1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  */
22 
23 #include <linux/device.h>
24 #include <linux/export.h>
25 #include <linux/err.h>
26 #include <linux/fs.h>
27 #include <linux/file.h>
28 #include <linux/sched.h>
29 #include <linux/slab.h>
30 #include <linux/uaccess.h>
31 #include <linux/compat.h>
32 #include <uapi/linux/kfd_ioctl.h>
33 #include <linux/time.h>
34 #include <linux/mm.h>
35 #include <linux/mman.h>
36 #include <linux/dma-buf.h>
37 #include <asm/processor.h>
38 #include "kfd_priv.h"
39 #include "kfd_device_queue_manager.h"
40 #include "kfd_dbgmgr.h"
41 #include "amdgpu_amdkfd.h"
42 #include "kfd_smi_events.h"
43 
44 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
45 static int kfd_open(struct inode *, struct file *);
46 static int kfd_release(struct inode *, struct file *);
47 static int kfd_mmap(struct file *, struct vm_area_struct *);
48 
49 static const char kfd_dev_name[] = "kfd";
50 
51 static const struct file_operations kfd_fops = {
52 	.owner = THIS_MODULE,
53 	.unlocked_ioctl = kfd_ioctl,
54 	.compat_ioctl = compat_ptr_ioctl,
55 	.open = kfd_open,
56 	.release = kfd_release,
57 	.mmap = kfd_mmap,
58 };
59 
60 static int kfd_char_dev_major = -1;
61 static struct class *kfd_class;
62 struct device *kfd_device;
63 
64 int kfd_chardev_init(void)
65 {
66 	int err = 0;
67 
68 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
69 	err = kfd_char_dev_major;
70 	if (err < 0)
71 		goto err_register_chrdev;
72 
73 	kfd_class = class_create(THIS_MODULE, kfd_dev_name);
74 	err = PTR_ERR(kfd_class);
75 	if (IS_ERR(kfd_class))
76 		goto err_class_create;
77 
78 	kfd_device = device_create(kfd_class, NULL,
79 					MKDEV(kfd_char_dev_major, 0),
80 					NULL, kfd_dev_name);
81 	err = PTR_ERR(kfd_device);
82 	if (IS_ERR(kfd_device))
83 		goto err_device_create;
84 
85 	return 0;
86 
87 err_device_create:
88 	class_destroy(kfd_class);
89 err_class_create:
90 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
91 err_register_chrdev:
92 	return err;
93 }
94 
95 void kfd_chardev_exit(void)
96 {
97 	device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
98 	class_destroy(kfd_class);
99 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
100 	kfd_device = NULL;
101 }
102 
103 struct device *kfd_chardev(void)
104 {
105 	return kfd_device;
106 }
107 
108 
109 static int kfd_open(struct inode *inode, struct file *filep)
110 {
111 	struct kfd_process *process;
112 	bool is_32bit_user_mode;
113 
114 	if (iminor(inode) != 0)
115 		return -ENODEV;
116 
117 	is_32bit_user_mode = in_compat_syscall();
118 
119 	if (is_32bit_user_mode) {
120 		dev_warn(kfd_device,
121 			"Process %d (32-bit) failed to open /dev/kfd\n"
122 			"32-bit processes are not supported by amdkfd\n",
123 			current->pid);
124 		return -EPERM;
125 	}
126 
127 	process = kfd_create_process(filep);
128 	if (IS_ERR(process))
129 		return PTR_ERR(process);
130 
131 	if (kfd_is_locked()) {
132 		dev_dbg(kfd_device, "kfd is locked!\n"
133 				"process %d unreferenced", process->pasid);
134 		kfd_unref_process(process);
135 		return -EAGAIN;
136 	}
137 
138 	/* filep now owns the reference returned by kfd_create_process */
139 	filep->private_data = process;
140 
141 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
142 		process->pasid, process->is_32bit_user_mode);
143 
144 	return 0;
145 }
146 
147 static int kfd_release(struct inode *inode, struct file *filep)
148 {
149 	struct kfd_process *process = filep->private_data;
150 
151 	if (process)
152 		kfd_unref_process(process);
153 
154 	return 0;
155 }
156 
157 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
158 					void *data)
159 {
160 	struct kfd_ioctl_get_version_args *args = data;
161 
162 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
163 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
164 
165 	return 0;
166 }
167 
168 static int set_queue_properties_from_user(struct queue_properties *q_properties,
169 				struct kfd_ioctl_create_queue_args *args)
170 {
171 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
172 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
173 		return -EINVAL;
174 	}
175 
176 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
177 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
178 		return -EINVAL;
179 	}
180 
181 	if ((args->ring_base_address) &&
182 		(!access_ok((const void __user *) args->ring_base_address,
183 			sizeof(uint64_t)))) {
184 		pr_err("Can't access ring base address\n");
185 		return -EFAULT;
186 	}
187 
188 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
189 		pr_err("Ring size must be a power of 2 or 0\n");
190 		return -EINVAL;
191 	}
192 
193 	if (!access_ok((const void __user *) args->read_pointer_address,
194 			sizeof(uint32_t))) {
195 		pr_err("Can't access read pointer\n");
196 		return -EFAULT;
197 	}
198 
199 	if (!access_ok((const void __user *) args->write_pointer_address,
200 			sizeof(uint32_t))) {
201 		pr_err("Can't access write pointer\n");
202 		return -EFAULT;
203 	}
204 
205 	if (args->eop_buffer_address &&
206 		!access_ok((const void __user *) args->eop_buffer_address,
207 			sizeof(uint32_t))) {
208 		pr_debug("Can't access eop buffer");
209 		return -EFAULT;
210 	}
211 
212 	if (args->ctx_save_restore_address &&
213 		!access_ok((const void __user *) args->ctx_save_restore_address,
214 			sizeof(uint32_t))) {
215 		pr_debug("Can't access ctx save restore buffer");
216 		return -EFAULT;
217 	}
218 
219 	q_properties->is_interop = false;
220 	q_properties->is_gws = false;
221 	q_properties->queue_percent = args->queue_percentage;
222 	q_properties->priority = args->queue_priority;
223 	q_properties->queue_address = args->ring_base_address;
224 	q_properties->queue_size = args->ring_size;
225 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
226 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
227 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
228 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
229 	q_properties->ctx_save_restore_area_address =
230 			args->ctx_save_restore_address;
231 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
232 	q_properties->ctl_stack_size = args->ctl_stack_size;
233 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
234 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
235 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
236 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
237 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
238 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
239 		q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
240 	else
241 		return -ENOTSUPP;
242 
243 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
244 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
245 	else
246 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
247 
248 	pr_debug("Queue Percentage: %d, %d\n",
249 			q_properties->queue_percent, args->queue_percentage);
250 
251 	pr_debug("Queue Priority: %d, %d\n",
252 			q_properties->priority, args->queue_priority);
253 
254 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
255 			q_properties->queue_address, args->ring_base_address);
256 
257 	pr_debug("Queue Size: 0x%llX, %u\n",
258 			q_properties->queue_size, args->ring_size);
259 
260 	pr_debug("Queue r/w Pointers: %px, %px\n",
261 			q_properties->read_ptr,
262 			q_properties->write_ptr);
263 
264 	pr_debug("Queue Format: %d\n", q_properties->format);
265 
266 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
267 
268 	pr_debug("Queue CTX save area: 0x%llX\n",
269 			q_properties->ctx_save_restore_area_address);
270 
271 	return 0;
272 }
273 
274 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
275 					void *data)
276 {
277 	struct kfd_ioctl_create_queue_args *args = data;
278 	struct kfd_dev *dev;
279 	int err = 0;
280 	unsigned int queue_id;
281 	struct kfd_process_device *pdd;
282 	struct queue_properties q_properties;
283 	uint32_t doorbell_offset_in_process = 0;
284 
285 	memset(&q_properties, 0, sizeof(struct queue_properties));
286 
287 	pr_debug("Creating queue ioctl\n");
288 
289 	err = set_queue_properties_from_user(&q_properties, args);
290 	if (err)
291 		return err;
292 
293 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
294 	dev = kfd_device_by_id(args->gpu_id);
295 	if (!dev) {
296 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
297 		return -EINVAL;
298 	}
299 
300 	mutex_lock(&p->mutex);
301 
302 	pdd = kfd_bind_process_to_device(dev, p);
303 	if (IS_ERR(pdd)) {
304 		err = -ESRCH;
305 		goto err_bind_process;
306 	}
307 
308 	pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
309 			p->pasid,
310 			dev->id);
311 
312 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id,
313 			&doorbell_offset_in_process);
314 	if (err != 0)
315 		goto err_create_queue;
316 
317 	args->queue_id = queue_id;
318 
319 
320 	/* Return gpu_id as doorbell offset for mmap usage */
321 	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
322 	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
323 	if (KFD_IS_SOC15(dev->device_info->asic_family))
324 		/* On SOC15 ASICs, include the doorbell offset within the
325 		 * process doorbell frame, which is 2 pages.
