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 
48 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
49 static int kfd_open(struct inode *, struct file *);
50 static int kfd_release(struct inode *, struct file *);
51 static int kfd_mmap(struct file *, struct vm_area_struct *);
52 
53 static const char kfd_dev_name[] = "kfd";
54 
55 static const struct file_operations kfd_fops = {
56 	.owner = THIS_MODULE,
57 	.unlocked_ioctl = kfd_ioctl,
58 	.compat_ioctl = compat_ptr_ioctl,
59 	.open = kfd_open,
60 	.release = kfd_release,
61 	.mmap = kfd_mmap,
62 };
63 
64 static int kfd_char_dev_major = -1;
65 static struct class *kfd_class;
66 struct device *kfd_device;
67 
68 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
69 {
70 	struct kfd_process_device *pdd;
71 
72 	mutex_lock(&p->mutex);
73 	pdd = kfd_process_device_data_by_id(p, gpu_id);
74 
75 	if (pdd)
76 		return pdd;
77 
78 	mutex_unlock(&p->mutex);
79 	return NULL;
80 }
81 
82 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
83 {
84 	mutex_unlock(&pdd->process->mutex);
85 }
86 
87 int kfd_chardev_init(void)
88 {
89 	int err = 0;
90 
91 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
92 	err = kfd_char_dev_major;
93 	if (err < 0)
94 		goto err_register_chrdev;
95 
96 	kfd_class = class_create(THIS_MODULE, kfd_dev_name);
97 	err = PTR_ERR(kfd_class);
98 	if (IS_ERR(kfd_class))
99 		goto err_class_create;
100 
101 	kfd_device = device_create(kfd_class, NULL,
102 					MKDEV(kfd_char_dev_major, 0),
103 					NULL, kfd_dev_name);
104 	err = PTR_ERR(kfd_device);
105 	if (IS_ERR(kfd_device))
106 		goto err_device_create;
107 
108 	return 0;
109 
110 err_device_create:
111 	class_destroy(kfd_class);
112 err_class_create:
113 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
114 err_register_chrdev:
115 	return err;
116 }
117 
118 void kfd_chardev_exit(void)
119 {
120 	device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
121 	class_destroy(kfd_class);
122 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
123 	kfd_device = NULL;
124 }
125 
126 
127 static int kfd_open(struct inode *inode, struct file *filep)
128 {
129 	struct kfd_process *process;
130 	bool is_32bit_user_mode;
131 
132 	if (iminor(inode) != 0)
133 		return -ENODEV;
134 
135 	is_32bit_user_mode = in_compat_syscall();
136 
137 	if (is_32bit_user_mode) {
138 		dev_warn(kfd_device,
139 			"Process %d (32-bit) failed to open /dev/kfd\n"
140 			"32-bit processes are not supported by amdkfd\n",
141 			current->pid);
142 		return -EPERM;
143 	}
144 
145 	process = kfd_create_process(filep);
146 	if (IS_ERR(process))
147 		return PTR_ERR(process);
148 
149 	if (kfd_is_locked()) {
150 		dev_dbg(kfd_device, "kfd is locked!\n"
151 				"process %d unreferenced", process->pasid);
152 		kfd_unref_process(process);
153 		return -EAGAIN;
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 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
190 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
191 		return -EINVAL;
192 	}
193 
194 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
195 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
196 		return -EINVAL;
197 	}
198 
199 	if ((args->ring_base_address) &&
200 		(!access_ok((const void __user *) args->ring_base_address,
201 			sizeof(uint64_t)))) {
202 		pr_err("Can't access ring base address\n");
203 		return -EFAULT;
204 	}
205 
206 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
207 		pr_err("Ring size must be a power of 2 or 0\n");
208 		return -EINVAL;
209 	}
210 
211 	if (!access_ok((const void __user *) args->read_pointer_address,
212 			sizeof(uint32_t))) {
213 		pr_err("Can't access read pointer\n");
214 		return -EFAULT;
215 	}
216 
217 	if (!access_ok((const void __user *) args->write_pointer_address,
218 			sizeof(uint32_t))) {
219 		pr_err("Can't access write pointer\n");
220 		return -EFAULT;
221 	}
222 
223 	if (args->eop_buffer_address &&
224 		!access_ok((const void __user *) args->eop_buffer_address,
225 			sizeof(uint32_t))) {
226 		pr_debug("Can't access eop buffer");
227 		return -EFAULT;
228 	}
229 
230 	if (args->ctx_save_restore_address &&
231 		!access_ok((const void __user *) args->ctx_save_restore_address,
232 			sizeof(uint32_t))) {
233 		pr_debug("Can't access ctx save restore buffer");
234 		return -EFAULT;
235 	}
236 
237 	q_properties->is_interop = false;
238 	q_properties->is_gws = false;
239 	q_properties->queue_percent = args->queue_percentage;
240 	q_properties->priority = args->queue_priority;
241 	q_properties->queue_address = args->ring_base_address;
242 	q_properties->queue_size = args->ring_size;
243 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
244 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
245 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
246 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
247 	q_properties->ctx_save_restore_area_address =
248 			args->ctx_save_restore_address;
249 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
250 	q_properties->ctl_stack_size = args->ctl_stack_size;
251 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
252 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
253 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
254 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
255 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
256 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
257 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
258 	else
259 		return -ENOTSUPP;
260 
261 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
262 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
263 	else
264 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
265 
266 	pr_debug("Queue Percentage: %d, %d\n",
267 			q_properties->queue_percent, args->queue_percentage);
268 
269 	pr_debug("Queue Priority: %d, %d\n",
270 			q_properties->priority, args->queue_priority);
271 
272 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
273 			q_properties->queue_address, args->ring_base_address);
274 
275 	pr_debug("Queue Size: 0x%llX, %u\n",
276 			q_properties->queue_size, args->ring_size);
277 
278 	pr_debug("Queue r/w Pointers: %px, %px\n",
279 			q_properties->read_ptr,
280 			q_properties->write_ptr);
281 
282 	pr_debug("Queue Format: %d\n", q_properties->format);
283 
284 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
285 
286 	pr_debug("Queue CTX save area: 0x%llX\n",
287 			q_properties->ctx_save_restore_area_address);
288 
289 	return 0;
290 }
291 
292 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
293 					void *data)
294 {
295 	struct kfd_ioctl_create_queue_args *args = data;
296 	struct kfd_dev *dev;
297 	int err = 0;
298 	unsigned int queue_id;
299 	struct kfd_process_device *pdd;
300 	struct queue_properties q_properties;
301 	uint32_t doorbell_offset_in_process = 0;
302 	struct amdgpu_bo *wptr_bo = NULL;
303 
304 	memset(&q_properties, 0, sizeof(struct queue_properties));
305 
306 	pr_debug("Creating queue ioctl\n");
307 
308 	err = set_queue_properties_from_user(&q_properties, args);
309 	if (err)
310 		return err;
311 
312 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
313 
314 	mutex_lock(&p->mutex);
315 
316 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
317 	if (!pdd) {
318 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
319 		err = -EINVAL;
320 		goto err_pdd;
321 	}
322 	dev = pdd->dev;
323 
324 	pdd = kfd_bind_process_to_device(dev, p);
325 	if (IS_ERR(pdd)) {
326 		err = -ESRCH;
327 		goto err_bind_process;
328 	}
329 
330 	if (!pdd->doorbell_index &&
331 	    kfd_alloc_process_doorbells(dev, &pdd->doorbell_index) < 0) {
332 		err = -ENOMEM;
333 		goto err_alloc_doorbells;
334 	}
335 
336 	/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
337 	 * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
338 	 */
339 	if (dev->shared_resources.enable_mes &&
340 			((dev->adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK)
341 			>> AMDGPU_MES_API_VERSION_SHIFT) >= 2) {
342 		struct amdgpu_bo_va_mapping *wptr_mapping;
343 		struct amdgpu_vm *wptr_vm;
344 
345 		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
346 		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
347 		if (err)
348 			goto err_wptr_map_gart;
349 
350 		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
351 				wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
352 		amdgpu_bo_unreserve(wptr_vm->root.bo);
353 		if (!wptr_mapping) {
354 			pr_err("Failed to lookup wptr bo\n");
355 			err = -EINVAL;
356 			goto err_wptr_map_gart;
357 		}
358 
359 		wptr_bo = wptr_mapping->bo_va->base.bo;
360 		if (wptr_bo->tbo.base.size > PAGE_SIZE) {
361 			pr_err("Requested GART mapping for wptr bo larger than one page\n");
362 			err = -EINVAL;
363 			goto err_wptr_map_gart;
364 		}
365 
366 		err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
367 		if (err) {
368 			pr_err("Failed to map wptr bo to GART\n");
369 			goto err_wptr_map_gart;
370 		}
371 	}
372 
373 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
374 			p->pasid,
375 			dev->id);
376 
377 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
378 			NULL, NULL, NULL, &doorbell_offset_in_process);
379 	if (err != 0)
380 		goto err_create_queue;
381 
382 	args->queue_id = queue_id;
383 
384 
385 	/* Return gpu_id as doorbell offset for mmap usage */
386 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
387 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
388 	if (KFD_IS_SOC15(dev))
389 		/* On SOC15 ASICs, include the doorbell offset within the
390 		 * process doorbell frame, which is 2 pages.
