd_ioctl_criu_args *args,
2496 				uint64_t *priv_offset,
2497 				uint64_t max_priv_data_size)
2498 {
2499 	int ret = 0;
2500 	uint32_t i;
2501 
2502 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2503 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2504 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2505 
2506 	for (i = 0; i < args->num_objects; i++) {
2507 		uint32_t object_type;
2508 
2509 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2510 			pr_err("Invalid private data size\n");
2511 			return -EINVAL;
2512 		}
2513 
2514 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2515 		if (ret) {
2516 			pr_err("Failed to copy private information from user\n");
2517 			goto exit;
2518 		}
2519 
2520 		switch (object_type) {
2521 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2522 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2523 						     priv_offset, max_priv_data_size);
2524 			if (ret)
2525 				goto exit;
2526 			break;
2527 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2528 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2529 						     priv_offset, max_priv_data_size);
2530 			if (ret)
2531 				goto exit;
2532 			break;
2533 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2534 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2535 						     priv_offset, max_priv_data_size);
2536 			if (ret)
2537 				goto exit;
2538 			break;
2539 		default:
2540 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2541 			ret = -EINVAL;
2542 			goto exit;
2543 		}
2544 	}
2545 exit:
2546 	return ret;
2547 }
2548 
2549 static int criu_restore(struct file *filep,
2550 			struct kfd_process *p,
2551 			struct kfd_ioctl_criu_args *args)
2552 {
2553 	uint64_t priv_offset = 0;
2554 	int ret = 0;
2555 
2556 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2557 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2558 
2559 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2560 	    !args->num_devices || !args->num_bos)
2561 		return -EINVAL;
2562 
2563 	mutex_lock(&p->mutex);
2564 
2565 	/*
2566 	 * Set the process to evicted state to avoid running any new queues before all the memory
2567 	 * mappings are ready.
2568 	 */
2569 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2570 	if (ret)
2571 		goto exit_unlock;
2572 
2573 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2574 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2575 	if (ret)
2576 		goto exit_unlock;
2577 
2578 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2579 	if (ret)
2580 		goto exit_unlock;
2581 
2582 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2583 	if (ret)
2584 		goto exit_unlock;
2585 
2586 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2587 	if (ret)
2588 		goto exit_unlock;
2589 
2590 	if (priv_offset != args->priv_data_size) {
2591 		pr_err("Invalid private data size\n");
2592 		ret = -EINVAL;
2593 	}
2594 
2595 exit_unlock:
2596 	mutex_unlock(&p->mutex);
2597 	if (ret)
2598 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2599 	else
2600 		pr_debug("CRIU restore successful\n");
2601 
2602 	return ret;
2603 }
2604 
2605 static int criu_unpause(struct file *filep,
2606 			struct kfd_process *p,
2607 			struct kfd_ioctl_criu_args *args)
2608 {
2609 	int ret;
2610 
2611 	mutex_lock(&p->mutex);
2612 
2613 	if (!p->queues_paused) {
2614 		mutex_unlock(&p->mutex);
2615 		return -EINVAL;
2616 	}
2617 
2618 	ret = kfd_process_restore_queues(p);
2619 	if (ret)
2620 		pr_err("Failed to unpause queues ret:%d\n", ret);
2621 	else
2622 		p->queues_paused = false;
2623 
2624 	mutex_unlock(&p->mutex);
2625 
2626 	return ret;
2627 }
2628 
2629 static int criu_resume(struct file *filep,
2630 			struct kfd_process *p,
