xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c (revision 2b1b838ea8e5437ef06a29818d16e9efdfaf0037)
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 (debug_largebar) {
1025 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1026 		return true;
1027 	}
1028 
1029 	if (dev->kfd->use_iommu_v2)
1030 		return false;
1031 
1032 	if (dev->local_mem_info.local_mem_size_private == 0 &&
1033 	    dev->local_mem_info.local_mem_size_public > 0)
1034 		return true;
1035 
1036 	if (dev->local_mem_info.local_mem_size_public == 0 &&
1037 	    dev->kfd->adev->gmc.is_app_apu) {
1038 		pr_debug("APP APU, Consider like a large bar system\n");
1039 		return true;
1040 	}
1041 
1042 	return false;
1043 }
1044 
1045 static int kfd_ioctl_get_available_memory(struct file *filep,
1046 					  struct kfd_process *p, void *data)
1047 {
1048 	struct kfd_ioctl_get_available_memory_args *args = data;
1049 	struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1050 
1051 	if (!pdd)
1052 		return -EINVAL;
1053 	args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1054 							pdd->dev->node_id);
1055 	kfd_unlock_pdd(pdd);
1056 	return 0;
1057 }
1058 
1059 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1060 					struct kfd_process *p, void *data)
1061 {
1062 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1063 	struct kfd_process_device *pdd;
1064 	void *mem;
1065 	struct kfd_node *dev;
1066 	int idr_handle;
1067 	long err;
1068 	uint64_t offset = args->mmap_offset;
1069 	uint32_t flags = args->flags;
1070 
1071 	if (args->size == 0)
1072 		return -EINVAL;
1073 
1074 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1075 	/* Flush pending deferred work to avoid racing with deferred actions
1076 	 * from previous memory map changes (e.g. munmap).
1077 	 */
1078 	svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1079 	mutex_lock(&p->svms.lock);
1080 	mmap_write_unlock(current->mm);
1081 	if (interval_tree_iter_first(&p->svms.objects,
1082 				     args->va_addr >> PAGE_SHIFT,
1083 				     (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1084 		pr_err("Address: 0x%llx already allocated by SVM\n",
1085 			args->va_addr);
1086 		mutex_unlock(&p->svms.lock);
1087 		return -EADDRINUSE;
1088 	}
1089 
1090 	/* When register user buffer check if it has been registered by svm by
1091 	 * buffer cpu virtual address.
1092 	 */
1093 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1094 	    interval_tree_iter_first(&p->svms.objects,
1095 				     args->mmap_offset >> PAGE_SHIFT,
1096 				     (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1097 		pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1098 			args->mmap_offset);
1099 		mutex_unlock(&p->svms.lock);
1100 		return -EADDRINUSE;
1101 	}
1102 
1103 	mutex_unlock(&p->svms.lock);
1104 #endif
1105 	mutex_lock(&p->mutex);
1106 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1107 	if (!pdd) {
1108 		err = -EINVAL;
1109 		goto err_pdd;
1110 	}
1111 
1112 	dev = pdd->dev;
1113 
1114 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1115 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1116 		!kfd_dev_is_large_bar(dev)) {
1117 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1118 		err = -EINVAL;
1119 		goto err_large_bar;
1120 	}
1121 
1122 	pdd = kfd_bind_process_to_device(dev, p);
1123 	if (IS_ERR(pdd)) {
1124 		err = PTR_ERR(pdd);
1125 		goto err_unlock;
1126 	}
1127 
1128 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1129 		if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1130 			err = -EINVAL;
1131 			goto err_unlock;
1132 		}
1133 		offset = kfd_get_process_doorbells(pdd);
1134 		if (!offset) {
1135 			err = -ENOMEM;
1136 			goto err_unlock;
1137 		}
1138 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1139 		if (args->size != PAGE_SIZE) {
1140 			err = -EINVAL;
1141 			goto err_unlock;
1142 		}
1143 		offset = dev->adev->rmmio_remap.bus_addr;
1144 		if (!offset) {
1145 			err = -ENOMEM;
1146 			goto err_unlock;
1147 		}
1148 	}
1149 
1150 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1151 		dev->adev, args->va_addr, args->size,
1152 		pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1153 		flags, false);
1154 
1155 	if (err)
1156 		goto err_unlock;
1157 
1158 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1159 	if (idr_handle < 0) {
1160 		err = -EFAULT;
1161 		goto err_free;
1162 	}
1163 
1164 	/* Update the VRAM usage count */
1165 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1166 		uint64_t size = args->size;
1167 
1168 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1169 			size >>= 1;
1170 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1171 	}
1172 
1173 	mutex_unlock(&p->mutex);
1174 
1175 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1176 	args->mmap_offset = offset;
1177 
1178 	/* MMIO is mapped through kfd device
1179 	 * Generate a kfd mmap offset
1180 	 */
1181 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1182 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1183 					| KFD_MMAP_GPU_ID(args->gpu_id);
1184 
1185 	return 0;
1186 
1187 err_free:
1188 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1189 					       pdd->drm_priv, NULL);
1190 err_unlock:
1191 err_pdd:
1192 err_large_bar:
1193 	mutex_unlock(&p->mutex);
1194 	return err;
1195 }
1196 
1197 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1198 					struct kfd_process *p, void *data)
1199 {
1200 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1201 	struct kfd_process_device *pdd;
1202 	void *mem;
1203 	int ret;
1204 	uint64_t size = 0;
1205 
1206 	mutex_lock(&p->mutex);
1207 	/*
1208 	 * Safeguard to prevent user space from freeing signal BO.
1209 	 * It will be freed at process termination.
1210 	 */
1211 	if (p->signal_handle && (p->signal_handle == args->handle)) {
1212 		pr_err("Free signal BO is not allowed\n");
1213 		ret = -EPERM;
1214 		goto err_unlock;
1215 	}
1216 
1217 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1218 	if (!pdd) {
1219 		pr_err("Process device data doesn't exist\n");
1220 		ret = -EINVAL;
1221 		goto err_pdd;
1222 	}
1223 
1224 	mem = kfd_process_device_translate_handle(
1225 		pdd, GET_IDR_HANDLE(args->handle));
1226 	if (!mem) {
1227 		ret = -EINVAL;
1228 		goto err_unlock;
1229 	}
1230 
1231 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1232 				(struct kgd_mem *)mem, pdd->drm_priv, &size);
1233 
1234 	/* If freeing the buffer failed, leave the handle in place for
1235 	 * clean-up during process tear-down.
1236 	 */
1237 	if (!ret)
1238 		kfd_process_device_remove_obj_handle(
1239 			pdd, GET_IDR_HANDLE(args->handle));
1240 
1241 	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1242 
1243 err_unlock:
1244 err_pdd:
1245 	mutex_unlock(&p->mutex);
1246 	return ret;
1247 }
1248 
1249 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1250 					struct kfd_process *p, void *data)
1251 {
1252 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1253 	struct kfd_process_device *pdd, *peer_pdd;
1254 	void *mem;
1255 	struct kfd_node *dev;
1256 	long err = 0;
1257 	int i;
1258 	uint32_t *devices_arr = NULL;
1259 
1260 	if (!args->n_devices) {
1261 		pr_debug("Device IDs array empty\n");
1262 		return -EINVAL;
1263 	}
1264 	if (args->n_success > args->n_devices) {
1265 		pr_debug("n_success exceeds n_devices\n");
1266 		return -EINVAL;
1267 	}
1268 
1269 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1270 				    GFP_KERNEL);
1271 	if (!devices_arr)
1272 		return -ENOMEM;
1273 
1274 	err = copy_from_user(devices_arr,
1275 			     (void __user *)args->device_ids_array_ptr,
1276 			     args->n_devices * sizeof(*devices_arr));
1277 	if (err != 0) {
1278 		err = -EFAULT;
1279 		goto copy_from_user_failed;
1280 	}
1281 
1282 	mutex_lock(&p->mutex);
1283 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1284 	if (!pdd) {
1285 		err = -EINVAL;
1286 		goto get_process_device_data_failed;
1287 	}
1288 	dev = pdd->dev;
1289 
1290 	pdd = kfd_bind_process_to_device(dev, p);
1291 	if (IS_ERR(pdd)) {
1292 		err = PTR_ERR(pdd);
1293 		goto bind_process_to_device_failed;
1294 	}
1295 
1296 	mem = kfd_process_device_translate_handle(pdd,
1297 						GET_IDR_HANDLE(args->handle));
1298 	if (!mem) {
1299 		err = -ENOMEM;
1300 		goto get_mem_obj_from_handle_failed;
1301 	}
1302 
1303 	for (i = args->n_success; i < args->n_devices; i++) {
1304 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1305 		if (!peer_pdd) {
1306 			pr_debug("Getting device by id failed for 0x%x\n",
1307 				 devices_arr[i]);
1308 			err = -EINVAL;
1309 			goto get_mem_obj_from_handle_failed;
1310 		}
1311 
1312 		peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1313 		if (IS_ERR(peer_pdd)) {
1314 			err = PTR_ERR(peer_pdd);
1315 			goto get_mem_obj_from_handle_failed;
1316 		}
1317 
1318 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1319 			peer_pdd->dev->adev, (struct kgd_mem *)mem,
1320 			peer_pdd->drm_priv);
1321 		if (err) {
1322 			struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1323 
1324 			dev_err(dev->adev->dev,
1325 			       "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1326 			       pci_domain_nr(pdev->bus),
