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 <uapi/asm-generic/mman-common.h>
35 #include <asm/processor.h>
36 #include "kfd_priv.h"
37 #include "kfd_device_queue_manager.h"
38 
39 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
40 static int kfd_open(struct inode *, struct file *);
41 static int kfd_mmap(struct file *, struct vm_area_struct *);
42 
43 static const char kfd_dev_name[] = "kfd";
44 
45 static const struct file_operations kfd_fops = {
46 	.owner = THIS_MODULE,
47 	.unlocked_ioctl = kfd_ioctl,
48 	.compat_ioctl = kfd_ioctl,
49 	.open = kfd_open,
50 	.mmap = kfd_mmap,
51 };
52 
53 static int kfd_char_dev_major = -1;
54 static struct class *kfd_class;
55 struct device *kfd_device;
56 
57 int kfd_chardev_init(void)
58 {
59 	int err = 0;
60 
61 	kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
62 	err = kfd_char_dev_major;
63 	if (err < 0)
64 		goto err_register_chrdev;
65 
66 	kfd_class = class_create(THIS_MODULE, kfd_dev_name);
67 	err = PTR_ERR(kfd_class);
68 	if (IS_ERR(kfd_class))
69 		goto err_class_create;
70 
71 	kfd_device = device_create(kfd_class, NULL,
72 					MKDEV(kfd_char_dev_major, 0),
73 					NULL, kfd_dev_name);
74 	err = PTR_ERR(kfd_device);
75 	if (IS_ERR(kfd_device))
76 		goto err_device_create;
77 
78 	return 0;
79 
80 err_device_create:
81 	class_destroy(kfd_class);
82 err_class_create:
83 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
84 err_register_chrdev:
85 	return err;
86 }
87 
88 void kfd_chardev_exit(void)
89 {
90 	device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
91 	class_destroy(kfd_class);
92 	unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
93 }
94 
95 struct device *kfd_chardev(void)
96 {
97 	return kfd_device;
98 }
99 
100 
101 static int kfd_open(struct inode *inode, struct file *filep)
102 {
103 	struct kfd_process *process;
104 	bool is_32bit_user_mode;
105 
106 	if (iminor(inode) != 0)
107 		return -ENODEV;
108 
109 	is_32bit_user_mode = is_compat_task();
110 
111 	if (is_32bit_user_mode == true) {
112 		dev_warn(kfd_device,
113 			"Process %d (32-bit) failed to open /dev/kfd\n"
114 			"32-bit processes are not supported by amdkfd\n",
115 			current->pid);
116 		return -EPERM;
117 	}
118 
119 	process = kfd_create_process(current);
120 	if (IS_ERR(process))
121 		return PTR_ERR(process);
122 
123 	dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
124 		process->pasid, process->is_32bit_user_mode);
125 
126 	return 0;
127 }
128 
129 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
130 					void *data)
131 {
132 	struct kfd_ioctl_get_version_args *args = data;
133 	int err = 0;
134 
135 	args->major_version = KFD_IOCTL_MAJOR_VERSION;
136 	args->minor_version = KFD_IOCTL_MINOR_VERSION;
137 
138 	return err;
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("kfd: 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("kfd: 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("kfd: 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("kfd: 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("kfd: 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("kfd: can't access write pointer\n");
178 		return -EFAULT;
179 	}
180 
181 	q_properties->is_interop = false;
182 	q_properties->queue_percent = args->queue_percentage;
183 	q_properties->priority = args->queue_priority;
184 	q_properties->queue_address = args->ring_base_address;
185 	q_properties->queue_size = args->ring_size;
186 	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
187 	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
188 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
