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/sched.h>
28 #include <linux/slab.h>
29 #include <linux/uaccess.h>
30 #include <linux/compat.h>
31 #include <uapi/linux/kfd_ioctl.h>
32 #include <linux/time.h>
33 #include <linux/mm.h>
34 #include <linux/mman.h>
35 #include <asm/processor.h>
36 #include "kfd_priv.h"
37 #include "kfd_device_queue_manager.h"
38 #include "kfd_dbgmgr.h"
39 
40 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
41 static int kfd_open(struct inode *, struct file *);
42 static int kfd_mmap(struct file *, struct vm_area_struct *);
43 
44 static const char kfd_dev_name[] = "kfd";
45 
46 static const struct file_operations kfd_fops = {
47 	.owner = THIS_MODULE,
48 	.unlocked_ioctl = kfd_ioctl,
49 	.compat_ioctl = kfd_ioctl,
50 	.open = kfd_open,
51 	.mmap = kfd_mmap,
52 };
53 
54 static int kfd_char_dev_major = -1;
55 static struct class *kfd_class;
56 struct device *kfd_device;
57 
58 int kfd_chardev_init(void)
59 {
60 	int err = 0;
61 
62 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
63 	err = kfd_char_dev_major;
64 	if (err < 0)
65 		goto err_register_chrdev;
66 
67 	kfd_class = class_create(THIS_MODULE, kfd_dev_name);
68 	err = PTR_ERR(kfd_class);
69 	if (IS_ERR(kfd_class))
70 		goto err_class_create;
71 
72 	kfd_device = device_create(kfd_class, NULL,
73 					MKDEV(kfd_char_dev_major, 0),
74 					NULL, kfd_dev_name);
75 	err = PTR_ERR(kfd_device);
76 	if (IS_ERR(kfd_device))
77 		goto err_device_create;
78 
79 	return 0;
80 
81 err_device_create:
82 	class_destroy(kfd_class);
83 err_class_create:
84 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
85 err_register_chrdev:
86 	return err;
87 }
88 
89 void kfd_chardev_exit(void)
90 {
91 	device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
92 	class_destroy(kfd_class);
93 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
94 }
95 
96 struct device *kfd_chardev(void)
97 {
98 	return kfd_device;
99 }
100 
101 
102 static int kfd_open(struct inode *inode, struct file *filep)
103 {
104 	struct kfd_process *process;
105 	bool is_32bit_user_mode;
106 
107 	if (iminor(inode) != 0)
108 		return -ENODEV;
109 
110 	is_32bit_user_mode = is_compat_task();
111 
112 	if (is_32bit_user_mode == true) {
113 		dev_warn(kfd_device,
114 			"Process %d (32-bit) failed to open /dev/kfd\n"
115 			"32-bit processes are not supported by amdkfd\n",
116 			current->pid);
117 		return -EPERM;
118 	}
119 
120 	process = kfd_create_process(current);
121 	if (IS_ERR(process))
122 		return PTR_ERR(process);
123 
124 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
125 		process->pasid, process->is_32bit_user_mode);
126 
127 	return 0;
128 }
129 
130 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
131 					void *data)
132 {
133 	struct kfd_ioctl_get_version_args *args = data;
134 	int err = 0;
135 
136 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
137 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
138 
139 	return err;
140 }
141 
142 static int set_queue_properties_from_user(struct queue_properties *q_properties,
143 				struct kfd_ioctl_create_queue_args *args)
144 {
145 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
146 		pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
147 		return -EINVAL;
148 	}
149 
150 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
151 		pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
152 		return -EINVAL;
153 	}
154 
155 	if ((args->ring_base_address) &&
156 		(!access_ok(VERIFY_WRITE,
157 			(const void __user *) args->ring_base_address,
158 			sizeof(uint64_t)))) {
159 		pr_err("kfd: can't access ring base address\n");
