xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c (revision e3b9f1e81de2083f359bacd2a94bf1c024f2ede0)
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(filep);
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 	q_properties->ctl_stack_size = args->ctl_stack_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: %p, %p\n",
236 			q_properties->read_ptr,
237 			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 area: 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("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("Looking for gpu id 0x%x\n", args->gpu_id);
268 	dev = kfd_device_by_id(args->gpu_id);
269 	if (!dev) {
270 		pr_debug("Could not find gpu id 0x%x\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("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, &queue_id);
287 	if (err != 0)
288 		goto err_create_queue;
289 
290 	args->queue_id = queue_id;
291 
292 
293 	/* Return gpu_id as doorbell offset for mmap usage */
294 	args->doorbell_offset = (KFD_MMAP_DOORBELL_MASK | args->gpu_id);
295 	args->doorbell_offset <<= PAGE_SHIFT;
296 
297 	mutex_unlock(&p->mutex);
298 
299 	pr_debug("Queue id %d was created successfully\n", args->queue_id);
300 
301 	pr_debug("Ring buffer address == 0x%016llX\n",
302 			args->ring_base_address);
303 
304 	pr_debug("Read ptr address    == 0x%016llX\n",
305 			args->read_pointer_address);
306 
307 	pr_debug("Write ptr address   == 0x%016llX\n",
308 			args->write_pointer_address);
309 
310 	return 0;
311 
312 err_create_queue:
313 err_bind_process:
314 	mutex_unlock(&p->mutex);
315 	return err;
316 }
317 
318 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
319 					void *data)
320 {
321 	int retval;
322 	struct kfd_ioctl_destroy_queue_args *args = data;
323 
324 	pr_debug("Destroying queue id %d for pasid %d\n",
325 				args->queue_id,
326 				p->pasid);
327 
328 	mutex_lock(&p->mutex);
329 
330 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
331 
332 	mutex_unlock(&p->mutex);
333 	return retval;
334 }
335 
336 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
337 					void *data)
338 {
339 	int retval;
340 	struct kfd_ioctl_update_queue_args *args = data;
341 	struct queue_properties properties;
342 
343 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
344 		pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
345 		return -EINVAL;
346 	}
347 
348 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
349 		pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
350 		return -EINVAL;
351 	}
352 
353 	if ((args->ring_base_address) &&
354 		(!access_ok(VERIFY_WRITE,
355 			(const void __user *) args->ring_base_address,
356 			sizeof(uint64_t)))) {
357 		pr_err("Can't access ring base address\n");
358 		return -EFAULT;
359 	}
360 
361 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
362 		pr_err("Ring size must be a power of 2 or 0\n");
363 		return -EINVAL;
364 	}
365 
366 	properties.queue_address = args->ring_base_address;
367 	properties.queue_size = args->ring_size;
368 	properties.queue_percent = args->queue_percentage;
369 	properties.priority = args->queue_priority;
370 
371 	pr_debug("Updating queue id %d for pasid %d\n",
372 			args->queue_id, p->pasid);
373 
374 	mutex_lock(&p->mutex);
375 
376 	retval = pqm_update_queue(&p->pqm, args->queue_id, &properties);
377 
378 	mutex_unlock(&p->mutex);
379 
380 	return retval;
381 }
382 
383 static int kfd_ioctl_set_memory_policy(struct file *filep,
384 					struct kfd_process *p, void *data)
385 {
386 	struct kfd_ioctl_set_memory_policy_args *args = data;
387 	struct kfd_dev *dev;
388 	int err = 0;
389 	struct kfd_process_device *pdd;
390 	enum cache_policy default_policy, alternate_policy;
391 
392 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
393 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
394 		return -EINVAL;
395 	}
396 
397 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
