xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c (revision e7d759f31ca295d589f7420719c311870bb3166f)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include <linux/device.h>
25 #include <linux/export.h>
26 #include <linux/err.h>
27 #include <linux/fs.h>
28 #include <linux/file.h>
29 #include <linux/sched.h>
30 #include <linux/slab.h>
31 #include <linux/uaccess.h>
32 #include <linux/compat.h>
33 #include <uapi/linux/kfd_ioctl.h>
34 #include <linux/time.h>
35 #include <linux/mm.h>
36 #include <linux/mman.h>
37 #include <linux/ptrace.h>
38 #include <linux/dma-buf.h>
39 #include <linux/fdtable.h>
40 #include <linux/processor.h>
41 #include "kfd_priv.h"
42 #include "kfd_device_queue_manager.h"
43 #include "kfd_svm.h"
44 #include "amdgpu_amdkfd.h"
45 #include "kfd_smi_events.h"
46 #include "amdgpu_dma_buf.h"
47 #include "kfd_debug.h"
48 
49 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
50 static int kfd_open(struct inode *, struct file *);
51 static int kfd_release(struct inode *, struct file *);
52 static int kfd_mmap(struct file *, struct vm_area_struct *);
53 
54 static const char kfd_dev_name[] = "kfd";
55 
56 static const struct file_operations kfd_fops = {
57 	.owner = THIS_MODULE,
58 	.unlocked_ioctl = kfd_ioctl,
59 	.compat_ioctl = compat_ptr_ioctl,
60 	.open = kfd_open,
61 	.release = kfd_release,
62 	.mmap = kfd_mmap,
63 };
64 
65 static int kfd_char_dev_major = -1;
66 static struct class *kfd_class;
67 struct device *kfd_device;
68 
69 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
70 {
71 	struct kfd_process_device *pdd;
72 
73 	mutex_lock(&p->mutex);
74 	pdd = kfd_process_device_data_by_id(p, gpu_id);
75 
76 	if (pdd)
77 		return pdd;
78 
79 	mutex_unlock(&p->mutex);
80 	return NULL;
81 }
82 
83 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
84 {
85 	mutex_unlock(&pdd->process->mutex);
86 }
87 
88 int kfd_chardev_init(void)
89 {
90 	int err = 0;
91 
92 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
93 	err = kfd_char_dev_major;
94 	if (err < 0)
95 		goto err_register_chrdev;
96 
97 	kfd_class = class_create(kfd_dev_name);
98 	err = PTR_ERR(kfd_class);
99 	if (IS_ERR(kfd_class))
100 		goto err_class_create;
101 
102 	kfd_device = device_create(kfd_class, NULL,
103 					MKDEV(kfd_char_dev_major, 0),
104 					NULL, kfd_dev_name);
105 	err = PTR_ERR(kfd_device);
106 	if (IS_ERR(kfd_device))
107 		goto err_device_create;
108 
109 	return 0;
110 
111 err_device_create:
112 	class_destroy(kfd_class);
113 err_class_create:
114 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
115 err_register_chrdev:
116 	return err;
117 }
118 
119 void kfd_chardev_exit(void)
120 {
121 	device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
122 	class_destroy(kfd_class);
123 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
124 	kfd_device = NULL;
125 }
126 
127 
128 static int kfd_open(struct inode *inode, struct file *filep)
129 {
130 	struct kfd_process *process;
131 	bool is_32bit_user_mode;
132 
133 	if (iminor(inode) != 0)
134 		return -ENODEV;
135 
136 	is_32bit_user_mode = in_compat_syscall();
137 
138 	if (is_32bit_user_mode) {
139 		dev_warn(kfd_device,
140 			"Process %d (32-bit) failed to open /dev/kfd\n"
141 			"32-bit processes are not supported by amdkfd\n",
142 			current->pid);
143 		return -EPERM;
144 	}
145 
146 	process = kfd_create_process(current);
147 	if (IS_ERR(process))
148 		return PTR_ERR(process);
149 
150 	if (kfd_process_init_cwsr_apu(process, filep)) {
151 		kfd_unref_process(process);
152 		return -EFAULT;
153 	}
154 
155 	/* filep now owns the reference returned by kfd_create_process */
156 	filep->private_data = process;
157 
158 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
159 		process->pasid, process->is_32bit_user_mode);
160 
161 	return 0;
162 }
163 
164 static int kfd_release(struct inode *inode, struct file *filep)
165 {
166 	struct kfd_process *process = filep->private_data;
167 
168 	if (process)
169 		kfd_unref_process(process);
170 
171 	return 0;
172 }
173 
174 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
175 					void *data)
176 {
177 	struct kfd_ioctl_get_version_args *args = data;
178 
179 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
180 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
181 
182 	return 0;
183 }
184 
185 static int set_queue_properties_from_user(struct queue_properties *q_properties,
186 				struct kfd_ioctl_create_queue_args *args)
187 {
188 	/*
189 	 * Repurpose queue percentage to accommodate new features:
190 	 * bit 0-7: queue percentage
191 	 * bit 8-15: pm4_target_xcc
192 	 */
193 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
194 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
195 		return -EINVAL;
196 	}
197 
198 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
199 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
200 		return -EINVAL;
201 	}
202 
203 	if ((args->ring_base_address) &&
204 		(!access_ok((const void __user *) args->ring_base_address,
205 			sizeof(uint64_t)))) {
206 		pr_err("Can't access ring base address\n");
207 		return -EFAULT;
208 	}
209 
210 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
211 		pr_err("Ring size must be a power of 2 or 0\n");
212 		return -EINVAL;
213 	}
214 
215 	if (!access_ok((const void __user *) args->read_pointer_address,
216 			sizeof(uint32_t))) {
217 		pr_err("Can't access read pointer\n");
218 		return -EFAULT;
219 	}
220 
221 	if (!access_ok((const void __user *) args->write_pointer_address,
222 			sizeof(uint32_t))) {
223 		pr_err("Can't access write pointer\n");
224 		return -EFAULT;
225 	}
226 
227 	if (args->eop_buffer_address &&
228 		!access_ok((const void __user *) args->eop_buffer_address,
229 			sizeof(uint32_t))) {
230 		pr_debug("Can't access eop buffer");
231 		return -EFAULT;
232 	}
233 
234 	if (args->ctx_save_restore_address &&
235 		!access_ok((const void __user *) args->ctx_save_restore_address,
236 			sizeof(uint32_t))) {
237 		pr_debug("Can't access ctx save restore buffer");
238 		return -EFAULT;
239 	}
240 
241 	q_properties->is_interop = false;
242 	q_properties->is_gws = false;
243 	q_properties->queue_percent = args->queue_percentage & 0xFF;
244 	/* bit 8-15 are repurposed to be PM4 target XCC */
245 	q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
246 	q_properties->priority = args->queue_priority;
247 	q_properties->queue_address = args->ring_base_address;
248 	q_properties->queue_size = args->ring_size;
249 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
250 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
251 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
252 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
253 	q_properties->ctx_save_restore_area_address =
254 			args->ctx_save_restore_address;
255 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
256 	q_properties->ctl_stack_size = args->ctl_stack_size;
257 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
258 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
259 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
260 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
261 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
262 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
263 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
264 	else
265 		return -ENOTSUPP;
266 
267 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
268 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
269 	else
270 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
271 
272 	pr_debug("Queue Percentage: %d, %d\n",
273 			q_properties->queue_percent, args->queue_percentage);
274 
275 	pr_debug("Queue Priority: %d, %d\n",
276 			q_properties->priority, args->queue_priority);
277 
278 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
279 			q_properties->queue_address, args->ring_base_address);
280 
281 	pr_debug("Queue Size: 0x%llX, %u\n",
282 			q_properties->queue_size, args->ring_size);
283 
284 	pr_debug("Queue r/w Pointers: %px, %px\n",
285 			q_properties->read_ptr,
286 			q_properties->write_ptr);
287 
288 	pr_debug("Queue Format: %d\n", q_properties->format);
289 
290 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
291 
292 	pr_debug("Queue CTX save area: 0x%llX\n",
293 			q_properties->ctx_save_restore_area_address);
294 
295 	return 0;
296 }
297 
298 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
299 					void *data)
300 {
301 	struct kfd_ioctl_create_queue_args *args = data;
302 	struct kfd_node *dev;
303 	int err = 0;
304 	unsigned int queue_id;
305 	struct kfd_process_device *pdd;
306 	struct queue_properties q_properties;
307 	uint32_t doorbell_offset_in_process = 0;
308 	struct amdgpu_bo *wptr_bo = NULL;
309 
310 	memset(&q_properties, 0, sizeof(struct queue_properties));
311 
312 	pr_debug("Creating queue ioctl\n");
313 
314 	err = set_queue_properties_from_user(&q_properties, args);
315 	if (err)
316 		return err;
317 
318 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
319 
320 	mutex_lock(&p->mutex);
321 
322 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
323 	if (!pdd) {
324 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
325 		err = -EINVAL;
326 		goto err_pdd;
327 	}
328 	dev = pdd->dev;
329 
330 	pdd = kfd_bind_process_to_device(dev, p);
331 	if (IS_ERR(pdd)) {
332 		err = -ESRCH;
333 		goto err_bind_process;
334 	}
335 
336 	if (!pdd->qpd.proc_doorbells) {
337 		err = kfd_alloc_process_doorbells(dev->kfd, pdd);
338 		if (err) {
339 			pr_debug("failed to allocate process doorbells\n");
340 			goto err_bind_process;
341 		}
342 	}
343 
344 	/* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
345 	 * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
346 	 */
347 	if (dev->kfd->shared_resources.enable_mes &&
348 			((dev->adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK)
349 			>> AMDGPU_MES_API_VERSION_SHIFT) >= 2) {
350 		struct amdgpu_bo_va_mapping *wptr_mapping;
351 		struct amdgpu_vm *wptr_vm;
352 
353 		wptr_vm = drm_priv_to_vm(pdd->drm_priv);
354 		err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
355 		if (err)
356 			goto err_wptr_map_gart;
357 
358 		wptr_mapping = amdgpu_vm_bo_lookup_mapping(
359 				wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
360 		amdgpu_bo_unreserve(wptr_vm->root.bo);
361 		if (!wptr_mapping) {
362 			pr_err("Failed to lookup wptr bo\n");
363 			err = -EINVAL;
364 			goto err_wptr_map_gart;
365 		}
366 
367 		wptr_bo = wptr_mapping->bo_va->base.bo;
368 		if (wptr_bo->tbo.base.size > PAGE_SIZE) {
369 			pr_err("Requested GART mapping for wptr bo larger than one page\n");
370 			err = -EINVAL;
371 			goto err_wptr_map_gart;
372 		}
373 
374 		err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
375 		if (err) {
376 			pr_err("Failed to map wptr bo to GART\n");
377 			goto err_wptr_map_gart;
378 		}
379 	}
380 
381 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
382 			p->pasid,
383 			dev->id);
384 
385 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
386 			NULL, NULL, NULL, &doorbell_offset_in_process);
387 	if (err != 0)
388 		goto err_create_queue;
389 
390 	args->queue_id = queue_id;
391 
392 
393 	/* Return gpu_id as doorbell offset for mmap usage */
394 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
395 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
396 	if (KFD_IS_SOC15(dev))
397 		/* On SOC15 ASICs, include the doorbell offset within the
398 		 * process doorbell frame, which is 2 pages.