160 		return -EFAULT;
161 	}
162 
163 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
164 		pr_err("kfd: ring size must be a power of 2 or 0\n");
165 		return -EINVAL;
166 	}
167 
168 	if (!access_ok(VERIFY_WRITE,
169 			(const void __user *) args->read_pointer_address,
170 			sizeof(uint32_t))) {
171 		pr_err("kfd: can't access read pointer\n");
172 		return -EFAULT;
173 	}
174 
175 	if (!access_ok(VERIFY_WRITE,
176 			(const void __user *) args->write_pointer_address,
177 			sizeof(uint32_t))) {
178 		pr_err("kfd: can't access write pointer\n");
179 		return -EFAULT;
180 	}
181 
182 	if (args->eop_buffer_address &&
183 		!access_ok(VERIFY_WRITE,
184 			(const void __user *) args->eop_buffer_address,
185 			sizeof(uint32_t))) {
186 		pr_debug("kfd: can't access eop buffer");
187 		return -EFAULT;
188 	}
189 
190 	if (args->ctx_save_restore_address &&
191 		!access_ok(VERIFY_WRITE,
192 			(const void __user *) args->ctx_save_restore_address,
193 			sizeof(uint32_t))) {
194 		pr_debug("kfd: can't access ctx save restore buffer");
195 		return -EFAULT;
196 	}
197 
198 	q_properties->is_interop = false;
199 	q_properties->queue_percent = args->queue_percentage;
200 	q_properties->priority = args->queue_priority;
201 	q_properties->queue_address = args->ring_base_address;
202 	q_properties->queue_size = args->ring_size;
203 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
204 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
205 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
206 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
207 	q_properties->ctx_save_restore_area_address =
208 			args->ctx_save_restore_address;
209 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
210 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
211 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
212 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
213 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
214 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
215 	else
216 		return -ENOTSUPP;
217 
218 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
219 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
220 	else
221 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
222 
223 	pr_debug("Queue Percentage (%d, %d)\n",
224 			q_properties->queue_percent, args->queue_percentage);
225 
226 	pr_debug("Queue Priority (%d, %d)\n",
227 			q_properties->priority, args->queue_priority);
228 
229 	pr_debug("Queue Address (0x%llX, 0x%llX)\n",
230 			q_properties->queue_address, args->ring_base_address);
231 
232 	pr_debug("Queue Size (0x%llX, %u)\n",
233 			q_properties->queue_size, args->ring_size);
234 
235 	pr_debug("Queue r/w Pointers (0x%llX, 0x%llX)\n",
236 			(uint64_t) q_properties->read_ptr,
237 			(uint64_t) q_properties->write_ptr);
238 
239 	pr_debug("Queue Format (%d)\n", q_properties->format);
240 
241 	pr_debug("Queue EOP (0x%llX)\n", q_properties->eop_ring_buffer_address);
242 
243 	pr_debug("Queue CTX save arex (0x%llX)\n",
244 			q_properties->ctx_save_restore_area_address);
245 
246 	return 0;
247 }
248 
249 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
250 					void *data)
251 {
252 	struct kfd_ioctl_create_queue_args *args = data;
253 	struct kfd_dev *dev;
254 	int err = 0;
255 	unsigned int queue_id;
256 	struct kfd_process_device *pdd;
257 	struct queue_properties q_properties;
258 
259 	memset(&q_properties, 0, sizeof(struct queue_properties));
260 
261 	pr_debug("kfd: creating queue ioctl\n");
262 
263 	err = set_queue_properties_from_user(&q_properties, args);
264 	if (err)
265 		return err;
266 