398 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
399 		return -EINVAL;
400 	}
401 
402 	dev = kfd_device_by_id(args->gpu_id);
403 	if (!dev)
404 		return -EINVAL;
405 
406 	mutex_lock(&p->mutex);
407 
408 	pdd = kfd_bind_process_to_device(dev, p);
409 	if (IS_ERR(pdd)) {
410 		err = -ESRCH;
411 		goto out;
412 	}
413 
414 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
415 			 ? cache_policy_coherent : cache_policy_noncoherent;
416 
417 	alternate_policy =
418 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
419 		   ? cache_policy_coherent : cache_policy_noncoherent;
420 
421 	if (!dev->dqm->ops.set_cache_memory_policy(dev->dqm,
422 				&pdd->qpd,
423 				default_policy,
424 				alternate_policy,
425 				(void __user *)args->alternate_aperture_base,
426 				args->alternate_aperture_size))
427 		err = -EINVAL;
428 
429 out:
430 	mutex_unlock(&p->mutex);
431 
432 	return err;
433 }
434 
435 static int kfd_ioctl_set_trap_handler(struct file *filep,
436 					struct kfd_process *p, void *data)
437 {
438 	struct kfd_ioctl_set_trap_handler_args *args = data;
439 	struct kfd_dev *dev;
440 	int err = 0;
441 	struct kfd_process_device *pdd;
442 
443 	dev = kfd_device_by_id(args->gpu_id);
444 	if (dev == NULL)
445 		return -EINVAL;
446 
447 	mutex_lock(&p->mutex);
448 
449 	pdd = kfd_bind_process_to_device(dev, p);
450 	if (IS_ERR(pdd)) {
451 		err = -ESRCH;
452 		goto out;
453 	}
454 
455 	if (dev->dqm->ops.set_trap_handler(dev->dqm,
456 					&pdd->qpd,
457 					args->tba_addr,
458 					args->tma_addr))
459 		err = -EINVAL;
460 
461 out:
462 	mutex_unlock(&p->mutex);
463 
464 	return err;
465 }
466 
467 static int kfd_ioctl_dbg_register(struct file *filep,
468 				struct kfd_process *p, void *data)
469 {
470 	struct kfd_ioctl_dbg_register_args *args = data;
471 	struct kfd_dev *dev;
472 	struct kfd_dbgmgr *dbgmgr_ptr;
473 	struct kfd_process_device *pdd;
474 	bool create_ok;
475 	long status = 0;
476 
477 	dev = kfd_device_by_id(args->gpu_id);
478 	if (!dev)
479 		return -EINVAL;
480 
481 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
482 		pr_debug("kfd_ioctl_dbg_register not supported on CZ\n");
483 		return -EINVAL;
484 	}
485 
486 	mutex_lock(&p->mutex);
487 	mutex_lock(kfd_get_dbgmgr_mutex());
488 
489 	/*
490 	 * make sure that we have pdd, if this the first queue created for
491 	 * this process
492 	 */
493 	pdd = kfd_bind_process_to_device(dev, p);
494 	if (IS_ERR(pdd)) {
495 		status = PTR_ERR(pdd);
496 		goto out;
497 	}
498 
499 	if (!dev->dbgmgr) {
500 		/* In case of a legal call, we have no dbgmgr yet */
501 		create_ok = kfd_dbgmgr_create(&dbgmgr_ptr, dev);
502 		if (create_ok) {
503 			status = kfd_dbgmgr_register(dbgmgr_ptr, p);
504 			if (status != 0)
505 				kfd_dbgmgr_destroy(dbgmgr_ptr);
506 			else
507 				dev->dbgmgr = dbgmgr_ptr;
508 		}
509 	} else {
510 		pr_debug("debugger already registered\n");
511 		status = -EINVAL;
512 	}
513 
514 out:
515 	mutex_unlock(kfd_get_dbgmgr_mutex());
516 	mutex_unlock(&p->mutex);
517 
518 	return status;
519 }
520 
521 static int kfd_ioctl_dbg_unregister(struct file *filep,
522 				struct kfd_process *p, void *data)
523 {
524 	struct kfd_ioctl_dbg_unregister_args *args = data;
525 	struct kfd_dev *dev;
526 	long status;
527 
528 	dev = kfd_device_by_id(args->gpu_id);
529 	if (!dev || !dev->dbgmgr)
530 		return -EINVAL;
531 
532 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
533 		pr_debug("kfd_ioctl_dbg_unregister not supported on CZ\n");
534 		return -EINVAL;
535 	}
536 
537 	mutex_lock(kfd_get_dbgmgr_mutex());
538 
539 	status = kfd_dbgmgr_unregister(dev->dbgmgr, p);
540 	if (!status) {
541 		kfd_dbgmgr_destroy(dev->dbgmgr);
542 		dev->dbgmgr = NULL;
543 	}
544 
545 	mutex_unlock(kfd_get_dbgmgr_mutex());
546 
547 	return status;
548 }
549 
550 /*
551  * Parse and generate variable size data structure for address watch.