326 		 */
327 		args->doorbell_offset |= doorbell_offset_in_process;
328 
329 	mutex_unlock(&p->mutex);
330 
331 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
332 
333 	pr_debug("Ring buffer address == 0x%016llX\n",
334 			args->ring_base_address);
335 
336 	pr_debug("Read ptr address    == 0x%016llX\n",
337 			args->read_pointer_address);
338 
339 	pr_debug("Write ptr address   == 0x%016llX\n",
340 			args->write_pointer_address);
341 
342 	return 0;
343 
344 err_create_queue:
345 err_bind_process:
346 	mutex_unlock(&p->mutex);
347 	return err;
348 }
349 
350 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
351 					void *data)
352 {
353 	int retval;
354 	struct kfd_ioctl_destroy_queue_args *args = data;
355 
356 	pr_debug("Destroying queue id %d for pasid 0x%x\n",
357 				args->queue_id,
358 				p->pasid);
359 
360 	mutex_lock(&p->mutex);
361 
362 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
363 
364 	mutex_unlock(&p->mutex);
365 	return retval;
366 }
367 
368 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
369 					void *data)
370 {
371 	int retval;
372 	struct kfd_ioctl_update_queue_args *args = data;
373 	struct queue_properties properties;
374 
375 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
376 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
377 		return -EINVAL;
378 	}
379 
380 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
381 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
382 		return -EINVAL;
383 	}
384 
385 	if ((args->ring_base_address) &&
386 		(!access_ok((const void __user *) args->ring_base_address,
387 			sizeof(uint64_t)))) {
388 		pr_err("Can't access ring base address\n");
389 		return -EFAULT;
390 	}
391 
392 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
393 		pr_err("Ring size must be a power of 2 or 0\n");
394 		return -EINVAL;
395 	}
396 
397 	properties.queue_address = args->ring_base_address;
398 	properties.queue_size = args->ring_size;
399 	properties.queue_percent = args->queue_percentage;
400 	properties.priority = args->queue_priority;
401 
402 	pr_debug("Updating queue id %d for pasid 0x%x\n",
403 			args->queue_id, p->pasid);
404 
405 	mutex_lock(&p->mutex);
406 
407 	retval = pqm_update_queue(&p->pqm, args->queue_id, &properties);
408 
409 	mutex_unlock(&p->mutex);
410 
411 	return retval;
412 }
413 
414 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
415 					void *data)
416 {
417 	int retval;
418 	const int max_num_cus = 1024;
419 	struct kfd_ioctl_set_cu_mask_args *args = data;
420 	struct queue_properties properties;
421 	uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
422 	size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
423 
424 	if ((args->num_cu_mask % 32) != 0) {
425 		pr_debug("num_cu_mask 0x%x must be a multiple of 32",
426 				args->num_cu_mask);
427 		return -EINVAL;
428 	}
429 
430 	properties.cu_mask_count = args->num_cu_mask;
431 	if (properties.cu_mask_count == 0) {
432 		pr_debug("CU mask cannot be 0");
433 		return -EINVAL;
434 	}
435 
436 	/* To prevent an unreasonably large CU mask size, set an arbitrary
437 	 * limit of max_num_cus bits.  We can then just drop any CU mask bits
438 	 * past max_num_cus bits and just use the first max_num_cus bits.
439 	 */
440 	if (properties.cu_mask_count > max_num_cus) {
441 		pr_debug("CU mask cannot be greater than 1024 bits");
442 		properties.cu_mask_count = max_num_cus;
443 		cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
444 	}
445 
446 	properties.cu_mask = kzalloc(cu_mask_size, GFP_KERNEL);
447 	if (!properties.cu_mask)
448 		return -ENOMEM;
449 
450 	retval = copy_from_user(properties.cu_mask, cu_mask_ptr, cu_mask_size);
451 	if (retval) {
452 		pr_debug("Could not copy CU mask from userspace");
453 		kfree(properties.cu_mask);
454 		return -EFAULT;
455 	}
456 
457 	mutex_lock(&p->mutex);
458 
459 	retval = pqm_set_cu_mask(&p->pqm, args->queue_id, &properties);
460 
461 	mutex_unlock(&p->mutex);
462 
463 	if (retval)
464 		kfree(properties.cu_mask);
465 
466 	return retval;
467 }
468 
469 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
470 					  struct kfd_process *p, void *data)
471 {
472 	struct kfd_ioctl_get_queue_wave_state_args *args = data;
473 	int r;
474 
475 	mutex_lock(&p->mutex);
476 
477 	r = pqm_get_wave_state(&p->pqm, args->queue_id,
478 			       (void __user *)args->ctl_stack_address,
479 			       &args->ctl_stack_used_size,
480 			       &args->save_area_used_size);
481 
482 	mutex_unlock(&p->mutex);
483 
484 	return r;
485 }
486 
487 static int kfd_ioctl_set_memory_policy(struct file *filep,
488 					struct kfd_process *p, void *data)
489 {
490 	struct kfd_ioctl_set_memory_policy_args *args = data;
491 	struct kfd_dev *dev;
492 	int err = 0;
493 	struct kfd_process_device *pdd;
494 	enum cache_policy default_policy, alternate_policy;
495 
496 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
497 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
498 		return -EINVAL;
499 	}
500 
501 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
502 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
503 		return -EINVAL;
504 	}
505 
506 	dev = kfd_device_by_id(args->gpu_id);
507 	if (!dev)
508 		return -EINVAL;
509 
510 	mutex_lock(&p->mutex);
511 
512 	pdd = kfd_bind_process_to_device(dev, p);
513 	if (IS_ERR(pdd)) {
514 		err = -ESRCH;
515 		goto out;
516 	}
517 
518 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
519 			 ? cache_policy_coherent : cache_policy_noncoherent;
520 
521 	alternate_policy =
522 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
523 		   ? cache_policy_coherent : cache_policy_noncoherent;
524 
525 	if (!dev->dqm->ops.set_cache_memory_policy(dev->dqm,
526 				&pdd->qpd,
527 				default_policy,
528 				alternate_policy,
529 				(void __user *)args->alternate_aperture_base,
530 				args->alternate_aperture_size))
531 		err = -EINVAL;
532 
533 out:
534 	mutex_unlock(&p->mutex);
535 
536 	return err;
537 }
538 
539 static int kfd_ioctl_set_trap_handler(struct file *filep,
540 					struct kfd_process *p, void *data)
541 {
542 	struct kfd_ioctl_set_trap_handler_args *args = data;
543 	struct kfd_dev *dev;
544 	int err = 0;
545 	struct kfd_process_device *pdd;
546 
547 	dev = kfd_device_by_id(args->gpu_id);
548 	if (!dev)
549 		return -EINVAL;
550 
551 	mutex_lock(&p->mutex);
552 
553 	pdd = kfd_bind_process_to_device(dev, p);
554 	if (IS_ERR(pdd)) {
555 		err = -ESRCH;
556 		goto out;
557 	}
558 
559 	kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
560 
561 out:
562 	mutex_unlock(&p->mutex);
563 
564 	return err;
565 }
566 
567 static int kfd_ioctl_dbg_register(struct file *filep,
568 				struct kfd_process *p, void *data)