267 	pr_debug("kfd: looking for gpu id 0x%x\n", args->gpu_id);
268 	dev = kfd_device_by_id(args->gpu_id);
269 	if (dev == NULL) {
270 		pr_debug("kfd: gpu id 0x%x was not found\n", args->gpu_id);
271 		return -EINVAL;
272 	}
273 
274 	mutex_lock(&p->mutex);
275 
276 	pdd = kfd_bind_process_to_device(dev, p);
277 	if (IS_ERR(pdd)) {
278 		err = -ESRCH;
279 		goto err_bind_process;
280 	}
281 
282 	pr_debug("kfd: creating queue for PASID %d on GPU 0x%x\n",
283 			p->pasid,
284 			dev->id);
285 
286 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties,
287 				0, q_properties.type, &queue_id);
288 	if (err != 0)
289 		goto err_create_queue;
290 
291 	args->queue_id = queue_id;
292 
293 
294 	/* Return gpu_id as doorbell offset for mmap usage */
295 	args->doorbell_offset = (KFD_MMAP_DOORBELL_MASK | args->gpu_id);
296 	args->doorbell_offset <<= PAGE_SHIFT;
297 
298 	mutex_unlock(&p->mutex);
299 
300 	pr_debug("kfd: queue id %d was created successfully\n", args->queue_id);
301 
302 	pr_debug("ring buffer address == 0x%016llX\n",
303 			args->ring_base_address);
304 
305 	pr_debug("read ptr address    == 0x%016llX\n",
306 			args->read_pointer_address);
307 
308 	pr_debug("write ptr address   == 0x%016llX\n",
309 			args->write_pointer_address);
310 
311 	return 0;
312 
313 err_create_queue:
314 err_bind_process:
315 	mutex_unlock(&p->mutex);
316 	return err;
317 }
318 
319 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
320 					void *data)
321 {
322 	int retval;
323 	struct kfd_ioctl_destroy_queue_args *args = data;
324 
325 	pr_debug("kfd: destroying queue id %d for PASID %d\n",
326 				args->queue_id,
327 				p->pasid);
328 
329 	mutex_lock(&p->mutex);
330 
331 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
332 
333 	mutex_unlock(&p->mutex);
334 	return retval;
335 }
336 
337 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
338 					void *data)
339 {
340 	int retval;
341 	struct kfd_ioctl_update_queue_args *args = data;
342 	struct queue_properties properties;
343 
344 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
345 		pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
346 		return -EINVAL;
347 	}
348 
349 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
350 		pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
351 		return -EINVAL;
352 	}
353 
354 	if ((args->ring_base_address) &&
355 		(!access_ok(VERIFY_WRITE,
356 			(const void __user *) args->ring_base_address,
357 			sizeof(uint64_t)))) {
358 		pr_err("kfd: can't access ring base address\n");
359 		return -EFAULT;
360 	}
361 
362 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
363 		pr_err("kfd: ring size must be a power of 2 or 0\n");
364 		return -EINVAL;
365 	}
366 
367 	properties.queue_address = args->ring_base_address;
368 	properties.queue_size = args->ring_size;
369 	properties.queue_percent = args->queue_percentage;
370 	properties.priority = args->queue_priority;
371 
372 	pr_debug("kfd: updating queue id %d for PASID %d\n",
373 			args->queue_id, p->pasid);
374 
375 	mutex_lock(&p->mutex);
376 
377 	retval = pqm_update_queue(&p->pqm, args->queue_id, &properties);
378 
379 	mutex_unlock(&p->mutex);
380 
381 	return retval;
382 }
383 
384 static int kfd_ioctl_set_memory_policy(struct file *filep,
385 					struct kfd_process *p, void *data)
386 {
387 	struct kfd_ioctl_set_memory_policy_args *args = data;
388 	struct kfd_dev *dev;
389 	int err = 0;
390 	struct kfd_process_device *pdd;
391 	enum cache_policy default_policy, alternate_policy;
392 
393 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
394 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
395 		return -EINVAL;