1327 			       pdev->bus->number,
1328 			       PCI_SLOT(pdev->devfn),
1329 			       PCI_FUNC(pdev->devfn),
1330 			       ((struct kgd_mem *)mem)->domain);
1331 			goto map_memory_to_gpu_failed;
1332 		}
1333 		args->n_success = i+1;
1334 	}
1335 
1336 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1337 	if (err) {
1338 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1339 		goto sync_memory_failed;
1340 	}
1341 
1342 	mutex_unlock(&p->mutex);
1343 
1344 	/* Flush TLBs after waiting for the page table updates to complete */
1345 	for (i = 0; i < args->n_devices; i++) {
1346 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1347 		if (WARN_ON_ONCE(!peer_pdd))
1348 			continue;
1349 		kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1350 	}
1351 	kfree(devices_arr);
1352 
1353 	return err;
1354 
1355 get_process_device_data_failed:
1356 bind_process_to_device_failed:
1357 get_mem_obj_from_handle_failed:
1358 map_memory_to_gpu_failed:
1359 sync_memory_failed:
1360 	mutex_unlock(&p->mutex);
1361 copy_from_user_failed:
1362 	kfree(devices_arr);
1363 
1364 	return err;
1365 }
1366 
1367 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1368 					struct kfd_process *p, void *data)
1369 {
1370 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1371 	struct kfd_process_device *pdd, *peer_pdd;
1372 	void *mem;
1373 	long err = 0;
1374 	uint32_t *devices_arr = NULL, i;
1375 	bool flush_tlb;
1376 
1377 	if (!args->n_devices) {
1378 		pr_debug("Device IDs array empty\n");
1379 		return -EINVAL;
1380 	}
1381 	if (args->n_success > args->n_devices) {
1382 		pr_debug("n_success exceeds n_devices\n");
1383 		return -EINVAL;
1384 	}
1385 
1386 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1387 				    GFP_KERNEL);
1388 	if (!devices_arr)
1389 		return -ENOMEM;
1390 
1391 	err = copy_from_user(devices_arr,
1392 			     (void __user *)args->device_ids_array_ptr,
1393 			     args->n_devices * sizeof(*devices_arr));
1394 	if (err != 0) {
1395 		err = -EFAULT;
1396 		goto copy_from_user_failed;
1397 	}
1398 
1399 	mutex_lock(&p->mutex);
1400 	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1401 	if (!pdd) {
1402 		err = -EINVAL;
1403 		goto bind_process_to_device_failed;
1404 	}
1405 
1406 	mem = kfd_process_device_translate_handle(pdd,
1407 						GET_IDR_HANDLE(args->handle));
1408 	if (!mem) {
1409 		err = -ENOMEM;
1410 		goto get_mem_obj_from_handle_failed;
1411 	}
1412 
1413 	for (i = args->n_success; i < args->n_devices; i++) {
1414 		peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1415 		if (!peer_pdd) {
1416 			err = -EINVAL;
1417 			goto get_mem_obj_from_handle_failed;
1418 		}
1419 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1420 			peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1421 		if (err) {
1422 			pr_err("Failed to unmap from gpu %d/%d\n",
1423 			       i, args->n_devices);
1424 			goto unmap_memory_from_gpu_failed;
1425 		}
1426 		args->n_success = i+1;
1427 	}
1428 
1429 	flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1430 	if (flush_tlb) {
1431 		err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1432 				(struct kgd_mem *) mem, true);
1433 		if (err) {
1434 			pr_debug("Sync memory failed, wait interrupted by user signal\n");
1435 			goto sync_memory_failed;
1436 		}
1437 	}
1438 	mutex_unlock(&p->mutex);
1439 
1440 	if (flush_tlb) {
1441 		/* Flush TLBs after waiting for the page table updates to complete */
1442 		for (i = 0; i < args->n_devices; i++) {
1443 			peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1444 			if (WARN_ON_ONCE(!peer_pdd))
1445 				continue;
1446 			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1447 		}
1448 	}
1449 	kfree(devices_arr);
1450 
1451 	return 0;
1452 
1453 bind_process_to_device_failed:
1454 get_mem_obj_from_handle_failed:
1455 unmap_memory_from_gpu_failed:
1456 sync_memory_failed:
1457 	mutex_unlock(&p->mutex);
1458 copy_from_user_failed:
1459 	kfree(devices_arr);
1460 	return err;
1461 }
1462 
1463 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1464 		struct kfd_process *p, void *data)
1465 {
1466 	int retval;
1467 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1468 	struct queue *q;
1469 	struct kfd_node *dev;
1470 
1471 	mutex_lock(&p->mutex);
1472 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1473 
1474 	if (q) {
1475 		dev = q->device;
1476 	} else {
1477 		retval = -EINVAL;
1478 		goto out_unlock;
1479 	}
1480 
1481 	if (!dev->gws) {
1482 		retval = -ENODEV;
1483 		goto out_unlock;
1484 	}
1485 
1486 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1487 		retval = -ENODEV;
1488 		goto out_unlock;
1489 	}
1490 
1491 	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1492 				      kfd_dbg_has_cwsr_workaround(dev))) {
1493 		retval = -EBUSY;
1494 		goto out_unlock;
1495 	}
1496 
1497 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1498 	mutex_unlock(&p->mutex);
1499 
1500 	args->first_gws = 0;
1501 	return retval;
1502 
1503 out_unlock:
1504 	mutex_unlock(&p->mutex);
1505 	return retval;
1506 }
1507 
1508 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1509 		struct kfd_process *p, void *data)
1510 {
1511 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1512 	struct kfd_node *dev = NULL;
1513 	struct amdgpu_device *dmabuf_adev;
1514 	void *metadata_buffer = NULL;
1515 	uint32_t flags;
1516 	int8_t xcp_id;
1517 	unsigned int i;
1518 	int r;
1519 
1520 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1521 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1522 		if (dev)
1523 			break;
1524 	if (!dev)
1525 		return -EINVAL;
1526 
1527 	if (args->metadata_ptr) {
1528 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1529 		if (!metadata_buffer)
1530 			return -ENOMEM;
1531 	}
1532 
1533 	/* Get dmabuf info from KGD */
1534 	r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1535 					  &dmabuf_adev, &args->size,
1536 					  metadata_buffer, args->metadata_size,
1537 					  &args->metadata_size, &flags, &xcp_id);
1538 	if (r)
1539 		goto exit;
1540 
1541 	if (xcp_id >= 0)
1542 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1543 	else
1544 		args->gpu_id = dmabuf_adev->kfd.dev->nodes[0]->id;
1545 	args->flags = flags;
1546 
1547 	/* Copy metadata buffer to user mode */
1548 	if (metadata_buffer) {
1549 		r = copy_to_user((void __user *)args->metadata_ptr,
1550 				 metadata_buffer, args->metadata_size);
1551 		if (r != 0)
1552 			r = -EFAULT;
1553 	}
1554 
1555 exit:
1556 	kfree(metadata_buffer);
1557 
1558 	return r;
1559 }
1560 
1561 static int kfd_ioctl_import_dmabuf(struct file *filep,
1562 				   struct kfd_process *p, void *data)
1563 {
1564 	struct kfd_ioctl_import_dmabuf_args *args = data;
1565 	struct kfd_process_device *pdd;
1566 	struct dma_buf *dmabuf;
1567 	int idr_handle;
1568 	uint64_t size;
1569 	void *mem;
1570 	int r;
1571 
1572 	dmabuf = dma_buf_get(args->dmabuf_fd);
1573 	if (IS_ERR(dmabuf))
1574 		return PTR_ERR(dmabuf);
1575 
1576 	mutex_lock(&p->mutex);
1577 	pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1578 	if (!pdd) {
1579 		r = -EINVAL;
1580 		goto err_unlock;
1581 	}
1582 
1583 	pdd = kfd_bind_process_to_device(pdd->dev, p);
1584 	if (IS_ERR(pdd)) {
1585 		r = PTR_ERR(pdd);
1586 		goto err_unlock;
1587 	}
1588 
1589 	r = amdgpu_amdkfd_gpuvm_import_dmabuf(pdd->dev->adev, dmabuf,
1590 					      args->va_addr, pdd->drm_priv,
1591 					      (struct kgd_mem **)&mem, &size,
1592 					      NULL);
1593 	if (r)
1594 		goto err_unlock;
1595 
1596 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1597 	if (idr_handle < 0) {
1598 		r = -EFAULT;
1599 		goto err_free;
1600 	}
1601 
1602 	mutex_unlock(&p->mutex);
1603 	dma_buf_put(dmabuf);
1604 
1605 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1606 
1607 	return 0;
1608 
1609 err_free:
1610 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1611 					       pdd->drm_priv, NULL);
1612 err_unlock:
1613 	mutex_unlock(&p->mutex);
1614 	dma_buf_put(dmabuf);
1615 	return r;
1616 }
1617 
1618 static int kfd_ioctl_export_dmabuf(struct file *filep,
1619 				   struct kfd_process *p, void *data)
1620 {
1621 	struct kfd_ioctl_export_dmabuf_args *args = data;
1622 	struct kfd_process_device *pdd;
1623 	struct dma_buf *dmabuf;
1624 	struct kfd_node *dev;
1625 	void *mem;
1626 	int ret = 0;
1627 
1628 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1629 	if (!dev)
1630 		return -EINVAL;
1631 
1632 	mutex_lock(&p->mutex);
1633 
1634 	pdd = kfd_get_process_device_data(dev, p);
1635 	if (!pdd) {
1636 		ret = -EINVAL;
1637 		goto err_unlock;
1638 	}
1639 
1640 	mem = kfd_process_device_translate_handle(pdd,
1641 						GET_IDR_HANDLE(args->handle));
1642 	if (!mem) {
1643 		ret = -EINVAL;
1644 		goto err_unlock;
1645 	}
1646 
1647 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1648 	mutex_unlock(&p->mutex);
1649 	if (ret)
1650 		goto err_out;
1651 
1652 	ret = dma_buf_fd(dmabuf, args->flags);
1653 	if (ret < 0) {
1654 		dma_buf_put(dmabuf);
1655 		goto err_out;
1656 	}
1657 	/* dma_buf_fd assigns the reference count to the fd, no need to
1658 	 * put the reference here.