391 		 */
392 		args->doorbell_offset |= doorbell_offset_in_process;
393 
394 	mutex_unlock(&p->mutex);
395 
396 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
397 
398 	pr_debug("Ring buffer address == 0x%016llX\n",
399 			args->ring_base_address);
400 
401 	pr_debug("Read ptr address    == 0x%016llX\n",
402 			args->read_pointer_address);
403 
404 	pr_debug("Write ptr address   == 0x%016llX\n",
405 			args->write_pointer_address);
406 
407 	return 0;
408 
409 err_create_queue:
410 	if (wptr_bo)
411 		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
412 err_wptr_map_gart:
413 err_alloc_doorbells:
414 err_bind_process:
415 err_pdd:
416 	mutex_unlock(&p->mutex);
417 	return err;
418 }
419 
420 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
421 					void *data)
422 {
423 	int retval;
424 	struct kfd_ioctl_destroy_queue_args *args = data;
425 
426 	pr_debug("Destroying queue id %d for pasid 0x%x\n",
427 				args->queue_id,
428 				p->pasid);
429 
430 	mutex_lock(&p->mutex);
431 
432 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
433 
434 	mutex_unlock(&p->mutex);
435 	return retval;
436 }
437 
438 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
439 					void *data)
440 {
441 	int retval;
442 	struct kfd_ioctl_update_queue_args *args = data;
443 	struct queue_properties properties;
444 
445 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
446 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
447 		return -EINVAL;
448 	}
449 
450 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
451 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
452 		return -EINVAL;
453 	}
454 
455 	if ((args->ring_base_address) &&
456 		(!access_ok((const void __user *) args->ring_base_address,
457 			sizeof(uint64_t)))) {
458 		pr_err("Can't access ring base address\n");
459 		return -EFAULT;
460 	}
461 
462 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
463 		pr_err("Ring size must be a power of 2 or 0\n");
464 		return -EINVAL;
465 	}
466 
467 	properties.queue_address = args->ring_base_address;
468 	properties.queue_size = args->ring_size;
469 	properties.queue_percent = args->queue_percentage;
470 	properties.priority = args->queue_priority;
471 
472 	pr_debug("Updating queue id %d for pasid 0x%x\n",
473 			args->queue_id, p->pasid);
474 
475 	mutex_lock(&p->mutex);
476 
477 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
478 
479 	mutex_unlock(&p->mutex);
480 
481 	return retval;
482 }
483 
484 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
485 					void *data)
486 {
487 	int retval;
488 	const int max_num_cus = 1024;
489 	struct kfd_ioctl_set_cu_mask_args *args = data;
490 	struct mqd_update_info minfo = {0};
491 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
492 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
493 
494 	if ((args->num_cu_mask % 32) != 0) {
495 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
496 				args->num_cu_mask);
497 		return -EINVAL;
498 	}
499 
500 	minfo.cu_mask.count = args->num_cu_mask;
501 	if (minfo.cu_mask.count == 0) {
502 		pr_debug("CU mask cannot be 0");
503 		return -EINVAL;
504 	}
505 
506 	/* To prevent an unreasonably large CU mask size, set an arbitrary
507 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
508 	 * past max_num_cus bits and just use the first max_num_cus bits.
509 	 */
510 	if (minfo.cu_mask.count > max_num_cus) {
511 		pr_debug("CU mask cannot be greater than 1024 bits");
512 		minfo.cu_mask.count = max_num_cus;
513 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
514 	}
515 
516 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
517 	if (!minfo.cu_mask.ptr)
518 		return -ENOMEM;
519 
520 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
521 	if (retval) {
522 		pr_debug("Could not copy CU mask from userspace");
523 		retval = -EFAULT;
524 		goto out;
525 	}
526 
527 	minfo.update_flag = UPDATE_FLAG_CU_MASK;
528 
529 	mutex_lock(&p->mutex);
530 
531 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
532 
533 	mutex_unlock(&p->mutex);
534 
535 out:
536 	kfree(minfo.cu_mask.ptr);
537 	return retval;
538 }
539 
540 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
541 					  struct kfd_process *p, void *data)
542 {
543 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
544 	int r;
545 
546 	mutex_lock(&p->mutex);
547 
548 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
549 			       (void __user *)args->ctl_stack_address,
550 			       &args->ctl_stack_used_size,
551 			       &args->save_area_used_size);
552 
553 	mutex_unlock(&p->mutex);
554 
555 	return r;
556 }
557 
558 static int kfd_ioctl_set_memory_policy(struct file *filep,
559 					struct kfd_process *p, void *data)
560 {
561 	struct kfd_ioctl_set_memory_policy_args *args = data;
562 	int err = 0;
563 	struct kfd_process_device *pdd;
564 	enum cache_policy default_policy, alternate_policy;
565 
566 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
567 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
568 		return -EINVAL;
569 	}
570 
571 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
572 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
573 		return -EINVAL;
574 	}
575 
576 	mutex_lock(&p->mutex);
577 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
578 	if (!pdd) {
579 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
580 		err = -EINVAL;
581 		goto err_pdd;
582 	}
583 
584 	pdd = kfd_bind_process_to_device(pdd->dev, p);
585 	if (IS_ERR(pdd)) {
586 		err = -ESRCH;
587 		goto out;
588 	}
589 
590 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
591 			 ? cache_policy_coherent : cache_policy_noncoherent;
592 
593 	alternate_policy =
594 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
595 		   ? cache_policy_coherent : cache_policy_noncoherent;
596 
597 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
598 				&pdd->qpd,
599 				default_policy,
600 				alternate_policy,
601 				(void __user *)args->alternate_aperture_base,
602 				args->alternate_aperture_size))
603 		err = -EINVAL;
604 
605 out:
606 err_pdd:
607 	mutex_unlock(&p->mutex);
608 
609 	return err;
610 }
611 
612 static int kfd_ioctl_set_trap_handler(struct file *filep,
613 					struct kfd_process *p, void *data)
614 {
615 	struct kfd_ioctl_set_trap_handler_args *args = data;
616 	int err = 0;
617 	struct kfd_process_device *pdd;
618 
619 	mutex_lock(&p->mutex);
620 
621 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
622 	if (!pdd) {
623 		err = -EINVAL;
624 		goto err_pdd;
625 	}
626 
627 	pdd = kfd_bind_process_to_device(pdd->dev, p);
628 	if (IS_ERR(pdd)) {
629 		err = -ESRCH;
630 		goto out;
631 	}
632 
633 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
634 
635 out:
636 err_pdd:
637 	mutex_unlock(&p->mutex);
638 
639 	return err;
640 }
641 
642 static int kfd_ioctl_dbg_register(struct file *filep,
643 				struct kfd_process *p, void *data)
644 {
645 	return -EPERM;
646 }
647 
648 static int kfd_ioctl_dbg_unregister(struct file *filep,
649 				struct kfd_process *p, void *data)
650 {
651 	return -EPERM;
652 }
653 
654 static int kfd_ioctl_dbg_address_watch(struct file *filep,
655 					struct kfd_process *p, void *data)
656 {
657 	return -EPERM;
658 }
659 
660 /* Parse and generate fixed size data structure for wave control */
661 static int kfd_ioctl_dbg_wave_control(struct file *filep,
662 					struct kfd_process *p, void *data)
663 {
664 	return -EPERM;
665 }
666 
667 static int kfd_ioctl_get_clock_counters(struct file *filep,
668 				struct kfd_process *p, void *data)
669 {
670 	struct kfd_ioctl_get_clock_counters_args *args = data;
671 	struct kfd_process_device *pdd;
672 
673 	mutex_lock(&p->mutex);
674 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
675 	mutex_unlock(&p->mutex);
676 	if (pdd)
677 		/* Reading GPU clock counter from KGD */
678 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
679 	else
680 		/* Node without GPU resource */
681 		args->gpu_clock_counter = 0;
682 
683 	/* No access to rdtsc. Using raw monotonic time */
684 	args->cpu_clock_counter = ktime_get_raw_ns();
685 	args->system_clock_counter = ktime_get_boottime_ns();
686 
687 	/* Since the counter is in nano-seconds we use 1GHz frequency */
688 	args->system_clock_freq = 1000000000;
689 
690 	return 0;
691 }
692 
693 
694 static int kfd_ioctl_get_process_apertures(struct file *filp,
695 				struct kfd_process *p, void *data)
696 {
697 	struct kfd_ioctl_get_process_apertures_args *args = data;
698 	struct kfd_process_device_apertures *pAperture;
699 	int i;
700 
701 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
702 
703 	args->num_of_nodes = 0;
704 
705 	mutex_lock(&p->mutex);
706 	/* Run over all pdd of the process */
707 	for (i = 0; i < p->n_pdds; i++) {
708 		struct kfd_process_device *pdd = p->pdds[i];
709 
710 		pAperture =
711 			&args->process_apertures[args->num_of_nodes];
712 		pAperture->gpu_id = pdd->dev->id;
713 		pAperture->lds_base = pdd->lds_base;
714 		pAperture->lds_limit = pdd->lds_limit;
715 		pAperture->gpuvm_base = pdd->gpuvm_base;
716 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
717 		pAperture->scratch_base = pdd->scratch_base;
718 		pAperture->scratch_limit = pdd->scratch_limit;
719 
720 		dev_dbg(kfd_device,
721 			"node id %u\n", args->num_of_nodes);
722 		dev_dbg(kfd_device,
723 			"gpu id %u\n", pdd->dev->id);
724 		dev_dbg(kfd_device,
725 			"lds_base %llX\n", pdd->lds_base);
726 		dev_dbg(kfd_device,
727 			"lds_limit %llX\n", pdd->lds_limit);
728 		dev_dbg(kfd_device,
729 			"gpuvm_base %llX\n", pdd->gpuvm_base);
730 		dev_dbg(kfd_device,
731 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
732 		dev_dbg(kfd_device,
733 			"scratch_base %llX\n", pdd->scratch_base);
734 		dev_dbg(kfd_device,
735 			"scratch_limit %llX\n", pdd->scratch_limit);
736 
737 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
738 			break;
739 	}
740 	mutex_unlock(&p->mutex);
741 
742 	return 0;
743 }
744 
745 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
746 				struct kfd_process *p, void *data)
747 {
748 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
749 	struct kfd_process_device_apertures *pa;
750 	int ret;
751 	int i;
752 
753 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
754 
755 	if (args->num_of_nodes == 0) {
756 		/* Return number of nodes, so that user space can alloacate
757 		 * sufficient memory
758 		 */
759 		mutex_lock(&p->mutex);
760 		args->num_of_nodes = p->n_pdds;
761 		goto out_unlock;
762 	}
763 
764 	/* Fill in process-aperture information for all available
765 	 * nodes, but not more than args->num_of_nodes as that is
766 	 * the amount of memory allocated by user
767 	 */
768 	pa = kzalloc((sizeof(struct kfd_process_device_apertures) *
769 				args->num_of_nodes), GFP_KERNEL);
770 	if (!pa)
771 		return -ENOMEM;
772 
773 	mutex_lock(&p->mutex);
774 
775 	if (!p->n_pdds) {
776 		args->num_of_nodes = 0;
777 		kfree(pa);
778 		goto out_unlock;
779 	}
780 
781 	/* Run over all pdd of the process */
782 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
783 		struct kfd_process_device *pdd = p->pdds[i];
784 
785 		pa[i].gpu_id = pdd->dev->id;
786 		pa[i].lds_base = pdd->lds_base;
787 		pa[i].lds_limit = pdd->lds_limit;
788 		pa[i].gpuvm_base = pdd->gpuvm_base;
789 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
790 		pa[i].scratch_base = pdd->scratch_base;
791 		pa[i].scratch_limit = pdd->scratch_limit;
792 
793 		dev_dbg(kfd_device,
794 			"gpu id %u\n", pdd->dev->id);
795 		dev_dbg(kfd_device,
796 			"lds_base %llX\n", pdd->lds_base);
797 		dev_dbg(kfd_device,
798 			"lds_limit %llX\n", pdd->lds_limit);
799 		dev_dbg(kfd_device,
800 			"gpuvm_base %llX\n", pdd->gpuvm_base);
801 		dev_dbg(kfd_device,
802 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
803 		dev_dbg(kfd_device,
804 			"scratch_base %llX\n", pdd->scratch_base);
805 		dev_dbg(kfd_device,
806 			"scratch_limit %llX\n", pdd->scratch_limit);
807 	}
808 	mutex_unlock(&p->mutex);
809 
810 	args->num_of_nodes = i;
811 	ret = copy_to_user(
812 			(void __user *)args->kfd_process_device_apertures_ptr,
813 			pa,
814 			(i * sizeof(struct kfd_process_device_apertures)));
815 	kfree(pa);
816 	return ret ? -EFAULT : 0;
817 
818 out_unlock:
819 	mutex_unlock(&p->mutex);
820 	return 0;
821 }
822 
823 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
824 					void *data)
825 {
826 	struct kfd_ioctl_create_event_args *args = data;
827 	int err;
828 
829 	/* For dGPUs the event page is allocated in user mode. The
830 	 * handle is passed to KFD with the first call to this IOCTL
831 	 * through the event_page_offset field.