399 		 */
400 		args->doorbell_offset |= doorbell_offset_in_process;
401 
402 	mutex_unlock(&p->mutex);
403 
404 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
405 
406 	pr_debug("Ring buffer address == 0x%016llX\n",
407 			args->ring_base_address);
408 
409 	pr_debug("Read ptr address    == 0x%016llX\n",
410 			args->read_pointer_address);
411 
412 	pr_debug("Write ptr address   == 0x%016llX\n",
413 			args->write_pointer_address);
414 
415 	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
416 	return 0;
417 
418 err_create_queue:
419 	if (wptr_bo)
420 		amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
421 err_wptr_map_gart:
422 err_bind_process:
423 err_pdd:
424 	mutex_unlock(&p->mutex);
425 	return err;
426 }
427 
428 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
429 					void *data)
430 {
431 	int retval;
432 	struct kfd_ioctl_destroy_queue_args *args = data;
433 
434 	pr_debug("Destroying queue id %d for pasid 0x%x\n",
435 				args->queue_id,
436 				p->pasid);
437 
438 	mutex_lock(&p->mutex);
439 
440 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
441 
442 	mutex_unlock(&p->mutex);
443 	return retval;
444 }
445 
446 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
447 					void *data)
448 {
449 	int retval;
450 	struct kfd_ioctl_update_queue_args *args = data;
451 	struct queue_properties properties;
452 
453 	/*
454 	 * Repurpose queue percentage to accommodate new features:
455 	 * bit 0-7: queue percentage
456 	 * bit 8-15: pm4_target_xcc
457 	 */
458 	if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
459 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
460 		return -EINVAL;
461 	}
462 
463 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
464 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
465 		return -EINVAL;
466 	}
467 
468 	if ((args->ring_base_address) &&
469 		(!access_ok((const void __user *) args->ring_base_address,
470 			sizeof(uint64_t)))) {
471 		pr_err("Can't access ring base address\n");
472 		return -EFAULT;
473 	}
474 
475 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
476 		pr_err("Ring size must be a power of 2 or 0\n");
477 		return -EINVAL;
478 	}
479 
480 	properties.queue_address = args->ring_base_address;
481 	properties.queue_size = args->ring_size;
482 	properties.queue_percent = args->queue_percentage & 0xFF;
483 	/* bit 8-15 are repurposed to be PM4 target XCC */
484 	properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
485 	properties.priority = args->queue_priority;
486 
487 	pr_debug("Updating queue id %d for pasid 0x%x\n",
488 			args->queue_id, p->pasid);
489 
490 	mutex_lock(&p->mutex);
491 
492 	retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
493 
494 	mutex_unlock(&p->mutex);
495 
496 	return retval;
497 }
498 
499 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
500 					void *data)
501 {
502 	int retval;
503 	const int max_num_cus = 1024;
504 	struct kfd_ioctl_set_cu_mask_args *args = data;
505 	struct mqd_update_info minfo = {0};
506 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
507 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
508 
509 	if ((args->num_cu_mask % 32) != 0) {
510 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
511 				args->num_cu_mask);
512 		return -EINVAL;
513 	}
514 
515 	minfo.cu_mask.count = args->num_cu_mask;
516 	if (minfo.cu_mask.count == 0) {
517 		pr_debug("CU mask cannot be 0");
518 		return -EINVAL;
519 	}
520 
521 	/* To prevent an unreasonably large CU mask size, set an arbitrary
522 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
523 	 * past max_num_cus bits and just use the first max_num_cus bits.
524 	 */
525 	if (minfo.cu_mask.count > max_num_cus) {
526 		pr_debug("CU mask cannot be greater than 1024 bits");
527 		minfo.cu_mask.count = max_num_cus;
528 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
529 	}
530 
531 	minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
532 	if (!minfo.cu_mask.ptr)
533 		return -ENOMEM;
534 
535 	retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
536 	if (retval) {
537 		pr_debug("Could not copy CU mask from userspace");
538 		retval = -EFAULT;
539 		goto out;
540 	}
541 
542 	mutex_lock(&p->mutex);
543 
544 	retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
545 
546 	mutex_unlock(&p->mutex);
547 
548 out:
549 	kfree(minfo.cu_mask.ptr);
550 	return retval;
551 }
552 
553 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
554 					  struct kfd_process *p, void *data)
555 {
556 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
557 	int r;
558 
559 	mutex_lock(&p->mutex);
560 
561 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
562 			       (void __user *)args->ctl_stack_address,
563 			       &args->ctl_stack_used_size,
564 			       &args->save_area_used_size);
565 
566 	mutex_unlock(&p->mutex);
567 
568 	return r;
569 }
570 
571 static int kfd_ioctl_set_memory_policy(struct file *filep,
572 					struct kfd_process *p, void *data)
573 {
574 	struct kfd_ioctl_set_memory_policy_args *args = data;
575 	int err = 0;
576 	struct kfd_process_device *pdd;
577 	enum cache_policy default_policy, alternate_policy;
578 
579 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
580 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
581 		return -EINVAL;
582 	}
583 
584 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
585 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
586 		return -EINVAL;
587 	}
588 
589 	mutex_lock(&p->mutex);
590 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
591 	if (!pdd) {
592 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
593 		err = -EINVAL;
594 		goto err_pdd;
595 	}
596 
597 	pdd = kfd_bind_process_to_device(pdd->dev, p);
598 	if (IS_ERR(pdd)) {
599 		err = -ESRCH;
600 		goto out;
601 	}
602 
603 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
604 			 ? cache_policy_coherent : cache_policy_noncoherent;
605 
606 	alternate_policy =
607 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
608 		   ? cache_policy_coherent : cache_policy_noncoherent;
609 
610 	if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
611 				&pdd->qpd,
612 				default_policy,
613 				alternate_policy,
614 				(void __user *)args->alternate_aperture_base,
615 				args->alternate_aperture_size))
616 		err = -EINVAL;
617 
618 out:
619 err_pdd:
620 	mutex_unlock(&p->mutex);
621 
622 	return err;
623 }
624 
625 static int kfd_ioctl_set_trap_handler(struct file *filep,
626 					struct kfd_process *p, void *data)
627 {
628 	struct kfd_ioctl_set_trap_handler_args *args = data;
629 	int err = 0;
630 	struct kfd_process_device *pdd;
631 
632 	mutex_lock(&p->mutex);
633 
634 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
635 	if (!pdd) {
636 		err = -EINVAL;
637 		goto err_pdd;
638 	}
639 
640 	pdd = kfd_bind_process_to_device(pdd->dev, p);
641 	if (IS_ERR(pdd)) {
642 		err = -ESRCH;
643 		goto out;
644 	}
645 
646 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
647 
648 out:
649 err_pdd:
650 	mutex_unlock(&p->mutex);
651 
652 	return err;
653 }
654 
655 static int kfd_ioctl_dbg_register(struct file *filep,
656 				struct kfd_process *p, void *data)
657 {
658 	return -EPERM;
659 }
660 
661 static int kfd_ioctl_dbg_unregister(struct file *filep,
662 				struct kfd_process *p, void *data)
663 {
664 	return -EPERM;
665 }
666 
667 static int kfd_ioctl_dbg_address_watch(struct file *filep,
668 					struct kfd_process *p, void *data)
669 {
670 	return -EPERM;
671 }
672 
673 /* Parse and generate fixed size data structure for wave control */
674 static int kfd_ioctl_dbg_wave_control(struct file *filep,
675 					struct kfd_process *p, void *data)
676 {
677 	return -EPERM;
678 }
679 
680 static int kfd_ioctl_get_clock_counters(struct file *filep,
681 				struct kfd_process *p, void *data)
682 {
683 	struct kfd_ioctl_get_clock_counters_args *args = data;
684 	struct kfd_process_device *pdd;
685 
686 	mutex_lock(&p->mutex);
687 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
688 	mutex_unlock(&p->mutex);
689 	if (pdd)
690 		/* Reading GPU clock counter from KGD */
691 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
692 	else
693 		/* Node without GPU resource */
694 		args->gpu_clock_counter = 0;
695 
696 	/* No access to rdtsc. Using raw monotonic time */
697 	args->cpu_clock_counter = ktime_get_raw_ns();
698 	args->system_clock_counter = ktime_get_boottime_ns();
699 
700 	/* Since the counter is in nano-seconds we use 1GHz frequency */
701 	args->system_clock_freq = 1000000000;
702 
703 	return 0;
704 }
705 
706 
707 static int kfd_ioctl_get_process_apertures(struct file *filp,
708 				struct kfd_process *p, void *data)
709 {
710 	struct kfd_ioctl_get_process_apertures_args *args = data;
711 	struct kfd_process_device_apertures *pAperture;
712 	int i;
713 
714 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
715 
716 	args->num_of_nodes = 0;
717 
718 	mutex_lock(&p->mutex);
719 	/* Run over all pdd of the process */
720 	for (i = 0; i < p->n_pdds; i++) {
721 		struct kfd_process_device *pdd = p->pdds[i];
722 
723 		pAperture =
724 			&args->process_apertures[args->num_of_nodes];
725 		pAperture->gpu_id = pdd->dev->id;
726 		pAperture->lds_base = pdd->lds_base;
727 		pAperture->lds_limit = pdd->lds_limit;
728 		pAperture->gpuvm_base = pdd->gpuvm_base;
729 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
730 		pAperture->scratch_base = pdd->scratch_base;
731 		pAperture->scratch_limit = pdd->scratch_limit;
732 
733 		dev_dbg(kfd_device,
734 			"node id %u\n", args->num_of_nodes);
735 		dev_dbg(kfd_device,
736 			"gpu id %u\n", pdd->dev->id);
737 		dev_dbg(kfd_device,
738 			"lds_base %llX\n", pdd->lds_base);
739 		dev_dbg(kfd_device,
740 			"lds_limit %llX\n", pdd->lds_limit);
741 		dev_dbg(kfd_device,
742 			"gpuvm_base %llX\n", pdd->gpuvm_base);
743 		dev_dbg(kfd_device,
744 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
745 		dev_dbg(kfd_device,
746 			"scratch_base %llX\n", pdd->scratch_base);
747 		dev_dbg(kfd_device,
748 			"scratch_limit %llX\n", pdd->scratch_limit);
749 
750 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
751 			break;
752 	}
753 	mutex_unlock(&p->mutex);
754 
755 	return 0;
756 }
757 
758 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
759 				struct kfd_process *p, void *data)
760 {
761 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
762 	struct kfd_process_device_apertures *pa;
763 	int ret;
764 	int i;
765 
766 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
767 
768 	if (args->num_of_nodes == 0) {
769 		/* Return number of nodes, so that user space can alloacate
770 		 * sufficient memory
771 		 */
772 		mutex_lock(&p->mutex);
773 		args->num_of_nodes = p->n_pdds;
774 		goto out_unlock;
775 	}
776 
777 	/* Fill in process-aperture information for all available
778 	 * nodes, but not more than args->num_of_nodes as that is
779 	 * the amount of memory allocated by user
780 	 */
781 	pa = kzalloc((sizeof(struct kfd_process_device_apertures) *
782 				args->num_of_nodes), GFP_KERNEL);
783 	if (!pa)
784 		return -ENOMEM;
785 
786 	mutex_lock(&p->mutex);
787 
788 	if (!p->n_pdds) {
789 		args->num_of_nodes = 0;
790 		kfree(pa);
791 		goto out_unlock;
792 	}
793 
794 	/* Run over all pdd of the process */
795 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
796 		struct kfd_process_device *pdd = p->pdds[i];
797 
798 		pa[i].gpu_id = pdd->dev->id;
799 		pa[i].lds_base = pdd->lds_base;
800 		pa[i].lds_limit = pdd->lds_limit;
801 		pa[i].gpuvm_base = pdd->gpuvm_base;
802 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
803 		pa[i].scratch_base = pdd->scratch_base;
804 		pa[i].scratch_limit = pdd->scratch_limit;
805 
806 		dev_dbg(kfd_device,
807 			"gpu id %u\n", pdd->dev->id);
808 		dev_dbg(kfd_device,
809 			"lds_base %llX\n", pdd->lds_base);
810 		dev_dbg(kfd_device,
811 			"lds_limit %llX\n", pdd->lds_limit);
812 		dev_dbg(kfd_device,
813 			"gpuvm_base %llX\n", pdd->gpuvm_base);
814 		dev_dbg(kfd_device,
815 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
816 		dev_dbg(kfd_device,
817 			"scratch_base %llX\n", pdd->scratch_base);
818 		dev_dbg(kfd_device,
819 			"scratch_limit %llX\n", pdd->scratch_limit);
820 	}
821 	mutex_unlock(&p->mutex);
822 
823 	args->num_of_nodes = i;
824 	ret = copy_to_user(
825 			(void __user *)args->kfd_process_device_apertures_ptr,
826 			pa,
827 			(i * sizeof(struct kfd_process_device_apertures)));
828 	kfree(pa);
829 	return ret ? -EFAULT : 0;
830 
831 out_unlock:
832 	mutex_unlock(&p->mutex);
833 	return 0;
834 }
835 
836 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
837 					void *data)
838 {
839 	struct kfd_ioctl_create_event_args *args = data;
840 	int err;
841 
842 	/* For dGPUs the event page is allocated in user mode. The
843 	 * handle is passed to KFD with the first call to this IOCTL
844 	 * through the event_page_offset field.
