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 = in_compat_syscall();
111 
112 	if (is_32bit_user_mode) {
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 
135 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
136 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
137 
138 	return 0;
139 }
140 
141 static int set_queue_properties_from_user(struct queue_properties *q_properties,
142 				struct kfd_ioctl_create_queue_args *args)
143 {
144 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
145 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
146 		return -EINVAL;
147 	}
148 
149 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
150 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
151 		return -EINVAL;
152 	}
153 
154 	if ((args->ring_base_address) &&
155 		(!access_ok(VERIFY_WRITE,
156 			(const void __user *) args->ring_base_address,
157 			sizeof(uint64_t)))) {
158 		pr_err("Can't access ring base address\n");
159 		return -EFAULT;
160 	}
161 
162 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
163 		pr_err("Ring size must be a power of 2 or 0\n");
164 		return -EINVAL;
165 	}
166 
167 	if (!access_ok(VERIFY_WRITE,
168 			(const void __user *) args->read_pointer_address,
169 			sizeof(uint32_t))) {
170 		pr_err("Can't access read pointer\n");
171 		return -EFAULT;
172 	}
173 
174 	if (!access_ok(VERIFY_WRITE,
175 			(const void __user *) args->write_pointer_address,
176 			sizeof(uint32_t))) {
177 		pr_err("Can't access write pointer\n");
178 		return -EFAULT;
179 	}
180 
181 	if (args->eop_buffer_address &&
182 		!access_ok(VERIFY_WRITE,
183 			(const void __user *) args->eop_buffer_address,
184 			sizeof(uint32_t))) {
185 		pr_debug("Can't access eop buffer");
186 		return -EFAULT;
187 	}
188 
189 	if (args->ctx_save_restore_address &&
190 		!access_ok(VERIFY_WRITE,
191 			(const void __user *) args->ctx_save_restore_address,
192 			sizeof(uint32_t))) {
193 		pr_debug("Can't access ctx save restore buffer");
194 		return -EFAULT;
195 	}
196 
197 	q_properties->is_interop = false;
198 	q_properties->queue_percent = args->queue_percentage;
199 	q_properties->priority = args->queue_priority;
200 	q_properties->queue_address = args->ring_base_address;
201 	q_properties->queue_size = args->ring_size;
202 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
203 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
204 	q_properties->eop_ring_buffer_address = args->eop_buffer_address;
205 	q_properties->eop_ring_buffer_size = args->eop_buffer_size;
206 	q_properties->ctx_save_restore_area_address =
207 			args->ctx_save_restore_address;
208 	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
209 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
210 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
211 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
212 	else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
213 		q_properties->type = KFD_QUEUE_TYPE_SDMA;
214 	else
215 		return -ENOTSUPP;
216 
217 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
218 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
219 	else
220 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
221 
222 	pr_debug("Queue Percentage: %d, %d\n",
223 			q_properties->queue_percent, args->queue_percentage);
224 
225 	pr_debug("Queue Priority: %d, %d\n",
226 			q_properties->priority, args->queue_priority);
227 
228 	pr_debug("Queue Address: 0x%llX, 0x%llX\n",
229 			q_properties->queue_address, args->ring_base_address);
230 
231 	pr_debug("Queue Size: 0x%llX, %u\n",
232 			q_properties->queue_size, args->ring_size);
233 
234 	pr_debug("Queue r/w Pointers: %p, %p\n",
235 			q_properties->read_ptr,
236 			q_properties->write_ptr);
237 