569 {
570 	struct kfd_ioctl_dbg_register_args *args = data;
571 	struct kfd_dev *dev;
572 	struct kfd_dbgmgr *dbgmgr_ptr;
573 	struct kfd_process_device *pdd;
574 	bool create_ok;
575 	long status = 0;
576 
577 	dev = kfd_device_by_id(args->gpu_id);
578 	if (!dev)
579 		return -EINVAL;
580 
581 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
582 		pr_debug("kfd_ioctl_dbg_register not supported on CZ\n");
583 		return -EINVAL;
584 	}
585 
586 	mutex_lock(&p->mutex);
587 	mutex_lock(kfd_get_dbgmgr_mutex());
588 
589 	/*
590 	 * make sure that we have pdd, if this the first queue created for
591 	 * this process
592 	 */
593 	pdd = kfd_bind_process_to_device(dev, p);
594 	if (IS_ERR(pdd)) {
595 		status = PTR_ERR(pdd);
596 		goto out;
597 	}
598 
599 	if (!dev->dbgmgr) {
600 		/* In case of a legal call, we have no dbgmgr yet */
601 		create_ok = kfd_dbgmgr_create(&dbgmgr_ptr, dev);
602 		if (create_ok) {
603 			status = kfd_dbgmgr_register(dbgmgr_ptr, p);
604 			if (status != 0)
605 				kfd_dbgmgr_destroy(dbgmgr_ptr);
606 			else
607 				dev->dbgmgr = dbgmgr_ptr;
608 		}
609 	} else {
610 		pr_debug("debugger already registered\n");
611 		status = -EINVAL;
612 	}
613 
614 out:
615 	mutex_unlock(kfd_get_dbgmgr_mutex());
616 	mutex_unlock(&p->mutex);
617 
618 	return status;
619 }
620 
621 static int kfd_ioctl_dbg_unregister(struct file *filep,
622 				struct kfd_process *p, void *data)
623 {
624 	struct kfd_ioctl_dbg_unregister_args *args = data;
625 	struct kfd_dev *dev;
626 	long status;
627 
628 	dev = kfd_device_by_id(args->gpu_id);
629 	if (!dev || !dev->dbgmgr)
630 		return -EINVAL;
631 
632 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
633 		pr_debug("kfd_ioctl_dbg_unregister not supported on CZ\n");
634 		return -EINVAL;
635 	}
636 
637 	mutex_lock(kfd_get_dbgmgr_mutex());
638 
639 	status = kfd_dbgmgr_unregister(dev->dbgmgr, p);
640 	if (!status) {
641 		kfd_dbgmgr_destroy(dev->dbgmgr);
642 		dev->dbgmgr = NULL;
643 	}
644 
645 	mutex_unlock(kfd_get_dbgmgr_mutex());
646 
647 	return status;
648 }
649 
650 /*
651  * Parse and generate variable size data structure for address watch.
652  * Total size of the buffer and # watch points is limited in order
653  * to prevent kernel abuse. (no bearing to the much smaller HW limitation
654  * which is enforced by dbgdev module)
655  * please also note that the watch address itself are not "copied from user",
656  * since it be set into the HW in user mode values.
657  *
658  */
659 static int kfd_ioctl_dbg_address_watch(struct file *filep,
660 					struct kfd_process *p, void *data)
661 {
662 	struct kfd_ioctl_dbg_address_watch_args *args = data;
663 	struct kfd_dev *dev;
664 	struct dbg_address_watch_info aw_info;
665 	unsigned char *args_buff;
666 	long status;
667 	void __user *cmd_from_user;
668 	uint64_t watch_mask_value = 0;
669 	unsigned int args_idx = 0;
670 
671 	memset((void *) &aw_info, 0, sizeof(struct dbg_address_watch_info));
672 
673 	dev = kfd_device_by_id(args->gpu_id);
674 	if (!dev)
675 		return -EINVAL;
676 
677 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
678 		pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
679 		return -EINVAL;
680 	}
681 
682 	cmd_from_user = (void __user *) args->content_ptr;
683 
684 	/* Validate arguments */
685 
686 	if ((args->buf_size_in_bytes > MAX_ALLOWED_AW_BUFF_SIZE) ||
687 		(args->buf_size_in_bytes <= sizeof(*args) + sizeof(int) * 2) ||
688 		(cmd_from_user == NULL))
689 		return -EINVAL;
690 
691 	/* this is the actual buffer to work with */
692 	args_buff = memdup_user(cmd_from_user,
693 				args->buf_size_in_bytes - sizeof(*args));
694 	if (IS_ERR(args_buff))
695 		return PTR_ERR(args_buff);
696 
697 	aw_info.process = p;
698 
699 	aw_info.num_watch_points = *((uint32_t *)(&args_buff[args_idx]));
700 	args_idx += sizeof(aw_info.num_watch_points);
701 
702 	aw_info.watch_mode = (enum HSA_DBG_WATCH_MODE *) &args_buff[args_idx];
703 	args_idx += sizeof(enum HSA_DBG_WATCH_MODE) * aw_info.num_watch_points;
704 
705 	/*
706 	 * set watch address base pointer to point on the array base
707 	 * within args_buff
708 	 */
709 	aw_info.watch_address = (uint64_t *) &args_buff[args_idx];
710 
711 	/* skip over the addresses buffer */
712 	args_idx += sizeof(aw_info.watch_address) * aw_info.num_watch_points;
713 
714 	if (args_idx >= args->buf_size_in_bytes - sizeof(*args)) {
715 		status = -EINVAL;
716 		goto out;
717 	}
718 
719 	watch_mask_value = (uint64_t) args_buff[args_idx];
720 
721 	if (watch_mask_value > 0) {
722 		/*
723 		 * There is an array of masks.
724 		 * set watch mask base pointer to point on the array base
725 		 * within args_buff
726 		 */
727 		aw_info.watch_mask = (uint64_t *) &args_buff[args_idx];
728 
729 		/* skip over the masks buffer */
730 		args_idx += sizeof(aw_info.watch_mask) *
731 				aw_info.num_watch_points;
732 	} else {
733 		/* just the NULL mask, set to NULL and skip over it */
734 		aw_info.watch_mask = NULL;
735 		args_idx += sizeof(aw_info.watch_mask);
736 	}
737 
738 	if (args_idx >= args->buf_size_in_bytes - sizeof(args)) {
739 		status = -EINVAL;
740 		goto out;
741 	}
742 
743 	/* Currently HSA Event is not supported for DBG */
744 	aw_info.watch_event = NULL;
745 
746 	mutex_lock(kfd_get_dbgmgr_mutex());
747 
748 	status = kfd_dbgmgr_address_watch(dev->dbgmgr, &aw_info);
749 
750 	mutex_unlock(kfd_get_dbgmgr_mutex());
751 
752 out:
753 	kfree(args_buff);
754 
755 	return status;
756 }
757 
758 /* Parse and generate fixed size data structure for wave control */
759 static int kfd_ioctl_dbg_wave_control(struct file *filep,
760 					struct kfd_process *p, void *data)
761 {
762 	struct kfd_ioctl_dbg_wave_control_args *args = data;
763 	struct kfd_dev *dev;
764 	struct dbg_wave_control_info wac_info;
765 	unsigned char *args_buff;
766 	uint32_t computed_buff_size;
767 	long status;
768 	void __user *cmd_from_user;
769 	unsigned int args_idx = 0;
770 
771 	memset((void *) &wac_info, 0, sizeof(struct dbg_wave_control_info));
772 
773 	/* we use compact form, independent of the packing attribute value */
774 	computed_buff_size = sizeof(*args) +
775 				sizeof(wac_info.mode) +
776 				sizeof(wac_info.operand) +
777 				sizeof(wac_info.dbgWave_msg.DbgWaveMsg) +
778 				sizeof(wac_info.dbgWave_msg.MemoryVA) +
779 				sizeof(wac_info.trapId);
780 
781 	dev = kfd_device_by_id(args->gpu_id);
782 	if (!dev)
783 		return -EINVAL;
784 
785 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
786 		pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
787 		return -EINVAL;
788 	}
789 
790 	/* input size must match the computed "compact" size */
791 	if (args->buf_size_in_bytes != computed_buff_size) {