845 	 */
846 	if (args->event_page_offset) {
847 		mutex_lock(&p->mutex);
848 		err = kfd_kmap_event_page(p, args->event_page_offset);
849 		mutex_unlock(&p->mutex);
850 		if (err)
851 			return err;
852 	}
853 
854 	err = kfd_event_create(filp, p, args->event_type,
855 				args->auto_reset != 0, args->node_id,
856 				&args->event_id, &args->event_trigger_data,
857 				&args->event_page_offset,
858 				&args->event_slot_index);
859 
860 	pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
861 	return err;
862 }
863 
864 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
865 					void *data)
866 {
867 	struct kfd_ioctl_destroy_event_args *args = data;
868 
869 	return kfd_event_destroy(p, args->event_id);
870 }
871 
872 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
873 				void *data)
874 {
875 	struct kfd_ioctl_set_event_args *args = data;
876 
877 	return kfd_set_event(p, args->event_id);
878 }
879 
880 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
881 				void *data)
882 {
883 	struct kfd_ioctl_reset_event_args *args = data;
884 
885 	return kfd_reset_event(p, args->event_id);
886 }
887 
888 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
889 				void *data)
890 {
891 	struct kfd_ioctl_wait_events_args *args = data;
892 
893 	return kfd_wait_on_events(p, args->num_events,
894 			(void __user *)args->events_ptr,
895 			(args->wait_for_all != 0),
896 			&args->timeout, &args->wait_result);
897 }
898 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
899 					struct kfd_process *p, void *data)
900 {
901 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
902 	struct kfd_process_device *pdd;
903 	struct kfd_node *dev;
904 	long err;
905 
906 	mutex_lock(&p->mutex);
907 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
908 	if (!pdd) {
909 		err = -EINVAL;
910 		goto err_pdd;
911 	}
912 	dev = pdd->dev;
913 
914 	pdd = kfd_bind_process_to_device(dev, p);
915 	if (IS_ERR(pdd)) {
916 		err = PTR_ERR(pdd);
917 		goto bind_process_to_device_fail;
918 	}
919 
920 	pdd->qpd.sh_hidden_private_base = args->va_addr;
921 
922 	mutex_unlock(&p->mutex);
923 
924 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
925 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
926 		dev->kfd2kgd->set_scratch_backing_va(
927 			dev->adev, args->va_addr, pdd->qpd.vmid);
928 
929 	return 0;
930 
931 bind_process_to_device_fail:
932 err_pdd:
933 	mutex_unlock(&p->mutex);
934 	return err;
935 }
936 
937 static int kfd_ioctl_get_tile_config(struct file *filep,
938 		struct kfd_process *p, void *data)
939 {
940 	struct kfd_ioctl_get_tile_config_args *args = data;
941 	struct kfd_process_device *pdd;
942 	struct tile_config config;
943 	int err = 0;
944 
945 	mutex_lock(&p->mutex);
946 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
947 	mutex_unlock(&p->mutex);
948 	if (!pdd)
949 		return -EINVAL;
950 
951 	amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
952 
953 	args->gb_addr_config = config.gb_addr_config;
954 	args->num_banks = config.num_banks;
955 	args->num_ranks = config.num_ranks;
956 
957 	if (args->num_tile_configs > config.num_tile_configs)
958 		args->num_tile_configs = config.num_tile_configs;
959 	err = copy_to_user((void __user *)args->tile_config_ptr,
960 			config.tile_config_ptr,
961 			args->num_tile_configs * sizeof(uint32_t));
962 	if (err) {
963 		args->num_tile_configs = 0;
964 		return -EFAULT;
965 	}
966 
967 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
968 		args->num_macro_tile_configs =
969 				config.num_macro_tile_configs;
970 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
971 			config.macro_tile_config_ptr,
972 			args->num_macro_tile_configs * sizeof(uint32_t));
973 	if (err) {
974 		args->num_macro_tile_configs = 0;
975 		return -EFAULT;
976 	}
977 
978 	return 0;
979 }
980 
981 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
982 				void *data)
983 {
984 	struct kfd_ioctl_acquire_vm_args *args = data;
985 	struct kfd_process_device *pdd;
986 	struct file *drm_file;
987 	int ret;
988 
989 	drm_file = fget(args->drm_fd);
990 	if (!drm_file)
991 		return -EINVAL;
992 
993 	mutex_lock(&p->mutex);
994 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
995 	if (!pdd) {
996 		ret = -EINVAL;
997 		goto err_pdd;
998 	}
999 
1000 	if (pdd->drm_file) {
1001 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
1002 		goto err_drm_file;
1003 	}
1004 
1005 	ret = kfd_process_device_init_vm(pdd, drm_file);
1006 	if (ret)
1007 		goto err_unlock;
1008 
1009 	/* On success, the PDD keeps the drm_file reference */
1010 	mutex_unlock(&p->mutex);
1011 
1012 	return 0;
1013 
1014 err_unlock:
1015 err_pdd:
1016 err_drm_file:
1017 	mutex_unlock(&p->mutex);
1018 	fput(drm_file);
1019 	return ret;
1020 }
1021 
1022 bool kfd_dev_is_large_bar(struct kfd_node *dev)
1023 {
1024 	if (dev->kfd->adev->debug_largebar) {
1025 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1026 		return true;
1027 	}
1028 
1029 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1030 	    dev->local_mem_info.local_mem_size_public > 0)
1031 		return true;
1032 
1033 	if (dev->local_mem_info.local_mem_size_public == 0 &&
1034 	    dev->kfd->adev->gmc.is_app_apu) {
1035 		pr_debug("APP APU, Consider like a large bar system\n");
1036 		return true;
1037 	}
1038 
1039 	return false;
1040 }
1041 
1042 static int kfd_ioctl_get_available_memory(struct file *filep,
1043 					  struct kfd_process *p, void *data)
1044 {
1045 	struct kfd_ioctl_get_available_memory_args *args = data;
1046 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1047 
1048 	if (!pdd)
1049 		return -EINVAL;
1050 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1051 							pdd->dev->node_id);
1052 	kfd_unlock_pdd(pdd);
1053 	return 0;
1054 }
1055 
1056 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1057 					struct kfd_process *p, void *data)
1058 {
1059 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1060 	struct kfd_process_device *pdd;
1061 	void *mem;
1062 	struct kfd_node *dev;
1063 	int idr_handle;
1064 	long err;
1065 	uint64_t offset = args->mmap_offset;
1066 	uint32_t flags = args->flags;
1067 
1068 	if (args->size == 0)
1069 		return -EINVAL;
1070 
1071 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1072 	/* Flush pending deferred work to avoid racing with deferred actions
1073 	 * from previous memory map changes (e.g. munmap).
1074 	 */
1075 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1076 	mutex_lock(&p->svms.lock);
1077 	mmap_write_unlock(current->mm);
1078 	if (interval_tree_iter_first(&p->svms.objects,
1079 				     args->va_addr >> PAGE_SHIFT,
1080 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1081 		pr_err("Address: 0x%llx already allocated by SVM\n",
1082 			args->va_addr);
1083 		mutex_unlock(&p->svms.lock);
1084 		return -EADDRINUSE;
1085 	}
1086 
1087 	/* When register user buffer check if it has been registered by svm by
1088 	 * buffer cpu virtual address.
1089 	 */
1090 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1091 	    interval_tree_iter_first(&p->svms.objects,
1092 				     args->mmap_offset >> PAGE_SHIFT,
1093 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1094 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1095 			args->mmap_offset);
1096 		mutex_unlock(&p->svms.lock);
1097 		return -EADDRINUSE;
1098 	}
1099 
1100 	mutex_unlock(&p->svms.lock);
1101 #endif
1102 	mutex_lock(&p->mutex);
1103 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1104 	if (!pdd) {
1105 		err = -EINVAL;
1106 		goto err_pdd;
1107 	}
1108 
1109 	dev = pdd->dev;
1110 
1111 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1112 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1113 		!kfd_dev_is_large_bar(dev)) {
1114 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1115 		err = -EINVAL;
1116 		goto err_large_bar;
1117 	}
1118 
1119 	pdd = kfd_bind_process_to_device(dev, p);
1120 	if (IS_ERR(pdd)) {
1121 		err = PTR_ERR(pdd);
1122 		goto err_unlock;
1123 	}
1124 
1125 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1126 		if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1127 			err = -EINVAL;
1128 			goto err_unlock;
1129 		}
1130 		offset = kfd_get_process_doorbells(pdd);
1131 		if (!offset) {
1132 			err = -ENOMEM;
1133 			goto err_unlock;
1134 		}
1135 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1136 		if (args->size != PAGE_SIZE) {
1137 			err = -EINVAL;
1138 			goto err_unlock;
1139 		}
1140 		offset = dev->adev->rmmio_remap.bus_addr;
1141 		if (!offset) {
1142 			err = -ENOMEM;
1143 			goto err_unlock;
1144 		}
1145 	}
1146 
1147 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1148 		dev->adev, args->va_addr, args->size,
1149 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1150 		flags, false);
1151 
1152 	if (err)
1153 		goto err_unlock;
1154 
1155 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1156 	if (idr_handle < 0) {
1157 		err = -EFAULT;
1158 		goto err_free;
1159 	}
1160 
1161 	/* Update the VRAM usage count */
1162 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1163 		uint64_t size = args->size;
1164 
1165 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1166 			size >>= 1;
1167 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1168 	}
1169 
1170 	mutex_unlock(&p->mutex);
1171 
1172 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1173 	args->mmap_offset = offset;
1174 
1175 	/* MMIO is mapped through kfd device
1176 	 * Generate a kfd mmap offset
1177 	 */
1178 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1179 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1180 					| KFD_MMAP_GPU_ID(args->gpu_id);
1181 
1182 	return 0;
1183 
1184 err_free:
1185 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1186 					       pdd->drm_priv, NULL);
1187 err_unlock:
1188 err_pdd:
1189 err_large_bar:
1190 	mutex_unlock(&p->mutex);
1191 	return err;
1192 }
1193 
1194 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1195 					struct kfd_process *p, void *data)
1196 {
1197 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1198 	struct kfd_process_device *pdd;
1199 	void *mem;
1200 	int ret;
1201 	uint64_t size = 0;
1202 
1203 	mutex_lock(&p->mutex);
1204 	/*
1205 	 * Safeguard to prevent user space from freeing signal BO.
1206 	 * It will be freed at process termination.
1207 	 */
1208 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1209 		pr_err("Free signal BO is not allowed\n");
1210 		ret = -EPERM;
1211 		goto err_unlock;
1212 	}
1213 
1214 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1215 	if (!pdd) {
1216 		pr_err("Process device data doesn't exist\n");
1217 		ret = -EINVAL;
1218 		goto err_pdd;
1219 	}
1220 
1221 	mem = kfd_process_device_translate_handle(
1222 		pdd, GET_IDR_HANDLE(args->handle));
1223 	if (!mem) {
1224 		ret = -EINVAL;
1225 		goto err_unlock;
1226 	}
1227 
1228 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1229 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1230 
1231 	/* If freeing the buffer failed, leave the handle in place for
1232 	 * clean-up during process tear-down.