2631 			struct kfd_ioctl_criu_args *args)
2632 {
2633 	struct kfd_process *target = NULL;
2634 	struct pid *pid = NULL;
2635 	int ret = 0;
2636 
2637 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2638 		 args->pid);
2639 
2640 	pid = find_get_pid(args->pid);
2641 	if (!pid) {
2642 		pr_err("Cannot find pid info for %i\n", args->pid);
2643 		return -ESRCH;
2644 	}
2645 
2646 	pr_debug("calling kfd_lookup_process_by_pid\n");
2647 	target = kfd_lookup_process_by_pid(pid);
2648 
2649 	put_pid(pid);
2650 
2651 	if (!target) {
2652 		pr_debug("Cannot find process info for %i\n", args->pid);
2653 		return -ESRCH;
2654 	}
2655 
2656 	mutex_lock(&target->mutex);
2657 	ret = kfd_criu_resume_svm(target);
2658 	if (ret) {
2659 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2660 		goto exit;
2661 	}
2662 
2663 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2664 	if (ret)
2665 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2666 
2667 exit:
2668 	mutex_unlock(&target->mutex);
2669 
2670 	kfd_unref_process(target);
2671 	return ret;
2672 }
2673 
2674 static int criu_process_info(struct file *filep,
2675 				struct kfd_process *p,
2676 				struct kfd_ioctl_criu_args *args)
2677 {
2678 	int ret = 0;
2679 
2680 	mutex_lock(&p->mutex);
2681 
2682 	if (!p->n_pdds) {
2683 		pr_err("No pdd for given process\n");
2684 		ret = -ENODEV;
2685 		goto err_unlock;
2686 	}
2687 
2688 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2689 	if (ret)
2690 		goto err_unlock;
2691 
2692 	p->queues_paused = true;
2693 
2694 	args->pid = task_pid_nr_ns(p->lead_thread,
2695 					task_active_pid_ns(p->lead_thread));
2696 
2697 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2698 					   &args->num_objects, &args->priv_data_size);
2699 	if (ret)
2700 		goto err_unlock;
2701 
2702 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2703 				args->num_devices, args->num_bos, args->num_objects,
2704 				args->priv_data_size);
2705 
2706 err_unlock:
2707 	if (ret) {
2708 		kfd_process_restore_queues(p);
2709 		p->queues_paused = false;
2710 	}
2711 	mutex_unlock(&p->mutex);
2712 	return ret;
2713 }
2714 
2715 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2716 {
2717 	struct kfd_ioctl_criu_args *args = data;
2718 	int ret;
2719 
2720 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2721 	switch (args->op) {
2722 	case KFD_CRIU_OP_PROCESS_INFO:
2723 		ret = criu_process_info(filep, p, args);
2724 		break;
2725 	case KFD_CRIU_OP_CHECKPOINT:
2726 		ret = criu_checkpoint(filep, p, args);
2727 		break;
2728 	case KFD_CRIU_OP_UNPAUSE:
2729 		ret = criu_unpause(filep, p, args);
2730 		break;
2731 	case KFD_CRIU_OP_RESTORE:
2732 		ret = criu_restore(filep, p, args);
2733 		break;
2734 	case KFD_CRIU_OP_RESUME:
2735 		ret = criu_resume(filep, p, args);
2736 		break;
2737 	default:
2738 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2739 		ret = -EINVAL;
2740 		break;
2741 	}
2742 
2743 	if (ret)
2744 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2745 
2746 	return ret;
2747 }
2748 
2749 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2750 			bool enable_ttmp_setup)
2751 {
2752 	int i = 0, ret = 0;
2753 
2754 	if (p->is_runtime_retry)
2755 		goto retry;
2756 
2757 	if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2758 		return -EBUSY;
2759 
2760 	for (i = 0; i < p->n_pdds; i++) {
2761 		struct kfd_process_device *pdd = p->pdds[i];
2762 
2763 		if (pdd->qpd.queue_count)
2764 			return -EEXIST;
2765 
2766 		/*
2767 		 * Setup TTMPs by default.
2768 		 * Note that this call must remain here for MES ADD QUEUE to
2769 		 * skip_process_ctx_clear unconditionally as the first call to
2770 		 * SET_SHADER_DEBUGGER clears any stale process context data
2771 		 * saved in MES.