1659 	 */
1660 	args->dmabuf_fd = ret;
1661 
1662 	return 0;
1663 
1664 err_unlock:
1665 	mutex_unlock(&p->mutex);
1666 err_out:
1667 	return ret;
1668 }
1669 
1670 /* Handle requests for watching SMI events */
1671 static int kfd_ioctl_smi_events(struct file *filep,
1672 				struct kfd_process *p, void *data)
1673 {
1674 	struct kfd_ioctl_smi_events_args *args = data;
1675 	struct kfd_process_device *pdd;
1676 
1677 	mutex_lock(&p->mutex);
1678 
1679 	pdd = kfd_process_device_data_by_id(p, args->gpuid);
1680 	mutex_unlock(&p->mutex);
1681 	if (!pdd)
1682 		return -EINVAL;
1683 
1684 	return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1685 }
1686 
1687 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1688 
1689 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1690 				    struct kfd_process *p, void *data)
1691 {
1692 	struct kfd_ioctl_set_xnack_mode_args *args = data;
1693 	int r = 0;
1694 
1695 	mutex_lock(&p->mutex);
1696 	if (args->xnack_enabled >= 0) {
1697 		if (!list_empty(&p->pqm.queues)) {
1698 			pr_debug("Process has user queues running\n");
1699 			r = -EBUSY;
1700 			goto out_unlock;
1701 		}
1702 
1703 		if (p->xnack_enabled == args->xnack_enabled)
1704 			goto out_unlock;
1705 
1706 		if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1707 			r = -EPERM;
1708 			goto out_unlock;
1709 		}
1710 
1711 		r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1712 	} else {
1713 		args->xnack_enabled = p->xnack_enabled;
1714 	}
1715 
1716 out_unlock:
1717 	mutex_unlock(&p->mutex);
1718 
1719 	return r;
1720 }
1721 
1722 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1723 {
1724 	struct kfd_ioctl_svm_args *args = data;
1725 	int r = 0;
1726 
1727 	pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1728 		 args->start_addr, args->size, args->op, args->nattr);
1729 
1730 	if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1731 		return -EINVAL;
1732 	if (!args->start_addr || !args->size)
1733 		return -EINVAL;
1734 
1735 	r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1736 		      args->attrs);
1737 
1738 	return r;
1739 }
1740 #else
1741 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1742 				    struct kfd_process *p, void *data)
1743 {
1744 	return -EPERM;
1745 }
1746 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1747 {
1748 	return -EPERM;
1749 }
1750 #endif
1751 
1752 static int criu_checkpoint_process(struct kfd_process *p,
1753 			     uint8_t __user *user_priv_data,
1754 			     uint64_t *priv_offset)
1755 {
1756 	struct kfd_criu_process_priv_data process_priv;
1757 	int ret;
1758 
1759 	memset(&process_priv, 0, sizeof(process_priv));
1760 
1761 	process_priv.version = KFD_CRIU_PRIV_VERSION;
1762 	/* For CR, we don't consider negative xnack mode which is used for
1763 	 * querying without changing it, here 0 simply means disabled and 1
1764 	 * means enabled so retry for finding a valid PTE.
1765 	 */
1766 	process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1767 
1768 	ret = copy_to_user(user_priv_data + *priv_offset,
1769 				&process_priv, sizeof(process_priv));
1770 
1771 	if (ret) {
1772 		pr_err("Failed to copy process information to user\n");
1773 		ret = -EFAULT;
1774 	}
1775 
1776 	*priv_offset += sizeof(process_priv);
1777 	return ret;
1778 }
1779 
1780 static int criu_checkpoint_devices(struct kfd_process *p,
1781 			     uint32_t num_devices,
1782 			     uint8_t __user *user_addr,
1783 			     uint8_t __user *user_priv_data,
1784 			     uint64_t *priv_offset)
1785 {
1786 	struct kfd_criu_device_priv_data *device_priv = NULL;
1787 	struct kfd_criu_device_bucket *device_buckets = NULL;
1788 	int ret = 0, i;
1789 
1790 	device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1791 	if (!device_buckets) {
1792 		ret = -ENOMEM;
1793 		goto exit;
1794 	}
1795 
1796 	device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1797 	if (!device_priv) {
1798 		ret = -ENOMEM;
1799 		goto exit;
1800 	}
1801 
1802 	for (i = 0; i < num_devices; i++) {
1803 		struct kfd_process_device *pdd = p->pdds[i];
1804 
1805 		device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1806 		device_buckets[i].actual_gpu_id = pdd->dev->id;
1807 
1808 		/*
1809 		 * priv_data does not contain useful information for now and is reserved for
1810 		 * future use, so we do not set its contents.
1811 		 */
1812 	}
1813 
1814 	ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1815 	if (ret) {
1816 		pr_err("Failed to copy device information to user\n");
1817 		ret = -EFAULT;
1818 		goto exit;
1819 	}
1820 
1821 	ret = copy_to_user(user_priv_data + *priv_offset,
1822 			   device_priv,
1823 			   num_devices * sizeof(*device_priv));
1824 	if (ret) {
1825 		pr_err("Failed to copy device information to user\n");
1826 		ret = -EFAULT;
1827 	}
1828 	*priv_offset += num_devices * sizeof(*device_priv);
1829 
1830 exit:
1831 	kvfree(device_buckets);
1832 	kvfree(device_priv);
1833 	return ret;
1834 }
1835 
1836 static uint32_t get_process_num_bos(struct kfd_process *p)
1837 {
1838 	uint32_t num_of_bos = 0;
1839 	int i;
1840 
1841 	/* Run over all PDDs of the process */
1842 	for (i = 0; i < p->n_pdds; i++) {
1843 		struct kfd_process_device *pdd = p->pdds[i];
1844 		void *mem;
1845 		int id;
1846 
1847 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1848 			struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1849 
1850 			if (!kgd_mem->va || kgd_mem->va > pdd->gpuvm_base)
1851 				num_of_bos++;
1852 		}
1853 	}
1854 	return num_of_bos;
1855 }
1856 
1857 static int criu_get_prime_handle(struct kgd_mem *mem, int flags,
1858 				      u32 *shared_fd)
1859 {
1860 	struct dma_buf *dmabuf;
1861 	int ret;
1862 
1863 	ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1864 	if (ret) {
1865 		pr_err("dmabuf export failed for the BO\n");
1866 		return ret;
1867 	}
1868 
1869 	ret = dma_buf_fd(dmabuf, flags);
1870 	if (ret < 0) {
1871 		pr_err("dmabuf create fd failed, ret:%d\n", ret);
1872 		goto out_free_dmabuf;
1873 	}
1874 
1875 	*shared_fd = ret;
1876 	return 0;
1877 
1878 out_free_dmabuf:
1879 	dma_buf_put(dmabuf);
1880 	return ret;
1881 }
1882 
1883 static int criu_checkpoint_bos(struct kfd_process *p,
1884 			       uint32_t num_bos,
1885 			       uint8_t __user *user_bos,
1886 			       uint8_t __user *user_priv_data,
1887 			       uint64_t *priv_offset)
1888 {
1889 	struct kfd_criu_bo_bucket *bo_buckets;
1890 	struct kfd_criu_bo_priv_data *bo_privs;
1891 	int ret = 0, pdd_index, bo_index = 0, id;
1892 	void *mem;
1893 
1894 	bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1895 	if (!bo_buckets)
1896 		return -ENOMEM;
1897 
1898 	bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1899 	if (!bo_privs) {
1900 		ret = -ENOMEM;
1901 		goto exit;
1902 	}
1903 
1904 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1905 		struct kfd_process_device *pdd = p->pdds[pdd_index];
1906 		struct amdgpu_bo *dumper_bo;
1907 		struct kgd_mem *kgd_mem;
1908 
1909 		idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1910 			struct kfd_criu_bo_bucket *bo_bucket;
1911 			struct kfd_criu_bo_priv_data *bo_priv;
1912 			int i, dev_idx = 0;
1913 
1914 			if (!mem) {
1915 				ret = -ENOMEM;
1916 				goto exit;
1917 			}
1918 
1919 			kgd_mem = (struct kgd_mem *)mem;
1920 			dumper_bo = kgd_mem->bo;
1921 
1922 			/* Skip checkpointing BOs that are used for Trap handler
1923 			 * code and state. Currently, these BOs have a VA that
1924 			 * is less GPUVM Base
1925 			 */
1926 			if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base)
1927 				continue;
1928 
1929 			bo_bucket = &bo_buckets[bo_index];
1930 			bo_priv = &bo_privs[bo_index];
1931 
1932 			bo_bucket->gpu_id = pdd->user_gpu_id;
1933 			bo_bucket->addr = (uint64_t)kgd_mem->va;
1934 			bo_bucket->size = amdgpu_bo_size(dumper_bo);
1935 			bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1936 			bo_priv->idr_handle = id;
1937 
1938 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1939 				ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1940 								&bo_priv->user_addr);
1941 				if (ret) {
1942 					pr_err("Failed to obtain user address for user-pointer bo\n");
1943 					goto exit;
1944 				}
1945 			}
1946 			if (bo_bucket->alloc_flags
1947 			    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1948 				ret = criu_get_prime_handle(kgd_mem,
1949 						bo_bucket->alloc_flags &
1950 						KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1951 						&bo_bucket->dmabuf_fd);
1952 				if (ret)
1953 					goto exit;
1954 			} else {
1955 				bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1956 			}
1957 
1958 			if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1959 				bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1960 					KFD_MMAP_GPU_ID(pdd->dev->id);
1961 			else if (bo_bucket->alloc_flags &
1962 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1963 				bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1964 					KFD_MMAP_GPU_ID(pdd->dev->id);
1965 			else