1233 	 */
1234 	if (!ret)
1235 		kfd_process_device_remove_obj_handle(
1236 			pdd, GET_IDR_HANDLE(args->handle));
1237 
1238 	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1239 
1240 err_unlock:
1241 err_pdd:
1242 	mutex_unlock(&p->mutex);
1243 	return ret;
1244 }
1245 
1246 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1247 					struct kfd_process *p, void *data)
1248 {
1249 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1250 	struct kfd_process_device *pdd, *peer_pdd;
1251 	void *mem;
1252 	struct kfd_node *dev;
1253 	long err = 0;
1254 	int i;
1255 	uint32_t *devices_arr = NULL;
1256 
1257 	if (!args->n_devices) {
1258 		pr_debug("Device IDs array empty\n");
1259 		return -EINVAL;
1260 	}
1261 	if (args->n_success > args->n_devices) {
1262 		pr_debug("n_success exceeds n_devices\n");
1263 		return -EINVAL;
1264 	}
1265 
1266 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1267 				    GFP_KERNEL);
1268 	if (!devices_arr)
1269 		return -ENOMEM;
1270 
1271 	err = copy_from_user(devices_arr,
1272 			     (void __user *)args->device_ids_array_ptr,
1273 			     args->n_devices * sizeof(*devices_arr));
1274 	if (err != 0) {
1275 		err = -EFAULT;
1276 		goto copy_from_user_failed;
1277 	}
1278 
1279 	mutex_lock(&p->mutex);
1280 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1281 	if (!pdd) {
1282 		err = -EINVAL;
1283 		goto get_process_device_data_failed;
1284 	}
1285 	dev = pdd->dev;
1286 
1287 	pdd = kfd_bind_process_to_device(dev, p);
1288 	if (IS_ERR(pdd)) {
1289 		err = PTR_ERR(pdd);
1290 		goto bind_process_to_device_failed;
1291 	}
1292 
1293 	mem = kfd_process_device_translate_handle(pdd,
1294 						GET_IDR_HANDLE(args->handle));
1295 	if (!mem) {
1296 		err = -ENOMEM;
1297 		goto get_mem_obj_from_handle_failed;
1298 	}
1299 
1300 	for (i = args->n_success; i < args->n_devices; i++) {
1301 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1302 		if (!peer_pdd) {
1303 			pr_debug("Getting device by id failed for 0x%x\n",
1304 				 devices_arr[i]);
1305 			err = -EINVAL;
1306 			goto get_mem_obj_from_handle_failed;
1307 		}
1308 
1309 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1310 		if (IS_ERR(peer_pdd)) {
1311 			err = PTR_ERR(peer_pdd);
1312 			goto get_mem_obj_from_handle_failed;
1313 		}
1314 
1315 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1316 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1317 			peer_pdd->drm_priv);
1318 		if (err) {
1319 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1320 
1321 			dev_err(dev->adev->dev,
1322 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1323 			       pci_domain_nr(pdev->bus),
1324 			       pdev->bus->number,
1325 			       PCI_SLOT(pdev->devfn),
1326 			       PCI_FUNC(pdev->devfn),
1327 			       ((struct kgd_mem *)mem)->domain);
1328 			goto map_memory_to_gpu_failed;
1329 		}
1330 		args->n_success = i+1;
1331 	}
1332 
1333 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1334 	if (err) {
1335 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1336 		goto sync_memory_failed;
1337 	}
1338 
1339 	mutex_unlock(&p->mutex);
1340 
1341 	/* Flush TLBs after waiting for the page table updates to complete */
1342 	for (i = 0; i < args->n_devices; i++) {
1343 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1344 		if (WARN_ON_ONCE(!peer_pdd))
1345 			continue;
1346 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1347 	}
1348 	kfree(devices_arr);
1349 
1350 	return err;
1351 
1352 get_process_device_data_failed:
1353 bind_process_to_device_failed:
1354 get_mem_obj_from_handle_failed:
1355 map_memory_to_gpu_failed:
1356 sync_memory_failed:
1357 	mutex_unlock(&p->mutex);
1358 copy_from_user_failed:
1359 	kfree(devices_arr);
1360 
1361 	return err;
1362 }
1363 
1364 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1365 					struct kfd_process *p, void *data)
1366 {
1367 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1368 	struct kfd_process_device *pdd, *peer_pdd;
1369 	void *mem;
1370 	long err = 0;
1371 	uint32_t *devices_arr = NULL, i;
1372 	bool flush_tlb;
1373 
1374 	if (!args->n_devices) {
1375 		pr_debug("Device IDs array empty\n");
1376 		return -EINVAL;
1377 	}
1378 	if (args->n_success > args->n_devices) {
1379 		pr_debug("n_success exceeds n_devices\n");
1380 		return -EINVAL;
1381 	}
1382 
1383 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1384 				    GFP_KERNEL);
1385 	if (!devices_arr)
1386 		return -ENOMEM;
1387 
1388 	err = copy_from_user(devices_arr,
1389 			     (void __user *)args->device_ids_array_ptr,
1390 			     args->n_devices * sizeof(*devices_arr));
1391 	if (err != 0) {
1392 		err = -EFAULT;
1393 		goto copy_from_user_failed;
1394 	}
1395 
1396 	mutex_lock(&p->mutex);
1397 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1398 	if (!pdd) {
1399 		err = -EINVAL;
1400 		goto bind_process_to_device_failed;
1401 	}
1402 
1403 	mem = kfd_process_device_translate_handle(pdd,
1404 						GET_IDR_HANDLE(args->handle));
1405 	if (!mem) {
1406 		err = -ENOMEM;
1407 		goto get_mem_obj_from_handle_failed;
1408 	}
1409 
1410 	for (i = args->n_success; i < args->n_devices; i++) {
1411 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1412 		if (!peer_pdd) {
1413 			err = -EINVAL;
1414 			goto get_mem_obj_from_handle_failed;
1415 		}
1416 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1417 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1418 		if (err) {
1419 			pr_err("Failed to unmap from gpu %d/%d\n",
1420 			       i, args->n_devices);
1421 			goto unmap_memory_from_gpu_failed;
1422 		}
1423 		args->n_success = i+1;
1424 	}
1425 
1426 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1427 	if (flush_tlb) {
1428 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1429 				(struct kgd_mem *) mem, true);
1430 		if (err) {
1431 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1432 			goto sync_memory_failed;
1433 		}
1434 	}
1435 
1436 	/* Flush TLBs after waiting for the page table updates to complete */
1437 	for (i = 0; i < args->n_devices; i++) {
1438 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1439 		if (WARN_ON_ONCE(!peer_pdd))
1440 			continue;
1441 		if (flush_tlb)
1442 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1443 
1444 		/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */
1445 		amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd->drm_priv);
1446 	}
1447 
1448 	mutex_unlock(&p->mutex);
1449 
1450 	kfree(devices_arr);
1451 
1452 	return 0;
1453 
1454 bind_process_to_device_failed:
1455 get_mem_obj_from_handle_failed:
1456 unmap_memory_from_gpu_failed:
1457 sync_memory_failed:
1458 	mutex_unlock(&p->mutex);
1459 copy_from_user_failed:
1460 	kfree(devices_arr);
1461 	return err;
1462 }
1463 
1464 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1465 		struct kfd_process *p, void *data)
1466 {
1467 	int retval;
1468 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1469 	struct queue *q;
1470 	struct kfd_node *dev;
1471 
1472 	mutex_lock(&p->mutex);
1473 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1474 
1475 	if (q) {
1476 		dev = q->device;
1477 	} else {
1478 		retval = -EINVAL;
1479 		goto out_unlock;
1480 	}
1481 
1482 	if (!dev->gws) {
1483 		retval = -ENODEV;
1484 		goto out_unlock;
1485 	}
1486 
1487 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1488 		retval = -ENODEV;
1489 		goto out_unlock;
1490 	}
1491 
1492 	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1493 				      kfd_dbg_has_cwsr_workaround(dev))) {
1494 		retval = -EBUSY;
1495 		goto out_unlock;
1496 	}
1497 
1498 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1499 	mutex_unlock(&p->mutex);
1500 
1501 	args->first_gws = 0;
1502 	return retval;
1503 
1504 out_unlock:
1505 	mutex_unlock(&p->mutex);
1506 	return retval;
1507 }
1508 
1509 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1510 		struct kfd_process *p, void *data)
1511 {
1512 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1513 	struct kfd_node *dev = NULL;
1514 	struct amdgpu_device *dmabuf_adev;
1515 	void *metadata_buffer = NULL;
1516 	uint32_t flags;
1517 	int8_t xcp_id;
1518 	unsigned int i;
1519 	int r;
1520 
1521 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1522 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1523 		if (dev)
1524 			break;
1525 	if (!dev)
1526 		return -EINVAL;
1527 
1528 	if (args->metadata_ptr) {
1529 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1530 		if (!metadata_buffer)
1531 			return -ENOMEM;
1532 	}
1533 
1534 	/* Get dmabuf info from KGD */
1535 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1536 					  &dmabuf_adev, &args->size,
1537 					  metadata_buffer, args->metadata_size,
1538 					  &args->metadata_size, &flags, &xcp_id);
1539 	if (r)
1540 		goto exit;
1541 
1542 	if (xcp_id >= 0)
1543 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1544 	else
1545 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[0]->id;
1546 	args->flags = flags;
1547 
1548 	/* Copy metadata buffer to user mode */
1549 	if (metadata_buffer) {
1550 		r = copy_to_user((void __user *)args->metadata_ptr,
1551 				 metadata_buffer, args->metadata_size);
1552 		if (r != 0)
1553 			r = -EFAULT;
1554 	}
1555 
1556 exit:
1557 	kfree(metadata_buffer);
1558 
1559 	return r;
1560 }
1561 
1562 static int kfd_ioctl_import_dmabuf(struct file *filep,
1563 				   struct kfd_process *p, void *data)
1564 {
1565 	struct kfd_ioctl_import_dmabuf_args *args = data;
1566 	struct kfd_process_device *pdd;
1567 	int idr_handle;
1568 	uint64_t size;
1569 	void *mem;
1570 	int r;
1571 
1572 	mutex_lock(&p->mutex);
1573 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1574 	if (!pdd) {
1575 		r = -EINVAL;
1576 		goto err_unlock;
1577 	}
1578 
1579 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1580 	if (IS_ERR(pdd)) {
1581 		r = PTR_ERR(pdd);
1582 		goto err_unlock;
1583 	}
1584 
1585 	r = amdgpu_amdkfd_gpuvm_import_dmabuf_fd(pdd->dev->adev, args->dmabuf_fd,
1586 						 args->va_addr, pdd->drm_priv,
1587 						 (struct kgd_mem **)&mem, &size,
1588 						 NULL);
1589 	if (r)
1590 		goto err_unlock;
1591 
1592 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1593 	if (idr_handle < 0) {
1594 		r = -EFAULT;
1595 		goto err_free;
1596 	}
1597 
1598 	mutex_unlock(&p->mutex);
1599 
1600 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1601 
1602 	return 0;
1603 
1604 err_free:
1605 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1606 					       pdd->drm_priv, NULL);
1607 err_unlock:
1608 	mutex_unlock(&p->mutex);
1609 	return r;
1610 }
1611 
1612 static int kfd_ioctl_export_dmabuf(struct file *filep,
1613 				   struct kfd_process *p, void *data)
1614 {
1615 	struct kfd_ioctl_export_dmabuf_args *args = data;
1616 	struct kfd_process_device *pdd;
1617 	struct dma_buf *dmabuf;
1618 	struct kfd_node *dev;
1619 	void *mem;
1620 	int ret = 0;
1621 
1622 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1623 	if (!dev)
1624 		return -EINVAL;
1625 
1626 	mutex_lock(&p->mutex);
1627 
1628 	pdd = kfd_get_process_device_data(dev, p);
1629 	if (!pdd) {
1630 		ret = -EINVAL;
1631 		goto err_unlock;
1632 	}
1633 
1634 	mem = kfd_process_device_translate_handle(pdd,
1635 						GET_IDR_HANDLE(args->handle));
1636 	if (!mem) {
1637 		ret = -EINVAL;
1638 		goto err_unlock;
1639 	}
1640 
1641 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1642 	mutex_unlock(&p->mutex);
1643 	if (ret)
1644 		goto err_out;
1645 
1646 	ret = dma_buf_fd(dmabuf, args->flags);
1647 	if (ret < 0) {
1648 		dma_buf_put(dmabuf);
1649 		goto err_out;
1650 	}
1651 	/* dma_buf_fd assigns the reference count to the fd, no need to
1652 	 * put the reference here.
1653 	 */
1654 	args->dmabuf_fd = ret;
1655 
1656 	return 0;
1657 
1658 err_unlock:
1659 	mutex_unlock(&p->mutex);
1660 err_out:
1661 	return ret;
1662 }
1663 
1664 /* Handle requests for watching SMI events */
1665 static int kfd_ioctl_smi_events(struct file *filep,
1666 				struct kfd_process *p, void *data)
1667 {
1668 	struct kfd_ioctl_smi_events_args *args = data;
1669 	struct kfd_process_device *pdd;
1670 
1671 	mutex_lock(&p->mutex);
1672 
1673 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1674 	mutex_unlock(&p->mutex);
1675 	if (!pdd)
1676 		return -EINVAL;
1677 
1678 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1679 }
1680 
1681 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1682 
1683 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1684 				    struct kfd_process *p, void *data)
1685 {
1686 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1687 	int r = 0;
1688 
1689 	mutex_lock(&p->mutex);
1690 	if (args->xnack_enabled >= 0) {
1691 		if (!list_empty(&p->pqm.queues)) {
1692 			pr_debug("Process has user queues running\n");
1693 			r = -EBUSY;
1694 			goto out_unlock;
1695 		}
1696 
1697 		if (p->xnack_enabled == args->xnack_enabled)
1698 			goto out_unlock;
1699 
1700 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1701 			r = -EPERM;
1702 			goto out_unlock;
1703 		}
1704 
1705 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1706 	} else {
1707 		args->xnack_enabled = p->xnack_enabled;
1708 	}
1709 
1710 out_unlock:
1711 	mutex_unlock(&p->mutex);
1712 
1713 	return r;
1714 }
1715 
1716 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1717 {
1718 	struct kfd_ioctl_svm_args *args = data;
1719 	int r = 0;
1720 
1721 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1722 		 args->start_addr, args->size, args->op, args->nattr);
1723 
1724 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1725 		return -EINVAL;
1726 	if (!args->start_addr || !args->size)
1727 		return -EINVAL;
1728 
1729 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1730 		      args->attrs);
1731 
1732 	return r;
1733 }
1734 #else
1735 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1736 				    struct kfd_process *p, void *data)
1737 {
1738 	return -EPERM;
1739 }
1740 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1741 {
1742 	return -EPERM;
1743 }
1744 #endif
1745 
1746 static int criu_checkpoint_process(struct kfd_process *p,
1747 			     uint8_t __user *user_priv_data,
1748 			     uint64_t *priv_offset)
1749 {
1750 	struct kfd_criu_process_priv_data process_priv;
1751 	int ret;
1752 
1753 	memset(&process_priv, 0, sizeof(process_priv));
1754 
1755 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1756 	/* For CR, we don't consider negative xnack mode which is used for
1757 	 * querying without changing it, here 0 simply means disabled and 1
1758 	 * means enabled so retry for finding a valid PTE.