396 	}
397 
398 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
399 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
400 		return -EINVAL;
401 	}
402 
403 	dev = kfd_device_by_id(args->gpu_id);
404 	if (dev == NULL)
405 		return -EINVAL;
406 
407 	mutex_lock(&p->mutex);
408 
409 	pdd = kfd_bind_process_to_device(dev, p);
410 	if (IS_ERR(pdd)) {
411 		err = -ESRCH;
412 		goto out;
413 	}
414 
415 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
416 			 ? cache_policy_coherent : cache_policy_noncoherent;
417 
418 	alternate_policy =
419 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
420 		   ? cache_policy_coherent : cache_policy_noncoherent;
421 
422 	if (!dev->dqm->ops.set_cache_memory_policy(dev->dqm,
423 				&pdd->qpd,
424 				default_policy,
425 				alternate_policy,
426 				(void __user *)args->alternate_aperture_base,
427 				args->alternate_aperture_size))
428 		err = -EINVAL;
429 
430 out:
431 	mutex_unlock(&p->mutex);
432 
433 	return err;
434 }
435 
436 static int kfd_ioctl_dbg_register(struct file *filep,
437 				struct kfd_process *p, void *data)
438 {
439 	struct kfd_ioctl_dbg_register_args *args = data;
440 	struct kfd_dev *dev;
441 	struct kfd_dbgmgr *dbgmgr_ptr;
442 	struct kfd_process_device *pdd;
443 	bool create_ok;
444 	long status = 0;
445 
446 	dev = kfd_device_by_id(args->gpu_id);
447 	if (dev == NULL)
448 		return -EINVAL;
449 
450 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
451 		pr_debug("kfd_ioctl_dbg_register not supported on CZ\n");
452 		return -EINVAL;
453 	}
454 
455 	mutex_lock(kfd_get_dbgmgr_mutex());
456 	mutex_lock(&p->mutex);
457 
458 	/*
459 	 * make sure that we have pdd, if this the first queue created for
460 	 * this process
461 	 */
462 	pdd = kfd_bind_process_to_device(dev, p);
463 	if (IS_ERR(pdd)) {
464 		mutex_unlock(&p->mutex);
465 		mutex_unlock(kfd_get_dbgmgr_mutex());
466 		return PTR_ERR(pdd);
467 	}
468 
469 	if (dev->dbgmgr == NULL) {
470 		/* In case of a legal call, we have no dbgmgr yet */
471 		create_ok = kfd_dbgmgr_create(&dbgmgr_ptr, dev);
472 		if (create_ok) {
473 			status = kfd_dbgmgr_register(dbgmgr_ptr, p);
474 			if (status != 0)
475 				kfd_dbgmgr_destroy(dbgmgr_ptr);
476 			else
477 				dev->dbgmgr = dbgmgr_ptr;
478 		}
479 	} else {
480 		pr_debug("debugger already registered\n");
481 		status = -EINVAL;
482 	}
483 
484 	mutex_unlock(&p->mutex);
485 	mutex_unlock(kfd_get_dbgmgr_mutex());
486 
487 	return status;
488 }
489 
490 static int kfd_ioctl_dbg_unrgesiter(struct file *filep,
491 				struct kfd_process *p, void *data)
492 {
493 	struct kfd_ioctl_dbg_unregister_args *args = data;
494 	struct kfd_dev *dev;
495 	long status;
496 
497 	dev = kfd_device_by_id(args->gpu_id);
498 	if (dev == NULL)
499 		return -EINVAL;
500 
501 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
502 		pr_debug("kfd_ioctl_dbg_unrgesiter not supported on CZ\n");
503 		return -EINVAL;
504 	}
505 
506 	mutex_lock(kfd_get_dbgmgr_mutex());
507 
508 	status = kfd_dbgmgr_unregister(dev->dbgmgr, p);
509 	if (status == 0) {
510 		kfd_dbgmgr_destroy(dev->dbgmgr);
511 		dev->dbgmgr = NULL;
512 	}
513 
514 	mutex_unlock(kfd_get_dbgmgr_mutex());
515 
516 	return status;
517 }
518 
519 /*
520  * Parse and generate variable size data structure for address watch.
521  * Total size of the buffer and # watch points is limited in order
522  * to prevent kernel abuse. (no bearing to the much smaller HW limitation
523  * which is enforced by dbgdev module)
524  * please also note that the watch address itself are not "copied from user",
525  * since it be set into the HW in user mode values.