792 		pr_debug("size mismatch, computed : actual %u : %u\n",
793 				args->buf_size_in_bytes, computed_buff_size);
794 		return -EINVAL;
795 	}
796 
797 	cmd_from_user = (void __user *) args->content_ptr;
798 
799 	if (cmd_from_user == NULL)
800 		return -EINVAL;
801 
802 	/* copy the entire buffer from user */
803 
804 	args_buff = memdup_user(cmd_from_user,
805 				args->buf_size_in_bytes - sizeof(*args));
806 	if (IS_ERR(args_buff))
807 		return PTR_ERR(args_buff);
808 
809 	/* move ptr to the start of the "pay-load" area */
810 	wac_info.process = p;
811 
812 	wac_info.operand = *((enum HSA_DBG_WAVEOP *)(&args_buff[args_idx]));
813 	args_idx += sizeof(wac_info.operand);
814 
815 	wac_info.mode = *((enum HSA_DBG_WAVEMODE *)(&args_buff[args_idx]));
816 	args_idx += sizeof(wac_info.mode);
817 
818 	wac_info.trapId = *((uint32_t *)(&args_buff[args_idx]));
819 	args_idx += sizeof(wac_info.trapId);
820 
821 	wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value =
822 					*((uint32_t *)(&args_buff[args_idx]));
823 	wac_info.dbgWave_msg.MemoryVA = NULL;
824 
825 	mutex_lock(kfd_get_dbgmgr_mutex());
826 
827 	pr_debug("Calling dbg manager process %p, operand %u, mode %u, trapId %u, message %u\n",
828 			wac_info.process, wac_info.operand,
829 			wac_info.mode, wac_info.trapId,
830 			wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value);
831 
832 	status = kfd_dbgmgr_wave_control(dev->dbgmgr, &wac_info);
833 
834 	pr_debug("Returned status of dbg manager is %ld\n", status);
835 
836 	mutex_unlock(kfd_get_dbgmgr_mutex());
837 
838 	kfree(args_buff);
839 
840 	return status;
841 }
842 
843 static int kfd_ioctl_get_clock_counters(struct file *filep,
844 				struct kfd_process *p, void *data)
845 {
846 	struct kfd_ioctl_get_clock_counters_args *args = data;
847 	struct kfd_dev *dev;
848 
849 	dev = kfd_device_by_id(args->gpu_id);
850 	if (dev)
851 		/* Reading GPU clock counter from KGD */
852 		args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(dev->kgd);
853 	else
854 		/* Node without GPU resource */
855 		args->gpu_clock_counter = 0;
856 
857 	/* No access to rdtsc. Using raw monotonic time */
858 	args->cpu_clock_counter = ktime_get_raw_ns();
859 	args->system_clock_counter = ktime_get_boottime_ns();
860 
861 	/* Since the counter is in nano-seconds we use 1GHz frequency */
862 	args->system_clock_freq = 1000000000;
863 
864 	return 0;
865 }
866 
867 
868 static int kfd_ioctl_get_process_apertures(struct file *filp,
869 				struct kfd_process *p, void *data)
870 {
871 	struct kfd_ioctl_get_process_apertures_args *args = data;
872 	struct kfd_process_device_apertures *pAperture;
873 	int i;
874 
875 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
876 
877 	args->num_of_nodes = 0;
878 
879 	mutex_lock(&p->mutex);
880 	/* Run over all pdd of the process */
881 	for (i = 0; i < p->n_pdds; i++) {
882 		struct kfd_process_device *pdd = p->pdds[i];
883 
884 		pAperture =
885 			&args->process_apertures[args->num_of_nodes];
886 		pAperture->gpu_id = pdd->dev->id;
887 		pAperture->lds_base = pdd->lds_base;
888 		pAperture->lds_limit = pdd->lds_limit;
889 		pAperture->gpuvm_base = pdd->gpuvm_base;
890 		pAperture->gpuvm_limit = pdd->gpuvm_limit;
891 		pAperture->scratch_base = pdd->scratch_base;
892 		pAperture->scratch_limit = pdd->scratch_limit;
893 
894 		dev_dbg(kfd_device,
895 			"node id %u\n", args->num_of_nodes);
896 		dev_dbg(kfd_device,
897 			"gpu id %u\n", pdd->dev->id);
898 		dev_dbg(kfd_device,
899 			"lds_base %llX\n", pdd->lds_base);
900 		dev_dbg(kfd_device,
901 			"lds_limit %llX\n", pdd->lds_limit);
902 		dev_dbg(kfd_device,
903 			"gpuvm_base %llX\n", pdd->gpuvm_base);
904 		dev_dbg(kfd_device,
905 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
906 		dev_dbg(kfd_device,
907 			"scratch_base %llX\n", pdd->scratch_base);
908 		dev_dbg(kfd_device,
909 			"scratch_limit %llX\n", pdd->scratch_limit);
910 
911 		if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
912 			break;
913 	}
914 	mutex_unlock(&p->mutex);
915 
916 	return 0;
917 }
918 
919 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
920 				struct kfd_process *p, void *data)
921 {
922 	struct kfd_ioctl_get_process_apertures_new_args *args = data;
923 	struct kfd_process_device_apertures *pa;
924 	int ret;
925 	int i;
926 
927 	dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
928 
929 	if (args->num_of_nodes == 0) {
930 		/* Return number of nodes, so that user space can alloacate
931 		 * sufficient memory
932 		 */
933 		mutex_lock(&p->mutex);
934 		args->num_of_nodes = p->n_pdds;
935 		goto out_unlock;
936 	}
937 
938 	/* Fill in process-aperture information for all available
939 	 * nodes, but not more than args->num_of_nodes as that is
940 	 * the amount of memory allocated by user
941 	 */
942 	pa = kzalloc((sizeof(struct kfd_process_device_apertures) *
943 				args->num_of_nodes), GFP_KERNEL);
944 	if (!pa)
945 		return -ENOMEM;
946 
947 	mutex_lock(&p->mutex);
948 
949 	if (!p->n_pdds) {
950 		args->num_of_nodes = 0;
951 		kfree(pa);
952 		goto out_unlock;
953 	}
954 
955 	/* Run over all pdd of the process */
956 	for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
957 		struct kfd_process_device *pdd = p->pdds[i];
958 
959 		pa[i].gpu_id = pdd->dev->id;
960 		pa[i].lds_base = pdd->lds_base;
961 		pa[i].lds_limit = pdd->lds_limit;
962 		pa[i].gpuvm_base = pdd->gpuvm_base;
963 		pa[i].gpuvm_limit = pdd->gpuvm_limit;
964 		pa[i].scratch_base = pdd->scratch_base;
965 		pa[i].scratch_limit = pdd->scratch_limit;
966 
967 		dev_dbg(kfd_device,
968 			"gpu id %u\n", pdd->dev->id);
969 		dev_dbg(kfd_device,
970 			"lds_base %llX\n", pdd->lds_base);
971 		dev_dbg(kfd_device,
972 			"lds_limit %llX\n", pdd->lds_limit);
973 		dev_dbg(kfd_device,
974 			"gpuvm_base %llX\n", pdd->gpuvm_base);
975 		dev_dbg(kfd_device,
976 			"gpuvm_limit %llX\n", pdd->gpuvm_limit);
977 		dev_dbg(kfd_device,
978 			"scratch_base %llX\n", pdd->scratch_base);
979 		dev_dbg(kfd_device,
980 			"scratch_limit %llX\n", pdd->scratch_limit);
981 	}
982 	mutex_unlock(&p->mutex);
983 
984 	args->num_of_nodes = i;
985 	ret = copy_to_user(
986 			(void __user *)args->kfd_process_device_apertures_ptr,
987 			pa,
988 			(i * sizeof(struct kfd_process_device_apertures)));
989 	kfree(pa);
990 	return ret ? -EFAULT : 0;
991 
992 out_unlock:
993 	mutex_unlock(&p->mutex);
994 	return 0;
995 }
996 
997 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
998 					void *data)
999 {
1000 	struct kfd_ioctl_create_event_args *args = data;
1001 	int err;
1002 
1003 	/* For dGPUs the event page is allocated in user mode. The
1004 	 * handle is passed to KFD with the first call to this IOCTL
1005 	 * through the event_page_offset field.