189 		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
190 		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
191 	else
192 		return -ENOTSUPP;
193 
194 	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
195 		q_properties->format = KFD_QUEUE_FORMAT_AQL;
196 	else
197 		q_properties->format = KFD_QUEUE_FORMAT_PM4;
198 
199 	pr_debug("Queue Percentage (%d, %d)\n",
200 			q_properties->queue_percent, args->queue_percentage);
201 
202 	pr_debug("Queue Priority (%d, %d)\n",
203 			q_properties->priority, args->queue_priority);
204 
205 	pr_debug("Queue Address (0x%llX, 0x%llX)\n",
206 			q_properties->queue_address, args->ring_base_address);
207 
208 	pr_debug("Queue Size (0x%llX, %u)\n",
209 			q_properties->queue_size, args->ring_size);
210 
211 	pr_debug("Queue r/w Pointers (0x%llX, 0x%llX)\n",
212 			(uint64_t) q_properties->read_ptr,
213 			(uint64_t) q_properties->write_ptr);
214 
215 	pr_debug("Queue Format (%d)\n", q_properties->format);
216 
217 	return 0;
218 }
219 
220 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
221 					void *data)
222 {
223 	struct kfd_ioctl_create_queue_args *args = data;
224 	struct kfd_dev *dev;
225 	int err = 0;
226 	unsigned int queue_id;
227 	struct kfd_process_device *pdd;
228 	struct queue_properties q_properties;
229 
230 	memset(&q_properties, 0, sizeof(struct queue_properties));
231 
232 	pr_debug("kfd: creating queue ioctl\n");
233 
234 	err = set_queue_properties_from_user(&q_properties, args);
235 	if (err)
236 		return err;
237 
238 	dev = kfd_device_by_id(args->gpu_id);
239 	if (dev == NULL)
240 		return -EINVAL;
241 
242 	mutex_lock(&p->mutex);
243 
244 	pdd = kfd_bind_process_to_device(dev, p);
245 	if (IS_ERR(pdd)) {
246 		err = -ESRCH;
247 		goto err_bind_process;
248 	}
249 
250 	pr_debug("kfd: creating queue for PASID %d on GPU 0x%x\n",
251 			p->pasid,
252 			dev->id);
253 
254 	err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, 0,
255 				KFD_QUEUE_TYPE_COMPUTE, &queue_id);
256 	if (err != 0)
257 		goto err_create_queue;
258 
259 	args->queue_id = queue_id;
260 
261 	/* Return gpu_id as doorbell offset for mmap usage */
262 	args->doorbell_offset = args->gpu_id << PAGE_SHIFT;
263 
264 	mutex_unlock(&p->mutex);
265 
266 	pr_debug("kfd: queue id %d was created successfully\n", args->queue_id);
267 
268 	pr_debug("ring buffer address == 0x%016llX\n",
269 			args->ring_base_address);
270 
271 	pr_debug("read ptr address    == 0x%016llX\n",
272 			args->read_pointer_address);
273 
274 	pr_debug("write ptr address   == 0x%016llX\n",
275 			args->write_pointer_address);
276 
277 	return 0;
278 
279 err_create_queue:
280 err_bind_process:
281 	mutex_unlock(&p->mutex);
282 	return err;
283 }
284 
285 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
286 					void *data)
287 {
288 	int retval;
289 	struct kfd_ioctl_destroy_queue_args *args = data;
290 
291 	pr_debug("kfd: destroying queue id %d for PASID %d\n",
292 				args->queue_id,
293 				p->pasid);
294 
295 	mutex_lock(&p->mutex);
296 
297 	retval = pqm_destroy_queue(&p->pqm, args->queue_id);
298 
299 	mutex_unlock(&p->mutex);
300 	return retval;
301 }
302 
303 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
304 					void *data)
305 {
306 	int retval;
307 	struct kfd_ioctl_update_queue_args *args = data;
308 	struct queue_properties properties;
309 
310 	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
311 		pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
312 		return -EINVAL;
313 	}
314 
315 	if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
316 		pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
317 		return -EINVAL;