238 	pr_debug("Queue Format: %d\n", q_properties->format);
239 
240 	pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
241 
242 	pr_debug("Queue CTX save area: 0x%llX\n",
243 			q_properties->ctx_save_restore_area_address);
244 
245 	return 0;
246 }
247 
248 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
249 					void *data)
250 {
251 	struct kfd_ioctl_create_queue_args *args = data;
252 	struct kfd_dev *dev;
253 	int err = 0;
254 	unsigned int queue_id;
255 	struct kfd_process_device *pdd;
256 	struct queue_properties q_properties;
257 
258 	memset(&q_properties, 0, sizeof(struct queue_properties));
259 
260 	pr_debug("Creating queue ioctl\n");
261 
262 	err = set_queue_properties_from_user(&q_properties, args);
263 	if (err)
264 		return err;
265 
266 	pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
267 	dev = kfd_device_by_id(args->gpu_id);
268 	if (!dev) {
269 		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
270 		return -EINVAL;
271 	}
272 
273 	mutex_lock(&p->mutex);
274 
275 	pdd = kfd_bind_process_to_device(dev, p);
276 	if (IS_ERR(pdd)) {
277 		err = -ESRCH;
278 		goto err_bind_process;
279 	}
280 
281 	pr_debug("Creating queue for PASID %d on gpu 0x%x\n",
282 			p->pasid,
283 			dev->id);
284 
285 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id);
286 	if (err != 0)
287 		goto err_create_queue;
288 
289 	args->queue_id = queue_id;
290 
291 
292 	/* Return gpu_id as doorbell offset for mmap usage */
293 	args->doorbell_offset = (KFD_MMAP_DOORBELL_MASK | args->gpu_id);
294 	args->doorbell_offset <<= PAGE_SHIFT;
295 
296 	mutex_unlock(&p->mutex);
297 
298 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
299 
300 	pr_debug("Ring buffer address == 0x%016llX\n",
301 			args->ring_base_address);
302 
303 	pr_debug("Read ptr address    == 0x%016llX\n",
304 			args->read_pointer_address);
305 
306 	pr_debug("Write ptr address   == 0x%016llX\n",
307 			args->write_pointer_address);
308 
309 	return 0;
310 
311 err_create_queue:
312 err_bind_process:
313 	mutex_unlock(&p->mutex);
314 	return err;
315 }
316 
317 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
318 					void *data)
319 {
320 	int retval;
321 	struct kfd_ioctl_destroy_queue_args *args = data;
322 
323 	pr_debug("Destroying queue id %d for pasid %d\n",
324 				args->queue_id,
325 				p->pasid);
326 
327 	mutex_lock(&p->mutex);
328 
329 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
330 
331 	mutex_unlock(&p->mutex);
332 	return retval;
333 }
334 
335 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
336 					void *data)
337 {
338 	int retval;
339 	struct kfd_ioctl_update_queue_args *args = data;
340 	struct queue_properties properties;
341 
342 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
343 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
344 		return -EINVAL;
345 	}
346 
347 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
348 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
349 		return -EINVAL;
350 	}
351 
352 	if ((args->ring_base_address) &&
353 		(!access_ok(VERIFY_WRITE,
354 			(const void __user *) args->ring_base_address,
355 			sizeof(uint64_t)))) {
356 		pr_err("Can't access ring base address\n");
357 		return -EFAULT;
358 	}
359 
360 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
361 		pr_err("Ring size must be a power of 2 or 0\n");
362 		return -EINVAL;
363 	}
364 
365 	properties.queue_address = args->ring_base_address;
366 	properties.queue_size = args->ring_size;
367 	properties.queue_percent = args->queue_percentage;
368 	properties.priority = args->queue_priority;
369 
370 	pr_debug("Updating queue id %d for pasid %d\n",
371 			args->queue_id, p->pasid);
372 
373 	mutex_lock(&p->mutex);
374 
375 	retval = pqm_update_queue(&p->pqm, args->queue_id, &properties);