832 	 */
833 	if (args->event_page_offset) {
834 		mutex_lock(&p->mutex);
835 		err = kfd_kmap_event_page(p, args->event_page_offset);
836 		mutex_unlock(&p->mutex);
837 		if (err)
838 			return err;
839 	}
840 
841 	err = kfd_event_create(filp, p, args->event_type,
842 				args->auto_reset != 0, args->node_id,
843 				&args->event_id, &args->event_trigger_data,
844 				&args->event_page_offset,
845 				&args->event_slot_index);
846 
847 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
848 	return err;
849 }
850 
851 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
852 					void *data)
853 {
854 	struct kfd_ioctl_destroy_event_args *args = data;
855 
856 	return kfd_event_destroy(p, args->event_id);
857 }
858 
859 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
860 				void *data)
861 {
862 	struct kfd_ioctl_set_event_args *args = data;
863 
864 	return kfd_set_event(p, args->event_id);
865 }
866 
867 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
868 				void *data)
869 {
870 	struct kfd_ioctl_reset_event_args *args = data;
871 
872 	return kfd_reset_event(p, args->event_id);
873 }
874 
875 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
876 				void *data)
877 {
878 	struct kfd_ioctl_wait_events_args *args = data;
879 
880 	return kfd_wait_on_events(p, args->num_events,
881 			(void __user *)args->events_ptr,
882 			(args->wait_for_all != 0),
883 			&args->timeout, &args->wait_result);
884 }
885 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
886 					struct kfd_process *p, void *data)
887 {
888 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
889 	struct kfd_process_device *pdd;
890 	struct kfd_dev *dev;
891 	long err;
892 
893 	mutex_lock(&p->mutex);
894 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
895 	if (!pdd) {
896 		err = -EINVAL;
897 		goto err_pdd;
898 	}
899 	dev = pdd->dev;
900 
901 	pdd = kfd_bind_process_to_device(dev, p);
902 	if (IS_ERR(pdd)) {
903 		err = PTR_ERR(pdd);
904 		goto bind_process_to_device_fail;
905 	}
906 
907 	pdd->qpd.sh_hidden_private_base = args->va_addr;
908 
909 	mutex_unlock(&p->mutex);
910 
911 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
912 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
913 		dev->kfd2kgd->set_scratch_backing_va(
914 			dev->adev, args->va_addr, pdd->qpd.vmid);
915 
916 	return 0;
917 
918 bind_process_to_device_fail:
919 err_pdd:
920 	mutex_unlock(&p->mutex);
921 	return err;
922 }
923 
924 static int kfd_ioctl_get_tile_config(struct file *filep,
925 		struct kfd_process *p, void *data)
926 {
927 	struct kfd_ioctl_get_tile_config_args *args = data;
928 	struct kfd_process_device *pdd;
929 	struct tile_config config;
930 	int err = 0;
931 
932 	mutex_lock(&p->mutex);
933 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
934 	mutex_unlock(&p->mutex);
935 	if (!pdd)
936 		return -EINVAL;
937 
938 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
939 
940 	args->gb_addr_config = config.gb_addr_config;
941 	args->num_banks = config.num_banks;
942 	args->num_ranks = config.num_ranks;
943 
944 	if (args->num_tile_configs > config.num_tile_configs)
945 		args->num_tile_configs = config.num_tile_configs;
946 	err = copy_to_user((void __user *)args->tile_config_ptr,
947 			config.tile_config_ptr,
948 			args->num_tile_configs * sizeof(uint32_t));
949 	if (err) {
950 		args->num_tile_configs = 0;
951 		return -EFAULT;
952 	}
953 
954 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
955 		args->num_macro_tile_configs =
956 				config.num_macro_tile_configs;
957 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
958 			config.macro_tile_config_ptr,
959 			args->num_macro_tile_configs * sizeof(uint32_t));
960 	if (err) {
961 		args->num_macro_tile_configs = 0;
962 		return -EFAULT;
963 	}
964 
965 	return 0;
966 }
967 
968 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
969 				void *data)
970 {
971 	struct kfd_ioctl_acquire_vm_args *args = data;
972 	struct kfd_process_device *pdd;
973 	struct file *drm_file;
974 	int ret;
975 
976 	drm_file = fget(args->drm_fd);
977 	if (!drm_file)
978 		return -EINVAL;
979 
980 	mutex_lock(&p->mutex);
981 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
982 	if (!pdd) {
983 		ret = -EINVAL;
984 		goto err_pdd;
985 	}
986 
987 	if (pdd->drm_file) {
988 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
989 		goto err_drm_file;
990 	}
991 
992 	ret = kfd_process_device_init_vm(pdd, drm_file);
993 	if (ret)
994 		goto err_unlock;
995 
996 	/* On success, the PDD keeps the drm_file reference */
997 	mutex_unlock(&p->mutex);
998 
999 	return 0;
1000 
1001 err_unlock:
1002 err_pdd:
1003 err_drm_file:
1004 	mutex_unlock(&p->mutex);
1005 	fput(drm_file);
1006 	return ret;
1007 }
1008 
1009 bool kfd_dev_is_large_bar(struct kfd_dev *dev)
1010 {
1011 	if (debug_largebar) {
1012 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1013 		return true;
1014 	}
1015 
1016 	if (dev->use_iommu_v2)
1017 		return false;
1018 
1019 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1020 			dev->local_mem_info.local_mem_size_public > 0)
1021 		return true;
1022 	return false;
1023 }
1024 
1025 static int kfd_ioctl_get_available_memory(struct file *filep,
1026 					  struct kfd_process *p, void *data)
1027 {
1028 	struct kfd_ioctl_get_available_memory_args *args = data;
1029 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1030 
1031 	if (!pdd)
1032 		return -EINVAL;
1033 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev);
1034 	kfd_unlock_pdd(pdd);
1035 	return 0;
1036 }
1037 
1038 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1039 					struct kfd_process *p, void *data)
1040 {
1041 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1042 	struct kfd_process_device *pdd;
1043 	void *mem;
1044 	struct kfd_dev *dev;
1045 	int idr_handle;
1046 	long err;
1047 	uint64_t offset = args->mmap_offset;
1048 	uint32_t flags = args->flags;
1049 
1050 	if (args->size == 0)
1051 		return -EINVAL;
1052 
1053 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1054 	/* Flush pending deferred work to avoid racing with deferred actions
1055 	 * from previous memory map changes (e.g. munmap).
1056 	 */
1057 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1058 	mutex_lock(&p->svms.lock);
1059 	mmap_write_unlock(current->mm);
1060 	if (interval_tree_iter_first(&p->svms.objects,
1061 				     args->va_addr >> PAGE_SHIFT,
1062 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1063 		pr_err("Address: 0x%llx already allocated by SVM\n",
1064 			args->va_addr);
1065 		mutex_unlock(&p->svms.lock);
1066 		return -EADDRINUSE;
1067 	}
1068 
1069 	/* When register user buffer check if it has been registered by svm by
1070 	 * buffer cpu virtual address.
1071 	 */
1072 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1073 	    interval_tree_iter_first(&p->svms.objects,
1074 				     args->mmap_offset >> PAGE_SHIFT,
1075 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1076 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1077 			args->mmap_offset);
1078 		mutex_unlock(&p->svms.lock);
1079 		return -EADDRINUSE;
1080 	}
1081 
1082 	mutex_unlock(&p->svms.lock);
1083 #endif
1084 	mutex_lock(&p->mutex);
1085 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1086 	if (!pdd) {
1087 		err = -EINVAL;
1088 		goto err_pdd;
1089 	}
1090 
1091 	dev = pdd->dev;
1092 
1093 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1094 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1095 		!kfd_dev_is_large_bar(dev)) {
1096 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1097 		err = -EINVAL;
1098 		goto err_large_bar;
1099 	}
1100 
1101 	pdd = kfd_bind_process_to_device(dev, p);
1102 	if (IS_ERR(pdd)) {
1103 		err = PTR_ERR(pdd);
1104 		goto err_unlock;
1105 	}
1106 
1107 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1108 		if (args->size != kfd_doorbell_process_slice(dev)) {
1109 			err = -EINVAL;
1110 			goto err_unlock;
1111 		}
1112 		offset = kfd_get_process_doorbells(pdd);
1113 		if (!offset) {
1114 			err = -ENOMEM;
1115 			goto err_unlock;
1116 		}
1117 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1118 		if (args->size != PAGE_SIZE) {
1119 			err = -EINVAL;
1120 			goto err_unlock;
1121 		}
1122 		offset = dev->adev->rmmio_remap.bus_addr;
1123 		if (!offset) {
1124 			err = -ENOMEM;
1125 			goto err_unlock;
1126 		}
1127 	}
1128 
1129 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1130 		dev->adev, args->va_addr, args->size,
1131 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1132 		flags, false);
1133 
1134 	if (err)
1135 		goto err_unlock;
1136 
1137 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1138 	if (idr_handle < 0) {
1139 		err = -EFAULT;
1140 		goto err_free;
1141 	}
1142 
1143 	/* Update the VRAM usage count */
1144 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1145 		uint64_t size = args->size;
1146 
1147 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1148 			size >>= 1;
1149 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1150 	}
1151 
1152 	mutex_unlock(&p->mutex);
1153 
1154 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1155 	args->mmap_offset = offset;
1156 
1157 	/* MMIO is mapped through kfd device
1158 	 * Generate a kfd mmap offset
1159 	 */
1160 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1161 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1162 					| KFD_MMAP_GPU_ID(args->gpu_id);
1163 
1164 	return 0;
1165 
1166 err_free:
1167 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1168 					       pdd->drm_priv, NULL);
1169 err_unlock:
1170 err_pdd:
1171 err_large_bar:
1172 	mutex_unlock(&p->mutex);
1173 	return err;
1174 }
1175 
1176 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1177 					struct kfd_process *p, void *data)
1178 {
1179 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1180 	struct kfd_process_device *pdd;
1181 	void *mem;
1182 	int ret;
1183 	uint64_t size = 0;
1184 
1185 	mutex_lock(&p->mutex);
1186 	/*
1187 	 * Safeguard to prevent user space from freeing signal BO.
