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/uaccess.h>
35 #include <uapi/asm-generic/mman-common.h>
36 #include <asm/processor.h>
37 #include "kfd_priv.h"
38 #include "kfd_device_queue_manager.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 	process->is_32bit_user_mode = is_32bit_user_mode;
125 
126 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
127 		process->pasid, process->is_32bit_user_mode);
128 
129 	kfd_init_apertures(process);
130 
131 	return 0;
132 }
133 
134 static long kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
135 					void __user *arg)
136 {
137 	struct kfd_ioctl_get_version_args args;
138 	int err = 0;
139 
140 	args.major_version = KFD_IOCTL_MAJOR_VERSION;
141 	args.minor_version = KFD_IOCTL_MINOR_VERSION;
142 
143 	if (copy_to_user(arg, &args, sizeof(args)))
144 		err = -EFAULT;
145 
146 	return err;
147 }
148 
149 static int set_queue_properties_from_user(struct queue_properties *q_properties,
150 				struct kfd_ioctl_create_queue_args *args)
151 {
152 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
153 		pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
154 		return -EINVAL;
155 	}
156 
157 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
158 		pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
159 		return -EINVAL;
160 	}
161 
162 	if ((args->ring_base_address) &&
163 		(!access_ok(VERIFY_WRITE,
164 			(const void __user *) args->ring_base_address,
165 			sizeof(uint64_t)))) {
166 		pr_err("kfd: can't access ring base address\n");
167 		return -EFAULT;
168 	}
169 
170 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
171 		pr_err("kfd: ring size must be a power of 2 or 0\n");
172 		return -EINVAL;
173 	}
174 
175 	if (!access_ok(VERIFY_WRITE,
176 			(const void __user *) args->read_pointer_address,
177 			sizeof(uint32_t))) {
178 		pr_err("kfd: can't access read pointer\n");
179 		return -EFAULT;
180 	}
181 
182 	if (!access_ok(VERIFY_WRITE,
183 			(const void __user *) args->write_pointer_address,
184 			sizeof(uint32_t))) {
185 		pr_err("kfd: can't access write pointer\n");
186 		return -EFAULT;
187 	}
188 
189 	q_properties->is_interop = false;
190 	q_properties->queue_percent = args->queue_percentage;
191 	q_properties->priority = args->queue_priority;
192 	q_properties->queue_address = args->ring_base_address;
193 	q_properties->queue_size = args->ring_size;
194 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
195 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
196 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
197 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
198 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
199 	else
200 		return -ENOTSUPP;
201 
202 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
203 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
204 	else
205 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
206 
207 	pr_debug("Queue Percentage (%d, %d)\n",
208 			q_properties->queue_percent, args->queue_percentage);
209 
210 	pr_debug("Queue Priority (%d, %d)\n",
211 			q_properties->priority, args->queue_priority);
212 
213 	pr_debug("Queue Address (0x%llX, 0x%llX)\n",
214 			q_properties->queue_address, args->ring_base_address);
215 
216 	pr_debug("Queue Size (0x%llX, %u)\n",
217 			q_properties->queue_size, args->ring_size);
218 
219 	pr_debug("Queue r/w Pointers (0x%llX, 0x%llX)\n",
220 			(uint64_t) q_properties->read_ptr,
221 			(uint64_t) q_properties->write_ptr);
222 