1006 	 */
1007 	if (args->event_page_offset) {
1008 		struct kfd_dev *kfd;
1009 		struct kfd_process_device *pdd;
1010 		void *mem, *kern_addr;
1011 		uint64_t size;
1012 
1013 		if (p->signal_page) {
1014 			pr_err("Event page is already set\n");
1015 			return -EINVAL;
1016 		}
1017 
1018 		kfd = kfd_device_by_id(GET_GPU_ID(args->event_page_offset));
1019 		if (!kfd) {
1020 			pr_err("Getting device by id failed in %s\n", __func__);
1021 			return -EINVAL;
1022 		}
1023 
1024 		mutex_lock(&p->mutex);
1025 		pdd = kfd_bind_process_to_device(kfd, p);
1026 		if (IS_ERR(pdd)) {
1027 			err = PTR_ERR(pdd);
1028 			goto out_unlock;
1029 		}
1030 
1031 		mem = kfd_process_device_translate_handle(pdd,
1032 				GET_IDR_HANDLE(args->event_page_offset));
1033 		if (!mem) {
1034 			pr_err("Can't find BO, offset is 0x%llx\n",
1035 			       args->event_page_offset);
1036 			err = -EINVAL;
1037 			goto out_unlock;
1038 		}
1039 		mutex_unlock(&p->mutex);
1040 
1041 		err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(kfd->kgd,
1042 						mem, &kern_addr, &size);
1043 		if (err) {
1044 			pr_err("Failed to map event page to kernel\n");
1045 			return err;
1046 		}
1047 
1048 		err = kfd_event_page_set(p, kern_addr, size);
1049 		if (err) {
1050 			pr_err("Failed to set event page\n");
1051 			return err;
1052 		}
1053 	}
1054 
1055 	err = kfd_event_create(filp, p, args->event_type,
1056 				args->auto_reset != 0, args->node_id,
1057 				&args->event_id, &args->event_trigger_data,
1058 				&args->event_page_offset,
1059 				&args->event_slot_index);
1060 
1061 	return err;
1062 
1063 out_unlock:
1064 	mutex_unlock(&p->mutex);
1065 	return err;
1066 }
1067 
1068 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
1069 					void *data)
1070 {
1071 	struct kfd_ioctl_destroy_event_args *args = data;
1072 
1073 	return kfd_event_destroy(p, args->event_id);
1074 }
1075 
1076 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
1077 				void *data)
1078 {
1079 	struct kfd_ioctl_set_event_args *args = data;
1080 
1081 	return kfd_set_event(p, args->event_id);
1082 }
1083 
1084 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
1085 				void *data)
1086 {
1087 	struct kfd_ioctl_reset_event_args *args = data;
1088 
1089 	return kfd_reset_event(p, args->event_id);
1090 }
1091 
1092 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
1093 				void *data)
1094 {
1095 	struct kfd_ioctl_wait_events_args *args = data;
1096 	int err;
1097 
1098 	err = kfd_wait_on_events(p, args->num_events,
1099 			(void __user *)args->events_ptr,
1100 			(args->wait_for_all != 0),
1101 			args->timeout, &args->wait_result);
1102 
1103 	return err;
1104 }
1105 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
1106 					struct kfd_process *p, void *data)
1107 {
1108 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
1109 	struct kfd_process_device *pdd;
1110 	struct kfd_dev *dev;
1111 	long err;
1112 
1113 	dev = kfd_device_by_id(args->gpu_id);
1114 	if (!dev)
1115 		return -EINVAL;
1116 
1117 	mutex_lock(&p->mutex);
1118 
1119 	pdd = kfd_bind_process_to_device(dev, p);
1120 	if (IS_ERR(pdd)) {
1121 		err = PTR_ERR(pdd);
1122 		goto bind_process_to_device_fail;
1123 	}
1124 
1125 	pdd->qpd.sh_hidden_private_base = args->va_addr;
1126 
1127 	mutex_unlock(&p->mutex);
1128 
1129 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
1130 	    pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
1131 		dev->kfd2kgd->set_scratch_backing_va(
1132 			dev->kgd, args->va_addr, pdd->qpd.vmid);
1133 
1134 	return 0;
1135 
1136 bind_process_to_device_fail:
1137 	mutex_unlock(&p->mutex);
1138 	return err;
1139 }
1140 
1141 static int kfd_ioctl_get_tile_config(struct file *filep,
1142 		struct kfd_process *p, void *data)
1143 {
1144 	struct kfd_ioctl_get_tile_config_args *args = data;
1145 	struct kfd_dev *dev;
1146 	struct tile_config config;
1147 	int err = 0;
1148 
1149 	dev = kfd_device_by_id(args->gpu_id);
1150 	if (!dev)
1151 		return -EINVAL;
1152 
1153 	amdgpu_amdkfd_get_tile_config(dev->kgd, &config);
1154 
1155 	args->gb_addr_config = config.gb_addr_config;
1156 	args->num_banks = config.num_banks;
1157 	args->num_ranks = config.num_ranks;
1158 
1159 	if (args->num_tile_configs > config.num_tile_configs)
1160 		args->num_tile_configs = config.num_tile_configs;
1161 	err = copy_to_user((void __user *)args->tile_config_ptr,
1162 			config.tile_config_ptr,
1163 			args->num_tile_configs * sizeof(uint32_t));
1164 	if (err) {
1165 		args->num_tile_configs = 0;
1166 		return -EFAULT;
1167 	}
1168 
1169 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
1170 		args->num_macro_tile_configs =
1171 				config.num_macro_tile_configs;
1172 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
1173 			config.macro_tile_config_ptr,
1174 			args->num_macro_tile_configs * sizeof(uint32_t));
1175 	if (err) {
1176 		args->num_macro_tile_configs = 0;
1177 		return -EFAULT;
1178 	}
1179 
1180 	return 0;
1181 }
1182 
1183 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
1184 				void *data)
1185 {
1186 	struct kfd_ioctl_acquire_vm_args *args = data;
1187 	struct kfd_process_device *pdd;
1188 	struct kfd_dev *dev;
1189 	struct file *drm_file;
1190 	int ret;
1191 
1192 	dev = kfd_device_by_id(args->gpu_id);
1193 	if (!dev)
1194 		return -EINVAL;
1195 
1196 	drm_file = fget(args->drm_fd);
1197 	if (!drm_file)
1198 		return -EINVAL;
1199 
1200 	mutex_lock(&p->mutex);
1201 
1202 	pdd = kfd_get_process_device_data(dev, p);
1203 	if (!pdd) {
1204 		ret = -EINVAL;
1205 		goto err_unlock;
1206 	}
1207 
1208 	if (pdd->drm_file) {
1209 		ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
1210 		goto err_unlock;
1211 	}
1212 
1213 	ret = kfd_process_device_init_vm(pdd, drm_file);
1214 	if (ret)
1215 		goto err_unlock;
1216 	/* On success, the PDD keeps the drm_file reference */
1217 	mutex_unlock(&p->mutex);
1218 
1219 	return 0;
1220 
1221 err_unlock:
1222 	mutex_unlock(&p->mutex);
1223 	fput(drm_file);
1224 	return ret;
1225 }
1226 
1227 bool kfd_dev_is_large_bar(struct kfd_dev *dev)
1228 {
1229 	struct kfd_local_mem_info mem_info;
1230 
1231 	if (debug_largebar) {
1232 		pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1233 		return true;
1234 	}
1235 
1236 	if (dev->use_iommu_v2)
1237 		return false;
1238 
1239 	amdgpu_amdkfd_get_local_mem_info(dev->kgd, &mem_info);
1240 	if (mem_info.local_mem_size_private == 0 &&
1241 			mem_info.local_mem_size_public > 0)
1242 		return true;
1243 	return false;
1244 }
1245 
1246 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1247 					struct kfd_process *p, void *data)
1248 {
1249 	struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1250 	struct kfd_process_device *pdd;
1251 	void *mem;
1252 	struct kfd_dev *dev;
1253 	int idr_handle;
1254 	long err;
1255 	uint64_t offset = args->mmap_offset;
1256 	uint32_t flags = args->flags;
1257 
1258 	if (args->size == 0)
1259 		return -EINVAL;
1260 
1261 	dev = kfd_device_by_id(args->gpu_id);
1262 	if (!dev)
1263 		return -EINVAL;
1264 
1265 	if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1266 		(flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1267 		!kfd_dev_is_large_bar(dev)) {