1188 	 * It will be freed at process termination.
1189 	 */
1190 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1191 		pr_err("Free signal BO is not allowed\n");
1192 		ret = -EPERM;
1193 		goto err_unlock;
1194 	}
1195 
1196 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1197 	if (!pdd) {
1198 		pr_err("Process device data doesn't exist\n");
1199 		ret = -EINVAL;
1200 		goto err_pdd;
1201 	}
1202 
1203 	mem = kfd_process_device_translate_handle(
1204 		pdd, GET_IDR_HANDLE(args->handle));
1205 	if (!mem) {
1206 		ret = -EINVAL;
1207 		goto err_unlock;
1208 	}
1209 
1210 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1211 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1212 
1213 	/* If freeing the buffer failed, leave the handle in place for
1214 	 * clean-up during process tear-down.
1215 	 */
1216 	if (!ret)
1217 		kfd_process_device_remove_obj_handle(
1218 			pdd, GET_IDR_HANDLE(args->handle));
1219 
1220 	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1221 
1222 err_unlock:
1223 err_pdd:
1224 	mutex_unlock(&p->mutex);
1225 	return ret;
1226 }
1227 
1228 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1229 					struct kfd_process *p, void *data)
1230 {
1231 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1232 	struct kfd_process_device *pdd, *peer_pdd;
1233 	void *mem;
1234 	struct kfd_dev *dev;
1235 	long err = 0;
1236 	int i;
1237 	uint32_t *devices_arr = NULL;
1238 
1239 	if (!args->n_devices) {
1240 		pr_debug("Device IDs array empty\n");
1241 		return -EINVAL;
1242 	}
1243 	if (args->n_success > args->n_devices) {
1244 		pr_debug("n_success exceeds n_devices\n");
1245 		return -EINVAL;
1246 	}
1247 
1248 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1249 				    GFP_KERNEL);
1250 	if (!devices_arr)
1251 		return -ENOMEM;
1252 
1253 	err = copy_from_user(devices_arr,
1254 			     (void __user *)args->device_ids_array_ptr,
1255 			     args->n_devices * sizeof(*devices_arr));
1256 	if (err != 0) {
1257 		err = -EFAULT;
1258 		goto copy_from_user_failed;
1259 	}
1260 
1261 	mutex_lock(&p->mutex);
1262 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1263 	if (!pdd) {
1264 		err = -EINVAL;
1265 		goto get_process_device_data_failed;
1266 	}
1267 	dev = pdd->dev;
1268 
1269 	pdd = kfd_bind_process_to_device(dev, p);
1270 	if (IS_ERR(pdd)) {
1271 		err = PTR_ERR(pdd);
1272 		goto bind_process_to_device_failed;
1273 	}
1274 
1275 	mem = kfd_process_device_translate_handle(pdd,
1276 						GET_IDR_HANDLE(args->handle));
1277 	if (!mem) {
1278 		err = -ENOMEM;
1279 		goto get_mem_obj_from_handle_failed;
1280 	}
1281 
1282 	for (i = args->n_success; i < args->n_devices; i++) {
1283 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1284 		if (!peer_pdd) {
1285 			pr_debug("Getting device by id failed for 0x%x\n",
1286 				 devices_arr[i]);
1287 			err = -EINVAL;
1288 			goto get_mem_obj_from_handle_failed;
1289 		}
1290 
1291 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1292 		if (IS_ERR(peer_pdd)) {
1293 			err = PTR_ERR(peer_pdd);
1294 			goto get_mem_obj_from_handle_failed;
1295 		}
1296 
1297 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1298 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1299 			peer_pdd->drm_priv);
1300 		if (err) {
1301 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1302 
1303 			dev_err(dev->adev->dev,
1304 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1305 			       pci_domain_nr(pdev->bus),
1306 			       pdev->bus->number,
1307 			       PCI_SLOT(pdev->devfn),
1308 			       PCI_FUNC(pdev->devfn),
1309 			       ((struct kgd_mem *)mem)->domain);
1310 			goto map_memory_to_gpu_failed;
1311 		}
1312 		args->n_success = i+1;
1313 	}
1314 
1315 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1316 	if (err) {
1317 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1318 		goto sync_memory_failed;
1319 	}
1320 
1321 	mutex_unlock(&p->mutex);
1322 
1323 	/* Flush TLBs after waiting for the page table updates to complete */
1324 	for (i = 0; i < args->n_devices; i++) {
1325 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1326 		if (WARN_ON_ONCE(!peer_pdd))
1327 			continue;
1328 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1329 	}
1330 	kfree(devices_arr);
1331 
1332 	return err;
1333 
1334 get_process_device_data_failed:
1335 bind_process_to_device_failed:
1336 get_mem_obj_from_handle_failed:
1337 map_memory_to_gpu_failed:
1338 sync_memory_failed:
1339 	mutex_unlock(&p->mutex);
1340 copy_from_user_failed:
1341 	kfree(devices_arr);
1342 
1343 	return err;
1344 }
1345 
1346 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1347 					struct kfd_process *p, void *data)
1348 {
1349 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1350 	struct kfd_process_device *pdd, *peer_pdd;
1351 	void *mem;
1352 	long err = 0;
1353 	uint32_t *devices_arr = NULL, i;
1354 	bool flush_tlb;
1355 
1356 	if (!args->n_devices) {
1357 		pr_debug("Device IDs array empty\n");
1358 		return -EINVAL;
1359 	}
1360 	if (args->n_success > args->n_devices) {
1361 		pr_debug("n_success exceeds n_devices\n");
1362 		return -EINVAL;
1363 	}
1364 
1365 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1366 				    GFP_KERNEL);
1367 	if (!devices_arr)
1368 		return -ENOMEM;
1369 
1370 	err = copy_from_user(devices_arr,
1371 			     (void __user *)args->device_ids_array_ptr,
1372 			     args->n_devices * sizeof(*devices_arr));
1373 	if (err != 0) {
1374 		err = -EFAULT;
1375 		goto copy_from_user_failed;
1376 	}
1377 
1378 	mutex_lock(&p->mutex);
1379 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1380 	if (!pdd) {
1381 		err = -EINVAL;
1382 		goto bind_process_to_device_failed;
1383 	}
1384 
1385 	mem = kfd_process_device_translate_handle(pdd,
1386 						GET_IDR_HANDLE(args->handle));
1387 	if (!mem) {
1388 		err = -ENOMEM;
1389 		goto get_mem_obj_from_handle_failed;
1390 	}
1391 
1392 	for (i = args->n_success; i < args->n_devices; i++) {
1393 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1394 		if (!peer_pdd) {
1395 			err = -EINVAL;
1396 			goto get_mem_obj_from_handle_failed;
1397 		}
1398 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1399 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1400 		if (err) {
1401 			pr_err("Failed to unmap from gpu %d/%d\n",
1402 			       i, args->n_devices);
1403 			goto unmap_memory_from_gpu_failed;
1404 		}
1405 		args->n_success = i+1;
1406 	}
1407 
1408 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev);
1409 	if (flush_tlb) {
1410 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1411 				(struct kgd_mem *) mem, true);
1412 		if (err) {
1413 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1414 			goto sync_memory_failed;
1415 		}
1416 	}
1417 	mutex_unlock(&p->mutex);
1418 
1419 	if (flush_tlb) {
1420 		/* Flush TLBs after waiting for the page table updates to complete */
1421 		for (i = 0; i < args->n_devices; i++) {
1422 			peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1423 			if (WARN_ON_ONCE(!peer_pdd))
1424 				continue;
1425 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1426 		}
1427 	}
1428 	kfree(devices_arr);
1429 
1430 	return 0;
1431 
1432 bind_process_to_device_failed:
1433 get_mem_obj_from_handle_failed:
1434 unmap_memory_from_gpu_failed:
1435 sync_memory_failed:
1436 	mutex_unlock(&p->mutex);
1437 copy_from_user_failed:
1438 	kfree(devices_arr);
1439 	return err;
1440 }
1441 
1442 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1443 		struct kfd_process *p, void *data)
1444 {
1445 	int retval;
1446 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1447 	struct queue *q;
1448 	struct kfd_dev *dev;
1449 
1450 	mutex_lock(&p->mutex);
1451 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1452 
1453 	if (q) {
1454 		dev = q->device;
1455 	} else {
1456 		retval = -EINVAL;
1457 		goto out_unlock;
1458 	}
1459 
1460 	if (!dev->gws) {
1461 		retval = -ENODEV;
1462 		goto out_unlock;
1463 	}
1464 
1465 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1466 		retval = -ENODEV;
1467 		goto out_unlock;
1468 	}
1469 
1470 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1471 	mutex_unlock(&p->mutex);
1472 
1473 	args->first_gws = 0;
1474 	return retval;
1475 
1476 out_unlock:
1477 	mutex_unlock(&p->mutex);
1478 	return retval;
1479 }
1480 
1481 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1482 		struct kfd_process *p, void *data)
1483 {
1484 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1485 	struct kfd_dev *dev = NULL;
1486 	struct amdgpu_device *dmabuf_adev;
1487 	void *metadata_buffer = NULL;
1488 	uint32_t flags;
1489 	unsigned int i;
1490 	int r;
1491 
1492 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1493 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1494 		if (dev)
1495 			break;
1496 	if (!dev)
1497 		return -EINVAL;
1498 
1499 	if (args->metadata_ptr) {
1500 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1501 		if (!metadata_buffer)
1502 			return -ENOMEM;
1503 	}
1504 
1505 	/* Get dmabuf info from KGD */
1506 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1507 					  &dmabuf_adev, &args->size,
1508 					  metadata_buffer, args->metadata_size,
1509 					  &args->metadata_size, &flags);
1510 	if (r)
1511 		goto exit;
1512 
1513 	/* Reverse-lookup gpu_id from kgd pointer */
1514 	dev = kfd_device_by_adev(dmabuf_adev);
1515 	if (!dev) {
1516 		r = -EINVAL;
1517 		goto exit;
1518 	}
1519 	args->gpu_id = dev->id;
1520 	args->flags = flags;
1521 
1522 	/* Copy metadata buffer to user mode */
1523 	if (metadata_buffer) {
1524 		r = copy_to_user((void __user *)args->metadata_ptr,
1525 				 metadata_buffer, args->metadata_size);
1526 		if (r != 0)
1527 			r = -EFAULT;
1528 	}
1529 
1530 exit:
1531 	kfree(metadata_buffer);