552  * Total size of the buffer and # watch points is limited in order
553  * to prevent kernel abuse. (no bearing to the much smaller HW limitation
554  * which is enforced by dbgdev module)
555  * please also note that the watch address itself are not "copied from user",
556  * since it be set into the HW in user mode values.
557  *
558  */
559 static int kfd_ioctl_dbg_address_watch(struct file *filep,
560 					struct kfd_process *p, void *data)
561 {
562 	struct kfd_ioctl_dbg_address_watch_args *args = data;
563 	struct kfd_dev *dev;
564 	struct dbg_address_watch_info aw_info;
565 	unsigned char *args_buff;
566 	long status;
567 	void __user *cmd_from_user;
568 	uint64_t watch_mask_value = 0;
569 	unsigned int args_idx = 0;
570 
571 	memset((void *) &aw_info, 0, sizeof(struct dbg_address_watch_info));
572 
573 	dev = kfd_device_by_id(args->gpu_id);
574 	if (!dev)
575 		return -EINVAL;
576 
577 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
578 		pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
579 		return -EINVAL;
580 	}
581 
582 	cmd_from_user = (void __user *) args->content_ptr;
583 
584 	/* Validate arguments */
585 
586 	if ((args->buf_size_in_bytes > MAX_ALLOWED_AW_BUFF_SIZE) ||
587 		(args->buf_size_in_bytes <= sizeof(*args) + sizeof(int) * 2) ||
588 		(cmd_from_user == NULL))
589 		return -EINVAL;
590 
591 	/* this is the actual buffer to work with */
592 	args_buff = memdup_user(cmd_from_user,
593 				args->buf_size_in_bytes - sizeof(*args));
594 	if (IS_ERR(args_buff))
595 		return PTR_ERR(args_buff);
596 
597 	aw_info.process = p;
598 
599 	aw_info.num_watch_points = *((uint32_t *)(&args_buff[args_idx]));
600 	args_idx += sizeof(aw_info.num_watch_points);
601 
602 	aw_info.watch_mode = (enum HSA_DBG_WATCH_MODE *) &args_buff[args_idx];
603 	args_idx += sizeof(enum HSA_DBG_WATCH_MODE) * aw_info.num_watch_points;
604 
605 	/*
606 	 * set watch address base pointer to point on the array base
607 	 * within args_buff
608 	 */
609 	aw_info.watch_address = (uint64_t *) &args_buff[args_idx];
610 
611 	/* skip over the addresses buffer */
612 	args_idx += sizeof(aw_info.watch_address) * aw_info.num_watch_points;
613 
614 	if (args_idx >= args->buf_size_in_bytes - sizeof(*args)) {
615 		status = -EINVAL;
616 		goto out;
617 	}
618 
619 	watch_mask_value = (uint64_t) args_buff[args_idx];
620 
621 	if (watch_mask_value > 0) {
622 		/*
623 		 * There is an array of masks.