223 	pr_debug("Queue Format (%d)\n", q_properties->format);
224 
225 	return 0;
226 }
227 
228 static long kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
229 					void __user *arg)
230 {
231 	struct kfd_ioctl_create_queue_args args;
232 	struct kfd_dev *dev;
233 	int err = 0;
234 	unsigned int queue_id;
235 	struct kfd_process_device *pdd;
236 	struct queue_properties q_properties;
237 
238 	memset(&q_properties, 0, sizeof(struct queue_properties));
239 
240 	if (copy_from_user(&args, arg, sizeof(args)))
241 		return -EFAULT;
242 
243 	pr_debug("kfd: creating queue ioctl\n");
244 
245 	err = set_queue_properties_from_user(&q_properties, &args);
246 	if (err)
247 		return err;
248 
249 	dev = kfd_device_by_id(args.gpu_id);
250 	if (dev == NULL)
251 		return -EINVAL;
252 
253 	mutex_lock(&p->mutex);
254 
255 	pdd = kfd_bind_process_to_device(dev, p);
256 	if (IS_ERR(pdd)) {
257 		err = PTR_ERR(pdd);
258 		goto err_bind_process;
259 	}
260 
261 	pr_debug("kfd: creating queue for PASID %d on GPU 0x%x\n",
262 			p->pasid,
263 			dev->id);
264 
265 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, 0,
266 				KFD_QUEUE_TYPE_COMPUTE, &queue_id);
267 	if (err != 0)
268 		goto err_create_queue;
269 
270 	args.queue_id = queue_id;
271 
272 	/* Return gpu_id as doorbell offset for mmap usage */
273 	args.doorbell_offset = args.gpu_id << PAGE_SHIFT;
274 
275 	if (copy_to_user(arg, &args, sizeof(args))) {
276 		err = -EFAULT;
277 		goto err_copy_args_out;
278 	}
279 
280 	mutex_unlock(&p->mutex);
281 
282 	pr_debug("kfd: queue id %d was created successfully\n", args.queue_id);
283 
284 	pr_debug("ring buffer address == 0x%016llX\n",
285 			args.ring_base_address);
286 
287 	pr_debug("read ptr address    == 0x%016llX\n",
288 			args.read_pointer_address);
289 
290 	pr_debug("write ptr address   == 0x%016llX\n",
291 			args.write_pointer_address);
292 
293 	return 0;
294 
295 err_copy_args_out:
296 	pqm_destroy_queue(&p->pqm, queue_id);
297 err_create_queue:
298 err_bind_process:
299 	mutex_unlock(&p->mutex);
300 	return err;
301 }
302 
303 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
304 					void __user *arg)
305 {
306 	int retval;
307 	struct kfd_ioctl_destroy_queue_args args;
308 
309 	if (copy_from_user(&args, arg, sizeof(args)))
310 		return -EFAULT;
311 
312 	pr_debug("kfd: destroying queue id %d for PASID %d\n",
313 				args.queue_id,
314 				p->pasid);
315 
316 	mutex_lock(&p->mutex);
317 
318 	retval = pqm_destroy_queue(&p->pqm, args.queue_id);
319 
320 	mutex_unlock(&p->mutex);
321 	return retval;
322 }
323 
324 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
325 					void __user *arg)
326 {
327 	int retval;
328 	struct kfd_ioctl_update_queue_args args;
329 	struct queue_properties properties;
330 
331 	if (copy_from_user(&args, arg, sizeof(args)))
332 		return -EFAULT;
333 
334 	if (args.queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
335 		pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
336 		return -EINVAL;
337 	}
338 
339 	if (args.queue_priority > KFD_MAX_QUEUE_PRIORITY) {
340 		pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
341 		return -EINVAL;
342 	}
343 
344 	if ((args.ring_base_address) &&
345 		(!access_ok(VERIFY_WRITE,
346 			(const void __user *) args.ring_base_address,
347 			sizeof(uint64_t)))) {
348 		pr_err("kfd: can't access ring base address\n");
349 		return -EFAULT;
350 	}
351 
352 	if (!is_power_of_2(args.ring_size) && (args.ring_size != 0)) {
353 		pr_err("kfd: ring size must be a power of 2 or 0\n");