1759 	 */
1760 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1761 
1762 	ret = copy_to_user(user_priv_data + *priv_offset,
1763 				&process_priv, sizeof(process_priv));
1764 
1765 	if (ret) {
1766 		pr_err("Failed to copy process information to user\n");
1767 		ret = -EFAULT;
1768 	}
1769 
1770 	*priv_offset += sizeof(process_priv);
1771 	return ret;
1772 }
1773 
1774 static int criu_checkpoint_devices(struct kfd_process *p,
1775 			     uint32_t num_devices,
1776 			     uint8_t __user *user_addr,
1777 			     uint8_t __user *user_priv_data,
1778 			     uint64_t *priv_offset)
1779 {
1780 	struct kfd_criu_device_priv_data *device_priv = NULL;
1781 	struct kfd_criu_device_bucket *device_buckets = NULL;
1782 	int ret = 0, i;
1783 
1784 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1785 	if (!device_buckets) {
1786 		ret = -ENOMEM;
1787 		goto exit;
1788 	}
1789 
1790 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1791 	if (!device_priv) {
1792 		ret = -ENOMEM;
1793 		goto exit;
1794 	}
1795 
1796 	for (i = 0; i < num_devices; i++) {
1797 		struct kfd_process_device *pdd = p->pdds[i];
1798 
1799 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1800 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1801 
1802 		/*
1803 		 * priv_data does not contain useful information for now and is reserved for
1804 		 * future use, so we do not set its contents.
1805 		 */
1806 	}
1807 
1808 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1809 	if (ret) {
1810 		pr_err("Failed to copy device information to user\n");
1811 		ret = -EFAULT;
1812 		goto exit;
1813 	}
1814 
1815 	ret = copy_to_user(user_priv_data + *priv_offset,
1816 			   device_priv,
1817 			   num_devices * sizeof(*device_priv));
1818 	if (ret) {
1819 		pr_err("Failed to copy device information to user\n");
1820 		ret = -EFAULT;
1821 	}
1822 	*priv_offset += num_devices * sizeof(*device_priv);
1823 
1824 exit:
1825 	kvfree(device_buckets);
1826 	kvfree(device_priv);
1827 	return ret;
1828 }
1829 
1830 static uint32_t get_process_num_bos(struct kfd_process *p)
1831 {
1832 	uint32_t num_of_bos = 0;
1833 	int i;
1834 
1835 	/* Run over all PDDs of the process */
1836 	for (i = 0; i < p->n_pdds; i++) {
1837 		struct kfd_process_device *pdd = p->pdds[i];
1838 		void *mem;
1839 		int id;
1840 
1841 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1842 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1843 
1844 			if (!kgd_mem->va || kgd_mem->va > pdd->gpuvm_base)
1845 				num_of_bos++;
1846 		}
1847 	}
1848 	return num_of_bos;
1849 }
1850 
1851 static int criu_get_prime_handle(struct kgd_mem *mem,
1852 				 int flags, u32 *shared_fd)
1853 {
1854 	struct dma_buf *dmabuf;
1855 	int ret;
1856 
1857 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1858 	if (ret) {
1859 		pr_err("dmabuf export failed for the BO\n");
1860 		return ret;
1861 	}
1862 
1863 	ret = dma_buf_fd(dmabuf, flags);
1864 	if (ret < 0) {
1865 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1866 		goto out_free_dmabuf;
1867 	}
1868 
1869 	*shared_fd = ret;
1870 	return 0;
1871 
1872 out_free_dmabuf:
1873 	dma_buf_put(dmabuf);
1874 	return ret;
1875 }
1876 
1877 static int criu_checkpoint_bos(struct kfd_process *p,
1878 			       uint32_t num_bos,
1879 			       uint8_t __user *user_bos,
1880 			       uint8_t __user *user_priv_data,
1881 			       uint64_t *priv_offset)
1882 {
1883 	struct kfd_criu_bo_bucket *bo_buckets;
1884 	struct kfd_criu_bo_priv_data *bo_privs;
1885 	int ret = 0, pdd_index, bo_index = 0, id;
1886 	void *mem;
1887 
1888 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1889 	if (!bo_buckets)
1890 		return -ENOMEM;
1891 
1892 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1893 	if (!bo_privs) {
1894 		ret = -ENOMEM;
1895 		goto exit;
1896 	}
1897 
1898 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1899 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1900 		struct amdgpu_bo *dumper_bo;
1901 		struct kgd_mem *kgd_mem;
1902 
1903 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1904 			struct kfd_criu_bo_bucket *bo_bucket;
1905 			struct kfd_criu_bo_priv_data *bo_priv;
1906 			int i, dev_idx = 0;
1907 
1908 			if (!mem) {
1909 				ret = -ENOMEM;
1910 				goto exit;
1911 			}
1912 
1913 			kgd_mem = (struct kgd_mem *)mem;
1914 			dumper_bo = kgd_mem->bo;
1915 
1916 			/* Skip checkpointing BOs that are used for Trap handler
1917 			 * code and state. Currently, these BOs have a VA that
1918 			 * is less GPUVM Base
1919 			 */
1920 			if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base)
1921 				continue;
1922 
1923 			bo_bucket = &bo_buckets[bo_index];
1924 			bo_priv = &bo_privs[bo_index];
1925 
1926 			bo_bucket->gpu_id = pdd->user_gpu_id;
1927 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1928 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1929 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1930 			bo_priv->idr_handle = id;
1931 
1932 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1933 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1934 								&bo_priv->user_addr);
1935 				if (ret) {
1936 					pr_err("Failed to obtain user address for user-pointer bo\n");
1937 					goto exit;
1938 				}
1939 			}
1940 			if (bo_bucket->alloc_flags
1941 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1942 				ret = criu_get_prime_handle(kgd_mem,
1943 						bo_bucket->alloc_flags &
1944 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1945 						&bo_bucket->dmabuf_fd);
1946 				if (ret)
1947 					goto exit;
1948 			} else {
1949 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1950 			}
1951 
1952 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1953 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1954 					KFD_MMAP_GPU_ID(pdd->dev->id);
1955 			else if (bo_bucket->alloc_flags &
1956 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1957 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1958 					KFD_MMAP_GPU_ID(pdd->dev->id);
1959 			else
1960 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1961 
1962 			for (i = 0; i < p->n_pdds; i++) {
1963 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->dev->adev, kgd_mem))
1964 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1965 			}
1966 
1967 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1968 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1969 					bo_bucket->size,
1970 					bo_bucket->addr,
1971 					bo_bucket->offset,
1972 					bo_bucket->gpu_id,
1973 					bo_bucket->alloc_flags,
1974 					bo_priv->idr_handle);
1975 			bo_index++;
1976 		}
1977 	}
1978 
1979 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1980 	if (ret) {
1981 		pr_err("Failed to copy BO information to user\n");
1982 		ret = -EFAULT;
1983 		goto exit;
1984 	}
1985 
1986 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1987 	if (ret) {
1988 		pr_err("Failed to copy BO priv information to user\n");
1989 		ret = -EFAULT;
1990 		goto exit;
1991 	}
1992 
1993 	*priv_offset += num_bos * sizeof(*bo_privs);
1994 
1995 exit:
1996 	while (ret && bo_index--) {
1997 		if (bo_buckets[bo_index].alloc_flags
1998 		    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
1999 			close_fd(bo_buckets[bo_index].dmabuf_fd);
2000 	}
2001 
2002 	kvfree(bo_buckets);
2003 	kvfree(bo_privs);
2004 	return ret;
2005 }
2006 
2007 static int criu_get_process_object_info(struct kfd_process *p,
2008 					uint32_t *num_devices,
2009 					uint32_t *num_bos,
2010 					uint32_t *num_objects,
2011 					uint64_t *objs_priv_size)
2012 {
2013 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2014 	uint32_t num_queues, num_events, num_svm_ranges;
2015 	int ret;
2016 
2017 	*num_devices = p->n_pdds;
2018 	*num_bos = get_process_num_bos(p);
2019 
2020 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2021 	if (ret)
2022 		return ret;
2023 
2024 	num_events = kfd_get_num_events(p);
2025 
2026 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2027 	if (ret)
2028 		return ret;
2029 
2030 	*num_objects = num_queues + num_events + num_svm_ranges;
2031 
2032 	if (objs_priv_size) {
2033 		priv_size = sizeof(struct kfd_criu_process_priv_data);
2034 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2035 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2036 		priv_size += queues_priv_data_size;
2037 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2038 		priv_size += svm_priv_data_size;
2039 		*objs_priv_size = priv_size;
2040 	}
2041 	return 0;
2042 }
2043 
2044 static int criu_checkpoint(struct file *filep,
2045 			   struct kfd_process *p,
2046 			   struct kfd_ioctl_criu_args *args)
2047 {
2048 	int ret;
2049 	uint32_t num_devices, num_bos, num_objects;
2050 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2051 
2052 	if (!args->devices || !args->bos || !args->priv_data)
2053 		return -EINVAL;
2054 
2055 	mutex_lock(&p->mutex);
2056 
2057 	if (!p->n_pdds) {
2058 		pr_err("No pdd for given process\n");
2059 		ret = -ENODEV;
2060 		goto exit_unlock;
2061 	}
2062 
2063 	/* Confirm all process queues are evicted */
2064 	if (!p->queues_paused) {
2065 		pr_err("Cannot dump process when queues are not in evicted state\n");
2066 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2067 		ret = -EINVAL;
2068 		goto exit_unlock;
2069 	}
2070 
2071 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2072 	if (ret)
2073 		goto exit_unlock;
2074 
2075 	if (num_devices != args->num_devices ||
2076 	    num_bos != args->num_bos ||
2077 	    num_objects != args->num_objects ||
2078 	    priv_size != args->priv_data_size) {
2079 
2080 		ret = -EINVAL;
2081 		goto exit_unlock;
2082 	}
2083 
2084 	/* each function will store private data inside priv_data and adjust priv_offset */
2085 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2086 	if (ret)
2087 		goto exit_unlock;
2088 
2089 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2090 				(uint8_t __user *)args->priv_data, &priv_offset);
2091 	if (ret)
2092 		goto exit_unlock;
2093 
2094 	/* Leave room for BOs in the private data. They need to be restored
2095 	 * before events, but we checkpoint them last to simplify the error
2096 	 * handling.
2097 	 */
2098 	bo_priv_offset = priv_offset;
2099 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2100 
2101 	if (num_objects) {
2102 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2103 						 &priv_offset);
2104 		if (ret)
2105 			goto exit_unlock;
2106 
2107 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2108 						 &priv_offset);
2109 		if (ret)
2110 			goto exit_unlock;
2111 
2112 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2113 		if (ret)
2114 			goto exit_unlock;
2115 	}
2116 
2117 	/* This must be the last thing in this function that can fail.
2118 	 * Otherwise we leak dmabuf file descriptors.