624 		 * set watch mask base pointer to point on the array base
625 		 * within args_buff
626 		 */
627 		aw_info.watch_mask = (uint64_t *) &args_buff[args_idx];
628 
629 		/* skip over the masks buffer */
630 		args_idx += sizeof(aw_info.watch_mask) *
631 				aw_info.num_watch_points;
632 	} else {
633 		/* just the NULL mask, set to NULL and skip over it */
634 		aw_info.watch_mask = NULL;
635 		args_idx += sizeof(aw_info.watch_mask);
636 	}
637 
638 	if (args_idx >= args->buf_size_in_bytes - sizeof(args)) {
639 		status = -EINVAL;
640 		goto out;
641 	}
642 
643 	/* Currently HSA Event is not supported for DBG */
644 	aw_info.watch_event = NULL;
645 
646 	mutex_lock(kfd_get_dbgmgr_mutex());
647 
648 	status = kfd_dbgmgr_address_watch(dev->dbgmgr, &aw_info);
649 
650 	mutex_unlock(kfd_get_dbgmgr_mutex());
651 
652 out:
653 	kfree(args_buff);
654 
655 	return status;
656 }
657 
658 /* Parse and generate fixed size data structure for wave control */
659 static int kfd_ioctl_dbg_wave_control(struct file *filep,
660 					struct kfd_process *p, void *data)
661 {
662 	struct kfd_ioctl_dbg_wave_control_args *args = data;
663 	struct kfd_dev *dev;
664 	struct dbg_wave_control_info wac_info;
665 	unsigned char *args_buff;
666 	uint32_t computed_buff_size;
667 	long status;
668 	void __user *cmd_from_user;
669 	unsigned int args_idx = 0;
670 
671 	memset((void *) &wac_info, 0, sizeof(struct dbg_wave_control_info));
672 
673 	/* we use compact form, independent of the packing attribute value */
674 	computed_buff_size = sizeof(*args) +
675 				sizeof(wac_info.mode) +
676 				sizeof(wac_info.operand) +
677 				sizeof(wac_info.dbgWave_msg.DbgWaveMsg) +
678 				sizeof(wac_info.dbgWave_msg.MemoryVA) +
679 				sizeof(wac_info.trapId);
680 
681 	dev = kfd_device_by_id(args->gpu_id);
682 	if (!dev)
683 		return -EINVAL;
684 
685 	if (dev->device_info->asic_family == CHIP_CARRIZO) {
686 		pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
687 		return -EINVAL;
688 	}
689 
690 	/* input size must match the computed "compact" size */
691 	if (args->buf_size_in_bytes != computed_buff_size) {
692 		pr_debug("size mismatch, computed : actual %u : %u\n",
693 				args->buf_size_in_bytes, computed_buff_size);
694 		return -EINVAL;
695 	}
696 
697 	cmd_from_user = (void __user *) args->content_ptr;
698 
699 	if (cmd_from_user == NULL)
700 		return -EINVAL;
701 
702 	/* copy the entire buffer from user */
703 
704 	args_buff = memdup_user(cmd_from_user,
705 				args->buf_size_in_bytes - sizeof(*args));
706 	if (IS_ERR(args_buff))
707 		return PTR_ERR(args_buff);
708 
709 	/* move ptr to the start of the "pay-load" area */
710 	wac_info.process = p;
711 
712 	wac_info.operand = *((enum HSA_DBG_WAVEOP *)(&args_buff[args_idx]));
713 	args_idx += sizeof(wac_info.operand);
714 
715 	wac_info.mode = *((enum HSA_DBG_WAVEMODE *)(&args_buff[args_idx]));
716 	args_idx += sizeof(wac_info.mode);
717 
718 	wac_info.trapId = *((uint32_t *)(&args_buff[args_idx]));
719 	args_idx += sizeof(wac_info.trapId);
720 
721 	wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value =
722 					*((uint32_t *)(&args_buff[args_idx]));
723 	wac_info.dbgWave_msg.MemoryVA = NULL;
724 
725 	mutex_lock(kfd_get_dbgmgr_mutex());
726 
727 	pr_debug("Calling dbg manager process %p, operand %u, mode %u, trapId %u, message %u\n",
728 			wac_info.process, wac_info.operand,
729 			wac_info.mode, wac_info.trapId,
730 			wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value);
731 
732 	status = kfd_dbgmgr_wave_control(dev->dbgmgr, &wac_info);
733 
734 	pr_debug("Returned status of dbg manager is %ld\n", status);
735 
736 	mutex_unlock(kfd_get_dbgmgr_mutex());
737 
738 	kfree(args_buff);
739 
740 	return status;
741 }
742 
743 static int kfd_ioctl_get_clock_counters(struct file *filep,
744 				struct kfd_process *p, void *data)
745 {