354 		return -EINVAL;
355 	}
356 
357 	properties.queue_address = args.ring_base_address;
358 	properties.queue_size = args.ring_size;
359 	properties.queue_percent = args.queue_percentage;
360 	properties.priority = args.queue_priority;
361 
362 	pr_debug("kfd: updating queue id %d for PASID %d\n",
363 			args.queue_id, p->pasid);
364 
365 	mutex_lock(&p->mutex);
366 
367 	retval = pqm_update_queue(&p->pqm, args.queue_id, &properties);
368 
369 	mutex_unlock(&p->mutex);
370 
371 	return retval;
372 }
373 
374 static long kfd_ioctl_set_memory_policy(struct file *filep,
375 				struct kfd_process *p, void __user *arg)
376 {
377 	struct kfd_ioctl_set_memory_policy_args args;
378 	struct kfd_dev *dev;
379 	int err = 0;
380 	struct kfd_process_device *pdd;
381 	enum cache_policy default_policy, alternate_policy;
382 
383 	if (copy_from_user(&args, arg, sizeof(args)))
384 		return -EFAULT;
385 
386 	if (args.default_policy != KFD_IOC_CACHE_POLICY_COHERENT
387 	    && args.default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
388 		return -EINVAL;
389 	}
390 
391 	if (args.alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
392 	    && args.alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
393 		return -EINVAL;
394 	}
395 
396 	dev = kfd_device_by_id(args.gpu_id);
397 	if (dev == NULL)
398 		return -EINVAL;
399 
400 	mutex_lock(&p->mutex);
401 
402 	pdd = kfd_bind_process_to_device(dev, p);
403 	if (IS_ERR(pdd)) {
404 		err = PTR_ERR(pdd);
405 		goto out;
406 	}
407 
408 	default_policy = (args.default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
409 			 ? cache_policy_coherent : cache_policy_noncoherent;
410 
411 	alternate_policy =
412 		(args.alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
413 		   ? cache_policy_coherent : cache_policy_noncoherent;
414 
415 	if (!dev->dqm->set_cache_memory_policy(dev->dqm,
416 				&pdd->qpd,
417 				default_policy,
418 				alternate_policy,
419 				(void __user *)args.alternate_aperture_base,
420 				args.alternate_aperture_size))
421 		err = -EINVAL;
422 
423 out:
424 	mutex_unlock(&p->mutex);
425 
426 	return err;
427 }
428 
429 static long kfd_ioctl_get_clock_counters(struct file *filep,
430 				struct kfd_process *p, void __user *arg)
431 {
432 	struct kfd_ioctl_get_clock_counters_args args;
433 	struct kfd_dev *dev;
434 	struct timespec time;
435 
436 	if (copy_from_user(&args, arg, sizeof(args)))
437 		return -EFAULT;
438 
439 	dev = kfd_device_by_id(args.gpu_id);
440 	if (dev == NULL)
441 		return -EINVAL;
442 
443 	/* Reading GPU clock counter from KGD */
444 	args.gpu_clock_counter = kfd2kgd->get_gpu_clock_counter(dev->kgd);
445 
446 	/* No access to rdtsc. Using raw monotonic time */
447 	getrawmonotonic(&time);
448 	args.cpu_clock_counter = (uint64_t)timespec_to_ns(&time);
449 
450 	get_monotonic_boottime(&time);
451 	args.system_clock_counter = (uint64_t)timespec_to_ns(&time);
452 
453 	/* Since the counter is in nano-seconds we use 1GHz frequency */
454 	args.system_clock_freq = 1000000000;
455 
456 	if (copy_to_user(arg, &args, sizeof(args)))
457 		return -EFAULT;
458 
459 	return 0;
460 }
461 
462 
463 static int kfd_ioctl_get_process_apertures(struct file *filp,
464 				struct kfd_process *p, void __user *arg)
465 {
466 	struct kfd_ioctl_get_process_apertures_args args;
467 	struct kfd_process_device_apertures *pAperture;
468 	struct kfd_process_device *pdd;
469 
470 	dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
471 