1966 				bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1967 
1968 			for (i = 0; i < p->n_pdds; i++) {
1969 				if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->dev->adev, kgd_mem))
1970 					bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1971 			}
1972 
1973 			pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1974 					"gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1975 					bo_bucket->size,
1976 					bo_bucket->addr,
1977 					bo_bucket->offset,
1978 					bo_bucket->gpu_id,
1979 					bo_bucket->alloc_flags,
1980 					bo_priv->idr_handle);
1981 			bo_index++;
1982 		}
1983 	}
1984 
1985 	ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1986 	if (ret) {
1987 		pr_err("Failed to copy BO information to user\n");
1988 		ret = -EFAULT;
1989 		goto exit;
1990 	}
1991 
1992 	ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1993 	if (ret) {
1994 		pr_err("Failed to copy BO priv information to user\n");
1995 		ret = -EFAULT;
1996 		goto exit;
1997 	}
1998 
1999 	*priv_offset += num_bos * sizeof(*bo_privs);
2000 
2001 exit:
2002 	while (ret && bo_index--) {
2003 		if (bo_buckets[bo_index].alloc_flags
2004 		    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2005 			close_fd(bo_buckets[bo_index].dmabuf_fd);
2006 	}
2007 
2008 	kvfree(bo_buckets);
2009 	kvfree(bo_privs);
2010 	return ret;
2011 }
2012 
2013 static int criu_get_process_object_info(struct kfd_process *p,
2014 					uint32_t *num_devices,
2015 					uint32_t *num_bos,
2016 					uint32_t *num_objects,
2017 					uint64_t *objs_priv_size)
2018 {
2019 	uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2020 	uint32_t num_queues, num_events, num_svm_ranges;
2021 	int ret;
2022 
2023 	*num_devices = p->n_pdds;
2024 	*num_bos = get_process_num_bos(p);
2025 
2026 	ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2027 	if (ret)
2028 		return ret;
2029 
2030 	num_events = kfd_get_num_events(p);
2031 
2032 	ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2033 	if (ret)
2034 		return ret;
2035 
2036 	*num_objects = num_queues + num_events + num_svm_ranges;
2037 
2038 	if (objs_priv_size) {
2039 		priv_size = sizeof(struct kfd_criu_process_priv_data);
2040 		priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2041 		priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2042 		priv_size += queues_priv_data_size;
2043 		priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2044 		priv_size += svm_priv_data_size;
2045 		*objs_priv_size = priv_size;
2046 	}
2047 	return 0;
2048 }
2049 
2050 static int criu_checkpoint(struct file *filep,
2051 			   struct kfd_process *p,
2052 			   struct kfd_ioctl_criu_args *args)
2053 {
2054 	int ret;
2055 	uint32_t num_devices, num_bos, num_objects;
2056 	uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2057 
2058 	if (!args->devices || !args->bos || !args->priv_data)
2059 		return -EINVAL;
2060 
2061 	mutex_lock(&p->mutex);
2062 
2063 	if (!p->n_pdds) {
2064 		pr_err("No pdd for given process\n");
2065 		ret = -ENODEV;
2066 		goto exit_unlock;
2067 	}
2068 
2069 	/* Confirm all process queues are evicted */
2070 	if (!p->queues_paused) {
2071 		pr_err("Cannot dump process when queues are not in evicted state\n");
2072 		/* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2073 		ret = -EINVAL;
2074 		goto exit_unlock;
2075 	}
2076 
2077 	ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2078 	if (ret)
2079 		goto exit_unlock;
2080 
2081 	if (num_devices != args->num_devices ||
2082 	    num_bos != args->num_bos ||
2083 	    num_objects != args->num_objects ||
2084 	    priv_size != args->priv_data_size) {
2085 
2086 		ret = -EINVAL;
2087 		goto exit_unlock;
2088 	}
2089 
2090 	/* each function will store private data inside priv_data and adjust priv_offset */
2091 	ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2092 	if (ret)
2093 		goto exit_unlock;
2094 
2095 	ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2096 				(uint8_t __user *)args->priv_data, &priv_offset);
2097 	if (ret)
2098 		goto exit_unlock;
2099 
2100 	/* Leave room for BOs in the private data. They need to be restored
2101 	 * before events, but we checkpoint them last to simplify the error
2102 	 * handling.
2103 	 */
2104 	bo_priv_offset = priv_offset;
2105 	priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2106 
2107 	if (num_objects) {
2108 		ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2109 						 &priv_offset);
2110 		if (ret)
2111 			goto exit_unlock;
2112 
2113 		ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2114 						 &priv_offset);
2115 		if (ret)
2116 			goto exit_unlock;
2117 
2118 		ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2119 		if (ret)
2120 			goto exit_unlock;
2121 	}
2122 
2123 	/* This must be the last thing in this function that can fail.
2124 	 * Otherwise we leak dmabuf file descriptors.
2125 	 */
2126 	ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2127 			   (uint8_t __user *)args->priv_data, &bo_priv_offset);
2128 
2129 exit_unlock:
2130 	mutex_unlock(&p->mutex);
2131 	if (ret)
2132 		pr_err("Failed to dump CRIU ret:%d\n", ret);
2133 	else
2134 		pr_debug("CRIU dump ret:%d\n", ret);
2135 
2136 	return ret;
2137 }
2138 
2139 static int criu_restore_process(struct kfd_process *p,
2140 				struct kfd_ioctl_criu_args *args,
2141 				uint64_t *priv_offset,
2142 				uint64_t max_priv_data_size)
2143 {
2144 	int ret = 0;
2145 	struct kfd_criu_process_priv_data process_priv;
2146 
2147 	if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2148 		return -EINVAL;
2149 
2150 	ret = copy_from_user(&process_priv,
2151 				(void __user *)(args->priv_data + *priv_offset),
2152 				sizeof(process_priv));
2153 	if (ret) {
2154 		pr_err("Failed to copy process private information from user\n");
2155 		ret = -EFAULT;
2156 		goto exit;
2157 	}
2158 	*priv_offset += sizeof(process_priv);
2159 
2160 	if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2161 		pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2162 			process_priv.version, KFD_CRIU_PRIV_VERSION);
2163 		return -EINVAL;
2164 	}
2165 
2166 	pr_debug("Setting XNACK mode\n");
2167 	if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2168 		pr_err("xnack mode cannot be set\n");
2169 		ret = -EPERM;
2170 		goto exit;
2171 	} else {
2172 		pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2173 		p->xnack_enabled = process_priv.xnack_mode;
2174 	}
2175 
2176 exit:
2177 	return ret;
2178 }
2179 
2180 static int criu_restore_devices(struct kfd_process *p,
2181 				struct kfd_ioctl_criu_args *args,
2182 				uint64_t *priv_offset,
2183 				uint64_t max_priv_data_size)
2184 {
2185 	struct kfd_criu_device_bucket *device_buckets;
2186 	struct kfd_criu_device_priv_data *device_privs;
2187 	int ret = 0;
2188 	uint32_t i;
2189 
2190 	if (args->num_devices != p->n_pdds)
2191 		return -EINVAL;
2192 
2193 	if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2194 		return -EINVAL;
2195 
2196 	device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2197 	if (!device_buckets)
2198 		return -ENOMEM;
2199 
2200 	ret = copy_from_user(device_buckets, (void __user *)args->devices,
2201 				args->num_devices * sizeof(*device_buckets));
2202 	if (ret) {
2203 		pr_err("Failed to copy devices buckets from user\n");
2204 		ret = -EFAULT;
2205 		goto exit;
2206 	}
2207 
2208 	for (i = 0; i < args->num_devices; i++) {
2209 		struct kfd_node *dev;
2210 		struct kfd_process_device *pdd;
2211 		struct file *drm_file;
2212 
2213 		/* device private data is not currently used */
2214 
2215 		if (!device_buckets[i].user_gpu_id) {
2216 			pr_err("Invalid user gpu_id\n");
2217 			ret = -EINVAL;
2218 			goto exit;
2219 		}
2220 
2221 		dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2222 		if (!dev) {
2223 			pr_err("Failed to find device with gpu_id = %x\n",
2224 				device_buckets[i].actual_gpu_id);
2225 			ret = -EINVAL;
2226 			goto exit;
2227 		}
2228 
2229 		pdd = kfd_get_process_device_data(dev, p);
2230 		if (!pdd) {
2231 			pr_err("Failed to get pdd for gpu_id = %x\n",
2232 					device_buckets[i].actual_gpu_id);
2233 			ret = -EINVAL;
2234 			goto exit;
2235 		}
2236 		pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2237 
2238 		drm_file = fget(device_buckets[i].drm_fd);
2239 		if (!drm_file) {
2240 			pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2241 				device_buckets[i].drm_fd);
2242 			ret = -EINVAL;
2243 			goto exit;
2244 		}
2245 
2246 		if (pdd->drm_file) {
2247 			ret = -EINVAL;
2248 			goto exit;
2249 		}
2250 
2251 		/* create the vm using render nodes for kfd pdd */
2252 		if (kfd_process_device_init_vm(pdd, drm_file)) {
2253 			pr_err("could not init vm for given pdd\n");
2254 			/* On success, the PDD keeps the drm_file reference */
2255 			fput(drm_file);
2256 			ret = -EINVAL;
2257 			goto exit;
2258 		}
2259 		/*
2260 		 * pdd now already has the vm bound to render node so below api won't create a new
2261 		 * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2262 		 * for iommu v2 binding  and runtime pm.