376 
377 	mutex_unlock(&p->mutex);
378 
379 	return retval;
380 }
381 
382 static int kfd_ioctl_set_memory_policy(struct file *filep,
383 					struct kfd_process *p, void *data)
384 {
385 	struct kfd_ioctl_set_memory_policy_args *args = data;
386 	struct kfd_dev *dev;
387 	int err = 0;
388 	struct kfd_process_device *pdd;
389 	enum cache_policy default_policy, alternate_policy;
390 
391 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
392 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
393 		return -EINVAL;
394 	}
395 
396 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
397 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
398 		return -EINVAL;
399 	}
400 
401 	dev = kfd_device_by_id(args->gpu_id);
402 	if (!dev)
403 		return -EINVAL;
404 
405 	mutex_lock(&p->mutex);
406 
407 	pdd = kfd_bind_process_to_device(dev, p);
408 	if (IS_ERR(pdd)) {
409 		err = -ESRCH;
410 		goto out;
411 	}
412 
413 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
414 			 ? cache_policy_coherent : cache_policy_noncoherent;
415 
416 	alternate_policy =
417 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
418 		   ? cache_policy_coherent : cache_policy_noncoherent;
419 
420 	if (!dev->dqm->ops.set_cache_memory_policy(dev->dqm,
421 				&pdd->qpd,
422 				default_policy,
423 				alternate_policy,
424 				(void __user *)args->alternate_aperture_base,
425 				args->alternate_aperture_size))
426 		err = -EINVAL;
427 
428 out:
429 	mutex_unlock(&p->mutex);
430 
431 	return err;
432 }
433 
434 static int kfd_ioctl_dbg_register(struct file *filep,
435 				struct kfd_process *p, void *data)
436 {
437 	struct kfd_ioctl_dbg_register_args *args = data;
438 	struct kfd_dev *dev;
439 	struct kfd_dbgmgr *dbgmgr_ptr;
440 	struct kfd_process_device *pdd;
441 	bool create_ok;
442 	long status = 0;
443 
444 	dev = kfd_device_by_id(args->gpu_id);
445 	if (!dev)
446 		return -EINVAL;
447 
448 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
449 		pr_debug("kfd_ioctl_dbg_register not supported on CZ\n");
450 		return -EINVAL;
451 	}
452 
453 	mutex_lock(&p->mutex);
454 	mutex_lock(kfd_get_dbgmgr_mutex());
455 
456 	/*
457 	 * make sure that we have pdd, if this the first queue created for
458 	 * this process
459 	 */
460 	pdd = kfd_bind_process_to_device(dev, p);
461 	if (IS_ERR(pdd)) {
462 		status = PTR_ERR(pdd);
463 		goto out;
464 	}
465 
466 	if (!dev->dbgmgr) {
467 		/* In case of a legal call, we have no dbgmgr yet */
468 		create_ok = kfd_dbgmgr_create(&dbgmgr_ptr, dev);
469 		if (create_ok) {
470 			status = kfd_dbgmgr_register(dbgmgr_ptr, p);
471 			if (status != 0)
472 				kfd_dbgmgr_destroy(dbgmgr_ptr);
473 			else
474 				dev->dbgmgr = dbgmgr_ptr;
475 		}
476 	} else {
477 		pr_debug("debugger already registered\n");
478 		status = -EINVAL;
479 	}
480 
481 out:
482 	mutex_unlock(kfd_get_dbgmgr_mutex());
483 	mutex_unlock(&p->mutex);
484 
485 	return status;
486 }
487 
488 static int kfd_ioctl_dbg_unregister(struct file *filep,
489 				struct kfd_process *p, void *data)
490 {
491 	struct kfd_ioctl_dbg_unregister_args *args = data;
492 	struct kfd_dev *dev;
493 	long status;
494 
495 	dev = kfd_device_by_id(args->gpu_id);
496 	if (!dev)
497 		return -EINVAL;
498 
499 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
500 		pr_debug("kfd_ioctl_dbg_unregister not supported on CZ\n");
501 		return -EINVAL;
502 	}
503 
504 	mutex_lock(kfd_get_dbgmgr_mutex());
505 
506 	status = kfd_dbgmgr_unregister(dev->dbgmgr, p);
507 	if (!status) {
508 		kfd_dbgmgr_destroy(dev->dbgmgr);
509 		dev->dbgmgr = NULL;
510 	}
511 
512 	mutex_unlock(kfd_get_dbgmgr_mutex());
513 
514 	return status;
515 }
516 
517 /*
518  * Parse and generate variable size data structure for address watch.