2772 		 */
2773 		if (pdd->dev->kfd->shared_resources.enable_mes)
2774 			kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
2775 	}
2776 
2777 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2778 	p->runtime_info.r_debug = r_debug;
2779 	p->runtime_info.ttmp_setup = enable_ttmp_setup;
2780 
2781 	if (p->runtime_info.ttmp_setup) {
2782 		for (i = 0; i < p->n_pdds; i++) {
2783 			struct kfd_process_device *pdd = p->pdds[i];
2784 
2785 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2786 				amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2787 				pdd->dev->kfd2kgd->enable_debug_trap(
2788 						pdd->dev->adev,
2789 						true,
2790 						pdd->dev->vm_info.last_vmid_kfd);
2791 			} else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2792 				pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2793 						pdd->dev->adev,
2794 						false,
2795 						0);
2796 			}
2797 		}
2798 	}
2799 
2800 retry:
2801 	if (p->debug_trap_enabled) {
2802 		if (!p->is_runtime_retry) {
2803 			kfd_dbg_trap_activate(p);
2804 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2805 					p, NULL, 0, false, NULL, 0);
2806 		}
2807 
2808 		mutex_unlock(&p->mutex);
2809 		ret = down_interruptible(&p->runtime_enable_sema);
2810 		mutex_lock(&p->mutex);
2811 
2812 		p->is_runtime_retry = !!ret;
2813 	}
2814 
2815 	return ret;
2816 }
2817 
2818 static int runtime_disable(struct kfd_process *p)
2819 {
2820 	int i = 0, ret;
2821 	bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2822 
2823 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2824 	p->runtime_info.r_debug = 0;
2825 
2826 	if (p->debug_trap_enabled) {
2827 		if (was_enabled)
2828 			kfd_dbg_trap_deactivate(p, false, 0);
2829 
2830 		if (!p->is_runtime_retry)
2831 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2832 					p, NULL, 0, false, NULL, 0);
2833 
2834 		mutex_unlock(&p->mutex);
2835 		ret = down_interruptible(&p->runtime_enable_sema);
2836 		mutex_lock(&p->mutex);
2837 
2838 		p->is_runtime_retry = !!ret;
2839 		if (ret)
2840 			return ret;
2841 	}
2842 
2843 	if (was_enabled && p->runtime_info.ttmp_setup) {
2844 		for (i = 0; i < p->n_pdds; i++) {
2845 			struct kfd_process_device *pdd = p->pdds[i];
2846 
2847 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2848 				amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2849 		}
2850 	}
2851 
2852 	p->runtime_info.ttmp_setup = false;
2853 
2854 	/* disable ttmp setup */
2855 	for (i = 0; i < p->n_pdds; i++) {
2856 		struct kfd_process_device *pdd = p->pdds[i];
2857 
2858 		if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2859 			pdd->spi_dbg_override =
2860 					pdd->dev->kfd2kgd->disable_debug_trap(
2861 					pdd->dev->adev,
2862 					false,
2863 					pdd->dev->vm_info.last_vmid_kfd);
2864 
2865 			if (!pdd->dev->kfd->shared_resources.enable_mes)
2866 				debug_refresh_runlist(pdd->dev->dqm);
2867 			else
2868 				kfd_dbg_set_mes_debug_mode(pdd,
2869 							   !kfd_dbg_has_cwsr_workaround(pdd->dev));
2870 		}
2871 	}
2872 
2873 	return 0;
2874 }
2875 
2876 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2877 {
2878 	struct kfd_ioctl_runtime_enable_args *args = data;
2879 	int r;
2880 
2881 	mutex_lock(&p->mutex);
2882 
2883 	if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2884 		r = runtime_enable(p, args->r_debug,
2885 				!!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2886 	else
2887 		r = runtime_disable(p);
2888 
2889 	mutex_unlock(&p->mutex);
2890 
2891 	return r;
2892 }
2893 
2894 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2895 {
2896 	struct kfd_ioctl_dbg_trap_args *args = data;
2897 	struct task_struct *thread = NULL;
2898 	struct mm_struct *mm = NULL;
2899 	struct pid *pid = NULL;
2900 	struct kfd_process *target = NULL;
2901 	struct kfd_process_device *pdd = NULL;
2902 	int r = 0;
2903 
2904 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2905 		pr_err("Debugging does not support sched_policy %i", sched_policy);
2906 		return -EINVAL;
2907 	}
2908 
2909 	pid = find_get_pid(args->pid);
2910 	if (!pid) {
2911 		pr_debug("Cannot find pid info for %i\n", args->pid);
2912 		r = -ESRCH;
2913 		goto out;
2914 	}
2915 
2916 	thread = get_pid_task(pid, PIDTYPE_PID);
2917 	if (!thread) {
2918 		r = -ESRCH;
2919 		goto out;
2920 	}