526  *
527  */
528 static int kfd_ioctl_dbg_address_watch(struct file *filep,
529 					struct kfd_process *p, void *data)
530 {
531 	struct kfd_ioctl_dbg_address_watch_args *args = data;
532 	struct kfd_dev *dev;
533 	struct dbg_address_watch_info aw_info;
534 	unsigned char *args_buff;
535 	long status;
536 	void __user *cmd_from_user;
537 	uint64_t watch_mask_value = 0;
538 	unsigned int args_idx = 0;
539 
540 	memset((void *) &aw_info, 0, sizeof(struct dbg_address_watch_info));
541 
542 	dev = kfd_device_by_id(args->gpu_id);
543 	if (dev == NULL)
544 		return -EINVAL;
545 
546 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
547 		pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
548 		return -EINVAL;
549 	}
550 
551 	cmd_from_user = (void __user *) args->content_ptr;
552 
553 	/* Validate arguments */
554 
555 	if ((args->buf_size_in_bytes > MAX_ALLOWED_AW_BUFF_SIZE) ||
556 		(args->buf_size_in_bytes <= sizeof(*args) + sizeof(int) * 2) ||
557 		(cmd_from_user == NULL))
558 		return -EINVAL;
559 
560 	/* this is the actual buffer to work with */
561 
562 	args_buff = kmalloc(args->buf_size_in_bytes -
563 					sizeof(*args), GFP_KERNEL);
564 	if (args_buff == NULL)
565 		return -ENOMEM;
566 
567 	status = copy_from_user(args_buff, cmd_from_user,
568 				args->buf_size_in_bytes - sizeof(*args));
569 
570 	if (status != 0) {
571 		pr_debug("Failed to copy address watch user data\n");
572 		kfree(args_buff);
573 		return -EINVAL;
574 	}
575 
576 	aw_info.process = p;
577 
578 	aw_info.num_watch_points = *((uint32_t *)(&args_buff[args_idx]));
579 	args_idx += sizeof(aw_info.num_watch_points);
580 
581 	aw_info.watch_mode = (enum HSA_DBG_WATCH_MODE *) &args_buff[args_idx];
582 	args_idx += sizeof(enum HSA_DBG_WATCH_MODE) * aw_info.num_watch_points;
583 
584 	/*
585 	 * set watch address base pointer to point on the array base
586 	 * within args_buff
587 	 */
588 	aw_info.watch_address = (uint64_t *) &args_buff[args_idx];
589 
590 	/* skip over the addresses buffer */
591 	args_idx += sizeof(aw_info.watch_address) * aw_info.num_watch_points;
592 
593 	if (args_idx >= args->buf_size_in_bytes - sizeof(*args)) {
594 		kfree(args_buff);
595 		return -EINVAL;
596 	}
597 
598 	watch_mask_value = (uint64_t) args_buff[args_idx];
599 
600 	if (watch_mask_value > 0) {
601 		/*
602 		 * There is an array of masks.
603 		 * set watch mask base pointer to point on the array base
604 		 * within args_buff
605 		 */
606 		aw_info.watch_mask = (uint64_t *) &args_buff[args_idx];
607 
608 		/* skip over the masks buffer */
609 		args_idx += sizeof(aw_info.watch_mask) *
610 				aw_info.num_watch_points;
611 	} else {
612 		/* just the NULL mask, set to NULL and skip over it */
613 		aw_info.watch_mask = NULL;
614 		args_idx += sizeof(aw_info.watch_mask);
615 	}
616 
617 	if (args_idx >= args->buf_size_in_bytes - sizeof(args)) {
618 		kfree(args_buff);
619 		return -EINVAL;
620 	}
621 
622 	/* Currently HSA Event is not supported for DBG */
623 	aw_info.watch_event = NULL;
624 
625 	mutex_lock(kfd_get_dbgmgr_mutex());
626 
627 	status = kfd_dbgmgr_address_watch(dev->dbgmgr, &aw_info);
628 
629 	mutex_unlock(kfd_get_dbgmgr_mutex());
630 
631 	kfree(args_buff);
632 
633 	return status;
634 }
635 
636 /* Parse and generate fixed size data structure for wave control */
637 static int kfd_ioctl_dbg_wave_control(struct file *filep,
638 					struct kfd_process *p, void *data)
639 {
640 	struct kfd_ioctl_dbg_wave_control_args *args = data;
641 	struct kfd_dev *dev;
642 	struct dbg_wave_control_info wac_info;
643 	unsigned char *args_buff;
644 	uint32_t computed_buff_size;
645 	long status;
646 	void __user *cmd_from_user;
647 	unsigned int args_idx = 0;
648 
649 	memset((void *) &wac_info, 0, sizeof(struct dbg_wave_control_info));
650 
651 	/* we use compact form, independent of the packing attribute value */