1268 		pr_err("Alloc host visible vram on small bar is not allowed\n");
1269 		return -EINVAL;
1270 	}
1271 
1272 	mutex_lock(&p->mutex);
1273 
1274 	pdd = kfd_bind_process_to_device(dev, p);
1275 	if (IS_ERR(pdd)) {
1276 		err = PTR_ERR(pdd);
1277 		goto err_unlock;
1278 	}
1279 
1280 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1281 		if (args->size != kfd_doorbell_process_slice(dev)) {
1282 			err = -EINVAL;
1283 			goto err_unlock;
1284 		}
1285 		offset = kfd_get_process_doorbells(pdd);
1286 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1287 		if (args->size != PAGE_SIZE) {
1288 			err = -EINVAL;
1289 			goto err_unlock;
1290 		}
1291 		offset = amdgpu_amdkfd_get_mmio_remap_phys_addr(dev->kgd);
1292 		if (!offset) {
1293 			err = -ENOMEM;
1294 			goto err_unlock;
1295 		}
1296 	}
1297 
1298 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1299 		dev->kgd, args->va_addr, args->size,
1300 		pdd->vm, (struct kgd_mem **) &mem, &offset,
1301 		flags);
1302 
1303 	if (err)
1304 		goto err_unlock;
1305 
1306 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1307 	if (idr_handle < 0) {
1308 		err = -EFAULT;
1309 		goto err_free;
1310 	}
1311 
1312 	/* Update the VRAM usage count */
1313 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
1314 		WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size);
1315 
1316 	mutex_unlock(&p->mutex);
1317 
1318 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1319 	args->mmap_offset = offset;
1320 
1321 	/* MMIO is mapped through kfd device
1322 	 * Generate a kfd mmap offset
1323 	 */
1324 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1325 		args->mmap_offset = KFD_MMAP_TYPE_MMIO
1326 					| KFD_MMAP_GPU_ID(args->gpu_id);
1327 
1328 	return 0;
1329 
1330 err_free:
1331 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem, NULL);
1332 err_unlock:
1333 	mutex_unlock(&p->mutex);
1334 	return err;
1335 }
1336 
1337 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1338 					struct kfd_process *p, void *data)
1339 {
1340 	struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1341 	struct kfd_process_device *pdd;
1342 	void *mem;
1343 	struct kfd_dev *dev;
1344 	int ret;
1345 	uint64_t size = 0;
1346 
1347 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1348 	if (!dev)
1349 		return -EINVAL;
1350 
1351 	mutex_lock(&p->mutex);
1352 
1353 	pdd = kfd_get_process_device_data(dev, p);
1354 	if (!pdd) {
1355 		pr_err("Process device data doesn't exist\n");
1356 		ret = -EINVAL;
1357 		goto err_unlock;
1358 	}
1359 
1360 	mem = kfd_process_device_translate_handle(
1361 		pdd, GET_IDR_HANDLE(args->handle));
1362 	if (!mem) {
1363 		ret = -EINVAL;
1364 		goto err_unlock;
1365 	}
1366 
1367 	ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd,
1368 						(struct kgd_mem *)mem, &size);
1369 
1370 	/* If freeing the buffer failed, leave the handle in place for
1371 	 * clean-up during process tear-down.
1372 	 */
1373 	if (!ret)
1374 		kfd_process_device_remove_obj_handle(
1375 			pdd, GET_IDR_HANDLE(args->handle));
1376 
1377 	WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1378 
1379 err_unlock:
1380 	mutex_unlock(&p->mutex);
1381 	return ret;
1382 }
1383 
1384 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1385 					struct kfd_process *p, void *data)
1386 {
1387 	struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1388 	struct kfd_process_device *pdd, *peer_pdd;
1389 	void *mem;
1390 	struct kfd_dev *dev, *peer;
1391 	long err = 0;
1392 	int i;
1393 	uint32_t *devices_arr = NULL;
1394 
1395 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1396 	if (!dev)
1397 		return -EINVAL;
1398 
1399 	if (!args->n_devices) {
1400 		pr_debug("Device IDs array empty\n");
1401 		return -EINVAL;
1402 	}
1403 	if (args->n_success > args->n_devices) {
1404 		pr_debug("n_success exceeds n_devices\n");
1405 		return -EINVAL;
1406 	}
1407 
1408 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1409 				    GFP_KERNEL);
1410 	if (!devices_arr)
1411 		return -ENOMEM;
1412 
1413 	err = copy_from_user(devices_arr,
1414 			     (void __user *)args->device_ids_array_ptr,
1415 			     args->n_devices * sizeof(*devices_arr));
1416 	if (err != 0) {
1417 		err = -EFAULT;
1418 		goto copy_from_user_failed;
1419 	}
1420 
1421 	mutex_lock(&p->mutex);
1422 
1423 	pdd = kfd_bind_process_to_device(dev, p);
1424 	if (IS_ERR(pdd)) {
1425 		err = PTR_ERR(pdd);
1426 		goto bind_process_to_device_failed;
1427 	}
1428 
1429 	mem = kfd_process_device_translate_handle(pdd,
1430 						GET_IDR_HANDLE(args->handle));
1431 	if (!mem) {
1432 		err = -ENOMEM;
1433 		goto get_mem_obj_from_handle_failed;
1434 	}
1435 
1436 	for (i = args->n_success; i < args->n_devices; i++) {
1437 		peer = kfd_device_by_id(devices_arr[i]);
1438 		if (!peer) {
1439 			pr_debug("Getting device by id failed for 0x%x\n",
1440 				 devices_arr[i]);
1441 			err = -EINVAL;
1442 			goto get_mem_obj_from_handle_failed;
1443 		}
1444 
1445 		peer_pdd = kfd_bind_process_to_device(peer, p);
1446 		if (IS_ERR(peer_pdd)) {
1447 			err = PTR_ERR(peer_pdd);
1448 			goto get_mem_obj_from_handle_failed;
1449 		}
1450 		err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1451 			peer->kgd, (struct kgd_mem *)mem, peer_pdd->vm);
1452 		if (err) {
1453 			pr_err("Failed to map to gpu %d/%d\n",
1454 			       i, args->n_devices);
1455 			goto map_memory_to_gpu_failed;
1456 		}
1457 		args->n_success = i+1;
1458 	}
1459 
1460 	mutex_unlock(&p->mutex);
1461 
1462 	err = amdgpu_amdkfd_gpuvm_sync_memory(dev->kgd, (struct kgd_mem *) mem, true);
1463 	if (err) {
1464 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
1465 		goto sync_memory_failed;
1466 	}
1467 
1468 	/* Flush TLBs after waiting for the page table updates to complete */
1469 	for (i = 0; i < args->n_devices; i++) {
1470 		peer = kfd_device_by_id(devices_arr[i]);
1471 		if (WARN_ON_ONCE(!peer))
1472 			continue;
1473 		peer_pdd = kfd_get_process_device_data(peer, p);
1474 		if (WARN_ON_ONCE(!peer_pdd))
1475 			continue;
1476 		kfd_flush_tlb(peer_pdd);
1477 	}
1478 
1479 	kfree(devices_arr);
1480 
1481 	return err;
1482 
1483 bind_process_to_device_failed:
1484 get_mem_obj_from_handle_failed:
1485 map_memory_to_gpu_failed:
1486 	mutex_unlock(&p->mutex);
1487 copy_from_user_failed:
1488 sync_memory_failed:
1489 	kfree(devices_arr);
1490 
1491 	return err;
1492 }
1493 
1494 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1495 					struct kfd_process *p, void *data)
1496 {
1497 	struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1498 	struct kfd_process_device *pdd, *peer_pdd;
1499 	void *mem;
1500 	struct kfd_dev *dev, *peer;
1501 	long err = 0;
1502 	uint32_t *devices_arr = NULL, i;
1503 
1504 	dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1505 	if (!dev)
1506 		return -EINVAL;
1507 
1508 	if (!args->n_devices) {
1509 		pr_debug("Device IDs array empty\n");
1510 		return -EINVAL;
1511 	}
1512 	if (args->n_success > args->n_devices) {
1513 		pr_debug("n_success exceeds n_devices\n");
1514 		return -EINVAL;
1515 	}
1516 