2921 
2922 	mm = get_task_mm(thread);
2923 	if (!mm) {
2924 		r = -ESRCH;
2925 		goto out;
2926 	}
2927 
2928 	if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2929 		bool create_process;
2930 
2931 		rcu_read_lock();
2932 		create_process = thread && thread != current && ptrace_parent(thread) == current;
2933 		rcu_read_unlock();
2934 
2935 		target = create_process ? kfd_create_process(thread) :
2936 					kfd_lookup_process_by_pid(pid);
2937 	} else {
2938 		target = kfd_lookup_process_by_pid(pid);
2939 	}
2940 
2941 	if (IS_ERR_OR_NULL(target)) {
2942 		pr_debug("Cannot find process PID %i to debug\n", args->pid);
2943 		r = target ? PTR_ERR(target) : -ESRCH;
2944 		target = NULL;
2945 		goto out;
2946 	}
2947 
2948 	/* Check if target is still PTRACED. */
2949 	rcu_read_lock();
2950 	if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2951 				&& ptrace_parent(target->lead_thread) != current) {
2952 		pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2953 		r = -EPERM;
2954 	}
2955 	rcu_read_unlock();
2956 
2957 	if (r)
2958 		goto out;
2959 
2960 	mutex_lock(&target->mutex);
2961 
2962 	if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2963 		pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2964 		r = -EINVAL;
2965 		goto unlock_out;
2966 	}
2967 
2968 	if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2969 			(args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2970 			 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2971 			 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2972 			 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2973 			 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2974 			 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2975 			 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2976 		r = -EPERM;
2977 		goto unlock_out;
2978 	}
2979 
2980 	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2981 	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2982 		int user_gpu_id = kfd_process_get_user_gpu_id(target,
2983 				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2984 					args->set_node_address_watch.gpu_id :
2985 					args->clear_node_address_watch.gpu_id);
2986 
2987 		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
2988 		if (user_gpu_id == -EINVAL || !pdd) {
2989 			r = -ENODEV;
2990 			goto unlock_out;
2991 		}
2992 	}
2993 
2994 	switch (args->op) {
2995 	case KFD_IOC_DBG_TRAP_ENABLE:
2996 		if (target != p)
2997 			target->debugger_process = p;
2998 
2999 		r = kfd_dbg_trap_enable(target,
3000 					args->enable.dbg_fd,
3001 					(void __user *)args->enable.rinfo_ptr,
3002 					&args->enable.rinfo_size);
3003 		if (!r)
3004 			target->exception_enable_mask = args->enable.exception_mask;
3005 
3006 		break;
3007 	case KFD_IOC_DBG_TRAP_DISABLE:
3008 		r = kfd_dbg_trap_disable(target);
3009 		break;
3010 	case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
3011 		r = kfd_dbg_send_exception_to_runtime(target,
3012 				args->send_runtime_event.gpu_id,
3013 				args->send_runtime_event.queue_id,
3014 				args->send_runtime_event.exception_mask);
3015 		break;
3016 	case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
3017 		kfd_dbg_set_enabled_debug_exception_mask(target,
3018 				args->set_exceptions_enabled.exception_mask);
3019 		break;
3020 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
3021 		r = kfd_dbg_trap_set_wave_launch_override(target,
3022 				args->launch_override.override_mode,
3023 				args->launch_override.enable_mask,
3024 				args->launch_override.support_request_mask,
3025 				&args->launch_override.enable_mask,
3026 				&args->launch_override.support_request_mask);
3027 		break;
3028 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
3029 		r = kfd_dbg_trap_set_wave_launch_mode(target,
3030 				args->launch_mode.launch_mode);
3031 		break;
3032 	case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
3033 		r = suspend_queues(target,
3034 				args->suspend_queues.num_queues,
3035 				args->suspend_queues.grace_period,
3036 				args->suspend_queues.exception_mask,
3037 				(uint32_t *)args->suspend_queues.queue_array_ptr);
3038 
3039 		break;
3040 	case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
3041 		r = resume_queues(target, args->resume_queues.num_queues,