2119 	 */
2120 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2121 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2122 
2123 exit_unlock:
2124 	mutex_unlock(&p->mutex);
2125 	if (ret)
2126 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2127 	else
2128 		pr_debug("CRIU dump ret:%d\n", ret);
2129 
2130 	return ret;
2131 }
2132 
2133 static int criu_restore_process(struct kfd_process *p,
2134 				struct kfd_ioctl_criu_args *args,
2135 				uint64_t *priv_offset,
2136 				uint64_t max_priv_data_size)
2137 {
2138 	int ret = 0;
2139 	struct kfd_criu_process_priv_data process_priv;
2140 
2141 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2142 		return -EINVAL;
2143 
2144 	ret = copy_from_user(&process_priv,
2145 				(void __user *)(args->priv_data + *priv_offset),
2146 				sizeof(process_priv));
2147 	if (ret) {
2148 		pr_err("Failed to copy process private information from user\n");
2149 		ret = -EFAULT;
2150 		goto exit;
2151 	}
2152 	*priv_offset += sizeof(process_priv);
2153 
2154 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2155 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2156 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2157 		return -EINVAL;
2158 	}
2159 
2160 	pr_debug("Setting XNACK mode\n");
2161 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2162 		pr_err("xnack mode cannot be set\n");
2163 		ret = -EPERM;
2164 		goto exit;
2165 	} else {
2166 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2167 		p->xnack_enabled = process_priv.xnack_mode;
2168 	}
2169 
2170 exit:
2171 	return ret;
2172 }
2173 
2174 static int criu_restore_devices(struct kfd_process *p,
2175 				struct kfd_ioctl_criu_args *args,
2176 				uint64_t *priv_offset,
2177 				uint64_t max_priv_data_size)
2178 {
2179 	struct kfd_criu_device_bucket *device_buckets;
2180 	struct kfd_criu_device_priv_data *device_privs;
2181 	int ret = 0;
2182 	uint32_t i;
2183 
2184 	if (args->num_devices != p->n_pdds)
2185 		return -EINVAL;
2186 
2187 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2188 		return -EINVAL;
2189 
2190 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2191 	if (!device_buckets)
2192 		return -ENOMEM;
2193 
2194 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2195 				args->num_devices * sizeof(*device_buckets));
2196 	if (ret) {
2197 		pr_err("Failed to copy devices buckets from user\n");
2198 		ret = -EFAULT;
2199 		goto exit;
2200 	}
2201 
2202 	for (i = 0; i < args->num_devices; i++) {
2203 		struct kfd_node *dev;
2204 		struct kfd_process_device *pdd;
2205 		struct file *drm_file;
2206 
2207 		/* device private data is not currently used */
2208 
2209 		if (!device_buckets[i].user_gpu_id) {
2210 			pr_err("Invalid user gpu_id\n");
2211 			ret = -EINVAL;
2212 			goto exit;
2213 		}
2214 
2215 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2216 		if (!dev) {
2217 			pr_err("Failed to find device with gpu_id = %x\n",
2218 				device_buckets[i].actual_gpu_id);
2219 			ret = -EINVAL;
2220 			goto exit;
2221 		}
2222 
2223 		pdd = kfd_get_process_device_data(dev, p);
2224 		if (!pdd) {
2225 			pr_err("Failed to get pdd for gpu_id = %x\n",
2226 					device_buckets[i].actual_gpu_id);
2227 			ret = -EINVAL;
2228 			goto exit;
2229 		}
2230 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2231 
2232 		drm_file = fget(device_buckets[i].drm_fd);
2233 		if (!drm_file) {
2234 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2235 				device_buckets[i].drm_fd);
2236 			ret = -EINVAL;
2237 			goto exit;
2238 		}
2239 
2240 		if (pdd->drm_file) {
2241 			ret = -EINVAL;
2242 			goto exit;
2243 		}
2244 
2245 		/* create the vm using render nodes for kfd pdd */
2246 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2247 			pr_err("could not init vm for given pdd\n");
2248 			/* On success, the PDD keeps the drm_file reference */
2249 			fput(drm_file);
2250 			ret = -EINVAL;
2251 			goto exit;
2252 		}
2253 		/*
2254 		 * pdd now already has the vm bound to render node so below api won't create a new
2255 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2256 		 * for iommu v2 binding  and runtime pm.
2257 		 */
2258 		pdd = kfd_bind_process_to_device(dev, p);
2259 		if (IS_ERR(pdd)) {
2260 			ret = PTR_ERR(pdd);
2261 			goto exit;
2262 		}
2263 
2264 		if (!pdd->qpd.proc_doorbells) {
2265 			ret = kfd_alloc_process_doorbells(dev->kfd, pdd);
2266 			if (ret)
2267 				goto exit;
2268 		}
2269 	}
2270 
2271 	/*
2272 	 * We are not copying device private data from user as we are not using the data for now,
2273 	 * but we still adjust for its private data.
2274 	 */
2275 	*priv_offset += args->num_devices * sizeof(*device_privs);
2276 
2277 exit:
2278 	kfree(device_buckets);
2279 	return ret;
2280 }
2281 
2282 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2283 				      struct kfd_criu_bo_bucket *bo_bucket,
2284 				      struct kfd_criu_bo_priv_data *bo_priv,
2285 				      struct kgd_mem **kgd_mem)
2286 {
2287 	int idr_handle;
2288 	int ret;
2289 	const bool criu_resume = true;
2290 	u64 offset;
2291 
2292 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2293 		if (bo_bucket->size !=
2294 				kfd_doorbell_process_slice(pdd->dev->kfd))
2295 			return -EINVAL;
2296 
2297 		offset = kfd_get_process_doorbells(pdd);
2298 		if (!offset)
2299 			return -ENOMEM;
2300 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2301 		/* MMIO BOs need remapped bus address */
2302 		if (bo_bucket->size != PAGE_SIZE) {
2303 			pr_err("Invalid page size\n");
2304 			return -EINVAL;
2305 		}
2306 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2307 		if (!offset) {
2308 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2309 			return -ENOMEM;
2310 		}
2311 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2312 		offset = bo_priv->user_addr;
2313 	}
2314 	/* Create the BO */
2315 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2316 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2317 						      &offset, bo_bucket->alloc_flags, criu_resume);
2318 	if (ret) {
2319 		pr_err("Could not create the BO\n");
2320 		return ret;
2321 	}
2322 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2323 		 bo_bucket->size, bo_bucket->addr, offset);
2324 
2325 	/* Restore previous IDR handle */
2326 	pr_debug("Restoring old IDR handle for the BO");
2327 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2328 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2329 
2330 	if (idr_handle < 0) {
2331 		pr_err("Could not allocate idr\n");
2332 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2333 						       NULL);
2334 		return -ENOMEM;
2335 	}
2336 
2337 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2338 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2339 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2340 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2341 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2342 		bo_bucket->restored_offset = offset;
2343 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2344 		bo_bucket->restored_offset = offset;
2345 		/* Update the VRAM usage count */
2346 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2347 	}
2348 	return 0;
2349 }
2350 
2351 static int criu_restore_bo(struct kfd_process *p,
2352 			   struct kfd_criu_bo_bucket *bo_bucket,
2353 			   struct kfd_criu_bo_priv_data *bo_priv)
2354 {
2355 	struct kfd_process_device *pdd;
2356 	struct kgd_mem *kgd_mem;
2357 	int ret;
2358 	int j;
2359 
2360 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2361 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2362 		 bo_priv->idr_handle);
2363 
2364 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2365 	if (!pdd) {
2366 		pr_err("Failed to get pdd\n");
2367 		return -ENODEV;
2368 	}
2369 
2370 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2371 	if (ret)
2372 		return ret;
2373 
2374 	/* now map these BOs to GPU/s */
2375 	for (j = 0; j < p->n_pdds; j++) {
2376 		struct kfd_node *peer;
2377 		struct kfd_process_device *peer_pdd;
2378 
2379 		if (!bo_priv->mapped_gpuids[j])
2380 			break;
2381 
2382 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2383 		if (!peer_pdd)
2384 			return -EINVAL;
2385 
2386 		peer = peer_pdd->dev;
2387 
2388 		peer_pdd = kfd_bind_process_to_device(peer, p);
2389 		if (IS_ERR(peer_pdd))
2390 			return PTR_ERR(peer_pdd);
2391 
2392 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2393 							    peer_pdd->drm_priv);
2394 		if (ret) {
2395 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2396 			return ret;
2397 		}
2398 	}
2399 
2400 	pr_debug("map memory was successful for the BO\n");
2401 	/* create the dmabuf object and export the bo */
2402 	if (bo_bucket->alloc_flags
2403 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2404 		ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2405 					    &bo_bucket->dmabuf_fd);
2406 		if (ret)
2407 			return ret;
2408 	} else {
2409 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2410 	}
2411 
2412 	return 0;
2413 }
2414 
2415 static int criu_restore_bos(struct kfd_process *p,
2416 			    struct kfd_ioctl_criu_args *args,
2417 			    uint64_t *priv_offset,
2418 			    uint64_t max_priv_data_size)
2419 {
2420 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2421 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2422 	int ret = 0;
2423 	uint32_t i = 0;
2424 
2425 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2426 		return -EINVAL;
2427 
2428 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2429 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2430 
2431 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2432 	if (!bo_buckets)
2433 		return -ENOMEM;
2434 
2435 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2436 			     args->num_bos * sizeof(*bo_buckets));
2437 	if (ret) {
2438 		pr_err("Failed to copy BOs information from user\n");
2439 		ret = -EFAULT;
2440 		goto exit;
2441 	}
2442 
2443 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2444 	if (!bo_privs) {
2445 		ret = -ENOMEM;
2446 		goto exit;
2447 	}
2448 
2449 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2450 			     args->num_bos * sizeof(*bo_privs));
2451 	if (ret) {
2452 		pr_err("Failed to copy BOs information from user\n");
2453 		ret = -EFAULT;
2454 		goto exit;
2455 	}
2456 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2457 
2458 	/* Create and map new BOs */
2459 	for (; i < args->num_bos; i++) {
2460 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2461 		if (ret) {
2462 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2463 			goto exit;
2464 		}
2465 	} /* done */
2466 
2467 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2468 	ret = copy_to_user((void __user *)args->bos,
2469 				bo_buckets,
2470 				(args->num_bos * sizeof(*bo_buckets)));
2471 	if (ret)
2472 		ret = -EFAULT;
2473 
2474 exit:
2475 	while (ret && i--) {
2476 		if (bo_buckets[i].alloc_flags
2477 		   & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2478 			close_fd(bo_buckets[i].dmabuf_fd);
2479 	}
2480 	kvfree(bo_buckets);
2481 	kvfree(bo_privs);
2482 	return ret;
2483 }
2484 
2485 static int criu_restore_objects(struct file *filep,
2486 				struct kfd_process *p,
2487 				struct kfd_ioctl_criu_args *args,
2488 				uint64_t *priv_offset,
2489 				uint64_t max_priv_data_size)
2490 {
2491 	int ret = 0;
2492 	uint32_t i;
2493 
2494 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2495 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2496 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2497 
2498 	for (i = 0; i < args->num_objects; i++) {
2499 		uint32_t object_type;
2500 
2501 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2502 			pr_err("Invalid private data size\n");
2503 			return -EINVAL;
2504 		}
2505 
2506 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2507 		if (ret) {
2508 			pr_err("Failed to copy private information from user\n");
2509 			goto exit;
2510 		}
2511 
2512 		switch (object_type) {
2513 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2514 			ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2515 						     priv_offset, max_priv_data_size);
2516 			if (ret)
2517 				goto exit;
2518 			break;
2519 		case KFD_CRIU_OBJECT_TYPE_EVENT:
2520 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2521 						     priv_offset, max_priv_data_size);
2522 			if (ret)
2523 				goto exit;
2524 			break;
2525 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2526 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2527 						     priv_offset, max_priv_data_size);
2528 			if (ret)
2529 				goto exit;
2530 			break;
2531 		default:
2532 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2533 			ret = -EINVAL;
2534 			goto exit;
2535 		}
2536 	}
2537 exit:
2538 	return ret;
2539 }
2540 
2541 static int criu_restore(struct file *filep,
2542 			struct kfd_process *p,
2543 			struct kfd_ioctl_criu_args *args)
2544 {
2545 	uint64_t priv_offset = 0;
2546 	int ret = 0;
2547 
2548 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2549 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2550 
2551 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2552 	    !args->num_devices || !args->num_bos)
2553 		return -EINVAL;
2554 
2555 	mutex_lock(&p->mutex);
2556 
2557 	/*
2558 	 * Set the process to evicted state to avoid running any new queues before all the memory
2559 	 * mappings are ready.