1532 
1533 	return r;
1534 }
1535 
1536 static int kfd_ioctl_import_dmabuf(struct file *filep,
1537 				   struct kfd_process *p, void *data)
1538 {
1539 	struct kfd_ioctl_import_dmabuf_args *args = data;
1540 	struct kfd_process_device *pdd;
1541 	struct dma_buf *dmabuf;
1542 	int idr_handle;
1543 	uint64_t size;
1544 	void *mem;
1545 	int r;
1546 
1547 	dmabuf = dma_buf_get(args->dmabuf_fd);
1548 	if (IS_ERR(dmabuf))
1549 		return PTR_ERR(dmabuf);
1550 
1551 	mutex_lock(&p->mutex);
1552 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1553 	if (!pdd) {
1554 		r = -EINVAL;
1555 		goto err_unlock;
1556 	}
1557 
1558 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1559 	if (IS_ERR(pdd)) {
1560 		r = PTR_ERR(pdd);
1561 		goto err_unlock;
1562 	}
1563 
1564 	r = amdgpu_amdkfd_gpuvm_import_dmabuf(pdd->dev->adev, dmabuf,
1565 					      args->va_addr, pdd->drm_priv,
1566 					      (struct kgd_mem **)&mem, &size,
1567 					      NULL);
1568 	if (r)
1569 		goto err_unlock;
1570 
1571 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1572 	if (idr_handle < 0) {
1573 		r = -EFAULT;
1574 		goto err_free;
1575 	}
1576 
1577 	mutex_unlock(&p->mutex);
1578 	dma_buf_put(dmabuf);
1579 
1580 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1581 
1582 	return 0;
1583 
1584 err_free:
1585 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1586 					       pdd->drm_priv, NULL);
1587 err_unlock:
1588 	mutex_unlock(&p->mutex);
1589 	dma_buf_put(dmabuf);
1590 	return r;
1591 }
1592 
1593 /* Handle requests for watching SMI events */
1594 static int kfd_ioctl_smi_events(struct file *filep,
1595 				struct kfd_process *p, void *data)
1596 {
1597 	struct kfd_ioctl_smi_events_args *args = data;
1598 	struct kfd_process_device *pdd;
1599 
1600 	mutex_lock(&p->mutex);
1601 
1602 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1603 	mutex_unlock(&p->mutex);
1604 	if (!pdd)
1605 		return -EINVAL;
1606 
1607 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1608 }
1609 
1610 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1611 
1612 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1613 				    struct kfd_process *p, void *data)
1614 {
1615 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1616 	int r = 0;
1617 
1618 	mutex_lock(&p->mutex);
1619 	if (args->xnack_enabled >= 0) {
1620 		if (!list_empty(&p->pqm.queues)) {
1621 			pr_debug("Process has user queues running\n");
1622 			r = -EBUSY;
1623 			goto out_unlock;
1624 		}
1625 
1626 		if (p->xnack_enabled == args->xnack_enabled)
1627 			goto out_unlock;
1628 
1629 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1630 			r = -EPERM;
1631 			goto out_unlock;
1632 		}
1633 
1634 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1635 	} else {
1636 		args->xnack_enabled = p->xnack_enabled;
1637 	}
1638 
1639 out_unlock:
1640 	mutex_unlock(&p->mutex);
1641 
1642 	return r;
1643 }
1644 
1645 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1646 {
1647 	struct kfd_ioctl_svm_args *args = data;
1648 	int r = 0;
1649 
1650 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1651 		 args->start_addr, args->size, args->op, args->nattr);
1652 
1653 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1654 		return -EINVAL;
1655 	if (!args->start_addr || !args->size)
1656 		return -EINVAL;
1657 
1658 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1659 		      args->attrs);
1660 
1661 	return r;
1662 }
1663 #else
1664 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1665 				    struct kfd_process *p, void *data)
1666 {
1667 	return -EPERM;
1668 }
1669 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1670 {
1671 	return -EPERM;
1672 }
1673 #endif
1674 
1675 static int criu_checkpoint_process(struct kfd_process *p,
1676 			     uint8_t __user *user_priv_data,
1677 			     uint64_t *priv_offset)
1678 {
1679 	struct kfd_criu_process_priv_data process_priv;
1680 	int ret;
1681 
1682 	memset(&process_priv, 0, sizeof(process_priv));
1683 
1684 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1685 	/* For CR, we don't consider negative xnack mode which is used for
1686 	 * querying without changing it, here 0 simply means disabled and 1
1687 	 * means enabled so retry for finding a valid PTE.
1688 	 */
1689 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1690 
1691 	ret = copy_to_user(user_priv_data + *priv_offset,
1692 				&process_priv, sizeof(process_priv));
1693 
1694 	if (ret) {
1695 		pr_err("Failed to copy process information to user\n");
1696 		ret = -EFAULT;
1697 	}
1698 
1699 	*priv_offset += sizeof(process_priv);
1700 	return ret;
1701 }
1702 
1703 static int criu_checkpoint_devices(struct kfd_process *p,
1704 			     uint32_t num_devices,
1705 			     uint8_t __user *user_addr,
1706 			     uint8_t __user *user_priv_data,
1707 			     uint64_t *priv_offset)
1708 {
1709 	struct kfd_criu_device_priv_data *device_priv = NULL;
1710 	struct kfd_criu_device_bucket *device_buckets = NULL;
1711 	int ret = 0, i;
1712 
1713 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1714 	if (!device_buckets) {
1715 		ret = -ENOMEM;
1716 		goto exit;
1717 	}
1718 
1719 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1720 	if (!device_priv) {
1721 		ret = -ENOMEM;
1722 		goto exit;
1723 	}
1724 
1725 	for (i = 0; i < num_devices; i++) {
1726 		struct kfd_process_device *pdd = p->pdds[i];
1727 
1728 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1729 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1730 
1731 		/*
1732 		 * priv_data does not contain useful information for now and is reserved for
1733 		 * future use, so we do not set its contents.
1734 		 */
1735 	}
1736 
1737 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1738 	if (ret) {
1739 		pr_err("Failed to copy device information to user\n");
1740 		ret = -EFAULT;
1741 		goto exit;
1742 	}
1743 
1744 	ret = copy_to_user(user_priv_data + *priv_offset,
1745 			   device_priv,
1746 			   num_devices * sizeof(*device_priv));
1747 	if (ret) {
1748 		pr_err("Failed to copy device information to user\n");
1749 		ret = -EFAULT;
1750 	}
1751 	*priv_offset += num_devices * sizeof(*device_priv);
1752 
1753 exit:
1754 	kvfree(device_buckets);
1755 	kvfree(device_priv);
1756 	return ret;
1757 }
1758 
1759 static uint32_t get_process_num_bos(struct kfd_process *p)
1760 {
1761 	uint32_t num_of_bos = 0;
1762 	int i;
1763 
1764 	/* Run over all PDDs of the process */
1765 	for (i = 0; i < p->n_pdds; i++) {
1766 		struct kfd_process_device *pdd = p->pdds[i];
1767 		void *mem;
1768 		int id;
1769 
1770 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1771 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1772 
1773 			if ((uint64_t)kgd_mem->va > pdd->gpuvm_base)
1774 				num_of_bos++;
1775 		}
1776 	}
1777 	return num_of_bos;
1778 }
1779 
1780 static int criu_get_prime_handle(struct drm_gem_object *gobj, int flags,
1781 				      u32 *shared_fd)
1782 {
1783 	struct dma_buf *dmabuf;
1784 	int ret;
1785 
1786 	dmabuf = amdgpu_gem_prime_export(gobj, flags);
1787 	if (IS_ERR(dmabuf)) {
1788 		ret = PTR_ERR(dmabuf);
1789 		pr_err("dmabuf export failed for the BO\n");
1790 		return ret;
1791 	}
1792 
1793 	ret = dma_buf_fd(dmabuf, flags);
1794 	if (ret < 0) {
1795 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1796 		goto out_free_dmabuf;
1797 	}
1798 
1799 	*shared_fd = ret;
1800 	return 0;
1801 
1802 out_free_dmabuf:
1803 	dma_buf_put(dmabuf);
1804 	return ret;
1805 }
1806 
1807 static int criu_checkpoint_bos(struct kfd_process *p,
1808 			       uint32_t num_bos,
1809 			       uint8_t __user *user_bos,
1810 			       uint8_t __user *user_priv_data,
1811 			       uint64_t *priv_offset)
1812 {
1813 	struct kfd_criu_bo_bucket *bo_buckets;
1814 	struct kfd_criu_bo_priv_data *bo_privs;
1815 	int ret = 0, pdd_index, bo_index = 0, id;
1816 	void *mem;
1817 
1818 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1819 	if (!bo_buckets)
1820 		return -ENOMEM;
1821 
1822 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1823 	if (!bo_privs) {
1824 		ret = -ENOMEM;
1825 		goto exit;
1826 	}
1827 
1828 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1829 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1830 		struct amdgpu_bo *dumper_bo;
1831 		struct kgd_mem *kgd_mem;
1832 
1833 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1834 			struct kfd_criu_bo_bucket *bo_bucket;
1835 			struct kfd_criu_bo_priv_data *bo_priv;
1836 			int i, dev_idx = 0;
1837 
1838 			if (!mem) {
1839 				ret = -ENOMEM;
1840 				goto exit;
1841 			}
1842 
1843 			kgd_mem = (struct kgd_mem *)mem;
1844 			dumper_bo = kgd_mem->bo;
1845 
1846 			if ((uint64_t)kgd_mem->va <= pdd->gpuvm_base)
1847 				continue;
1848 
1849 			bo_bucket = &bo_buckets[bo_index];
1850 			bo_priv = &bo_privs[bo_index];
1851 
1852 			bo_bucket->gpu_id = pdd->user_gpu_id;
1853 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1854 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1855 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1856 			bo_priv->idr_handle = id;
1857 
1858 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1859 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1860 								&bo_priv->user_addr);
1861 				if (ret) {
1862 					pr_err("Failed to obtain user address for user-pointer bo\n");
1863 					goto exit;
1864 				}
1865 			}
1866 			if (bo_bucket->alloc_flags
1867 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1868 				ret = criu_get_prime_handle(&dumper_bo->tbo.base,
1869 						bo_bucket->alloc_flags &
1870 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1871 						&bo_bucket->dmabuf_fd);
1872 				if (ret)
1873 					goto exit;
1874 			} else {
1875 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1876 			}
1877 