2263 		 */
2264 		pdd = kfd_bind_process_to_device(dev, p);
2265 		if (IS_ERR(pdd)) {
2266 			ret = PTR_ERR(pdd);
2267 			goto exit;
2268 		}
2269 
2270 		if (!pdd->qpd.proc_doorbells) {
2271 			ret = kfd_alloc_process_doorbells(dev->kfd, pdd);
2272 			if (ret)
2273 				goto exit;
2274 		}
2275 	}
2276 
2277 	/*
2278 	 * We are not copying device private data from user as we are not using the data for now,
2279 	 * but we still adjust for its private data.
2280 	 */
2281 	*priv_offset += args->num_devices * sizeof(*device_privs);
2282 
2283 exit:
2284 	kfree(device_buckets);
2285 	return ret;
2286 }
2287 
2288 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2289 				      struct kfd_criu_bo_bucket *bo_bucket,
2290 				      struct kfd_criu_bo_priv_data *bo_priv,
2291 				      struct kgd_mem **kgd_mem)
2292 {
2293 	int idr_handle;
2294 	int ret;
2295 	const bool criu_resume = true;
2296 	u64 offset;
2297 
2298 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2299 		if (bo_bucket->size !=
2300 				kfd_doorbell_process_slice(pdd->dev->kfd))
2301 			return -EINVAL;
2302 
2303 		offset = kfd_get_process_doorbells(pdd);
2304 		if (!offset)
2305 			return -ENOMEM;
2306 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2307 		/* MMIO BOs need remapped bus address */
2308 		if (bo_bucket->size != PAGE_SIZE) {
2309 			pr_err("Invalid page size\n");
2310 			return -EINVAL;
2311 		}
2312 		offset = pdd->dev->adev->rmmio_remap.bus_addr;
2313 		if (!offset) {
2314 			pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2315 			return -ENOMEM;
2316 		}
2317 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2318 		offset = bo_priv->user_addr;
2319 	}
2320 	/* Create the BO */
2321 	ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2322 						      bo_bucket->size, pdd->drm_priv, kgd_mem,
2323 						      &offset, bo_bucket->alloc_flags, criu_resume);
2324 	if (ret) {
2325 		pr_err("Could not create the BO\n");
2326 		return ret;
2327 	}
2328 	pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2329 		 bo_bucket->size, bo_bucket->addr, offset);
2330 
2331 	/* Restore previous IDR handle */
2332 	pr_debug("Restoring old IDR handle for the BO");
2333 	idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2334 			       bo_priv->idr_handle + 1, GFP_KERNEL);
2335 
2336 	if (idr_handle < 0) {
2337 		pr_err("Could not allocate idr\n");
2338 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2339 						       NULL);
2340 		return -ENOMEM;
2341 	}
2342 
2343 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2344 		bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2345 	if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2346 		bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2347 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2348 		bo_bucket->restored_offset = offset;
2349 	} else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2350 		bo_bucket->restored_offset = offset;
2351 		/* Update the VRAM usage count */
2352 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2353 	}
2354 	return 0;
2355 }
2356 
2357 static int criu_restore_bo(struct kfd_process *p,
2358 			   struct kfd_criu_bo_bucket *bo_bucket,
2359 			   struct kfd_criu_bo_priv_data *bo_priv)
2360 {
2361 	struct kfd_process_device *pdd;
2362 	struct kgd_mem *kgd_mem;
2363 	int ret;
2364 	int j;
2365 
2366 	pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2367 		 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2368 		 bo_priv->idr_handle);
2369 
2370 	pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2371 	if (!pdd) {
2372 		pr_err("Failed to get pdd\n");
2373 		return -ENODEV;
2374 	}
2375 
2376 	ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2377 	if (ret)
2378 		return ret;
2379 
2380 	/* now map these BOs to GPU/s */
2381 	for (j = 0; j < p->n_pdds; j++) {
2382 		struct kfd_node *peer;
2383 		struct kfd_process_device *peer_pdd;
2384 
2385 		if (!bo_priv->mapped_gpuids[j])
2386 			break;
2387 
2388 		peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2389 		if (!peer_pdd)
2390 			return -EINVAL;
2391 
2392 		peer = peer_pdd->dev;
2393 
2394 		peer_pdd = kfd_bind_process_to_device(peer, p);
2395 		if (IS_ERR(peer_pdd))
2396 			return PTR_ERR(peer_pdd);
2397 
2398 		ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2399 							    peer_pdd->drm_priv);
2400 		if (ret) {
2401 			pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2402 			return ret;
2403 		}
2404 	}
2405 
2406 	pr_debug("map memory was successful for the BO\n");
2407 	/* create the dmabuf object and export the bo */
2408 	if (bo_bucket->alloc_flags
2409 	    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2410 		ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2411 					    &bo_bucket->dmabuf_fd);
2412 		if (ret)
2413 			return ret;
2414 	} else {
2415 		bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2416 	}
2417 
2418 	return 0;
2419 }
2420 
2421 static int criu_restore_bos(struct kfd_process *p,
2422 			    struct kfd_ioctl_criu_args *args,
2423 			    uint64_t *priv_offset,
2424 			    uint64_t max_priv_data_size)
2425 {
2426 	struct kfd_criu_bo_bucket *bo_buckets = NULL;
2427 	struct kfd_criu_bo_priv_data *bo_privs = NULL;
2428 	int ret = 0;
2429 	uint32_t i = 0;
2430 
2431 	if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2432 		return -EINVAL;
2433 
2434 	/* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2435 	amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2436 
2437 	bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2438 	if (!bo_buckets)
2439 		return -ENOMEM;
2440 
2441 	ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2442 			     args->num_bos * sizeof(*bo_buckets));
2443 	if (ret) {
2444 		pr_err("Failed to copy BOs information from user\n");
2445 		ret = -EFAULT;
2446 		goto exit;
2447 	}
2448 
2449 	bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2450 	if (!bo_privs) {
2451 		ret = -ENOMEM;
2452 		goto exit;
2453 	}
2454 
2455 	ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2456 			     args->num_bos * sizeof(*bo_privs));
2457 	if (ret) {
2458 		pr_err("Failed to copy BOs information from user\n");
2459 		ret = -EFAULT;
2460 		goto exit;
2461 	}
2462 	*priv_offset += args->num_bos * sizeof(*bo_privs);
2463 
2464 	/* Create and map new BOs */
2465 	for (; i < args->num_bos; i++) {
2466 		ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2467 		if (ret) {
2468 			pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2469 			goto exit;
2470 		}
2471 	} /* done */
2472 
2473 	/* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2474 	ret = copy_to_user((void __user *)args->bos,
2475 				bo_buckets,
2476 				(args->num_bos * sizeof(*bo_buckets)));
2477 	if (ret)
2478 		ret = -EFAULT;
2479 
2480 exit:
2481 	while (ret && i--) {
2482 		if (bo_buckets[i].alloc_flags
2483 		   & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2484 			close_fd(bo_buckets[i].dmabuf_fd);
2485 	}
2486 	kvfree(bo_buckets);
2487 	kvfree(bo_privs);
2488 	return ret;
2489 }
2490 
2491 static int criu_restore_objects(struct file *filep,
2492 				struct kfd_process *p,
2493 				struct kfd_ioctl_criu_args *args,
2494 				uint64_t *priv_offset,
2495 				uint64_t max_priv_data_size)
2496 {
2497 	int ret = 0;
2498 	uint32_t i;
2499 
2500 	BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2501 	BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2502 	BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2503 
2504 	for (i = 0; i < args->num_objects; i++) {
2505 		uint32_t object_type;
2506 
2507 		if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2508 			pr_err("Invalid private data size\n");
2509 			return -EINVAL;
2510 		}
2511 
2512 		ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2513 		if (ret) {
2514 			pr_err("Failed to copy private information from user\n");
2515 			goto exit;
2516 		}
2517 
2518 		switch (object_type) {
2519 		case KFD_CRIU_OBJECT_TYPE_QUEUE:
2520 			ret = kfd_criu_restore_queue(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_EVENT:
2526 			ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2527 						     priv_offset, max_priv_data_size);
2528 			if (ret)
2529 				goto exit;
2530 			break;
2531 		case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2532 			ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2533 						     priv_offset, max_priv_data_size);
2534 			if (ret)
2535 				goto exit;
2536 			break;
2537 		default:
2538 			pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2539 			ret = -EINVAL;
2540 			goto exit;
2541 		}
2542 	}
2543 exit:
2544 	return ret;
2545 }
2546 
2547 static int criu_restore(struct file *filep,
2548 			struct kfd_process *p,
2549 			struct kfd_ioctl_criu_args *args)
2550 {
2551 	uint64_t priv_offset = 0;
2552 	int ret = 0;
2553 
2554 	pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2555 		 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2556 
2557 	if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2558 	    !args->num_devices || !args->num_bos)
2559 		return -EINVAL;
2560 
2561 	mutex_lock(&p->mutex);
2562 
2563 	/*
2564 	 * Set the process to evicted state to avoid running any new queues before all the memory
2565 	 * mappings are ready.