652 	computed_buff_size = sizeof(*args) +
653 				sizeof(wac_info.mode) +
654 				sizeof(wac_info.operand) +
655 				sizeof(wac_info.dbgWave_msg.DbgWaveMsg) +
656 				sizeof(wac_info.dbgWave_msg.MemoryVA) +
657 				sizeof(wac_info.trapId);
658 
659 	dev = kfd_device_by_id(args->gpu_id);
660 	if (dev == NULL)
661 		return -EINVAL;
662 
663 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
664 		pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
665 		return -EINVAL;
666 	}
667 
668 	/* input size must match the computed "compact" size */
669 	if (args->buf_size_in_bytes != computed_buff_size) {
670 		pr_debug("size mismatch, computed : actual %u : %u\n",
671 				args->buf_size_in_bytes, computed_buff_size);
672 		return -EINVAL;
673 	}
674 
675 	cmd_from_user = (void __user *) args->content_ptr;
676 
677 	if (cmd_from_user == NULL)
678 		return -EINVAL;
679 
680 	/* this is the actual buffer to work with */
681 
682 	args_buff = kmalloc(args->buf_size_in_bytes - sizeof(*args),
683 			GFP_KERNEL);
684 
685 	if (args_buff == NULL)
686 		return -ENOMEM;
687 
688 	/* Now copy the entire buffer from user */
689 	status = copy_from_user(args_buff, cmd_from_user,
690 				args->buf_size_in_bytes - sizeof(*args));
691 	if (status != 0) {
692 		pr_debug("Failed to copy wave control user data\n");
693 		kfree(args_buff);
694 		return -EINVAL;
695 	}
696 
697 	/* move ptr to the start of the "pay-load" area */
698 	wac_info.process = p;
699 
700 	wac_info.operand = *((enum HSA_DBG_WAVEOP *)(&args_buff[args_idx]));
701 	args_idx += sizeof(wac_info.operand);
702 
703 	wac_info.mode = *((enum HSA_DBG_WAVEMODE *)(&args_buff[args_idx]));
704 	args_idx += sizeof(wac_info.mode);
705 
706 	wac_info.trapId = *((uint32_t *)(&args_buff[args_idx]));
707 	args_idx += sizeof(wac_info.trapId);
708 
709 	wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value =
710 					*((uint32_t *)(&args_buff[args_idx]));
711 	wac_info.dbgWave_msg.MemoryVA = NULL;
712 
713 	mutex_lock(kfd_get_dbgmgr_mutex());
714 
715 	pr_debug("Calling dbg manager process %p, operand %u, mode %u, trapId %u, message %u\n",
716 			wac_info.process, wac_info.operand,
717 			wac_info.mode, wac_info.trapId,
718 			wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value);
719 
720 	status = kfd_dbgmgr_wave_control(dev->dbgmgr, &wac_info);
721 
722 	pr_debug("Returned status of dbg manager is %ld\n", status);
723 
724 	mutex_unlock(kfd_get_dbgmgr_mutex());
725 
726 	kfree(args_buff);
727 
728 	return status;
729 }
730 
731 static int kfd_ioctl_get_clock_counters(struct file *filep,
732 				struct kfd_process *p, void *data)
733 {
734 	struct kfd_ioctl_get_clock_counters_args *args = data;
735 	struct kfd_dev *dev;
736 	struct timespec64 time;
737 
738 	dev = kfd_device_by_id(args->gpu_id);
739 	if (dev == NULL)
740 		return -EINVAL;
741 
742 	/* Reading GPU clock counter from KGD */
743 	args->gpu_clock_counter =
744 		dev->kfd2kgd->get_gpu_clock_counter(dev->kgd);
745 
746 	/* No access to rdtsc. Using raw monotonic time */
747 	getrawmonotonic64(&time);
748 	args->cpu_clock_counter = (uint64_t)timespec64_to_ns(&time);
749 
750 	get_monotonic_boottime64(&time);
751 	args->system_clock_counter = (uint64_t)timespec64_to_ns(&time);
752 
753 	/* Since the counter is in nano-seconds we use 1GHz frequency */
754 	args->system_clock_freq = 1000000000;
755 
756 	return 0;
757 }
758 
759 
760 static int kfd_ioctl_get_process_apertures(struct file *filp,
761 				struct kfd_process *p, void *data)
762 {
763 	struct kfd_ioctl_get_process_apertures_args *args = data;
764 	struct kfd_process_device_apertures *pAperture;
765 	struct kfd_process_device *pdd;
766 
767 	dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
768 
769 	args->num_of_nodes = 0;
770 
771 	mutex_lock(&p->mutex);
772 
773 	/*if the process-device list isn't empty*/