519  * Total size of the buffer and # watch points is limited in order
520  * to prevent kernel abuse. (no bearing to the much smaller HW limitation
521  * which is enforced by dbgdev module)
522  * please also note that the watch address itself are not "copied from user",
523  * since it be set into the HW in user mode values.
524  *
525  */
526 static int kfd_ioctl_dbg_address_watch(struct file *filep,
527 					struct kfd_process *p, void *data)
528 {
529 	struct kfd_ioctl_dbg_address_watch_args *args = data;
530 	struct kfd_dev *dev;
531 	struct dbg_address_watch_info aw_info;
532 	unsigned char *args_buff;
533 	long status;
534 	void __user *cmd_from_user;
535 	uint64_t watch_mask_value = 0;
536 	unsigned int args_idx = 0;
537 
538 	memset((void *) &aw_info, 0, sizeof(struct dbg_address_watch_info));
539 
540 	dev = kfd_device_by_id(args->gpu_id);
541 	if (!dev)
542 		return -EINVAL;
543 
544 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
545 		pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
546 		return -EINVAL;
547 	}
548 
549 	cmd_from_user = (void __user *) args->content_ptr;
550 
551 	/* Validate arguments */
552 
553 	if ((args->buf_size_in_bytes > MAX_ALLOWED_AW_BUFF_SIZE) ||
554 		(args->buf_size_in_bytes <= sizeof(*args) + sizeof(int) * 2) ||
555 		(cmd_from_user == NULL))
556 		return -EINVAL;
557 
558 	/* this is the actual buffer to work with */
559 	args_buff = memdup_user(cmd_from_user,
560 				args->buf_size_in_bytes - sizeof(*args));
561 	if (IS_ERR(args_buff))
562 		return PTR_ERR(args_buff);
563 
564 	aw_info.process = p;
565 
566 	aw_info.num_watch_points = *((uint32_t *)(&args_buff[args_idx]));
567 	args_idx += sizeof(aw_info.num_watch_points);
568 
569 	aw_info.watch_mode = (enum HSA_DBG_WATCH_MODE *) &args_buff[args_idx];
570 	args_idx += sizeof(enum HSA_DBG_WATCH_MODE) * aw_info.num_watch_points;
571 
572 	/*
573 	 * set watch address base pointer to point on the array base
574 	 * within args_buff
575 	 */
576 	aw_info.watch_address = (uint64_t *) &args_buff[args_idx];
577 
578 	/* skip over the addresses buffer */
579 	args_idx += sizeof(aw_info.watch_address) * aw_info.num_watch_points;
580 
581 	if (args_idx >= args->buf_size_in_bytes - sizeof(*args)) {
582 		status = -EINVAL;
583 		goto out;
584 	}
585 
586 	watch_mask_value = (uint64_t) args_buff[args_idx];
587 
588 	if (watch_mask_value > 0) {
589 		/*
590 		 * There is an array of masks.