2566 	 */
2567 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2568 	if (ret)
2569 		goto exit_unlock;
2570 
2571 	/* Each function will adjust priv_offset based on how many bytes they consumed */
2572 	ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2573 	if (ret)
2574 		goto exit_unlock;
2575 
2576 	ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2577 	if (ret)
2578 		goto exit_unlock;
2579 
2580 	ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2581 	if (ret)
2582 		goto exit_unlock;
2583 
2584 	ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2585 	if (ret)
2586 		goto exit_unlock;
2587 
2588 	if (priv_offset != args->priv_data_size) {
2589 		pr_err("Invalid private data size\n");
2590 		ret = -EINVAL;
2591 	}
2592 
2593 exit_unlock:
2594 	mutex_unlock(&p->mutex);
2595 	if (ret)
2596 		pr_err("Failed to restore CRIU ret:%d\n", ret);
2597 	else
2598 		pr_debug("CRIU restore successful\n");
2599 
2600 	return ret;
2601 }
2602 
2603 static int criu_unpause(struct file *filep,
2604 			struct kfd_process *p,
2605 			struct kfd_ioctl_criu_args *args)
2606 {
2607 	int ret;
2608 
2609 	mutex_lock(&p->mutex);
2610 
2611 	if (!p->queues_paused) {
2612 		mutex_unlock(&p->mutex);
2613 		return -EINVAL;
2614 	}
2615 
2616 	ret = kfd_process_restore_queues(p);
2617 	if (ret)
2618 		pr_err("Failed to unpause queues ret:%d\n", ret);
2619 	else
2620 		p->queues_paused = false;
2621 
2622 	mutex_unlock(&p->mutex);
2623 
2624 	return ret;
2625 }
2626 
2627 static int criu_resume(struct file *filep,
2628 			struct kfd_process *p,
2629 			struct kfd_ioctl_criu_args *args)
2630 {
2631 	struct kfd_process *target = NULL;
2632 	struct pid *pid = NULL;
2633 	int ret = 0;
2634 
2635 	pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2636 		 args->pid);
2637 
2638 	pid = find_get_pid(args->pid);
2639 	if (!pid) {
2640 		pr_err("Cannot find pid info for %i\n", args->pid);
2641 		return -ESRCH;
2642 	}
2643 
2644 	pr_debug("calling kfd_lookup_process_by_pid\n");
2645 	target = kfd_lookup_process_by_pid(pid);
2646 
2647 	put_pid(pid);
2648 
2649 	if (!target) {
2650 		pr_debug("Cannot find process info for %i\n", args->pid);
2651 		return -ESRCH;
2652 	}
2653 
2654 	mutex_lock(&target->mutex);
2655 	ret = kfd_criu_resume_svm(target);
2656 	if (ret) {
2657 		pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2658 		goto exit;
2659 	}
2660 
2661 	ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2662 	if (ret)
2663 		pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2664 
2665 exit:
2666 	mutex_unlock(&target->mutex);
2667 
2668 	kfd_unref_process(target);
2669 	return ret;
2670 }
2671 
2672 static int criu_process_info(struct file *filep,
2673 				struct kfd_process *p,
2674 				struct kfd_ioctl_criu_args *args)
2675 {
2676 	int ret = 0;
2677 
2678 	mutex_lock(&p->mutex);
2679 
2680 	if (!p->n_pdds) {
2681 		pr_err("No pdd for given process\n");
2682 		ret = -ENODEV;
2683 		goto err_unlock;
2684 	}
2685 
2686 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2687 	if (ret)
2688 		goto err_unlock;
2689 
2690 	p->queues_paused = true;
2691 
2692 	args->pid = task_pid_nr_ns(p->lead_thread,
2693 					task_active_pid_ns(p->lead_thread));
2694 
2695 	ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2696 					   &args->num_objects, &args->priv_data_size);
2697 	if (ret)
2698 		goto err_unlock;
2699 
2700 	dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2701 				args->num_devices, args->num_bos, args->num_objects,
2702 				args->priv_data_size);
2703 
2704 err_unlock:
2705 	if (ret) {
2706 		kfd_process_restore_queues(p);
2707 		p->queues_paused = false;
2708 	}
2709 	mutex_unlock(&p->mutex);
2710 	return ret;
2711 }
2712 
2713 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2714 {
2715 	struct kfd_ioctl_criu_args *args = data;
2716 	int ret;
2717 
2718 	dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2719 	switch (args->op) {
2720 	case KFD_CRIU_OP_PROCESS_INFO:
2721 		ret = criu_process_info(filep, p, args);
2722 		break;
2723 	case KFD_CRIU_OP_CHECKPOINT:
2724 		ret = criu_checkpoint(filep, p, args);
2725 		break;
2726 	case KFD_CRIU_OP_UNPAUSE:
2727 		ret = criu_unpause(filep, p, args);
2728 		break;
2729 	case KFD_CRIU_OP_RESTORE:
2730 		ret = criu_restore(filep, p, args);
2731 		break;
2732 	case KFD_CRIU_OP_RESUME:
2733 		ret = criu_resume(filep, p, args);
2734 		break;
2735 	default:
2736 		dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2737 		ret = -EINVAL;
2738 		break;
2739 	}
2740 
2741 	if (ret)
2742 		dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2743 
2744 	return ret;
2745 }
2746 
2747 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2748 			bool enable_ttmp_setup)
2749 {
2750 	int i = 0, ret = 0;
2751 
2752 	if (p->is_runtime_retry)
2753 		goto retry;
2754 
2755 	if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2756 		return -EBUSY;
2757 
2758 	for (i = 0; i < p->n_pdds; i++) {
2759 		struct kfd_process_device *pdd = p->pdds[i];
2760 
2761 		if (pdd->qpd.queue_count)
2762 			return -EEXIST;
2763 
2764 		/*
2765 		 * Setup TTMPs by default.
2766 		 * Note that this call must remain here for MES ADD QUEUE to
2767 		 * skip_process_ctx_clear unconditionally as the first call to
2768 		 * SET_SHADER_DEBUGGER clears any stale process context data
2769 		 * saved in MES.