746 	struct kfd_ioctl_get_clock_counters_args *args = data;
747 	struct kfd_dev *dev;
748 	struct timespec64 time;
749 
750 	dev = kfd_device_by_id(args->gpu_id);
751 	if (dev == NULL)
752 		return -EINVAL;
753 
754 	/* Reading GPU clock counter from KGD */
755 	args->gpu_clock_counter =
756 		dev->kfd2kgd->get_gpu_clock_counter(dev->kgd);
757 
758 	/* No access to rdtsc. Using raw monotonic time */
759 	getrawmonotonic64(&time);
760 	args->cpu_clock_counter = (uint64_t)timespec64_to_ns(&time);
761 
762 	get_monotonic_boottime64(&time);
763 	args->system_clock_counter = (uint64_t)timespec64_to_ns(&time);
764 
765 	/* Since the counter is in nano-seconds we use 1GHz frequency */
766 	args->system_clock_freq = 1000000000;
767 
768 	return 0;
769 }
770 
771 
772 static int kfd_ioctl_get_process_apertures(struct file *filp,
773 				struct kfd_process *p, void *data)
774 {
775 	struct kfd_ioctl_get_process_apertures_args *args = data;
776 	struct kfd_process_device_apertures *pAperture;
777 	struct kfd_process_device *pdd;
778 
779 	dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
780 
781 	args->num_of_nodes = 0;
782 
783 	mutex_lock(&p->mutex);
784 
785 	/*if the process-device list isn't empty*/
786 	if (kfd_has_process_device_data(p)) {
787 		/* Run over all pdd of the process */
788 		pdd = kfd_get_first_process_device_data(p);
789 		do {
790 			pAperture =
791 				&args->process_apertures[args->num_of_nodes];
792 			pAperture->gpu_id = pdd->dev->id;
793 			pAperture->lds_base = pdd->lds_base;
794 			pAperture->lds_limit = pdd->lds_limit;
795 			pAperture->gpuvm_base = pdd->gpuvm_base;
796 			pAperture->gpuvm_limit = pdd->gpuvm_limit;
797 			pAperture->scratch_base = pdd->scratch_base;
798 			pAperture->scratch_limit = pdd->scratch_limit;
799 
800 			dev_dbg(kfd_device,
801 				"node id %u\n", args->num_of_nodes);
802 			dev_dbg(kfd_device,
803 				"gpu id %u\n", pdd->dev->id);
804 			dev_dbg(kfd_device,
805 				"lds_base %llX\n", pdd->lds_base);
806 			dev_dbg(kfd_device,
807 				"lds_limit %llX\n", pdd->lds_limit);
808 			dev_dbg(kfd_device,
809 				"gpuvm_base %llX\n", pdd->gpuvm_base);
810 			dev_dbg(kfd_device,
811 				"gpuvm_limit %llX\n", pdd->gpuvm_limit);
812 			dev_dbg(kfd_device,
813 				"scratch_base %llX\n", pdd->scratch_base);
814 			dev_dbg(kfd_device,
815 				"scratch_limit %llX\n", pdd->scratch_limit);
816 
817 			args->num_of_nodes++;
818 
819 			pdd = kfd_get_next_process_device_data(p, pdd);
820 		} while (pdd && (args->num_of_nodes < NUM_OF_SUPPORTED_GPUS));
821 	}
822 
823 	mutex_unlock(&p->mutex);
824 
825 	return 0;
826 }
827 
828 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
829 					void *data)
830 {
831 	struct kfd_ioctl_create_event_args *args = data;
832 	int err;
833 
834 	err = kfd_event_create(filp, p, args->event_type,
835 				args->auto_reset != 0, args->node_id,
836 				&args->event_id, &args->event_trigger_data,
837 				&args->event_page_offset,
838 				&args->event_slot_index);
839 
840 	return err;
841 }
842 
843 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
844 					void *data)
845 {
846 	struct kfd_ioctl_destroy_event_args *args = data;
847 
848 	return kfd_event_destroy(p, args->event_id);
849 }
850 
851 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
852 				void *data)
853 {
854 	struct kfd_ioctl_set_event_args *args = data;
855 
856 	return kfd_set_event(p, args->event_id);
857 }
858 
859 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
860 				void *data)
861 {
862 	struct kfd_ioctl_reset_event_args *args = data;
863 
864 	return kfd_reset_event(p, args->event_id);
865 }
866 
867 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
868 				void *data)
869 {
870 	struct kfd_ioctl_wait_events_args *args = data;
871 	int err;
872 
873 	err = kfd_wait_on_events(p, args->num_events,
874 			(void __user *)args->events_ptr,
875 			(args->wait_for_all != 0),