591 		 * set watch mask base pointer to point on the array base
592 		 * within args_buff
593 		 */
594 		aw_info.watch_mask = (uint64_t *) &args_buff[args_idx];
595 
596 		/* skip over the masks buffer */
597 		args_idx += sizeof(aw_info.watch_mask) *
598 				aw_info.num_watch_points;
599 	} else {
600 		/* just the NULL mask, set to NULL and skip over it */
601 		aw_info.watch_mask = NULL;
602 		args_idx += sizeof(aw_info.watch_mask);
603 	}
604 
605 	if (args_idx >= args->buf_size_in_bytes - sizeof(args)) {
606 		status = -EINVAL;
607 		goto out;
608 	}
609 
610 	/* Currently HSA Event is not supported for DBG */
611 	aw_info.watch_event = NULL;
612 
613 	mutex_lock(kfd_get_dbgmgr_mutex());
614 
615 	status = kfd_dbgmgr_address_watch(dev->dbgmgr, &aw_info);
616 
617 	mutex_unlock(kfd_get_dbgmgr_mutex());
618 
619 out:
620 	kfree(args_buff);
621 
622 	return status;
623 }
624 
625 /* Parse and generate fixed size data structure for wave control */
626 static int kfd_ioctl_dbg_wave_control(struct file *filep,
627 					struct kfd_process *p, void *data)
628 {
629 	struct kfd_ioctl_dbg_wave_control_args *args = data;
630 	struct kfd_dev *dev;
631 	struct dbg_wave_control_info wac_info;
632 	unsigned char *args_buff;
633 	uint32_t computed_buff_size;
634 	long status;
635 	void __user *cmd_from_user;
636 	unsigned int args_idx = 0;
637 
638 	memset((void *) &wac_info, 0, sizeof(struct dbg_wave_control_info));
639 
640 	/* we use compact form, independent of the packing attribute value */
641 	computed_buff_size = sizeof(*args) +
642 				sizeof(wac_info.mode) +
643 				sizeof(wac_info.operand) +
644 				sizeof(wac_info.dbgWave_msg.DbgWaveMsg) +
645 				sizeof(wac_info.dbgWave_msg.MemoryVA) +
646 				sizeof(wac_info.trapId);
647 
648 	dev = kfd_device_by_id(args->gpu_id);
649 	if (!dev)
650 		return -EINVAL;
651 
652 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
653 		pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
654 		return -EINVAL;
655 	}
656 
657 	/* input size must match the computed "compact" size */
658 	if (args->buf_size_in_bytes != computed_buff_size) {
659 		pr_debug("size mismatch, computed : actual %u : %u\n",
660 				args->buf_size_in_bytes, computed_buff_size);
661 		return -EINVAL;
662 	}
663 
664 	cmd_from_user = (void __user *) args->content_ptr;
665 
666 	if (cmd_from_user == NULL)
667 		return -EINVAL;
668 
669 	/* copy the entire buffer from user */
670 
671 	args_buff = memdup_user(cmd_from_user,
672 				args->buf_size_in_bytes - sizeof(*args));
673 	if (IS_ERR(args_buff))
674 		return PTR_ERR(args_buff);
675 
676 	/* move ptr to the start of the "pay-load" area */
677 	wac_info.process = p;
678 
679 	wac_info.operand = *((enum HSA_DBG_WAVEOP *)(&args_buff[args_idx]));
680 	args_idx += sizeof(wac_info.operand);
681 
682 	wac_info.mode = *((enum HSA_DBG_WAVEMODE *)(&args_buff[args_idx]));
683 	args_idx += sizeof(wac_info.mode);
684 
685 	wac_info.trapId = *((uint32_t *)(&args_buff[args_idx]));
686 	args_idx += sizeof(wac_info.trapId);
687 
688 	wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value =
689 					*((uint32_t *)(&args_buff[args_idx]));
690 	wac_info.dbgWave_msg.MemoryVA = NULL;
691 
692 	mutex_lock(kfd_get_dbgmgr_mutex());
693 
694 	pr_debug("Calling dbg manager process %p, operand %u, mode %u, trapId %u, message %u\n",
695 			wac_info.process, wac_info.operand,
696 			wac_info.mode, wac_info.trapId,
697 			wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value);
698 
699 	status = kfd_dbgmgr_wave_control(dev->dbgmgr, &wac_info);
700 
701 	pr_debug("Returned status of dbg manager is %ld\n", status);
702 
703 	mutex_unlock(kfd_get_dbgmgr_mutex());
704 
705 	kfree(args_buff);
706 
707 	return status;
708 }
709 
710 static int kfd_ioctl_get_clock_counters(struct file *filep,