774 	if (kfd_has_process_device_data(p)) {
775 		/* Run over all pdd of the process */
776 		pdd = kfd_get_first_process_device_data(p);
777 		do {
778 			pAperture =
779 				&args->process_apertures[args->num_of_nodes];
780 			pAperture->gpu_id = pdd->dev->id;
781 			pAperture->lds_base = pdd->lds_base;
782 			pAperture->lds_limit = pdd->lds_limit;
783 			pAperture->gpuvm_base = pdd->gpuvm_base;
784 			pAperture->gpuvm_limit = pdd->gpuvm_limit;
785 			pAperture->scratch_base = pdd->scratch_base;
786 			pAperture->scratch_limit = pdd->scratch_limit;
787 
788 			dev_dbg(kfd_device,
789 				"node id %u\n", args->num_of_nodes);
790 			dev_dbg(kfd_device,
791 				"gpu id %u\n", pdd->dev->id);
792 			dev_dbg(kfd_device,
793 				"lds_base %llX\n", pdd->lds_base);
794 			dev_dbg(kfd_device,
795 				"lds_limit %llX\n", pdd->lds_limit);
796 			dev_dbg(kfd_device,
797 				"gpuvm_base %llX\n", pdd->gpuvm_base);
798 			dev_dbg(kfd_device,
799 				"gpuvm_limit %llX\n", pdd->gpuvm_limit);
800 			dev_dbg(kfd_device,
801 				"scratch_base %llX\n", pdd->scratch_base);
802 			dev_dbg(kfd_device,
803 				"scratch_limit %llX\n", pdd->scratch_limit);
804 
805 			args->num_of_nodes++;
806 		} while ((pdd = kfd_get_next_process_device_data(p, pdd)) != NULL &&
807 				(args->num_of_nodes < NUM_OF_SUPPORTED_GPUS));
808 	}
809 
810 	mutex_unlock(&p->mutex);
811 
812 	return 0;
813 }
814 
815 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
816 					void *data)
817 {
818 	struct kfd_ioctl_create_event_args *args = data;
819 	int err;
820 
821 	err = kfd_event_create(filp, p, args->event_type,
822 				args->auto_reset != 0, args->node_id,
823 				&args->event_id, &args->event_trigger_data,
824 				&args->event_page_offset,
825 				&args->event_slot_index);
826 
827 	return err;
828 }
829 
830 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
831 					void *data)
832 {
833 	struct kfd_ioctl_destroy_event_args *args = data;
834 
835 	return kfd_event_destroy(p, args->event_id);
836 }
837 
838 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
839 				void *data)
840 {
841 	struct kfd_ioctl_set_event_args *args = data;
842 
843 	return kfd_set_event(p, args->event_id);
844 }
845 
846 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
847 				void *data)
848 {
849 	struct kfd_ioctl_reset_event_args *args = data;
850 
851 	return kfd_reset_event(p, args->event_id);
852 }
853 
854 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
855 				void *data)
856 {
857 	struct kfd_ioctl_wait_events_args *args = data;
858 	enum kfd_event_wait_result wait_result;
859 	int err;
860 
861 	err = kfd_wait_on_events(p, args->num_events,
862 			(void __user *)args->events_ptr,
863 			(args->wait_for_all != 0),
864 			args->timeout, &wait_result);
865 
866 	args->wait_result = wait_result;
867 
868 	return err;
869 }
870 
871 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
872 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, .cmd_drv = 0, .name = #ioctl}
873 
874 /** Ioctl table */
875 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
876 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
877 			kfd_ioctl_get_version, 0),
878 
879 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
880 			kfd_ioctl_create_queue, 0),
881 
882 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
883 			kfd_ioctl_destroy_queue, 0),
884 
885 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
886 			kfd_ioctl_set_memory_policy, 0),
887 
888 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
889 			kfd_ioctl_get_clock_counters, 0),
890 
891 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
892 			kfd_ioctl_get_process_apertures, 0),
893 
894 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
895 			kfd_ioctl_update_queue, 0),
896 
897 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