472 	if (copy_from_user(&args, arg, sizeof(args)))
473 		return -EFAULT;
474 
475 	args.num_of_nodes = 0;
476 
477 	mutex_lock(&p->mutex);
478 
479 	/*if the process-device list isn't empty*/
480 	if (kfd_has_process_device_data(p)) {
481 		/* Run over all pdd of the process */
482 		pdd = kfd_get_first_process_device_data(p);
483 		do {
484 			pAperture = &args.process_apertures[args.num_of_nodes];
485 			pAperture->gpu_id = pdd->dev->id;
486 			pAperture->lds_base = pdd->lds_base;
487 			pAperture->lds_limit = pdd->lds_limit;
488 			pAperture->gpuvm_base = pdd->gpuvm_base;
489 			pAperture->gpuvm_limit = pdd->gpuvm_limit;
490 			pAperture->scratch_base = pdd->scratch_base;
491 			pAperture->scratch_limit = pdd->scratch_limit;
492 
493 			dev_dbg(kfd_device,
494 				"node id %u\n", args.num_of_nodes);
495 			dev_dbg(kfd_device,
496 				"gpu id %u\n", pdd->dev->id);
497 			dev_dbg(kfd_device,
498 				"lds_base %llX\n", pdd->lds_base);
499 			dev_dbg(kfd_device,
500 				"lds_limit %llX\n", pdd->lds_limit);
501 			dev_dbg(kfd_device,
502 				"gpuvm_base %llX\n", pdd->gpuvm_base);
503 			dev_dbg(kfd_device,
504 				"gpuvm_limit %llX\n", pdd->gpuvm_limit);
505 			dev_dbg(kfd_device,
506 				"scratch_base %llX\n", pdd->scratch_base);
507 			dev_dbg(kfd_device,
508 				"scratch_limit %llX\n", pdd->scratch_limit);
509 
510 			args.num_of_nodes++;
511 		} while ((pdd = kfd_get_next_process_device_data(p, pdd)) != NULL &&
512 				(args.num_of_nodes < NUM_OF_SUPPORTED_GPUS));
513 	}
514 
515 	mutex_unlock(&p->mutex);
516 
517 	if (copy_to_user(arg, &args, sizeof(args)))
518 		return -EFAULT;
519 
520 	return 0;
521 }
522 
523 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
524 {
525 	struct kfd_process *process;
526 	long err = -EINVAL;
527 
528 	dev_dbg(kfd_device,
529 		"ioctl cmd 0x%x (#%d), arg 0x%lx\n",
530 		cmd, _IOC_NR(cmd), arg);
531 
532 	process = kfd_get_process(current);
533 	if (IS_ERR(process))
534 		return PTR_ERR(process);
535 
536 	switch (cmd) {
537 	case KFD_IOC_GET_VERSION:
538 		err = kfd_ioctl_get_version(filep, process, (void __user *)arg);
539 		break;
540 	case KFD_IOC_CREATE_QUEUE:
541 		err = kfd_ioctl_create_queue(filep, process,
542 						(void __user *)arg);
543 		break;
544 
545 	case KFD_IOC_DESTROY_QUEUE:
546 		err = kfd_ioctl_destroy_queue(filep, process,
547 						(void __user *)arg);
548 		break;
549 
550 	case KFD_IOC_SET_MEMORY_POLICY:
551 		err = kfd_ioctl_set_memory_policy(filep, process,
552 						(void __user *)arg);
553 		break;
554 
555 	case KFD_IOC_GET_CLOCK_COUNTERS:
556 		err = kfd_ioctl_get_clock_counters(filep, process,
557 						(void __user *)arg);
558 		break;
559 
560 	case KFD_IOC_GET_PROCESS_APERTURES:
561 		err = kfd_ioctl_get_process_apertures(filep, process,
562 						(void __user *)arg);
563 		break;
564 
565 	case KFD_IOC_UPDATE_QUEUE:
566 		err = kfd_ioctl_update_queue(filep, process,
567 						(void __user *)arg);
568 		break;
569 
570 	default:
571 		dev_err(kfd_device,
572 			"unknown ioctl cmd 0x%x, arg 0x%lx)\n",
573 			cmd, arg);
574 		err = -EINVAL;
575 		break;
576 	}
577 
578 	if (err < 0)
579 		dev_err(kfd_device,
580 			"ioctl error %ld for ioctl cmd 0x%x (#%d)\n",
581 			err, cmd, _IOC_NR(cmd));
582 
583 	return err;
584 }
585 
586 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
587 {
588 	struct kfd_process *process;
589 
590 	process = kfd_get_process(current);
591 	if (IS_ERR(process))
592 		return PTR_ERR(process);
593 
594 	return kfd_doorbell_mmap(process, vma);
595 }
596