2560 	 */
2561 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2562 	if (ret)
2563 		goto exit_unlock;
2564 
2565 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2566 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2567 	if (ret)
2568 		goto exit_unlock;
2569 
2570 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2571 	if (ret)
2572 		goto exit_unlock;
2573 
2574 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2575 	if (ret)
2576 		goto exit_unlock;
2577 
2578 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2579 	if (ret)
2580 		goto exit_unlock;
2581 
2582 	if (priv_offset != args->priv_data_size) {
2583 		pr_err("Invalid private data size\n");
2584 		ret = -EINVAL;
2585 	}
2586 
2587 exit_unlock:
2588 	mutex_unlock(&p->mutex);
2589 	if (ret)
2590 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2591 	else
2592 		pr_debug("CRIU restore successful\n");
2593 
2594 	return ret;
2595 }
2596 
2597 static int criu_unpause(struct file *filep,
2598 			struct kfd_process *p,
2599 			struct kfd_ioctl_criu_args *args)
2600 {
2601 	int ret;
2602 
2603 	mutex_lock(&p->mutex);
2604 
2605 	if (!p->queues_paused) {
2606 		mutex_unlock(&p->mutex);
2607 		return -EINVAL;
2608 	}
2609 
2610 	ret = kfd_process_restore_queues(p);
2611 	if (ret)
2612 		pr_err("Failed to unpause queues ret:%d\n", ret);
2613 	else
2614 		p->queues_paused = false;
2615 
2616 	mutex_unlock(&p->mutex);
2617 
2618 	return ret;
2619 }
2620 
2621 static int criu_resume(struct file *filep,
2622 			struct kfd_process *p,
2623 			struct kfd_ioctl_criu_args *args)
2624 {
2625 	struct kfd_process *target = NULL;
2626 	struct pid *pid = NULL;
2627 	int ret = 0;
2628 
2629 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2630 		 args->pid);
2631 
2632 	pid = find_get_pid(args->pid);
2633 	if (!pid) {
2634 		pr_err("Cannot find pid info for %i\n", args->pid);
2635 		return -ESRCH;
2636 	}
2637 
2638 	pr_debug("calling kfd_lookup_process_by_pid\n");
2639 	target = kfd_lookup_process_by_pid(pid);
2640 
2641 	put_pid(pid);
2642 
2643 	if (!target) {
2644 		pr_debug("Cannot find process info for %i\n", args->pid);
2645 		return -ESRCH;
2646 	}
2647 
2648 	mutex_lock(&target->mutex);
2649 	ret = kfd_criu_resume_svm(target);
2650 	if (ret) {
2651 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2652 		goto exit;
2653 	}
2654 
2655 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2656 	if (ret)
2657 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2658 
2659 exit:
2660 	mutex_unlock(&target->mutex);
2661 
2662 	kfd_unref_process(target);
2663 	return ret;
2664 }
2665 
2666 static int criu_process_info(struct file *filep,
2667 				struct kfd_process *p,
2668 				struct kfd_ioctl_criu_args *args)
2669 {
2670 	int ret = 0;
2671 
2672 	mutex_lock(&p->mutex);
2673 
2674 	if (!p->n_pdds) {
2675 		pr_err("No pdd for given process\n");
2676 		ret = -ENODEV;
2677 		goto err_unlock;
2678 	}
2679 
2680 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2681 	if (ret)
2682 		goto err_unlock;
2683 
2684 	p->queues_paused = true;
2685 
2686 	args->pid = task_pid_nr_ns(p->lead_thread,
2687 					task_active_pid_ns(p->lead_thread));
2688 
2689 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2690 					   &args->num_objects, &args->priv_data_size);
2691 	if (ret)
2692 		goto err_unlock;
2693 
2694 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2695 				args->num_devices, args->num_bos, args->num_objects,
2696 				args->priv_data_size);
2697 
2698 err_unlock:
2699 	if (ret) {
2700 		kfd_process_restore_queues(p);
2701 		p->queues_paused = false;
2702 	}
2703 	mutex_unlock(&p->mutex);
2704 	return ret;
2705 }
2706 
2707 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2708 {
2709 	struct kfd_ioctl_criu_args *args = data;
2710 	int ret;
2711 
2712 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2713 	switch (args->op) {
2714 	case KFD_CRIU_OP_PROCESS_INFO:
2715 		ret = criu_process_info(filep, p, args);
2716 		break;
2717 	case KFD_CRIU_OP_CHECKPOINT:
2718 		ret = criu_checkpoint(filep, p, args);
2719 		break;
2720 	case KFD_CRIU_OP_UNPAUSE:
2721 		ret = criu_unpause(filep, p, args);
2722 		break;
2723 	case KFD_CRIU_OP_RESTORE:
2724 		ret = criu_restore(filep, p, args);
2725 		break;
2726 	case KFD_CRIU_OP_RESUME:
2727 		ret = criu_resume(filep, p, args);
2728 		break;
2729 	default:
2730 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2731 		ret = -EINVAL;
2732 		break;
2733 	}
2734 
2735 	if (ret)
2736 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2737 
2738 	return ret;
2739 }
2740 
2741 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2742 			bool enable_ttmp_setup)
2743 {
2744 	int i = 0, ret = 0;
2745 
2746 	if (p->is_runtime_retry)
2747 		goto retry;
2748 
2749 	if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2750 		return -EBUSY;
2751 
2752 	for (i = 0; i < p->n_pdds; i++) {
2753 		struct kfd_process_device *pdd = p->pdds[i];
2754 
2755 		if (pdd->qpd.queue_count)
2756 			return -EEXIST;
2757 
2758 		/*
2759 		 * Setup TTMPs by default.
2760 		 * Note that this call must remain here for MES ADD QUEUE to
2761 		 * skip_process_ctx_clear unconditionally as the first call to
2762 		 * SET_SHADER_DEBUGGER clears any stale process context data
2763 		 * saved in MES.
2764 		 */
2765 		if (pdd->dev->kfd->shared_resources.enable_mes)
2766 			kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
2767 	}
2768 
2769 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2770 	p->runtime_info.r_debug = r_debug;
2771 	p->runtime_info.ttmp_setup = enable_ttmp_setup;
2772 
2773 	if (p->runtime_info.ttmp_setup) {
2774 		for (i = 0; i < p->n_pdds; i++) {
2775 			struct kfd_process_device *pdd = p->pdds[i];
2776 
2777 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2778 				amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2779 				pdd->dev->kfd2kgd->enable_debug_trap(
2780 						pdd->dev->adev,
2781 						true,
2782 						pdd->dev->vm_info.last_vmid_kfd);
2783 			} else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2784 				pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2785 						pdd->dev->adev,
2786 						false,
2787 						0);
2788 			}
2789 		}
2790 	}
2791 
2792 retry:
2793 	if (p->debug_trap_enabled) {
2794 		if (!p->is_runtime_retry) {
2795 			kfd_dbg_trap_activate(p);
2796 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2797 					p, NULL, 0, false, NULL, 0);
2798 		}
2799 
2800 		mutex_unlock(&p->mutex);
2801 		ret = down_interruptible(&p->runtime_enable_sema);
2802 		mutex_lock(&p->mutex);
2803 
2804 		p->is_runtime_retry = !!ret;
2805 	}
2806 
2807 	return ret;
2808 }
2809 
2810 static int runtime_disable(struct kfd_process *p)
2811 {
2812 	int i = 0, ret;
2813 	bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2814 
2815 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2816 	p->runtime_info.r_debug = 0;
2817 
2818 	if (p->debug_trap_enabled) {
2819 		if (was_enabled)
2820 			kfd_dbg_trap_deactivate(p, false, 0);
2821 
2822 		if (!p->is_runtime_retry)
2823 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2824 					p, NULL, 0, false, NULL, 0);
2825 
2826 		mutex_unlock(&p->mutex);
2827 		ret = down_interruptible(&p->runtime_enable_sema);
2828 		mutex_lock(&p->mutex);
2829 
2830 		p->is_runtime_retry = !!ret;
2831 		if (ret)
2832 			return ret;
2833 	}
2834 
2835 	if (was_enabled && p->runtime_info.ttmp_setup) {
2836 		for (i = 0; i < p->n_pdds; i++) {
2837 			struct kfd_process_device *pdd = p->pdds[i];
2838 
2839 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2840 				amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2841 		}
2842 	}
2843 
2844 	p->runtime_info.ttmp_setup = false;
2845 
2846 	/* disable ttmp setup */
2847 	for (i = 0; i < p->n_pdds; i++) {
2848 		struct kfd_process_device *pdd = p->pdds[i];
2849 
2850 		if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2851 			pdd->spi_dbg_override =
2852 					pdd->dev->kfd2kgd->disable_debug_trap(
2853 					pdd->dev->adev,
2854 					false,
2855 					pdd->dev->vm_info.last_vmid_kfd);
2856 
2857 			if (!pdd->dev->kfd->shared_resources.enable_mes)
2858 				debug_refresh_runlist(pdd->dev->dqm);
2859 			else
2860 				kfd_dbg_set_mes_debug_mode(pdd,
2861 							   !kfd_dbg_has_cwsr_workaround(pdd->dev));
2862 		}
2863 	}
2864 
2865 	return 0;
2866 }
2867 
2868 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2869 {
2870 	struct kfd_ioctl_runtime_enable_args *args = data;
2871 	int r;
2872 
2873 	mutex_lock(&p->mutex);
2874 
2875 	if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2876 		r = runtime_enable(p, args->r_debug,
2877 				!!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2878 	else
2879 		r = runtime_disable(p);
2880 
2881 	mutex_unlock(&p->mutex);
2882 
2883 	return r;
2884 }
2885 
2886 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2887 {
2888 	struct kfd_ioctl_dbg_trap_args *args = data;
2889 	struct task_struct *thread = NULL;
2890 	struct mm_struct *mm = NULL;
2891 	struct pid *pid = NULL;
2892 	struct kfd_process *target = NULL;
2893 	struct kfd_process_device *pdd = NULL;
2894 	int r = 0;
2895 
2896 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2897 		pr_err("Debugging does not support sched_policy %i", sched_policy);
2898 		return -EINVAL;
2899 	}
2900 
2901 	pid = find_get_pid(args->pid);
2902 	if (!pid) {
2903 		pr_debug("Cannot find pid info for %i\n", args->pid);
2904 		r = -ESRCH;
2905 		goto out;
2906 	}
2907 
2908 	thread = get_pid_task(pid, PIDTYPE_PID);
2909 	if (!thread) {
2910 		r = -ESRCH;
2911 		goto out;
2912 	}
2913 
2914 	mm = get_task_mm(thread);
2915 	if (!mm) {
2916 		r = -ESRCH;
2917 		goto out;
2918 	}
2919 
2920 	if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2921 		bool create_process;
2922 
2923 		rcu_read_lock();
2924 		create_process = thread && thread != current && ptrace_parent(thread) == current;
2925 		rcu_read_unlock();
2926 
2927 		target = create_process ? kfd_create_process(thread) :
2928 					kfd_lookup_process_by_pid(pid);
2929 	} else {
2930 		target = kfd_lookup_process_by_pid(pid);
2931 	}
2932 
2933 	if (IS_ERR_OR_NULL(target)) {
2934 		pr_debug("Cannot find process PID %i to debug\n", args->pid);
2935 		r = target ? PTR_ERR(target) : -ESRCH;
2936 		goto out;
2937 	}
2938 
2939 	/* Check if target is still PTRACED. */
2940 	rcu_read_lock();
2941 	if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2942 				&& ptrace_parent(target->lead_thread) != current) {
2943 		pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2944 		r = -EPERM;
2945 	}
2946 	rcu_read_unlock();
2947 
2948 	if (r)
2949 		goto out;
2950 
2951 	mutex_lock(&target->mutex);
2952 
2953 	if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2954 		pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2955 		r = -EINVAL;
2956 		goto unlock_out;
2957 	}
2958 
2959 	if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2960 			(args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2961 			 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2962 			 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2963 			 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2964 			 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2965 			 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2966 			 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2967 		r = -EPERM;
2968 		goto unlock_out;
2969 	}
2970 
2971 	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2972 	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2973 		int user_gpu_id = kfd_process_get_user_gpu_id(target,
2974 				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2975 					args->set_node_address_watch.gpu_id :
2976 					args->clear_node_address_watch.gpu_id);
2977 
2978 		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
2979 		if (user_gpu_id == -EINVAL || !pdd) {
2980 			r = -ENODEV;
2981 			goto unlock_out;
2982 		}
2983 	}
2984 
2985 	switch (args->op) {
2986 	case KFD_IOC_DBG_TRAP_ENABLE:
2987 		if (target != p)
2988 			target->debugger_process = p;
2989 
2990 		r = kfd_dbg_trap_enable(target,
2991 					args->enable.dbg_fd,
2992 					(void __user *)args->enable.rinfo_ptr,
2993 					&args->enable.rinfo_size);
2994 		if (!r)
2995 			target->exception_enable_mask = args->enable.exception_mask;
2996 
2997 		break;
2998 	case KFD_IOC_DBG_TRAP_DISABLE:
2999 		r = kfd_dbg_trap_disable(target);
3000 		break;
3001 	case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
3002 		r = kfd_dbg_send_exception_to_runtime(target,
3003 				args->send_runtime_event.gpu_id,
3004 				args->send_runtime_event.queue_id,
3005 				args->send_runtime_event.exception_mask);
3006 		break;
3007 	case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
3008 		kfd_dbg_set_enabled_debug_exception_mask(target,
3009 				args->set_exceptions_enabled.exception_mask);
3010 		break;