898 			kfd_ioctl_create_event, 0),
899 
900 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
901 			kfd_ioctl_destroy_event, 0),
902 
903 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
904 			kfd_ioctl_set_event, 0),
905 
906 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
907 			kfd_ioctl_reset_event, 0),
908 
909 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
910 			kfd_ioctl_wait_events, 0),
911 
912 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER,
913 			kfd_ioctl_dbg_register, 0),
914 
915 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER,
916 			kfd_ioctl_dbg_unrgesiter, 0),
917 
918 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH,
919 			kfd_ioctl_dbg_address_watch, 0),
920 
921 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL,
922 			kfd_ioctl_dbg_wave_control, 0),
923 };
924 
925 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
926 
927 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
928 {
929 	struct kfd_process *process;
930 	amdkfd_ioctl_t *func;
931 	const struct amdkfd_ioctl_desc *ioctl = NULL;
932 	unsigned int nr = _IOC_NR(cmd);
933 	char stack_kdata[128];
934 	char *kdata = NULL;
935 	unsigned int usize, asize;
936 	int retcode = -EINVAL;
937 
938 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
939 		goto err_i1;
940 
941 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
942 		u32 amdkfd_size;
943 
944 		ioctl = &amdkfd_ioctls[nr];
945 
946 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
947 		usize = asize = _IOC_SIZE(cmd);
948 		if (amdkfd_size > asize)
949 			asize = amdkfd_size;
950 
951 		cmd = ioctl->cmd;
952 	} else
953 		goto err_i1;
954 
955 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#%d), arg 0x%lx\n", cmd, nr, arg);
956 
957 	process = kfd_get_process(current);
958 	if (IS_ERR(process)) {
959 		dev_dbg(kfd_device, "no process\n");
960 		goto err_i1;
961 	}
962 
963 	/* Do not trust userspace, use our own definition */
964 	func = ioctl->func;
965 
966 	if (unlikely(!func)) {
967 		dev_dbg(kfd_device, "no function\n");
968 		retcode = -EINVAL;
969 		goto err_i1;
970 	}
971 
972 	if (cmd & (IOC_IN | IOC_OUT)) {
973 		if (asize <= sizeof(stack_kdata)) {
974 			kdata = stack_kdata;
975 		} else {
976 			kdata = kmalloc(asize, GFP_KERNEL);
977 			if (!kdata) {
978 				retcode = -ENOMEM;
979 				goto err_i1;
980 			}
981 		}
982 		if (asize > usize)
983 			memset(kdata + usize, 0, asize - usize);
984 	}
985 
986 	if (cmd & IOC_IN) {
987 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
988 			retcode = -EFAULT;
989 			goto err_i1;
990 		}
991 	} else if (cmd & IOC_OUT) {
992 		memset(kdata, 0, usize);
993 	}
994 
995 	retcode = func(filep, process, kdata);
996 
997 	if (cmd & IOC_OUT)
998 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
999 			retcode = -EFAULT;
1000 
1001 err_i1:
1002 	if (!ioctl)
1003 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
1004 			  task_pid_nr(current), cmd, nr);
1005 
1006 	if (kdata != stack_kdata)
1007 		kfree(kdata);
1008 
1009 	if (retcode)
1010 		dev_dbg(kfd_device, "ret = %d\n", retcode);
1011 
1012 	return retcode;
1013 }
1014 
1015 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
1016 {
1017 	struct kfd_process *process;
1018 
1019 	process = kfd_get_process(current);
1020 	if (IS_ERR(process))
1021 		return PTR_ERR(process);
1022 
1023 	if ((vma->vm_pgoff & KFD_MMAP_DOORBELL_MASK) ==
1024 			KFD_MMAP_DOORBELL_MASK) {
1025 		vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_DOORBELL_MASK;
1026 		return kfd_doorbell_mmap(process, vma);
1027 	} else if ((vma->vm_pgoff & KFD_MMAP_EVENTS_MASK) ==
1028 			KFD_MMAP_EVENTS_MASK) {
1029 		vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_EVENTS_MASK;
1030 		return kfd_event_mmap(process, vma);
1031 	}
1032 
1033 	return -EFAULT;
1034 }
1035