2770 		 */
2771 		if (pdd->dev->kfd->shared_resources.enable_mes)
2772 			kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
2773 	}
2774 
2775 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2776 	p->runtime_info.r_debug = r_debug;
2777 	p->runtime_info.ttmp_setup = enable_ttmp_setup;
2778 
2779 	if (p->runtime_info.ttmp_setup) {
2780 		for (i = 0; i < p->n_pdds; i++) {
2781 			struct kfd_process_device *pdd = p->pdds[i];
2782 
2783 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2784 				amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2785 				pdd->dev->kfd2kgd->enable_debug_trap(
2786 						pdd->dev->adev,
2787 						true,
2788 						pdd->dev->vm_info.last_vmid_kfd);
2789 			} else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2790 				pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2791 						pdd->dev->adev,
2792 						false,
2793 						0);
2794 			}
2795 		}
2796 	}
2797 
2798 retry:
2799 	if (p->debug_trap_enabled) {
2800 		if (!p->is_runtime_retry) {
2801 			kfd_dbg_trap_activate(p);
2802 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2803 					p, NULL, 0, false, NULL, 0);
2804 		}
2805 
2806 		mutex_unlock(&p->mutex);
2807 		ret = down_interruptible(&p->runtime_enable_sema);
2808 		mutex_lock(&p->mutex);
2809 
2810 		p->is_runtime_retry = !!ret;
2811 	}
2812 
2813 	return ret;
2814 }
2815 
2816 static int runtime_disable(struct kfd_process *p)
2817 {
2818 	int i = 0, ret;
2819 	bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2820 
2821 	p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2822 	p->runtime_info.r_debug = 0;
2823 
2824 	if (p->debug_trap_enabled) {
2825 		if (was_enabled)
2826 			kfd_dbg_trap_deactivate(p, false, 0);
2827 
2828 		if (!p->is_runtime_retry)
2829 			kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2830 					p, NULL, 0, false, NULL, 0);
2831 
2832 		mutex_unlock(&p->mutex);
2833 		ret = down_interruptible(&p->runtime_enable_sema);
2834 		mutex_lock(&p->mutex);
2835 
2836 		p->is_runtime_retry = !!ret;
2837 		if (ret)
2838 			return ret;
2839 	}
2840 
2841 	if (was_enabled && p->runtime_info.ttmp_setup) {
2842 		for (i = 0; i < p->n_pdds; i++) {
2843 			struct kfd_process_device *pdd = p->pdds[i];
2844 
2845 			if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2846 				amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2847 		}
2848 	}
2849 
2850 	p->runtime_info.ttmp_setup = false;
2851 
2852 	/* disable ttmp setup */
2853 	for (i = 0; i < p->n_pdds; i++) {
2854 		struct kfd_process_device *pdd = p->pdds[i];
2855 
2856 		if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2857 			pdd->spi_dbg_override =
2858 					pdd->dev->kfd2kgd->disable_debug_trap(
2859 					pdd->dev->adev,
2860 					false,
2861 					pdd->dev->vm_info.last_vmid_kfd);
2862 
2863 			if (!pdd->dev->kfd->shared_resources.enable_mes)
2864 				debug_refresh_runlist(pdd->dev->dqm);
2865 			else
2866 				kfd_dbg_set_mes_debug_mode(pdd,
2867 							   !kfd_dbg_has_cwsr_workaround(pdd->dev));
2868 		}
2869 	}
2870 
2871 	return 0;
2872 }
2873 
2874 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2875 {
2876 	struct kfd_ioctl_runtime_enable_args *args = data;
2877 	int r;
2878 
2879 	mutex_lock(&p->mutex);
2880 
2881 	if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2882 		r = runtime_enable(p, args->r_debug,
2883 				!!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2884 	else
2885 		r = runtime_disable(p);
2886 
2887 	mutex_unlock(&p->mutex);
2888 
2889 	return r;
2890 }
2891 
2892 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2893 {
2894 	struct kfd_ioctl_dbg_trap_args *args = data;
2895 	struct task_struct *thread = NULL;
2896 	struct mm_struct *mm = NULL;
2897 	struct pid *pid = NULL;
2898 	struct kfd_process *target = NULL;
2899 	struct kfd_process_device *pdd = NULL;
2900 	int r = 0;
2901 
2902 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2903 		pr_err("Debugging does not support sched_policy %i", sched_policy);
2904 		return -EINVAL;
2905 	}
2906 
2907 	pid = find_get_pid(args->pid);
2908 	if (!pid) {
2909 		pr_debug("Cannot find pid info for %i\n", args->pid);
2910 		r = -ESRCH;
2911 		goto out;
2912 	}
2913 
2914 	thread = get_pid_task(pid, PIDTYPE_PID);
2915 	if (!thread) {
2916 		r = -ESRCH;
2917 		goto out;
2918 	}
2919 
2920 	mm = get_task_mm(thread);
2921 	if (!mm) {
2922 		r = -ESRCH;
2923 		goto out;
2924 	}
2925 
2926 	if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2927 		bool create_process;
2928 
2929 		rcu_read_lock();
2930 		create_process = thread && thread != current && ptrace_parent(thread) == current;
2931 		rcu_read_unlock();
2932 
2933 		target = create_process ? kfd_create_process(thread) :
2934 					kfd_lookup_process_by_pid(pid);
2935 	} else {
2936 		target = kfd_lookup_process_by_pid(pid);
2937 	}
2938 
2939 	if (IS_ERR_OR_NULL(target)) {
2940 		pr_debug("Cannot find process PID %i to debug\n", args->pid);
2941 		r = target ? PTR_ERR(target) : -ESRCH;
2942 		goto out;
2943 	}
2944 
2945 	/* Check if target is still PTRACED. */
2946 	rcu_read_lock();
2947 	if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2948 				&& ptrace_parent(target->lead_thread) != current) {
2949 		pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2950 		r = -EPERM;
2951 	}
2952 	rcu_read_unlock();
2953 
2954 	if (r)
2955 		goto out;
2956 
2957 	mutex_lock(&target->mutex);
2958 
2959 	if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2960 		pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2961 		r = -EINVAL;
2962 		goto unlock_out;
2963 	}
2964 
2965 	if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2966 			(args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2967 			 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2968 			 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2969 			 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2970 			 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2971 			 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2972 			 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2973 		r = -EPERM;
2974 		goto unlock_out;
2975 	}
2976 
2977 	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2978 	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2979 		int user_gpu_id = kfd_process_get_user_gpu_id(target,
2980 				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2981 					args->set_node_address_watch.gpu_id :
2982 					args->clear_node_address_watch.gpu_id);
2983 
2984 		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
2985 		if (user_gpu_id == -EINVAL || !pdd) {
2986 			r = -ENODEV;
2987 			goto unlock_out;
2988 		}
2989 	}
2990 
2991 	switch (args->op) {
2992 	case KFD_IOC_DBG_TRAP_ENABLE:
2993 		if (target != p)
2994 			target->debugger_process = p;
2995 
2996 		r = kfd_dbg_trap_enable(target,
2997 					args->enable.dbg_fd,
2998 					(void __user *)args->enable.rinfo_ptr,
2999 					&args->enable.rinfo_size);
3000 		if (!r)
3001 			target->exception_enable_mask = args->enable.exception_mask;
3002 
3003 		break;
3004 	case KFD_IOC_DBG_TRAP_DISABLE:
3005 		r = kfd_dbg_trap_disable(target);
3006 		break;
3007 	case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
3008 		r = kfd_dbg_send_exception_to_runtime(target,
3009 				args->send_runtime_event.gpu_id,
3010 				args->send_runtime_event.queue_id,
3011 				args->send_runtime_event.exception_mask);
3012 		break;
3013 	case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
3014 		kfd_dbg_set_enabled_debug_exception_mask(target,
3015 				args->set_exceptions_enabled.exception_mask);
3016 		break;
3017 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
3018 		r = kfd_dbg_trap_set_wave_launch_override(target,
3019 				args->launch_override.override_mode,
3020 				args->launch_override.enable_mask,
3021 				args->launch_override.support_request_mask,
3022 				&args->launch_override.enable_mask,
3023 				&args->launch_override.support_request_mask);
3024 		break;
3025 	case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
3026 		r = kfd_dbg_trap_set_wave_launch_mode(target,
3027 				args->launch_mode.launch_mode);
3028 		break;
3029 	case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
3030 		r = suspend_queues(target,
3031 				args->suspend_queues.num_queues,
3032 				args->suspend_queues.grace_period,
3033 				args->suspend_queues.exception_mask,
3034 				(uint32_t *)args->suspend_queues.queue_array_ptr);
3035 
3036 		break;
3037 	case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
3038 		r = resume_queues(target, args->resume_queues.num_queues,
3039 				(uint32_t *)args->resume_queues.queue_array_ptr);
3040 		break;
3041 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3042 		r = kfd_dbg_trap_set_dev_address_watch(pdd,
3043 				args->set_node_address_watch.address,
3044 				args->set_node_address_watch.mask,
3045 				&args->set_node_address_watch.id,
3046 				args->set_node_address_watch.mode);
3047 		break;
3048 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3049 		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3050 				args->clear_node_address_watch.id);
3051 		break;
3052 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
3053 		r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3054 		break;
3055 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3056 		r = kfd_dbg_ev_query_debug_event(target,
3057 				&args->query_debug_event.queue_id,
3058 				&args->query_debug_event.gpu_id,
3059 				args->query_debug_event.exception_mask,
3060 				&args->query_debug_event.exception_mask);
3061 		break;
3062 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3063 		r = kfd_dbg_trap_query_exception_info(target,
3064 				args->query_exception_info.source_id,
3065 				args->query_exception_info.exception_code,
3066 				args->query_exception_info.clear_exception,
3067 				(void __user *)args->query_exception_info.info_ptr,
3068 				&args->query_exception_info.info_size);
3069 		break;
3070 	case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3071 		r = pqm_get_queue_snapshot(&target->pqm,
3072 				args->queue_snapshot.exception_mask,
3073 				(void __user *)args->queue_snapshot.snapshot_buf_ptr,
3074 				&args->queue_snapshot.num_queues,