3011 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
3012 		r = kfd_dbg_trap_set_wave_launch_override(target,
3013 				args->launch_override.override_mode,
3014 				args->launch_override.enable_mask,
3015 				args->launch_override.support_request_mask,
3016 				&args->launch_override.enable_mask,
3017 				&args->launch_override.support_request_mask);
3018 		break;
3019 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
3020 		r = kfd_dbg_trap_set_wave_launch_mode(target,
3021 				args->launch_mode.launch_mode);
3022 		break;
3023 	case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
3024 		r = suspend_queues(target,
3025 				args->suspend_queues.num_queues,
3026 				args->suspend_queues.grace_period,
3027 				args->suspend_queues.exception_mask,
3028 				(uint32_t *)args->suspend_queues.queue_array_ptr);
3029 
3030 		break;
3031 	case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
3032 		r = resume_queues(target, args->resume_queues.num_queues,
3033 				(uint32_t *)args->resume_queues.queue_array_ptr);
3034 		break;
3035 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3036 		r = kfd_dbg_trap_set_dev_address_watch(pdd,
3037 				args->set_node_address_watch.address,
3038 				args->set_node_address_watch.mask,
3039 				&args->set_node_address_watch.id,
3040 				args->set_node_address_watch.mode);
3041 		break;
3042 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3043 		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3044 				args->clear_node_address_watch.id);
3045 		break;
3046 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
3047 		r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3048 		break;
3049 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3050 		r = kfd_dbg_ev_query_debug_event(target,
3051 				&args->query_debug_event.queue_id,
3052 				&args->query_debug_event.gpu_id,
3053 				args->query_debug_event.exception_mask,
3054 				&args->query_debug_event.exception_mask);
3055 		break;
3056 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3057 		r = kfd_dbg_trap_query_exception_info(target,
3058 				args->query_exception_info.source_id,
3059 				args->query_exception_info.exception_code,
3060 				args->query_exception_info.clear_exception,
3061 				(void __user *)args->query_exception_info.info_ptr,
3062 				&args->query_exception_info.info_size);
3063 		break;
3064 	case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3065 		r = pqm_get_queue_snapshot(&target->pqm,
3066 				args->queue_snapshot.exception_mask,
3067 				(void __user *)args->queue_snapshot.snapshot_buf_ptr,
3068 				&args->queue_snapshot.num_queues,
3069 				&args->queue_snapshot.entry_size);
3070 		break;
3071 	case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3072 		r = kfd_dbg_trap_device_snapshot(target,
3073 				args->device_snapshot.exception_mask,
3074 				(void __user *)args->device_snapshot.snapshot_buf_ptr,
3075 				&args->device_snapshot.num_devices,
3076 				&args->device_snapshot.entry_size);
3077 		break;
3078 	default:
3079 		pr_err("Invalid option: %i\n", args->op);
3080 		r = -EINVAL;
3081 	}
3082 
3083 unlock_out:
3084 	mutex_unlock(&target->mutex);
3085 
3086 out:
3087 	if (thread)
3088 		put_task_struct(thread);
3089 
3090 	if (mm)
3091 		mmput(mm);
3092 
3093 	if (pid)
3094 		put_pid(pid);
3095 
3096 	if (target)
3097 		kfd_unref_process(target);
3098 
3099 	return r;
3100 }
3101 
3102 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3103 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3104 			    .cmd_drv = 0, .name = #ioctl}
3105 
3106 /** Ioctl table */
3107 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3108 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3109 			kfd_ioctl_get_version, 0),
3110 
3111 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3112 			kfd_ioctl_create_queue, 0),
3113 
3114 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3115 			kfd_ioctl_destroy_queue, 0),
3116 
3117 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3118 			kfd_ioctl_set_memory_policy, 0),
3119 
3120 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3121 			kfd_ioctl_get_clock_counters, 0),
3122 
3123 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3124 			kfd_ioctl_get_process_apertures, 0),
3125 
3126 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3127 			kfd_ioctl_update_queue, 0),
3128 
3129 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3130 			kfd_ioctl_create_event, 0),
3131 
3132 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3133 			kfd_ioctl_destroy_event, 0),
3134 
3135 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3136 			kfd_ioctl_set_event, 0),
3137 
3138 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3139 			kfd_ioctl_reset_event, 0),
3140 
3141 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3142 			kfd_ioctl_wait_events, 0),
3143 
3144 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3145 			kfd_ioctl_dbg_register, 0),
3146 
3147 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3148 			kfd_ioctl_dbg_unregister, 0),
3149 
3150 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3151 			kfd_ioctl_dbg_address_watch, 0),
3152 
3153 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3154 			kfd_ioctl_dbg_wave_control, 0),
3155 
3156 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3157 			kfd_ioctl_set_scratch_backing_va, 0),
3158 
3159 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3160 			kfd_ioctl_get_tile_config, 0),
3161 
3162 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3163 			kfd_ioctl_set_trap_handler, 0),
3164 
3165 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3166 			kfd_ioctl_get_process_apertures_new, 0),
3167 
3168 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3169 			kfd_ioctl_acquire_vm, 0),
3170 
3171 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3172 			kfd_ioctl_alloc_memory_of_gpu, 0),
3173 
3174 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3175 			kfd_ioctl_free_memory_of_gpu, 0),
3176 
3177 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3178 			kfd_ioctl_map_memory_to_gpu, 0),
3179 
3180 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3181 			kfd_ioctl_unmap_memory_from_gpu, 0),
3182 
3183 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3184 			kfd_ioctl_set_cu_mask, 0),
3185 
3186 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3187 			kfd_ioctl_get_queue_wave_state, 0),
3188 
3189 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3190 				kfd_ioctl_get_dmabuf_info, 0),
3191 
3192 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3193 				kfd_ioctl_import_dmabuf, 0),
3194 
3195 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3196 			kfd_ioctl_alloc_queue_gws, 0),
3197 
3198 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3199 			kfd_ioctl_smi_events, 0),
3200 
3201 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3202 
3203 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3204 			kfd_ioctl_set_xnack_mode, 0),
3205 
3206 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3207 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3208 
3209 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3210 			kfd_ioctl_get_available_memory, 0),
3211 
3212 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3213 				kfd_ioctl_export_dmabuf, 0),
3214 
3215 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3216 			kfd_ioctl_runtime_enable, 0),
3217 
3218 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3219 			kfd_ioctl_set_debug_trap, 0),
3220 };
3221 
3222 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
3223 
3224 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3225 {
3226 	struct kfd_process *process;
3227 	amdkfd_ioctl_t *func;
3228 	const struct amdkfd_ioctl_desc *ioctl = NULL;
3229 	unsigned int nr = _IOC_NR(cmd);
3230 	char stack_kdata[128];
3231 	char *kdata = NULL;
3232 	unsigned int usize, asize;
3233 	int retcode = -EINVAL;
3234 	bool ptrace_attached = false;
3235 
3236 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3237 		goto err_i1;
3238 
3239 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3240 		u32 amdkfd_size;
3241 
3242 		ioctl = &amdkfd_ioctls[nr];
3243 
3244 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
3245 		usize = asize = _IOC_SIZE(cmd);
3246 		if (amdkfd_size > asize)
3247 			asize = amdkfd_size;
3248 
3249 		cmd = ioctl->cmd;
3250 	} else
3251 		goto err_i1;
3252 
3253 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3254 
3255 	/* Get the process struct from the filep. Only the process
3256 	 * that opened /dev/kfd can use the file descriptor. Child
3257 	 * processes need to create their own KFD device context.
3258 	 */
3259 	process = filep->private_data;
3260 
3261 	rcu_read_lock();
3262 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3263 	    ptrace_parent(process->lead_thread) == current)
3264 		ptrace_attached = true;
3265 	rcu_read_unlock();
3266 
3267 	if (process->lead_thread != current->group_leader
3268 	    && !ptrace_attached) {
3269 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3270 		retcode = -EBADF;
3271 		goto err_i1;
3272 	}
3273 
3274 	/* Do not trust userspace, use our own definition */
3275 	func = ioctl->func;
3276 
3277 	if (unlikely(!func)) {
3278 		dev_dbg(kfd_device, "no function\n");
3279 		retcode = -EINVAL;
3280 		goto err_i1;
3281 	}
3282 
3283 	/*
3284 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3285 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3286 	 * more priviledged access.
3287 	 */
3288 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3289 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
3290 						!capable(CAP_SYS_ADMIN)) {
3291 			retcode = -EACCES;
3292 			goto err_i1;
3293 		}
3294 	}
3295 
3296 	if (cmd & (IOC_IN | IOC_OUT)) {
3297 		if (asize <= sizeof(stack_kdata)) {
3298 			kdata = stack_kdata;
3299 		} else {
3300 			kdata = kmalloc(asize, GFP_KERNEL);
3301 			if (!kdata) {
3302 				retcode = -ENOMEM;
3303 				goto err_i1;
3304 			}
3305 		}
3306 		if (asize > usize)
3307 			memset(kdata + usize, 0, asize - usize);
3308 	}
3309 
3310 	if (cmd & IOC_IN) {
3311 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3312 			retcode = -EFAULT;
3313 			goto err_i1;
3314 		}
3315 	} else if (cmd & IOC_OUT) {
3316 		memset(kdata, 0, usize);
3317 	}
3318 
3319 	retcode = func(filep, process, kdata);
3320 
3321 	if (cmd & IOC_OUT)
3322 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3323 			retcode = -EFAULT;
3324 
3325 err_i1:
3326 	if (!ioctl)
3327 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3328 			  task_pid_nr(current), cmd, nr);
3329 
3330 	if (kdata != stack_kdata)
3331 		kfree(kdata);
3332 
3333 	if (retcode)
3334 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3335 				nr, arg, retcode);
3336 
3337 	return retcode;
3338 }
3339 
3340 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3341 		      struct vm_area_struct *vma)
3342 {
3343 	phys_addr_t address;
3344 
3345 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3346 		return -EINVAL;
3347 
3348 	address = dev->adev->rmmio_remap.bus_addr;
3349 
3350 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3351 				VM_DONTDUMP | VM_PFNMAP);
3352 
3353 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3354 
3355 	pr_debug("pasid 0x%x mapping mmio page\n"
3356 		 "     target user address == 0x%08llX\n"
3357 		 "     physical address    == 0x%08llX\n"
3358 		 "     vm_flags            == 0x%04lX\n"
3359 		 "     size                == 0x%04lX\n",
3360 		 process->pasid, (unsigned long long) vma->vm_start,
3361 		 address, vma->vm_flags, PAGE_SIZE);
3362 
3363 	return io_remap_pfn_range(vma,
3364 				vma->vm_start,
3365 				address >> PAGE_SHIFT,
3366 				PAGE_SIZE,
3367 				vma->vm_page_prot);
3368 }
3369 
3370 
3371 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3372 {
3373 	struct kfd_process *process;
3374 	struct kfd_node *dev = NULL;
3375 	unsigned long mmap_offset;
3376 	unsigned int gpu_id;
3377 
3378 	process = kfd_get_process(current);
3379 	if (IS_ERR(process))
3380 		return PTR_ERR(process);
3381 
3382 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3383 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3384 	if (gpu_id)
3385 		dev = kfd_device_by_id(gpu_id);
3386 
3387 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3388 	case KFD_MMAP_TYPE_DOORBELL:
3389 		if (!dev)
3390 			return -ENODEV;
3391 		return kfd_doorbell_mmap(dev, process, vma);
3392 
3393 	case KFD_MMAP_TYPE_EVENTS:
3394 		return kfd_event_mmap(process, vma);
3395 
3396 	case KFD_MMAP_TYPE_RESERVED_MEM:
3397 		if (!dev)
3398 			return -ENODEV;
3399 		return kfd_reserved_mem_mmap(dev, process, vma);
3400 	case KFD_MMAP_TYPE_MMIO:
3401 		if (!dev)
3402 			return -ENODEV;
3403 		return kfd_mmio_mmap(dev, process, vma);
3404 	}
3405 
3406 	return -EFAULT;
3407 }
3408