1878 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1879 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1880 					KFD_MMAP_GPU_ID(pdd->dev->id);
1881 			else if (bo_bucket->alloc_flags &
1882 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1883 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1884 					KFD_MMAP_GPU_ID(pdd->dev->id);
1885 			else
1886 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1887 
1888 			for (i = 0; i < p->n_pdds; i++) {
1889 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->dev->adev, kgd_mem))
1890 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1891 			}
1892 
1893 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1894 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1895 					bo_bucket->size,
1896 					bo_bucket->addr,
1897 					bo_bucket->offset,
1898 					bo_bucket->gpu_id,
1899 					bo_bucket->alloc_flags,
1900 					bo_priv->idr_handle);
1901 			bo_index++;
1902 		}
1903 	}
1904 
1905 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1906 	if (ret) {
1907 		pr_err("Failed to copy BO information to user\n");
1908 		ret = -EFAULT;
1909 		goto exit;
1910 	}
1911 
1912 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1913 	if (ret) {
1914 		pr_err("Failed to copy BO priv information to user\n");
1915 		ret = -EFAULT;
1916 		goto exit;
1917 	}
1918 
1919 	*priv_offset += num_bos * sizeof(*bo_privs);
1920 
1921 exit:
1922 	while (ret && bo_index--) {
1923 		if (bo_buckets[bo_index].alloc_flags
1924 		    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
1925 			close_fd(bo_buckets[bo_index].dmabuf_fd);
1926 	}
1927 
1928 	kvfree(bo_buckets);
1929 	kvfree(bo_privs);
1930 	return ret;
1931 }
1932 
1933 static int criu_get_process_object_info(struct kfd_process *p,
1934 					uint32_t *num_devices,
1935 					uint32_t *num_bos,
1936 					uint32_t *num_objects,
1937 					uint64_t *objs_priv_size)
1938 {
1939 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
1940 	uint32_t num_queues, num_events, num_svm_ranges;
1941 	int ret;
1942 
1943 	*num_devices = p->n_pdds;
1944 	*num_bos = get_process_num_bos(p);
1945 
1946 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
1947 	if (ret)
1948 		return ret;
1949 
1950 	num_events = kfd_get_num_events(p);
1951 
1952 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
1953 	if (ret)
1954 		return ret;
1955 
1956 	*num_objects = num_queues + num_events + num_svm_ranges;
1957 
1958 	if (objs_priv_size) {
1959 		priv_size = sizeof(struct kfd_criu_process_priv_data);
1960 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
1961 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
1962 		priv_size += queues_priv_data_size;
1963 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
1964 		priv_size += svm_priv_data_size;
1965 		*objs_priv_size = priv_size;
1966 	}
1967 	return 0;
1968 }
1969 
1970 static int criu_checkpoint(struct file *filep,
1971 			   struct kfd_process *p,
1972 			   struct kfd_ioctl_criu_args *args)
1973 {
1974 	int ret;
1975 	uint32_t num_devices, num_bos, num_objects;
1976 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
1977 
1978 	if (!args->devices || !args->bos || !args->priv_data)
1979 		return -EINVAL;
1980 
1981 	mutex_lock(&p->mutex);
1982 
1983 	if (!p->n_pdds) {
1984 		pr_err("No pdd for given process\n");
1985 		ret = -ENODEV;
1986 		goto exit_unlock;
1987 	}
1988 
1989 	/* Confirm all process queues are evicted */
1990 	if (!p->queues_paused) {
1991 		pr_err("Cannot dump process when queues are not in evicted state\n");
1992 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
1993 		ret = -EINVAL;
1994 		goto exit_unlock;
1995 	}
1996 
1997 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
1998 	if (ret)
1999 		goto exit_unlock;
2000 
2001 	if (num_devices != args->num_devices ||
2002 	    num_bos != args->num_bos ||
2003 	    num_objects != args->num_objects ||
2004 	    priv_size != args->priv_data_size) {
2005 
2006 		ret = -EINVAL;
2007 		goto exit_unlock;
2008 	}
2009 
2010 	/* each function will store private data inside priv_data and adjust priv_offset */
2011 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2012 	if (ret)
2013 		goto exit_unlock;
2014 
2015 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2016 				(uint8_t __user *)args->priv_data, &priv_offset);
2017 	if (ret)
2018 		goto exit_unlock;
2019 
2020 	/* Leave room for BOs in the private data. They need to be restored
2021 	 * before events, but we checkpoint them last to simplify the error
2022 	 * handling.
2023 	 */
2024 	bo_priv_offset = priv_offset;
2025 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2026 
2027 	if (num_objects) {
2028 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2029 						 &priv_offset);
2030 		if (ret)
2031 			goto exit_unlock;
2032 
2033 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2034 						 &priv_offset);
2035 		if (ret)
2036 			goto exit_unlock;
2037 
2038 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2039 		if (ret)
2040 			goto exit_unlock;
2041 	}
2042 
2043 	/* This must be the last thing in this function that can fail.
2044 	 * Otherwise we leak dmabuf file descriptors.
2045 	 */
2046 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2047 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2048 
2049 exit_unlock:
2050 	mutex_unlock(&p->mutex);
2051 	if (ret)
2052 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2053 	else
2054 		pr_debug("CRIU dump ret:%d\n", ret);
2055 
2056 	return ret;
2057 }
2058 
2059 static int criu_restore_process(struct kfd_process *p,
2060 				struct kfd_ioctl_criu_args *args,
2061 				uint64_t *priv_offset,
2062 				uint64_t max_priv_data_size)
2063 {
2064 	int ret = 0;
2065 	struct kfd_criu_process_priv_data process_priv;
2066 
2067 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2068 		return -EINVAL;
2069 
2070 	ret = copy_from_user(&process_priv,
2071 				(void __user *)(args->priv_data + *priv_offset),
2072 				sizeof(process_priv));
2073 	if (ret) {
2074 		pr_err("Failed to copy process private information from user\n");
2075 		ret = -EFAULT;
2076 		goto exit;
2077 	}
2078 	*priv_offset += sizeof(process_priv);
2079 
2080 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2081 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2082 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2083 		return -EINVAL;
2084 	}
2085 
2086 	pr_debug("Setting XNACK mode\n");
2087 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2088 		pr_err("xnack mode cannot be set\n");
2089 		ret = -EPERM;
2090 		goto exit;
2091 	} else {
2092 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2093 		p->xnack_enabled = process_priv.xnack_mode;
2094 	}
2095 
2096 exit:
2097 	return ret;
2098 }
2099 
2100 static int criu_restore_devices(struct kfd_process *p,
2101 				struct kfd_ioctl_criu_args *args,
2102 				uint64_t *priv_offset,
2103 				uint64_t max_priv_data_size)
2104 {
2105 	struct kfd_criu_device_bucket *device_buckets;
2106 	struct kfd_criu_device_priv_data *device_privs;
2107 	int ret = 0;
2108 	uint32_t i;
2109 
2110 	if (args->num_devices != p->n_pdds)
2111 		return -EINVAL;
2112 
2113 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2114 		return -EINVAL;
2115 
2116 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2117 	if (!device_buckets)
2118 		return -ENOMEM;
2119 
2120 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2121 				args->num_devices * sizeof(*device_buckets));
2122 	if (ret) {
2123 		pr_err("Failed to copy devices buckets from user\n");
2124 		ret = -EFAULT;
2125 		goto exit;
2126 	}
2127 
2128 	for (i = 0; i < args->num_devices; i++) {
2129 		struct kfd_dev *dev;
2130 		struct kfd_process_device *pdd;
2131 		struct file *drm_file;
2132 
2133 		/* device private data is not currently used */
2134 
2135 		if (!device_buckets[i].user_gpu_id) {
2136 			pr_err("Invalid user gpu_id\n");
2137 			ret = -EINVAL;
2138 			goto exit;
2139 		}
2140 
2141 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2142 		if (!dev) {
2143 			pr_err("Failed to find device with gpu_id = %x\n",
2144 				device_buckets[i].actual_gpu_id);
2145 			ret = -EINVAL;
2146 			goto exit;
2147 		}
2148 
2149 		pdd = kfd_get_process_device_data(dev, p);
2150 		if (!pdd) {
2151 			pr_err("Failed to get pdd for gpu_id = %x\n",
2152 					device_buckets[i].actual_gpu_id);
2153 			ret = -EINVAL;
2154 			goto exit;
2155 		}
2156 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2157 
2158 		drm_file = fget(device_buckets[i].drm_fd);
2159 		if (!drm_file) {
2160 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2161 				device_buckets[i].drm_fd);
2162 			ret = -EINVAL;
2163 			goto exit;
2164 		}
2165 
2166 		if (pdd->drm_file) {
2167 			ret = -EINVAL;
2168 			goto exit;
2169 		}
2170 
2171 		/* create the vm using render nodes for kfd pdd */
2172 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2173 			pr_err("could not init vm for given pdd\n");
2174 			/* On success, the PDD keeps the drm_file reference */
2175 			fput(drm_file);
2176 			ret = -EINVAL;
2177 			goto exit;
2178 		}
2179 		/*
2180 		 * pdd now already has the vm bound to render node so below api won't create a new
2181 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2182 		 * for iommu v2 binding  and runtime pm.
2183 		 */
2184 		pdd = kfd_bind_process_to_device(dev, p);
2185 		if (IS_ERR(pdd)) {
2186 			ret = PTR_ERR(pdd);
2187 			goto exit;
2188 		}
2189 
2190 		if (!pdd->doorbell_index &&
2191 		    kfd_alloc_process_doorbells(pdd->dev, &pdd->doorbell_index) < 0) {
2192 			ret = -ENOMEM;
2193 			goto exit;
2194 		}
2195 	}
2196 
2197 	/*
2198 	 * We are not copying device private data from user as we are not using the data for now,
2199 	 * but we still adjust for its private data.