3075 				&args->queue_snapshot.entry_size);
3076 		break;
3077 	case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3078 		r = kfd_dbg_trap_device_snapshot(target,
3079 				args->device_snapshot.exception_mask,
3080 				(void __user *)args->device_snapshot.snapshot_buf_ptr,
3081 				&args->device_snapshot.num_devices,
3082 				&args->device_snapshot.entry_size);
3083 		break;
3084 	default:
3085 		pr_err("Invalid option: %i\n", args->op);
3086 		r = -EINVAL;
3087 	}
3088 
3089 unlock_out:
3090 	mutex_unlock(&target->mutex);
3091 
3092 out:
3093 	if (thread)
3094 		put_task_struct(thread);
3095 
3096 	if (mm)
3097 		mmput(mm);
3098 
3099 	if (pid)
3100 		put_pid(pid);
3101 
3102 	if (target)
3103 		kfd_unref_process(target);
3104 
3105 	return r;
3106 }
3107 
3108 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3109 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3110 			    .cmd_drv = 0, .name = #ioctl}
3111 
3112 /** Ioctl table */
3113 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3114 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3115 			kfd_ioctl_get_version, 0),
3116 
3117 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3118 			kfd_ioctl_create_queue, 0),
3119 
3120 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3121 			kfd_ioctl_destroy_queue, 0),
3122 
3123 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3124 			kfd_ioctl_set_memory_policy, 0),
3125 
3126 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3127 			kfd_ioctl_get_clock_counters, 0),
3128 
3129 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3130 			kfd_ioctl_get_process_apertures, 0),
3131 
3132 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3133 			kfd_ioctl_update_queue, 0),
3134 
3135 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3136 			kfd_ioctl_create_event, 0),
3137 
3138 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3139 			kfd_ioctl_destroy_event, 0),
3140 
3141 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3142 			kfd_ioctl_set_event, 0),
3143 
3144 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3145 			kfd_ioctl_reset_event, 0),
3146 
3147 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3148 			kfd_ioctl_wait_events, 0),
3149 
3150 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3151 			kfd_ioctl_dbg_register, 0),
3152 
3153 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3154 			kfd_ioctl_dbg_unregister, 0),
3155 
3156 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3157 			kfd_ioctl_dbg_address_watch, 0),
3158 
3159 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3160 			kfd_ioctl_dbg_wave_control, 0),
3161 
3162 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3163 			kfd_ioctl_set_scratch_backing_va, 0),
3164 
3165 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3166 			kfd_ioctl_get_tile_config, 0),
3167 
3168 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3169 			kfd_ioctl_set_trap_handler, 0),
3170 
3171 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3172 			kfd_ioctl_get_process_apertures_new, 0),
3173 
3174 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3175 			kfd_ioctl_acquire_vm, 0),
3176 
3177 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3178 			kfd_ioctl_alloc_memory_of_gpu, 0),
3179 
3180 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3181 			kfd_ioctl_free_memory_of_gpu, 0),
3182 
3183 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3184 			kfd_ioctl_map_memory_to_gpu, 0),
3185 
3186 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3187 			kfd_ioctl_unmap_memory_from_gpu, 0),
3188 
3189 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3190 			kfd_ioctl_set_cu_mask, 0),
3191 
3192 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3193 			kfd_ioctl_get_queue_wave_state, 0),
3194 
3195 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3196 				kfd_ioctl_get_dmabuf_info, 0),
3197 
3198 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3199 				kfd_ioctl_import_dmabuf, 0),
3200 
3201 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3202 			kfd_ioctl_alloc_queue_gws, 0),
3203 
3204 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3205 			kfd_ioctl_smi_events, 0),
3206 
3207 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3208 
3209 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3210 			kfd_ioctl_set_xnack_mode, 0),
3211 
3212 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3213 			kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3214 
3215 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3216 			kfd_ioctl_get_available_memory, 0),
3217 
3218 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3219 				kfd_ioctl_export_dmabuf, 0),
3220 
3221 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3222 			kfd_ioctl_runtime_enable, 0),
3223 
3224 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3225 			kfd_ioctl_set_debug_trap, 0),
3226 };
3227 
3228 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
3229 
3230 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3231 {
3232 	struct kfd_process *process;
3233 	amdkfd_ioctl_t *func;
3234 	const struct amdkfd_ioctl_desc *ioctl = NULL;
3235 	unsigned int nr = _IOC_NR(cmd);
3236 	char stack_kdata[128];
3237 	char *kdata = NULL;
3238 	unsigned int usize, asize;
3239 	int retcode = -EINVAL;
3240 	bool ptrace_attached = false;
3241 
3242 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3243 		goto err_i1;
3244 
3245 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3246 		u32 amdkfd_size;
3247 
3248 		ioctl = &amdkfd_ioctls[nr];
3249 
3250 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
3251 		usize = asize = _IOC_SIZE(cmd);
3252 		if (amdkfd_size > asize)
3253 			asize = amdkfd_size;
3254 
3255 		cmd = ioctl->cmd;
3256 	} else
3257 		goto err_i1;
3258 
3259 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3260 
3261 	/* Get the process struct from the filep. Only the process
3262 	 * that opened /dev/kfd can use the file descriptor. Child
3263 	 * processes need to create their own KFD device context.
3264 	 */
3265 	process = filep->private_data;
3266 
3267 	rcu_read_lock();
3268 	if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3269 	    ptrace_parent(process->lead_thread) == current)
3270 		ptrace_attached = true;
3271 	rcu_read_unlock();
3272 
3273 	if (process->lead_thread != current->group_leader
3274 	    && !ptrace_attached) {
3275 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3276 		retcode = -EBADF;
3277 		goto err_i1;
3278 	}
3279 
3280 	/* Do not trust userspace, use our own definition */
3281 	func = ioctl->func;
3282 
3283 	if (unlikely(!func)) {
3284 		dev_dbg(kfd_device, "no function\n");
3285 		retcode = -EINVAL;
3286 		goto err_i1;
3287 	}
3288 
3289 	/*
3290 	 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3291 	 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3292 	 * more priviledged access.
3293 	 */
3294 	if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3295 		if (!capable(CAP_CHECKPOINT_RESTORE) &&
3296 						!capable(CAP_SYS_ADMIN)) {
3297 			retcode = -EACCES;
3298 			goto err_i1;
3299 		}
3300 	}
3301 
3302 	if (cmd & (IOC_IN | IOC_OUT)) {
3303 		if (asize <= sizeof(stack_kdata)) {
3304 			kdata = stack_kdata;
3305 		} else {
3306 			kdata = kmalloc(asize, GFP_KERNEL);
3307 			if (!kdata) {
3308 				retcode = -ENOMEM;
3309 				goto err_i1;
3310 			}
3311 		}
3312 		if (asize > usize)
3313 			memset(kdata + usize, 0, asize - usize);
3314 	}
3315 
3316 	if (cmd & IOC_IN) {
3317 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3318 			retcode = -EFAULT;
3319 			goto err_i1;
3320 		}
3321 	} else if (cmd & IOC_OUT) {
3322 		memset(kdata, 0, usize);
3323 	}
3324 
3325 	retcode = func(filep, process, kdata);
3326 
3327 	if (cmd & IOC_OUT)
3328 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3329 			retcode = -EFAULT;
3330 
3331 err_i1:
3332 	if (!ioctl)
3333 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3334 			  task_pid_nr(current), cmd, nr);
3335 
3336 	if (kdata != stack_kdata)
3337 		kfree(kdata);
3338 
3339 	if (retcode)
3340 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3341 				nr, arg, retcode);
3342 
3343 	return retcode;
3344 }
3345 
3346 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3347 		      struct vm_area_struct *vma)
3348 {
3349 	phys_addr_t address;
3350 
3351 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3352 		return -EINVAL;
3353 
3354 	address = dev->adev->rmmio_remap.bus_addr;
3355 
3356 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3357 				VM_DONTDUMP | VM_PFNMAP);
3358 
3359 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3360 
3361 	pr_debug("pasid 0x%x mapping mmio page\n"
3362 		 "     target user address == 0x%08llX\n"
3363 		 "     physical address    == 0x%08llX\n"
3364 		 "     vm_flags            == 0x%04lX\n"
3365 		 "     size                == 0x%04lX\n",
3366 		 process->pasid, (unsigned long long) vma->vm_start,
3367 		 address, vma->vm_flags, PAGE_SIZE);
3368 
3369 	return io_remap_pfn_range(vma,
3370 				vma->vm_start,
3371 				address >> PAGE_SHIFT,
3372 				PAGE_SIZE,
3373 				vma->vm_page_prot);
3374 }
3375 
3376 
3377 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3378 {
3379 	struct kfd_process *process;
3380 	struct kfd_node *dev = NULL;
3381 	unsigned long mmap_offset;
3382 	unsigned int gpu_id;
3383 
3384 	process = kfd_get_process(current);
3385 	if (IS_ERR(process))
3386 		return PTR_ERR(process);
3387 
3388 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3389 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3390 	if (gpu_id)
3391 		dev = kfd_device_by_id(gpu_id);
3392 
3393 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3394 	case KFD_MMAP_TYPE_DOORBELL:
3395 		if (!dev)
3396 			return -ENODEV;
3397 		return kfd_doorbell_mmap(dev, process, vma);
3398 
3399 	case KFD_MMAP_TYPE_EVENTS:
3400 		return kfd_event_mmap(process, vma);
3401 
3402 	case KFD_MMAP_TYPE_RESERVED_MEM:
3403 		if (!dev)
3404 			return -ENODEV;
3405 		return kfd_reserved_mem_mmap(dev, process, vma);
3406 	case KFD_MMAP_TYPE_MMIO:
3407 		if (!dev)
3408 			return -ENODEV;
3409 		return kfd_mmio_mmap(dev, process, vma);
3410 	}
3411 
3412 	return -EFAULT;
3413 }
3414