711 				struct kfd_process *p, void *data)
712 {
713 	struct kfd_ioctl_get_clock_counters_args *args = data;
714 	struct kfd_dev *dev;
715 	struct timespec64 time;
716 
717 	dev = kfd_device_by_id(args->gpu_id);
718 	if (dev == NULL)
719 		return -EINVAL;
720 
721 	/* Reading GPU clock counter from KGD */
722 	args->gpu_clock_counter =
723 		dev->kfd2kgd->get_gpu_clock_counter(dev->kgd);
724 
725 	/* No access to rdtsc. Using raw monotonic time */
726 	getrawmonotonic64(&time);
727 	args->cpu_clock_counter = (uint64_t)timespec64_to_ns(&time);
728 
729 	get_monotonic_boottime64(&time);
730 	args->system_clock_counter = (uint64_t)timespec64_to_ns(&time);
731 
732 	/* Since the counter is in nano-seconds we use 1GHz frequency */
733 	args->system_clock_freq = 1000000000;
734 
735 	return 0;
736 }
737 
738 
739 static int kfd_ioctl_get_process_apertures(struct file *filp,
740 				struct kfd_process *p, void *data)
741 {
742 	struct kfd_ioctl_get_process_apertures_args *args = data;
743 	struct kfd_process_device_apertures *pAperture;
744 	struct kfd_process_device *pdd;
745 
746 	dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
747 
748 	args->num_of_nodes = 0;
749 
750 	mutex_lock(&p->mutex);
751 
752 	/*if the process-device list isn't empty*/
753 	if (kfd_has_process_device_data(p)) {
754 		/* Run over all pdd of the process */
755 		pdd = kfd_get_first_process_device_data(p);
756 		do {
757 			pAperture =
758 				&args->process_apertures[args->num_of_nodes];
759 			pAperture->gpu_id = pdd->dev->id;
760 			pAperture->lds_base = pdd->lds_base;
761 			pAperture->lds_limit = pdd->lds_limit;
762 			pAperture->gpuvm_base = pdd->gpuvm_base;
763 			pAperture->gpuvm_limit = pdd->gpuvm_limit;
764 			pAperture->scratch_base = pdd->scratch_base;
765 			pAperture->scratch_limit = pdd->scratch_limit;
766 
767 			dev_dbg(kfd_device,
768 				"node id %u\n", args->num_of_nodes);
769 			dev_dbg(kfd_device,
770 				"gpu id %u\n", pdd->dev->id);
771 			dev_dbg(kfd_device,
772 				"lds_base %llX\n", pdd->lds_base);
773 			dev_dbg(kfd_device,
774 				"lds_limit %llX\n", pdd->lds_limit);
775 			dev_dbg(kfd_device,
776 				"gpuvm_base %llX\n", pdd->gpuvm_base);
777 			dev_dbg(kfd_device,
778 				"gpuvm_limit %llX\n", pdd->gpuvm_limit);
779 			dev_dbg(kfd_device,
780 				"scratch_base %llX\n", pdd->scratch_base);
781 			dev_dbg(kfd_device,
782 				"scratch_limit %llX\n", pdd->scratch_limit);
783 
784 			args->num_of_nodes++;
785 
786 			pdd = kfd_get_next_process_device_data(p, pdd);
787 		} while (pdd && (args->num_of_nodes < NUM_OF_SUPPORTED_GPUS));
788 	}
789 
790 	mutex_unlock(&p->mutex);
791 
792 	return 0;
793 }
794 
795 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
796 					void *data)
797 {
798 	struct kfd_ioctl_create_event_args *args = data;
799 	int err;
800 
801 	err = kfd_event_create(filp, p, args->event_type,
802 				args->auto_reset != 0, args->node_id,
803 				&args->event_id, &args->event_trigger_data,
804 				&args->event_page_offset,
805 				&args->event_slot_index);
806 
807 	return err;
808 }
809 
810 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
811 					void *data)
812 {
813 	struct kfd_ioctl_destroy_event_args *args = data;
814 
815 	return kfd_event_destroy(p, args->event_id);
816 }
817 
818 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
819 				void *data)
820 {
821 	struct kfd_ioctl_set_event_args *args = data;
822 
823 	return kfd_set_event(p, args->event_id);
824 }
825 
826 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
827 				void *data)
828 {
829 	struct kfd_ioctl_reset_event_args *args = data;
830 
831 	return kfd_reset_event(p, args->event_id);
832 }
833 
834 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
835 				void *data)
836 {