1517 	devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1518 				    GFP_KERNEL);
1519 	if (!devices_arr)
1520 		return -ENOMEM;
1521 
1522 	err = copy_from_user(devices_arr,
1523 			     (void __user *)args->device_ids_array_ptr,
1524 			     args->n_devices * sizeof(*devices_arr));
1525 	if (err != 0) {
1526 		err = -EFAULT;
1527 		goto copy_from_user_failed;
1528 	}
1529 
1530 	mutex_lock(&p->mutex);
1531 
1532 	pdd = kfd_get_process_device_data(dev, p);
1533 	if (!pdd) {
1534 		err = -EINVAL;
1535 		goto bind_process_to_device_failed;
1536 	}
1537 
1538 	mem = kfd_process_device_translate_handle(pdd,
1539 						GET_IDR_HANDLE(args->handle));
1540 	if (!mem) {
1541 		err = -ENOMEM;
1542 		goto get_mem_obj_from_handle_failed;
1543 	}
1544 
1545 	for (i = args->n_success; i < args->n_devices; i++) {
1546 		peer = kfd_device_by_id(devices_arr[i]);
1547 		if (!peer) {
1548 			err = -EINVAL;
1549 			goto get_mem_obj_from_handle_failed;
1550 		}
1551 
1552 		peer_pdd = kfd_get_process_device_data(peer, p);
1553 		if (!peer_pdd) {
1554 			err = -ENODEV;
1555 			goto get_mem_obj_from_handle_failed;
1556 		}
1557 		err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1558 			peer->kgd, (struct kgd_mem *)mem, peer_pdd->vm);
1559 		if (err) {
1560 			pr_err("Failed to unmap from gpu %d/%d\n",
1561 			       i, args->n_devices);
1562 			goto unmap_memory_from_gpu_failed;
1563 		}
1564 		args->n_success = i+1;
1565 	}
1566 	kfree(devices_arr);
1567 
1568 	mutex_unlock(&p->mutex);
1569 
1570 	return 0;
1571 
1572 bind_process_to_device_failed:
1573 get_mem_obj_from_handle_failed:
1574 unmap_memory_from_gpu_failed:
1575 	mutex_unlock(&p->mutex);
1576 copy_from_user_failed:
1577 	kfree(devices_arr);
1578 	return err;
1579 }
1580 
1581 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1582 		struct kfd_process *p, void *data)
1583 {
1584 	int retval;
1585 	struct kfd_ioctl_alloc_queue_gws_args *args = data;
1586 	struct queue *q;
1587 	struct kfd_dev *dev;
1588 
1589 	mutex_lock(&p->mutex);
1590 	q = pqm_get_user_queue(&p->pqm, args->queue_id);
1591 
1592 	if (q) {
1593 		dev = q->device;
1594 	} else {
1595 		retval = -EINVAL;
1596 		goto out_unlock;
1597 	}
1598 
1599 	if (!dev->gws) {
1600 		retval = -ENODEV;
1601 		goto out_unlock;
1602 	}
1603 
1604 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1605 		retval = -ENODEV;
1606 		goto out_unlock;
1607 	}
1608 
1609 	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1610 	mutex_unlock(&p->mutex);
1611 
1612 	args->first_gws = 0;
1613 	return retval;
1614 
1615 out_unlock:
1616 	mutex_unlock(&p->mutex);
1617 	return retval;
1618 }
1619 
1620 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1621 		struct kfd_process *p, void *data)
1622 {
1623 	struct kfd_ioctl_get_dmabuf_info_args *args = data;
1624 	struct kfd_dev *dev = NULL;
1625 	struct kgd_dev *dma_buf_kgd;
1626 	void *metadata_buffer = NULL;
1627 	uint32_t flags;
1628 	unsigned int i;
1629 	int r;
1630 
1631 	/* Find a KFD GPU device that supports the get_dmabuf_info query */
1632 	for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1633 		if (dev)
1634 			break;
1635 	if (!dev)
1636 		return -EINVAL;
1637 
1638 	if (args->metadata_ptr) {
1639 		metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1640 		if (!metadata_buffer)
1641 			return -ENOMEM;
1642 	}
1643 
1644 	/* Get dmabuf info from KGD */
1645 	r = amdgpu_amdkfd_get_dmabuf_info(dev->kgd, args->dmabuf_fd,
1646 					  &dma_buf_kgd, &args->size,
1647 					  metadata_buffer, args->metadata_size,
1648 					  &args->metadata_size, &flags);
1649 	if (r)
1650 		goto exit;
1651 
1652 	/* Reverse-lookup gpu_id from kgd pointer */
1653 	dev = kfd_device_by_kgd(dma_buf_kgd);
1654 	if (!dev) {
1655 		r = -EINVAL;
1656 		goto exit;
1657 	}
1658 	args->gpu_id = dev->id;
1659 	args->flags = flags;
1660 
1661 	/* Copy metadata buffer to user mode */
1662 	if (metadata_buffer) {
1663 		r = copy_to_user((void __user *)args->metadata_ptr,
1664 				 metadata_buffer, args->metadata_size);
1665 		if (r != 0)
1666 			r = -EFAULT;
1667 	}
1668 
1669 exit:
1670 	kfree(metadata_buffer);
1671 
1672 	return r;
1673 }
1674 
1675 static int kfd_ioctl_import_dmabuf(struct file *filep,
1676 				   struct kfd_process *p, void *data)
1677 {
1678 	struct kfd_ioctl_import_dmabuf_args *args = data;
1679 	struct kfd_process_device *pdd;
1680 	struct dma_buf *dmabuf;
1681 	struct kfd_dev *dev;
1682 	int idr_handle;
1683 	uint64_t size;
1684 	void *mem;
1685 	int r;
1686 
1687 	dev = kfd_device_by_id(args->gpu_id);
1688 	if (!dev)
1689 		return -EINVAL;
1690 
1691 	dmabuf = dma_buf_get(args->dmabuf_fd);
1692 	if (IS_ERR(dmabuf))
1693 		return PTR_ERR(dmabuf);
1694 
1695 	mutex_lock(&p->mutex);
1696 
1697 	pdd = kfd_bind_process_to_device(dev, p);
1698 	if (IS_ERR(pdd)) {
1699 		r = PTR_ERR(pdd);
1700 		goto err_unlock;
1701 	}
1702 
1703 	r = amdgpu_amdkfd_gpuvm_import_dmabuf(dev->kgd, dmabuf,
1704 					      args->va_addr, pdd->vm,
1705 					      (struct kgd_mem **)&mem, &size,
1706 					      NULL);
1707 	if (r)
1708 		goto err_unlock;
1709 
1710 	idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1711 	if (idr_handle < 0) {
1712 		r = -EFAULT;
1713 		goto err_free;
1714 	}
1715 
1716 	mutex_unlock(&p->mutex);
1717 	dma_buf_put(dmabuf);
1718 
1719 	args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1720 
1721 	return 0;
1722 
1723 err_free:
1724 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem, NULL);
1725 err_unlock:
1726 	mutex_unlock(&p->mutex);
1727 	dma_buf_put(dmabuf);
1728 	return r;
1729 }
1730 
1731 /* Handle requests for watching SMI events */
1732 static int kfd_ioctl_smi_events(struct file *filep,
1733 				struct kfd_process *p, void *data)
1734 {
1735 	struct kfd_ioctl_smi_events_args *args = data;
1736 	struct kfd_dev *dev;
1737 
1738 	dev = kfd_device_by_id(args->gpuid);
1739 	if (!dev)
1740 		return -EINVAL;
1741 
1742 	return kfd_smi_event_open(dev, &args->anon_fd);
1743 }
1744 
1745 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
1746 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
1747 			    .cmd_drv = 0, .name = #ioctl}
1748 
1749 /** Ioctl table */
1750 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
1751 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
1752 			kfd_ioctl_get_version, 0),
1753 
1754 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
1755 			kfd_ioctl_create_queue, 0),
1756 
1757 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
1758 			kfd_ioctl_destroy_queue, 0),
1759 
1760 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
1761 			kfd_ioctl_set_memory_policy, 0),
1762 
1763 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
1764 			kfd_ioctl_get_clock_counters, 0),
1765 
1766 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
1767 			kfd_ioctl_get_process_apertures, 0),
1768 
1769 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
1770 			kfd_ioctl_update_queue, 0),
1771 
1772 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
1773 			kfd_ioctl_create_event, 0),
1774 
1775 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