3042 				(uint32_t *)args->resume_queues.queue_array_ptr);
3043 		break;
3044 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3045 		r = kfd_dbg_trap_set_dev_address_watch(pdd,
3046 				args->set_node_address_watch.address,
3047 				args->set_node_address_watch.mask,
3048 				&args->set_node_address_watch.id,
3049 				args->set_node_address_watch.mode);
3050 		break;
3051 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3052 		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3053 				args->clear_node_address_watch.id);
3054 		break;
3055 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
3056 		r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3057 		break;
3058 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3059 		r = kfd_dbg_ev_query_debug_event(target,
3060 				&args->query_debug_event.queue_id,
3061 				&args->query_debug_event.gpu_id,
3062 				args->query_debug_event.exception_mask,
3063 				&args->query_debug_event.exception_mask);
3064 		break;
3065 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3066 		r = kfd_dbg_trap_query_exception_info(target,
3067 				args->query_exception_info.source_id,
3068 				args->query_exception_info.exception_code,
3069 				args->query_exception_info.clear_exception,
3070 				(void __user *)args->query_exception_info.info_ptr,
3071 				&args->query_exception_info.info_size);
3072 		break;
3073 	case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3074 		r = pqm_get_queue_snapshot(&target->pqm,
3075 				args->queue_snapshot.exception_mask,
3076 				(void __user *)args->queue_snapshot.snapshot_buf_ptr,
3077 				&args->queue_snapshot.num_queues,
3078 				&args->queue_snapshot.entry_size);
3079 		break;
3080 	case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3081 		r = kfd_dbg_trap_device_snapshot(target,
3082 				args->device_snapshot.exception_mask,
3083 				(void __user *)args->device_snapshot.snapshot_buf_ptr,
3084 				&args->device_snapshot.num_devices,
3085 				&args->device_snapshot.entry_size);
3086 		break;
3087 	default:
3088 		pr_err("Invalid option: %i\n", args->op);
3089 		r = -EINVAL;
3090 	}
3091 
3092 unlock_out:
3093 	mutex_unlock(&target->mutex);
3094 
3095 out:
3096 	if (thread)
3097 		put_task_struct(thread);
3098 
3099 	if (mm)
3100 		mmput(mm);
3101 
3102 	if (pid)
3103 		put_pid(pid);
3104 
3105 	if (target)
3106 		kfd_unref_process(target);
3107 
3108 	return r;
3109 }
3110 
3111 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3112 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3113 			    .cmd_drv = 0, .name = #ioctl}
3114 
3115 /** Ioctl table */
3116 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3117 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3118 			kfd_ioctl_get_version, 0),
3119 
3120 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3121 			kfd_ioctl_create_queue, 0),
3122 
3123 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3124 			kfd_ioctl_destroy_queue, 0),
3125 
3126 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3127 			kfd_ioctl_set_memory_policy, 0),
3128 
3129 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3130 			kfd_ioctl_get_clock_counters, 0),
3131 
3132 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3133 			kfd_ioctl_get_process_apertures, 0),
3134 
3135 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3136 			kfd_ioctl_update_queue, 0),
3137 
3138 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3139 			kfd_ioctl_create_event, 0),
3140 
3141 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3142 			kfd_ioctl_destroy_event, 0),
3143 
3144 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3145 			kfd_ioctl_set_event, 0),
3146 
3147 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3148 			kfd_ioctl_reset_event, 0),
3149 
3150 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3151 			kfd_ioctl_wait_events, 0),
3152 
3153 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3154 			kfd_ioctl_dbg_register, 0),
3155 
3156 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3157 			kfd_ioctl_dbg_unregister, 0),
3158 
3159 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3160 			kfd_ioctl_dbg_address_watch, 0),
3161 
3162 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3163 			kfd_ioctl_dbg_wave_control, 0),