837 	struct kfd_ioctl_wait_events_args *args = data;
838 	int err;
839 
840 	err = kfd_wait_on_events(p, args->num_events,
841 			(void __user *)args->events_ptr,
842 			(args->wait_for_all != 0),
843 			args->timeout, &args->wait_result);
844 
845 	return err;
846 }
847 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
848 					struct kfd_process *p, void *data)
849 {
850 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
851 	struct kfd_process_device *pdd;
852 	struct kfd_dev *dev;
853 	long err;
854 
855 	dev = kfd_device_by_id(args->gpu_id);
856 	if (!dev)
857 		return -EINVAL;
858 
859 	mutex_lock(&p->mutex);
860 
861 	pdd = kfd_bind_process_to_device(dev, p);
862 	if (IS_ERR(pdd)) {
863 		err = PTR_ERR(pdd);
864 		goto bind_process_to_device_fail;
865 	}
866 
867 	pdd->qpd.sh_hidden_private_base = args->va_addr;
868 
869 	mutex_unlock(&p->mutex);
870 
871 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS && pdd->qpd.vmid != 0)
872 		dev->kfd2kgd->set_scratch_backing_va(
873 			dev->kgd, args->va_addr, pdd->qpd.vmid);
874 
875 	return 0;
876 
877 bind_process_to_device_fail:
878 	mutex_unlock(&p->mutex);
879 	return err;
880 }
881 
882 static int kfd_ioctl_get_tile_config(struct file *filep,
883 		struct kfd_process *p, void *data)
884 {
885 	struct kfd_ioctl_get_tile_config_args *args = data;
886 	struct kfd_dev *dev;
887 	struct tile_config config;
888 	int err = 0;
889 
890 	dev = kfd_device_by_id(args->gpu_id);
891 	if (!dev)
892 		return -EINVAL;
893 
894 	dev->kfd2kgd->get_tile_config(dev->kgd, &config);
895 
896 	args->gb_addr_config = config.gb_addr_config;
897 	args->num_banks = config.num_banks;
898 	args->num_ranks = config.num_ranks;
899 
900 	if (args->num_tile_configs > config.num_tile_configs)
901 		args->num_tile_configs = config.num_tile_configs;
902 	err = copy_to_user((void __user *)args->tile_config_ptr,
903 			config.tile_config_ptr,
904 			args->num_tile_configs * sizeof(uint32_t));
905 	if (err) {
906 		args->num_tile_configs = 0;
907 		return -EFAULT;
908 	}
909 
910 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
911 		args->num_macro_tile_configs =
912 				config.num_macro_tile_configs;
913 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
914 			config.macro_tile_config_ptr,
915 			args->num_macro_tile_configs * sizeof(uint32_t));
916 	if (err) {
917 		args->num_macro_tile_configs = 0;
918 		return -EFAULT;
919 	}
920 
921 	return 0;
922 }
923 
924 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
925 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
926 			    .cmd_drv = 0, .name = #ioctl}
927 
928 /** Ioctl table */
929 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
930 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
931 			kfd_ioctl_get_version, 0),
932 
933 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
934 			kfd_ioctl_create_queue, 0),
935 
936 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
937 			kfd_ioctl_destroy_queue, 0),
938 
939 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
940 			kfd_ioctl_set_memory_policy, 0),
941 
942 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
943 			kfd_ioctl_get_clock_counters, 0),
944 
945 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
946 			kfd_ioctl_get_process_apertures, 0),
947 
948 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
949 			kfd_ioctl_update_queue, 0),
950 
951 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
952 			kfd_ioctl_create_event, 0),
953 
954 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
955 			kfd_ioctl_destroy_event, 0),
956 
957 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
958 			kfd_ioctl_set_event, 0),
959 
960 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
961 			kfd_ioctl_reset_event, 0),