1776 			kfd_ioctl_destroy_event, 0),
1777 
1778 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
1779 			kfd_ioctl_set_event, 0),
1780 
1781 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
1782 			kfd_ioctl_reset_event, 0),
1783 
1784 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
1785 			kfd_ioctl_wait_events, 0),
1786 
1787 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER,
1788 			kfd_ioctl_dbg_register, 0),
1789 
1790 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER,
1791 			kfd_ioctl_dbg_unregister, 0),
1792 
1793 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH,
1794 			kfd_ioctl_dbg_address_watch, 0),
1795 
1796 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL,
1797 			kfd_ioctl_dbg_wave_control, 0),
1798 
1799 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
1800 			kfd_ioctl_set_scratch_backing_va, 0),
1801 
1802 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
1803 			kfd_ioctl_get_tile_config, 0),
1804 
1805 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
1806 			kfd_ioctl_set_trap_handler, 0),
1807 
1808 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
1809 			kfd_ioctl_get_process_apertures_new, 0),
1810 
1811 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
1812 			kfd_ioctl_acquire_vm, 0),
1813 
1814 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
1815 			kfd_ioctl_alloc_memory_of_gpu, 0),
1816 
1817 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
1818 			kfd_ioctl_free_memory_of_gpu, 0),
1819 
1820 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
1821 			kfd_ioctl_map_memory_to_gpu, 0),
1822 
1823 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
1824 			kfd_ioctl_unmap_memory_from_gpu, 0),
1825 
1826 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
1827 			kfd_ioctl_set_cu_mask, 0),
1828 
1829 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
1830 			kfd_ioctl_get_queue_wave_state, 0),
1831 
1832 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
1833 				kfd_ioctl_get_dmabuf_info, 0),
1834 
1835 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
1836 				kfd_ioctl_import_dmabuf, 0),
1837 
1838 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
1839 			kfd_ioctl_alloc_queue_gws, 0),
1840 
1841 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
1842 			kfd_ioctl_smi_events, 0),
1843 };
1844 
1845 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
1846 
1847 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
1848 {
1849 	struct kfd_process *process;
1850 	amdkfd_ioctl_t *func;
1851 	const struct amdkfd_ioctl_desc *ioctl = NULL;
1852 	unsigned int nr = _IOC_NR(cmd);
1853 	char stack_kdata[128];
1854 	char *kdata = NULL;
1855 	unsigned int usize, asize;
1856 	int retcode = -EINVAL;
1857 
1858 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
1859 		goto err_i1;
1860 
1861 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
1862 		u32 amdkfd_size;
1863 
1864 		ioctl = &amdkfd_ioctls[nr];
1865 
1866 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
1867 		usize = asize = _IOC_SIZE(cmd);
1868 		if (amdkfd_size > asize)
1869 			asize = amdkfd_size;
1870 
1871 		cmd = ioctl->cmd;
1872 	} else
1873 		goto err_i1;
1874 
1875 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
1876 
1877 	/* Get the process struct from the filep. Only the process
1878 	 * that opened /dev/kfd can use the file descriptor. Child
1879 	 * processes need to create their own KFD device context.
1880 	 */
1881 	process = filep->private_data;
1882 	if (process->lead_thread != current->group_leader) {
1883 		dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
1884 		retcode = -EBADF;
1885 		goto err_i1;
1886 	}
1887 
1888 	/* Do not trust userspace, use our own definition */
1889 	func = ioctl->func;
1890 
1891 	if (unlikely(!func)) {
1892 		dev_dbg(kfd_device, "no function\n");
1893 		retcode = -EINVAL;
1894 		goto err_i1;
1895 	}
1896 
1897 	if (cmd & (IOC_IN | IOC_OUT)) {
1898 		if (asize <= sizeof(stack_kdata)) {
1899 			kdata = stack_kdata;
1900 		} else {
1901 			kdata = kmalloc(asize, GFP_KERNEL);
1902 			if (!kdata) {
1903 				retcode = -ENOMEM;
1904 				goto err_i1;
1905 			}
1906 		}
1907 		if (asize > usize)
1908 			memset(kdata + usize, 0, asize - usize);
1909 	}
1910 
1911 	if (cmd & IOC_IN) {
1912 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
1913 			retcode = -EFAULT;
1914 			goto err_i1;
1915 		}
1916 	} else if (cmd & IOC_OUT) {
1917 		memset(kdata, 0, usize);
1918 	}
1919 
1920 	retcode = func(filep, process, kdata);
1921 
1922 	if (cmd & IOC_OUT)
1923 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
1924 			retcode = -EFAULT;
1925 
1926 err_i1:
1927 	if (!ioctl)
1928 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
1929 			  task_pid_nr(current), cmd, nr);
1930 
1931 	if (kdata != stack_kdata)
1932 		kfree(kdata);
1933 
1934 	if (retcode)
1935 		dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
1936 				nr, arg, retcode);
1937 
1938 	return retcode;
1939 }
1940 
1941 static int kfd_mmio_mmap(struct kfd_dev *dev, struct kfd_process *process,
1942 		      struct vm_area_struct *vma)
1943 {
1944 	phys_addr_t address;
1945 	int ret;
1946 
1947 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1948 		return -EINVAL;
1949 
1950 	address = amdgpu_amdkfd_get_mmio_remap_phys_addr(dev->kgd);
1951 
1952 	vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
1953 				VM_DONTDUMP | VM_PFNMAP;
1954 
1955 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1956 
1957 	pr_debug("pasid 0x%x mapping mmio page\n"
1958 		 "     target user address == 0x%08llX\n"
1959 		 "     physical address    == 0x%08llX\n"
1960 		 "     vm_flags            == 0x%04lX\n"
1961 		 "     size                == 0x%04lX\n",
1962 		 process->pasid, (unsigned long long) vma->vm_start,
1963 		 address, vma->vm_flags, PAGE_SIZE);
1964 
1965 	ret = io_remap_pfn_range(vma,
1966 				vma->vm_start,
1967 				address >> PAGE_SHIFT,
1968 				PAGE_SIZE,
1969 				vma->vm_page_prot);
1970 	return ret;
1971 }
1972 
1973 
1974 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
1975 {
1976 	struct kfd_process *process;
1977 	struct kfd_dev *dev = NULL;
1978 	unsigned long mmap_offset;
1979 	unsigned int gpu_id;
1980 
1981 	process = kfd_get_process(current);
1982 	if (IS_ERR(process))
1983 		return PTR_ERR(process);
1984 
1985 	mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
1986 	gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
1987 	if (gpu_id)
1988 		dev = kfd_device_by_id(gpu_id);
1989 
1990 	switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
1991 	case KFD_MMAP_TYPE_DOORBELL:
1992 		if (!dev)
1993 			return -ENODEV;
1994 		return kfd_doorbell_mmap(dev, process, vma);
1995 
1996 	case KFD_MMAP_TYPE_EVENTS:
1997 		return kfd_event_mmap(process, vma);
1998 
1999 	case KFD_MMAP_TYPE_RESERVED_MEM:
2000 		if (!dev)
2001 			return -ENODEV;
2002 		return kfd_reserved_mem_mmap(dev, process, vma);
2003 	case KFD_MMAP_TYPE_MMIO:
2004 		if (!dev)
2005 			return -ENODEV;
2006 		return kfd_mmio_mmap(dev, process, vma);
2007 	}
2008 
2009 	return -EFAULT;
2010 }
2011