2200 	 */
2201 	*priv_offset += args->num_devices * sizeof(*device_privs);
2202 
2203 exit:
2204 	kfree(device_buckets);
2205 	return ret;
2206 }
2207 
2208 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2209 				      struct kfd_criu_bo_bucket *bo_bucket,
2210 				      struct kfd_criu_bo_priv_data *bo_priv,
2211 				      struct kgd_mem **kgd_mem)
2212 {
2213 	int idr_handle;
2214 	int ret;
2215 	const bool criu_resume = true;
2216 	u64 offset;
2217 
2218 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2219 		if (bo_bucket->size != kfd_doorbell_process_slice(pdd->dev))
2220 			return -EINVAL;
2221 
2222 		offset = kfd_get_process_doorbells(pdd);
2223 		if (!offset)
2224 			return -ENOMEM;
2225 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2226 		/* MMIO BOs need remapped bus address */
2227 		if (bo_bucket->size != PAGE_SIZE) {
2228 			pr_err("Invalid page size\n");
2229 			return -EINVAL;
2230 		}
2231 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2232 		if (!offset) {
2233 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2234 			return -ENOMEM;
2235 		}
2236 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2237 		offset = bo_priv->user_addr;
2238 	}
2239 	/* Create the BO */
2240 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2241 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2242 						      &offset, bo_bucket->alloc_flags, criu_resume);
2243 	if (ret) {
2244 		pr_err("Could not create the BO\n");
2245 		return ret;
2246 	}
2247 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2248 		 bo_bucket->size, bo_bucket->addr, offset);
2249 
2250 	/* Restore previous IDR handle */
2251 	pr_debug("Restoring old IDR handle for the BO");
2252 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2253 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2254 
2255 	if (idr_handle < 0) {
2256 		pr_err("Could not allocate idr\n");
2257 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2258 						       NULL);
2259 		return -ENOMEM;
2260 	}
2261 
2262 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2263 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2264 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2265 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2266 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2267 		bo_bucket->restored_offset = offset;
2268 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2269 		bo_bucket->restored_offset = offset;
2270 		/* Update the VRAM usage count */
2271 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2272 	}
2273 	return 0;
2274 }
2275 
2276 static int criu_restore_bo(struct kfd_process *p,
2277 			   struct kfd_criu_bo_bucket *bo_bucket,
2278 			   struct kfd_criu_bo_priv_data *bo_priv)
2279 {
2280 	struct kfd_process_device *pdd;
2281 	struct kgd_mem *kgd_mem;
2282 	int ret;
2283 	int j;
2284 
2285 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2286 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2287 		 bo_priv->idr_handle);
2288 
2289 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2290 	if (!pdd) {
2291 		pr_err("Failed to get pdd\n");
2292 		return -ENODEV;
2293 	}
2294 
2295 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2296 	if (ret)
2297 		return ret;
2298 
2299 	/* now map these BOs to GPU/s */
2300 	for (j = 0; j < p->n_pdds; j++) {
2301 		struct kfd_dev *peer;
2302 		struct kfd_process_device *peer_pdd;
2303 
2304 		if (!bo_priv->mapped_gpuids[j])
2305 			break;
2306 
2307 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2308 		if (!peer_pdd)
2309 			return -EINVAL;
2310 
2311 		peer = peer_pdd->dev;
2312 
2313 		peer_pdd = kfd_bind_process_to_device(peer, p);
2314 		if (IS_ERR(peer_pdd))
2315 			return PTR_ERR(peer_pdd);
2316 
2317 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2318 							    peer_pdd->drm_priv);
2319 		if (ret) {
2320 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2321 			return ret;
2322 		}
2323 	}
2324 
2325 	pr_debug("map memory was successful for the BO\n");
2326 	/* create the dmabuf object and export the bo */
2327 	if (bo_bucket->alloc_flags
2328 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2329 		ret = criu_get_prime_handle(&kgd_mem->bo->tbo.base, DRM_RDWR,
2330 					    &bo_bucket->dmabuf_fd);
2331 		if (ret)
2332 			return ret;
2333 	} else {
2334 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2335 	}
2336 
2337 	return 0;
2338 }
2339 
2340 static int criu_restore_bos(struct kfd_process *p,
2341 			    struct kfd_ioctl_criu_args *args,
2342 			    uint64_t *priv_offset,
2343 			    uint64_t max_priv_data_size)
2344 {
2345 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2346 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2347 	int ret = 0;
2348 	uint32_t i = 0;
2349 
2350 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2351 		return -EINVAL;
2352 
2353 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2354 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2355 
2356 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2357 	if (!bo_buckets)
2358 		return -ENOMEM;
2359 
2360 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2361 			     args->num_bos * sizeof(*bo_buckets));
2362 	if (ret) {
2363 		pr_err("Failed to copy BOs information from user\n");
2364 		ret = -EFAULT;
2365 		goto exit;
2366 	}
2367 
2368 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2369 	if (!bo_privs) {
2370 		ret = -ENOMEM;
2371 		goto exit;
2372 	}
2373 
2374 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2375 			     args->num_bos * sizeof(*bo_privs));
2376 	if (ret) {
2377 		pr_err("Failed to copy BOs information from user\n");
2378 		ret = -EFAULT;
2379 		goto exit;
2380 	}
2381 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2382 
2383 	/* Create and map new BOs */
2384 	for (; i < args->num_bos; i++) {
2385 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2386 		if (ret) {
2387 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2388 			goto exit;
2389 		}
2390 	} /* done */
2391 
2392 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2393 	ret = copy_to_user((void __user *)args->bos,
2394 				bo_buckets,
2395 				(args->num_bos * sizeof(*bo_buckets)));
2396 	if (ret)
2397 		ret = -EFAULT;
2398 
2399 exit:
2400 	while (ret && i--) {
2401 		if (bo_buckets[i].alloc_flags
2402 		   & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2403 			close_fd(bo_buckets[i].dmabuf_fd);
2404 	}
2405 	kvfree(bo_buckets);
2406 	kvfree(bo_privs);
2407 	return ret;
2408 }
2409 
2410 static int criu_restore_objects(struct file *filep,
2411 				struct kfd_process *p,
2412 				struct kfd_ioctl_criu_args *args,
2413 				uint64_t *priv_offset,
2414 				uint64_t max_priv_data_size)
2415 {
2416 	int ret = 0;
2417 	uint32_t i;
2418 
2419 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2420 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2421 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2422 
2423 	for (i = 0; i < args->num_objects; i++) {
2424 		uint32_t object_type;
2425 
2426 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2427 			pr_err("Invalid private data size\n");
2428 			return -EINVAL;
2429 		}
2430 
2431 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2432 		if (ret) {
2433 			pr_err("Failed to copy private information from user\n");
2434 			goto exit;
2435 		}
2436 
2437 		switch (object_type) {
2438 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2439 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2440 						     priv_offset, max_priv_data_size);
2441 			if (ret)
2442 				goto exit;
2443 			break;
2444 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2445 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2446 						     priv_offset, max_priv_data_size);
2447 			if (ret)
2448 				goto exit;
2449 			break;
2450 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2451 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2452 						     priv_offset, max_priv_data_size);
2453 			if (ret)
2454 				goto exit;
2455 			break;
2456 		default:
2457 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2458 			ret = -EINVAL;
2459 			goto exit;
2460 		}
2461 	}
2462 exit:
2463 	return ret;
2464 }
2465 
2466 static int criu_restore(struct file *filep,
2467 			struct kfd_process *p,
2468 			struct kfd_ioctl_criu_args *args)
2469 {
2470 	uint64_t priv_offset = 0;
2471 	int ret = 0;
2472 
2473 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2474 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2475 
2476 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2477 	    !args->num_devices || !args->num_bos)
2478 		return -EINVAL;
2479 
2480 	mutex_lock(&p->mutex);
2481 
2482 	/*
2483 	 * Set the process to evicted state to avoid running any new queues before all the memory
2484 	 * mappings are ready.
2485 	 */
2486 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2487 	if (ret)
2488 		goto exit_unlock;
2489 
2490 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2491 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2492 	if (ret)
2493 		goto exit_unlock;
2494 
2495 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2496 	if (ret)
2497 		goto exit_unlock;
2498 
2499 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2500 	if (ret)
2501 		goto exit_unlock;
2502 
2503 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2504 	if (ret)
2505 		goto exit_unlock;
2506 
2507 	if (priv_offset != args->priv_data_size) {
2508 		pr_err("Invalid private data size\n");
2509 		ret = -EINVAL;
2510 	}
2511 
2512 exit_unlock:
2513 	mutex_unlock(&p->mutex);
2514 	if (ret)
2515 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2516 	else
2517 		pr_debug("CRIU restore successful\n");
2518 
2519 	return ret;
2520 }
2521 
2522 static int criu_unpause(struct file *filep,
2523 			struct kfd_process *p,
2524 			struct kfd_ioctl_criu_args *args)
2525 {
2526 	int ret;
2527 
2528 	mutex_lock(&p->mutex);
2529 
2530 	if (!p->queues_paused) {
2531 		mutex_unlock(&p->mutex);
2532 		return -EINVAL;
2533 	}
2534 
2535 	ret = kfd_process_restore_queues(p);
2536 	if (ret)
2537 		pr_err("Failed to unpause queues ret:%d\n", ret);
2538 	else
2539 		p->queues_paused = false;
2540 
2541 	mutex_unlock(&p->mutex);
2542 
2543 	return ret;
2544 }
2545 
2546 static int criu_resume(struct file *filep,
2547 			struct kfd_process *p,
2548 			struct kfd_ioctl_criu_args *args)
2549 {
2550 	struct kfd_process *target = NULL;
2551 	struct pid *pid = NULL;
2552 	int ret = 0;
2553 
2554 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2555 		 args->pid);
2556 
2557 	pid = find_get_pid(args->pid);
2558 	if (!pid) {
2559 		pr_err("Cannot find pid info for %i\n", args->pid);
2560 		return -ESRCH;
2561 	}
2562 
2563 	pr_debug("calling kfd_lookup_process_by_pid\n");
2564 	target = kfd_lookup_process_by_pid(pid);
2565 
2566 	put_pid(pid);
2567 
2568 	if (!target) {
2569 		pr_debug("Cannot find process info for %i\n", args->pid);
2570 		return -ESRCH;
2571 	}
2572 
2573 	mutex_lock(&target->mutex);
2574 	ret = kfd_criu_resume_svm(target);
2575 	if (ret) {
2576 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2577 		goto exit;
2578 	}
2579 
2580 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2581 	if (ret)
2582 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2583 
2584 exit:
2585 	mutex_unlock(&target->mutex);
2586 
2587 	kfd_unref_process(target);
2588 	return ret;
2589 }
2590 
2591 static int criu_process_info(struct file *filep,
2592 				struct kfd_process *p,
2593 				struct kfd_ioctl_criu_args *args)
2594 {
2595 	int ret = 0;
2596 
2597 	mutex_lock(&p->mutex);
2598 
2599 	if (!p->n_pdds) {
2600 		pr_err("No pdd for given process\n");
2601 		ret = -ENODEV;
2602 		goto err_unlock;
2603 	}
2604 
2605 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2606 	if (ret)
2607 		goto err_unlock;
2608 
2609 	p->queues_paused = true;
2610 
2611 	args->pid = task_pid_nr_ns(p->lead_thread,
2612 					task_active_pid_ns(p->lead_thread));
2613 
2614 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2615 					   &args->num_objects, &args->priv_data_size);
2616 	if (ret)
2617 		goto err_unlock;
2618 
2619 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2620 				args->num_devices, args->num_bos, args->num_objects,
2621 				args->priv_data_size);
2622 
2623 err_unlock:
2624 	if (ret) {
2625 		kfd_process_restore_queues(p);
2626 		p->queues_paused = false;
2627 	}
2628 	mutex_unlock(&p->mutex);
2629 	return ret;
2630 }
2631 
2632 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2633 {
2634 	struct kfd_ioctl_criu_args *args = data;
2635 	int ret;
2636 
2637 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2638 	switch (args->op) {
2639 	case KFD_CRIU_OP_PROCESS_INFO:
2640 		ret = criu_process_info(filep, p, args);
2641 		break;
2642 	case KFD_CRIU_OP_CHECKPOINT:
2643 		ret = criu_checkpoint(filep, p, args);
2644 		break;
2645 	case KFD_CRIU_OP_UNPAUSE:
2646 		ret = criu_unpause(filep, p, args);
2647 		break;
2648 	case KFD_CRIU_OP_RESTORE:
2649 		ret = criu_restore(filep, p, args);
2650 		break;
2651 	case KFD_CRIU_OP_RESUME:
2652 		ret = criu_resume(filep, p, args);
2653 		break;
2654 	default:
2655 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2656 		ret = -EINVAL;
2657 		break;
2658 	}
2659 
2660 	if (ret)
2661 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2662 
2663 	return ret;
2664 }
2665 
2666 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
2667 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
2668 			    .cmd_drv = 0, .name = #ioctl}
2669 
2670 /** Ioctl table */
2671 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
2672 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
2673 			kfd_ioctl_get_version, 0),
2674 
2675 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
2676 			kfd_ioctl_create_queue, 0),
2677 
2678 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
2679 			kfd_ioctl_destroy_queue, 0),
2680 
2681 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
2682 			kfd_ioctl_set_memory_policy, 0),
2683 
2684 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
2685 			kfd_ioctl_get_clock_counters, 0),
2686 
2687 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
2688 			kfd_ioctl_get_process_apertures, 0),
2689 
2690 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
2691 			kfd_ioctl_update_queue, 0),
2692 
2693 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
2694 			kfd_ioctl_create_event, 0),
2695 
2696 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
2697 			kfd_ioctl_destroy_event, 0),
2698 
2699 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
2700 			kfd_ioctl_set_event, 0),
2701 
2702 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
2703 			kfd_ioctl_reset_event, 0),
2704 
2705 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
2706 			kfd_ioctl_wait_events, 0),
2707 
2708 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
2709 			kfd_ioctl_dbg_register, 0),
2710 
2711 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
2712 			kfd_ioctl_dbg_unregister, 0),
2713 
2714 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
2715 			kfd_ioctl_dbg_address_watch, 0),
2716 
2717 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
2718 			kfd_ioctl_dbg_wave_control, 0),
2719 
2720 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
2721 			kfd_ioctl_set_scratch_backing_va, 0),
2722 
2723 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
2724 			kfd_ioctl_get_tile_config, 0),
2725 
2726 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
2727 			kfd_ioctl_set_trap_handler, 0),
2728 
2729 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
2730 			kfd_ioctl_get_process_apertures_new, 0),
2731 
2732 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
2733 			kfd_ioctl_acquire_vm, 0),
2734 
2735 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
2736 			kfd_ioctl_alloc_memory_of_gpu, 0),
2737 
2738 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
2739 			kfd_ioctl_free_memory_of_gpu, 0),
2740 
2741 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
2742 			kfd_ioctl_map_memory_to_gpu, 0),
2743 
2744 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
2745 			kfd_ioctl_unmap_memory_from_gpu, 0),
2746 
2747 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
2748 			kfd_ioctl_set_cu_mask, 0),
2749 
2750 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
2751 			kfd_ioctl_get_queue_wave_state, 0),
2752 
2753 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
2754 				kfd_ioctl_get_dmabuf_info, 0),
2755 
2756 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
2757 				kfd_ioctl_import_dmabuf, 0),
2758 
2759 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
2760 			kfd_ioctl_alloc_queue_gws, 0),
2761 
2762 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
2763 			kfd_ioctl_smi_events, 0),
2764 
2765 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
2766 
2767 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
2768 			kfd_ioctl_set_xnack_mode, 0),
2769 
2770 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
2771 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
2772 
2773 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
2774 			kfd_ioctl_get_available_memory, 0),
2775 };
2776 
2777 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
2778 
2779 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
2780 {
2781 	struct kfd_process *process;
2782 	amdkfd_ioctl_t *func;
2783 	const struct amdkfd_ioctl_desc *ioctl = NULL;
2784 	unsigned int nr = _IOC_NR(cmd);
2785 	char stack_kdata[128];
2786 	char *kdata = NULL;
2787 	unsigned int usize, asize;
2788 	int retcode = -EINVAL;
2789 	bool ptrace_attached = false;
2790 
2791 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
2792 		goto err_i1;
2793 
2794 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
2795 		u32 amdkfd_size;
2796 
2797 		ioctl = &amdkfd_ioctls[nr];
2798 
2799 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
2800 		usize = asize = _IOC_SIZE(cmd);
2801 		if (amdkfd_size > asize)
2802 			asize = amdkfd_size;
2803 
2804 		cmd = ioctl->cmd;
2805 	} else
2806 		goto err_i1;
2807 
2808 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
2809 
2810 	/* Get the process struct from the filep. Only the process
2811 	 * that opened /dev/kfd can use the file descriptor. Child
2812 	 * processes need to create their own KFD device context.
2813 	 */
2814 	process = filep->private_data;
2815 
2816 	rcu_read_lock();
2817 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
2818 	    ptrace_parent(process->lead_thread) == current)
2819 		ptrace_attached = true;
2820 	rcu_read_unlock();
2821 
2822 	if (process->lead_thread != current->group_leader
2823 	    && !ptrace_attached) {
2824 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
2825 		retcode = -EBADF;
2826 		goto err_i1;
2827 	}
2828 
2829 	/* Do not trust userspace, use our own definition */
2830 	func = ioctl->func;
2831 
2832 	if (unlikely(!func)) {
2833 		dev_dbg(kfd_device, "no function\n");
2834 		retcode = -EINVAL;
2835 		goto err_i1;
2836 	}
2837 
2838 	/*
2839 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
2840 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
2841 	 * more priviledged access.
2842 	 */
2843 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
2844 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
2845 						!capable(CAP_SYS_ADMIN)) {
2846 			retcode = -EACCES;
2847 			goto err_i1;
2848 		}
2849 	}
2850 
2851 	if (cmd & (IOC_IN | IOC_OUT)) {
2852 		if (asize <= sizeof(stack_kdata)) {
2853 			kdata = stack_kdata;
2854 		} else {
2855 			kdata = kmalloc(asize, GFP_KERNEL);
2856 			if (!kdata) {
2857 				retcode = -ENOMEM;
2858 				goto err_i1;
2859 			}
2860 		}
2861 		if (asize > usize)
2862 			memset(kdata + usize, 0, asize - usize);
2863 	}
2864 
2865 	if (cmd & IOC_IN) {
2866 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
2867 			retcode = -EFAULT;
2868 			goto err_i1;
2869 		}
2870 	} else if (cmd & IOC_OUT) {
2871 		memset(kdata, 0, usize);
2872 	}
2873 
2874 	retcode = func(filep, process, kdata);
2875 
2876 	if (cmd & IOC_OUT)
2877 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
2878 			retcode = -EFAULT;
2879 
2880 err_i1:
2881 	if (!ioctl)
2882 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
2883 			  task_pid_nr(current), cmd, nr);
2884 
2885 	if (kdata != stack_kdata)
2886 		kfree(kdata);
2887 
2888 	if (retcode)
2889 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
2890 				nr, arg, retcode);
2891 
2892 	return retcode;
2893 }
2894 
2895 static int kfd_mmio_mmap(struct kfd_dev *dev, struct kfd_process *process,
2896 		      struct vm_area_struct *vma)
2897 {
2898 	phys_addr_t address;
2899 
2900 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2901 		return -EINVAL;
2902 
2903 	address = dev->adev->rmmio_remap.bus_addr;
2904 
2905 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
2906 				VM_DONTDUMP | VM_PFNMAP);
2907 
2908 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
2909 
2910 	pr_debug("pasid 0x%x mapping mmio page\n"
2911 		 "     target user address == 0x%08llX\n"
2912 		 "     physical address    == 0x%08llX\n"
2913 		 "     vm_flags            == 0x%04lX\n"
2914 		 "     size                == 0x%04lX\n",
2915 		 process->pasid, (unsigned long long) vma->vm_start,
2916 		 address, vma->vm_flags, PAGE_SIZE);
2917 
2918 	return io_remap_pfn_range(vma,
2919 				vma->vm_start,
2920 				address >> PAGE_SHIFT,
2921 				PAGE_SIZE,
2922 				vma->vm_page_prot);
2923 }
2924 
2925 
2926 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
2927 {
2928 	struct kfd_process *process;
2929 	struct kfd_dev *dev = NULL;
2930 	unsigned long mmap_offset;
2931 	unsigned int gpu_id;
2932 
2933 	process = kfd_get_process(current);
2934 	if (IS_ERR(process))
2935 		return PTR_ERR(process);
2936 
2937 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
2938 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
2939 	if (gpu_id)
2940 		dev = kfd_device_by_id(gpu_id);
2941 
2942 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
2943 	case KFD_MMAP_TYPE_DOORBELL:
2944 		if (!dev)
2945 			return -ENODEV;
2946 		return kfd_doorbell_mmap(dev, process, vma);
2947 
2948 	case KFD_MMAP_TYPE_EVENTS:
2949 		return kfd_event_mmap(process, vma);
2950 
2951 	case KFD_MMAP_TYPE_RESERVED_MEM:
2952 		if (!dev)
2953 			return -ENODEV;
2954 		return kfd_reserved_mem_mmap(dev, process, vma);
2955 	case KFD_MMAP_TYPE_MMIO:
2956 		if (!dev)
2957 			return -ENODEV;
2958 		return kfd_mmio_mmap(dev, process, vma);
2959 	}
2960 
2961 	return -EFAULT;
2962 }
2963