876 			args->timeout, &args->wait_result);
877 
878 	return err;
879 }
880 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
881 					struct kfd_process *p, void *data)
882 {
883 	struct kfd_ioctl_set_scratch_backing_va_args *args = data;
884 	struct kfd_process_device *pdd;
885 	struct kfd_dev *dev;
886 	long err;
887 
888 	dev = kfd_device_by_id(args->gpu_id);
889 	if (!dev)
890 		return -EINVAL;
891 
892 	mutex_lock(&p->mutex);
893 
894 	pdd = kfd_bind_process_to_device(dev, p);
895 	if (IS_ERR(pdd)) {
896 		err = PTR_ERR(pdd);
897 		goto bind_process_to_device_fail;
898 	}
899 
900 	pdd->qpd.sh_hidden_private_base = args->va_addr;
901 
902 	mutex_unlock(&p->mutex);
903 
904 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS && pdd->qpd.vmid != 0)
905 		dev->kfd2kgd->set_scratch_backing_va(
906 			dev->kgd, args->va_addr, pdd->qpd.vmid);
907 
908 	return 0;
909 
910 bind_process_to_device_fail:
911 	mutex_unlock(&p->mutex);
912 	return err;
913 }
914 
915 static int kfd_ioctl_get_tile_config(struct file *filep,
916 		struct kfd_process *p, void *data)
917 {
918 	struct kfd_ioctl_get_tile_config_args *args = data;
919 	struct kfd_dev *dev;
920 	struct tile_config config;
921 	int err = 0;
922 
923 	dev = kfd_device_by_id(args->gpu_id);
924 	if (!dev)
925 		return -EINVAL;
926 
927 	dev->kfd2kgd->get_tile_config(dev->kgd, &config);
928 
929 	args->gb_addr_config = config.gb_addr_config;
930 	args->num_banks = config.num_banks;
931 	args->num_ranks = config.num_ranks;
932 
933 	if (args->num_tile_configs > config.num_tile_configs)
934 		args->num_tile_configs = config.num_tile_configs;
935 	err = copy_to_user((void __user *)args->tile_config_ptr,
936 			config.tile_config_ptr,
937 			args->num_tile_configs * sizeof(uint32_t));
938 	if (err) {
939 		args->num_tile_configs = 0;
940 		return -EFAULT;
941 	}
942 
943 	if (args->num_macro_tile_configs > config.num_macro_tile_configs)
944 		args->num_macro_tile_configs =
945 				config.num_macro_tile_configs;
946 	err = copy_to_user((void __user *)args->macro_tile_config_ptr,
947 			config.macro_tile_config_ptr,
948 			args->num_macro_tile_configs * sizeof(uint32_t));
949 	if (err) {
950 		args->num_macro_tile_configs = 0;
951 		return -EFAULT;
952 	}
953 
954 	return 0;
955 }
956 
957 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
958 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
959 			    .cmd_drv = 0, .name = #ioctl}
960 
961 /** Ioctl table */
962 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
963 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
964 			kfd_ioctl_get_version, 0),
965 
966 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
967 			kfd_ioctl_create_queue, 0),
968 
969 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
970 			kfd_ioctl_destroy_queue, 0),
971 
972 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
973 			kfd_ioctl_set_memory_policy, 0),
974 
975 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
976 			kfd_ioctl_get_clock_counters, 0),
977 
978 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
979 			kfd_ioctl_get_process_apertures, 0),
980 
981 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
982 			kfd_ioctl_update_queue, 0),
983 
984 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
985 			kfd_ioctl_create_event, 0),
986 
987 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
988 			kfd_ioctl_destroy_event, 0),
989 
990 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
991 			kfd_ioctl_set_event, 0),
992 
993 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
994 			kfd_ioctl_reset_event, 0),
995 
996 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
997 			kfd_ioctl_wait_events, 0),
998 
999 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER,
1000 			kfd_ioctl_dbg_register, 0),
1001 
1002 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER,