3164 
3165 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3166 			kfd_ioctl_set_scratch_backing_va, 0),
3167 
3168 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3169 			kfd_ioctl_get_tile_config, 0),
3170 
3171 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3172 			kfd_ioctl_set_trap_handler, 0),
3173 
3174 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3175 			kfd_ioctl_get_process_apertures_new, 0),
3176 
3177 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3178 			kfd_ioctl_acquire_vm, 0),
3179 
3180 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3181 			kfd_ioctl_alloc_memory_of_gpu, 0),
3182 
3183 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3184 			kfd_ioctl_free_memory_of_gpu, 0),
3185 
3186 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3187 			kfd_ioctl_map_memory_to_gpu, 0),
3188 
3189 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3190 			kfd_ioctl_unmap_memory_from_gpu, 0),
3191 
3192 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3193 			kfd_ioctl_set_cu_mask, 0),
3194 
3195 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3196 			kfd_ioctl_get_queue_wave_state, 0),
3197 
3198 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3199 				kfd_ioctl_get_dmabuf_info, 0),
3200 
3201 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3202 				kfd_ioctl_import_dmabuf, 0),
3203 
3204 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3205 			kfd_ioctl_alloc_queue_gws, 0),
3206 
3207 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3208 			kfd_ioctl_smi_events, 0),
3209 
3210 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3211 
3212 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3213 			kfd_ioctl_set_xnack_mode, 0),
3214 
3215 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3216 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3217 
3218 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3219 			kfd_ioctl_get_available_memory, 0),
3220 
3221 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3222 				kfd_ioctl_export_dmabuf, 0),
3223 
3224 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3225 			kfd_ioctl_runtime_enable, 0),
3226 
3227 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3228 			kfd_ioctl_set_debug_trap, 0),
3229 };
3230 
3231 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
3232 
3233 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3234 {
3235 	struct kfd_process *process;
3236 	amdkfd_ioctl_t *func;
3237 	const struct amdkfd_ioctl_desc *ioctl = NULL;
3238 	unsigned int nr = _IOC_NR(cmd);
3239 	char stack_kdata[128];
3240 	char *kdata = NULL;
3241 	unsigned int usize, asize;
3242 	int retcode = -EINVAL;
3243 	bool ptrace_attached = false;
3244 
3245 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3246 		goto err_i1;
3247 
3248 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3249 		u32 amdkfd_size;
3250 
3251 		ioctl = &amdkfd_ioctls[nr];
3252 
3253 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
3254 		usize = asize = _IOC_SIZE(cmd);
3255 		if (amdkfd_size > asize)
3256 			asize = amdkfd_size;
3257 
3258 		cmd = ioctl->cmd;
3259 	} else
3260 		goto err_i1;
3261 
3262 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3263 
3264 	/* Get the process struct from the filep. Only the process
3265 	 * that opened /dev/kfd can use the file descriptor. Child
3266 	 * processes need to create their own KFD device context.
3267 	 */
3268 	process = filep->private_data;
3269 
3270 	rcu_read_lock();
3271 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3272 	    ptrace_parent(process->lead_thread) == current)
3273 		ptrace_attached = true;
3274 	rcu_read_unlock();
3275 
3276 	if (process->lead_thread != current->group_leader
3277 	    && !ptrace_attached) {
3278 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3279 		retcode = -EBADF;
3280 		goto err_i1;
3281 	}
3282 
3283 	/* Do not trust userspace, use our own definition */
3284 	func = ioctl->func;
3285 
3286 	if (unlikely(!func)) {
3287 		dev_dbg(kfd_device, "no function\n");
3288 		retcode = -EINVAL;
3289 		goto err_i1;
3290 	}
3291 
3292 	/*
3293 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3294 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3295 	 * more priviledged access.