962 
963 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
964 			kfd_ioctl_wait_events, 0),
965 
966 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER,
967 			kfd_ioctl_dbg_register, 0),
968 
969 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER,
970 			kfd_ioctl_dbg_unregister, 0),
971 
972 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH,
973 			kfd_ioctl_dbg_address_watch, 0),
974 
975 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL,
976 			kfd_ioctl_dbg_wave_control, 0),
977 
978 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
979 			kfd_ioctl_set_scratch_backing_va, 0),
980 
981 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
982 			kfd_ioctl_get_tile_config, 0)
983 };
984 
985 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
986 
987 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
988 {
989 	struct kfd_process *process;
990 	amdkfd_ioctl_t *func;
991 	const struct amdkfd_ioctl_desc *ioctl = NULL;
992 	unsigned int nr = _IOC_NR(cmd);
993 	char stack_kdata[128];
994 	char *kdata = NULL;
995 	unsigned int usize, asize;
996 	int retcode = -EINVAL;
997 
998 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
999 		goto err_i1;
1000 
1001 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
1002 		u32 amdkfd_size;
1003 
1004 		ioctl = &amdkfd_ioctls[nr];
1005 
1006 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
1007 		usize = asize = _IOC_SIZE(cmd);
1008 		if (amdkfd_size > asize)
1009 			asize = amdkfd_size;
1010 
1011 		cmd = ioctl->cmd;
1012 	} else
1013 		goto err_i1;
1014 
1015 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#%d), arg 0x%lx\n", cmd, nr, arg);
1016 
1017 	process = kfd_get_process(current);
1018 	if (IS_ERR(process)) {
1019 		dev_dbg(kfd_device, "no process\n");
1020 		goto err_i1;
1021 	}
1022 
1023 	/* Do not trust userspace, use our own definition */
1024 	func = ioctl->func;
1025 
1026 	if (unlikely(!func)) {
1027 		dev_dbg(kfd_device, "no function\n");
1028 		retcode = -EINVAL;
1029 		goto err_i1;
1030 	}
1031 
1032 	if (cmd & (IOC_IN | IOC_OUT)) {
1033 		if (asize <= sizeof(stack_kdata)) {
1034 			kdata = stack_kdata;
1035 		} else {
1036 			kdata = kmalloc(asize, GFP_KERNEL);
1037 			if (!kdata) {
1038 				retcode = -ENOMEM;
1039 				goto err_i1;
1040 			}
1041 		}
1042 		if (asize > usize)
1043 			memset(kdata + usize, 0, asize - usize);
1044 	}
1045 
1046 	if (cmd & IOC_IN) {
1047 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
1048 			retcode = -EFAULT;
1049 			goto err_i1;
1050 		}
1051 	} else if (cmd & IOC_OUT) {
1052 		memset(kdata, 0, usize);
1053 	}
1054 
1055 	retcode = func(filep, process, kdata);
1056 
1057 	if (cmd & IOC_OUT)
1058 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
1059 			retcode = -EFAULT;
1060 
1061 err_i1:
1062 	if (!ioctl)
1063 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
1064 			  task_pid_nr(current), cmd, nr);
1065 
1066 	if (kdata != stack_kdata)
1067 		kfree(kdata);
1068 
1069 	if (retcode)
1070 		dev_dbg(kfd_device, "ret = %d\n", retcode);
1071 
1072 	return retcode;
1073 }
1074 
1075 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
1076 {
1077 	struct kfd_process *process;
1078 
1079 	process = kfd_get_process(current);
1080 	if (IS_ERR(process))
1081 		return PTR_ERR(process);
1082 
1083 	if ((vma->vm_pgoff & KFD_MMAP_DOORBELL_MASK) ==
1084 			KFD_MMAP_DOORBELL_MASK) {
1085 		vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_DOORBELL_MASK;
1086 		return kfd_doorbell_mmap(process, vma);
1087 	} else if ((vma->vm_pgoff & KFD_MMAP_EVENTS_MASK) ==
1088 			KFD_MMAP_EVENTS_MASK) {
1089 		vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_EVENTS_MASK;
1090 		return kfd_event_mmap(process, vma);
1091 	}
1092 
1093 	return -EFAULT;
1094 }
1095