1003 			kfd_ioctl_dbg_unregister, 0),
1004 
1005 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH,
1006 			kfd_ioctl_dbg_address_watch, 0),
1007 
1008 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL,
1009 			kfd_ioctl_dbg_wave_control, 0),
1010 
1011 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
1012 			kfd_ioctl_set_scratch_backing_va, 0),
1013 
1014 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
1015 			kfd_ioctl_get_tile_config, 0),
1016 
1017 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
1018 			kfd_ioctl_set_trap_handler, 0),
1019 };
1020 
1021 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
1022 
1023 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
1024 {
1025 	struct kfd_process *process;
1026 	amdkfd_ioctl_t *func;
1027 	const struct amdkfd_ioctl_desc *ioctl = NULL;
1028 	unsigned int nr = _IOC_NR(cmd);
1029 	char stack_kdata[128];
1030 	char *kdata = NULL;
1031 	unsigned int usize, asize;
1032 	int retcode = -EINVAL;
1033 
1034 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
1035 		goto err_i1;
1036 
1037 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
1038 		u32 amdkfd_size;
1039 
1040 		ioctl = &amdkfd_ioctls[nr];
1041 
1042 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
1043 		usize = asize = _IOC_SIZE(cmd);
1044 		if (amdkfd_size > asize)
1045 			asize = amdkfd_size;
1046 
1047 		cmd = ioctl->cmd;
1048 	} else
1049 		goto err_i1;
1050 
1051 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#%d), arg 0x%lx\n", cmd, nr, arg);
1052 
1053 	process = kfd_get_process(current);
1054 	if (IS_ERR(process)) {
1055 		dev_dbg(kfd_device, "no process\n");
1056 		goto err_i1;
1057 	}
1058 
1059 	/* Do not trust userspace, use our own definition */
1060 	func = ioctl->func;
1061 
1062 	if (unlikely(!func)) {
1063 		dev_dbg(kfd_device, "no function\n");
1064 		retcode = -EINVAL;
1065 		goto err_i1;
1066 	}
1067 
1068 	if (cmd & (IOC_IN | IOC_OUT)) {
1069 		if (asize <= sizeof(stack_kdata)) {
1070 			kdata = stack_kdata;
1071 		} else {
1072 			kdata = kmalloc(asize, GFP_KERNEL);
1073 			if (!kdata) {
1074 				retcode = -ENOMEM;
1075 				goto err_i1;
1076 			}
1077 		}
1078 		if (asize > usize)
1079 			memset(kdata + usize, 0, asize - usize);
1080 	}
1081 
1082 	if (cmd & IOC_IN) {
1083 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
1084 			retcode = -EFAULT;
1085 			goto err_i1;
1086 		}
1087 	} else if (cmd & IOC_OUT) {
1088 		memset(kdata, 0, usize);
1089 	}
1090 
1091 	retcode = func(filep, process, kdata);
1092 
1093 	if (cmd & IOC_OUT)
1094 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
1095 			retcode = -EFAULT;
1096 
1097 err_i1:
1098 	if (!ioctl)
1099 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
1100 			  task_pid_nr(current), cmd, nr);
1101 
1102 	if (kdata != stack_kdata)
1103 		kfree(kdata);
1104 
1105 	if (retcode)
1106 		dev_dbg(kfd_device, "ret = %d\n", retcode);
1107 
1108 	return retcode;
1109 }
1110 
1111 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
1112 {
1113 	struct kfd_process *process;
1114 
1115 	process = kfd_get_process(current);
1116 	if (IS_ERR(process))
1117 		return PTR_ERR(process);
1118 
1119 	if ((vma->vm_pgoff & KFD_MMAP_DOORBELL_MASK) ==
1120 			KFD_MMAP_DOORBELL_MASK) {
1121 		vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_DOORBELL_MASK;
1122 		return kfd_doorbell_mmap(process, vma);
1123 	} else if ((vma->vm_pgoff & KFD_MMAP_EVENTS_MASK) ==
1124 			KFD_MMAP_EVENTS_MASK) {
1125 		vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_EVENTS_MASK;
1126 		return kfd_event_mmap(process, vma);
1127 	} else if ((vma->vm_pgoff & KFD_MMAP_RESERVED_MEM_MASK) ==
1128 			KFD_MMAP_RESERVED_MEM_MASK) {
1129 		vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_RESERVED_MEM_MASK;
1130 		return kfd_reserved_mem_mmap(process, vma);
1131 	}
1132 
1133 	return -EFAULT;
1134 }
1135