3296 	 */
3297 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3298 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
3299 						!capable(CAP_SYS_ADMIN)) {
3300 			retcode = -EACCES;
3301 			goto err_i1;
3302 		}
3303 	}
3304 
3305 	if (cmd & (IOC_IN | IOC_OUT)) {
3306 		if (asize <= sizeof(stack_kdata)) {
3307 			kdata = stack_kdata;
3308 		} else {
3309 			kdata = kmalloc(asize, GFP_KERNEL);
3310 			if (!kdata) {
3311 				retcode = -ENOMEM;
3312 				goto err_i1;
3313 			}
3314 		}
3315 		if (asize > usize)
3316 			memset(kdata + usize, 0, asize - usize);
3317 	}
3318 
3319 	if (cmd & IOC_IN) {
3320 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3321 			retcode = -EFAULT;
3322 			goto err_i1;
3323 		}
3324 	} else if (cmd & IOC_OUT) {
3325 		memset(kdata, 0, usize);
3326 	}
3327 
3328 	retcode = func(filep, process, kdata);
3329 
3330 	if (cmd & IOC_OUT)
3331 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3332 			retcode = -EFAULT;
3333 
3334 err_i1:
3335 	if (!ioctl)
3336 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3337 			  task_pid_nr(current), cmd, nr);
3338 
3339 	if (kdata != stack_kdata)
3340 		kfree(kdata);
3341 
3342 	if (retcode)
3343 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3344 				nr, arg, retcode);
3345 
3346 	return retcode;
3347 }
3348 
3349 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3350 		      struct vm_area_struct *vma)
3351 {
3352 	phys_addr_t address;
3353 
3354 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3355 		return -EINVAL;
3356 
3357 	address = dev->adev->rmmio_remap.bus_addr;
3358 
3359 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3360 				VM_DONTDUMP | VM_PFNMAP);
3361 
3362 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3363 
3364 	pr_debug("pasid 0x%x mapping mmio page\n"
3365 		 "     target user address == 0x%08llX\n"
3366 		 "     physical address    == 0x%08llX\n"
3367 		 "     vm_flags            == 0x%04lX\n"
3368 		 "     size                == 0x%04lX\n",
3369 		 process->pasid, (unsigned long long) vma->vm_start,
3370 		 address, vma->vm_flags, PAGE_SIZE);
3371 
3372 	return io_remap_pfn_range(vma,
3373 				vma->vm_start,
3374 				address >> PAGE_SHIFT,
3375 				PAGE_SIZE,
3376 				vma->vm_page_prot);
3377 }
3378 
3379 
3380 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3381 {
3382 	struct kfd_process *process;
3383 	struct kfd_node *dev = NULL;
3384 	unsigned long mmap_offset;
3385 	unsigned int gpu_id;
3386 
3387 	process = kfd_get_process(current);
3388 	if (IS_ERR(process))
3389 		return PTR_ERR(process);
3390 
3391 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3392 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3393 	if (gpu_id)
3394 		dev = kfd_device_by_id(gpu_id);
3395 
3396 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3397 	case KFD_MMAP_TYPE_DOORBELL:
3398 		if (!dev)
3399 			return -ENODEV;
3400 		return kfd_doorbell_mmap(dev, process, vma);
3401 
3402 	case KFD_MMAP_TYPE_EVENTS:
3403 		return kfd_event_mmap(process, vma);
3404 
3405 	case KFD_MMAP_TYPE_RESERVED_MEM:
3406 		if (!dev)
3407 			return -ENODEV;
3408 		return kfd_reserved_mem_mmap(dev, process, vma);
3409 	case KFD_MMAP_TYPE_MMIO:
3410 		if (!dev)
3411 			return -ENODEV;
3412 		return kfd_mmio_mmap(dev, process, vma);
3413 	}
3414 
3415 	return -EFAULT;
3416 }
3417