318 	}
319 
320 	if ((args->ring_base_address) &&
321 		(!access_ok(VERIFY_WRITE,
322 			(const void __user *) args->ring_base_address,
323 			sizeof(uint64_t)))) {
324 		pr_err("kfd: can't access ring base address\n");
325 		return -EFAULT;
326 	}
327 
328 	if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
329 		pr_err("kfd: ring size must be a power of 2 or 0\n");
330 		return -EINVAL;
331 	}
332 
333 	properties.queue_address = args->ring_base_address;
334 	properties.queue_size = args->ring_size;
335 	properties.queue_percent = args->queue_percentage;
336 	properties.priority = args->queue_priority;
337 
338 	pr_debug("kfd: updating queue id %d for PASID %d\n",
339 			args->queue_id, p->pasid);
340 
341 	mutex_lock(&p->mutex);
342 
343 	retval = pqm_update_queue(&p->pqm, args->queue_id, &properties);
344 
345 	mutex_unlock(&p->mutex);
346 
347 	return retval;
348 }
349 
350 static int kfd_ioctl_set_memory_policy(struct file *filep,
351 					struct kfd_process *p, void *data)
352 {
353 	struct kfd_ioctl_set_memory_policy_args *args = data;
354 	struct kfd_dev *dev;
355 	int err = 0;
356 	struct kfd_process_device *pdd;
357 	enum cache_policy default_policy, alternate_policy;
358 
359 	if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
360 	    && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
361 		return -EINVAL;
362 	}
363 
364 	if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
365 	    && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
366 		return -EINVAL;
367 	}
368 
369 	dev = kfd_device_by_id(args->gpu_id);
370 	if (dev == NULL)
371 		return -EINVAL;
372 
373 	mutex_lock(&p->mutex);
374 
375 	pdd = kfd_bind_process_to_device(dev, p);
376 	if (IS_ERR(pdd)) {
377 		err = -ESRCH;
378 		goto out;
379 	}
380 
381 	default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
382 			 ? cache_policy_coherent : cache_policy_noncoherent;
383 
384 	alternate_policy =
385 		(args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
386 		   ? cache_policy_coherent : cache_policy_noncoherent;
387 
388 	if (!dev->dqm->set_cache_memory_policy(dev->dqm,
389 				&pdd->qpd,
390 				default_policy,
391 				alternate_policy,
392 				(void __user *)args->alternate_aperture_base,
393 				args->alternate_aperture_size))
394 		err = -EINVAL;
395 
396 out:
397 	mutex_unlock(&p->mutex);
398 
399 	return err;
400 }
401 
402 static int kfd_ioctl_get_clock_counters(struct file *filep,
403 				struct kfd_process *p, void *data)
404 {
405 	struct kfd_ioctl_get_clock_counters_args *args = data;
406 	struct kfd_dev *dev;
407 	struct timespec time;
408 
409 	dev = kfd_device_by_id(args->gpu_id);
410 	if (dev == NULL)
411 		return -EINVAL;
412 
413 	/* Reading GPU clock counter from KGD */
414 	args->gpu_clock_counter = kfd2kgd->get_gpu_clock_counter(dev->kgd);
415 
416 	/* No access to rdtsc. Using raw monotonic time */
417 	getrawmonotonic(&time);
418 	args->cpu_clock_counter = (uint64_t)timespec_to_ns(&time);
419 
420 	get_monotonic_boottime(&time);
421 	args->system_clock_counter = (uint64_t)timespec_to_ns(&time);
422 
423 	/* Since the counter is in nano-seconds we use 1GHz frequency */
424 	args->system_clock_freq = 1000000000;
425 
426 	return 0;
427 }
428 
429 
430 static int kfd_ioctl_get_process_apertures(struct file *filp,
431 				struct kfd_process *p, void *data)
432 {
433 	struct kfd_ioctl_get_process_apertures_args *args = data;
434 	struct kfd_process_device_apertures *pAperture;
435 	struct kfd_process_device *pdd;
436 
437 	dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
438 
439 	args->num_of_nodes = 0;
440 
441 	mutex_lock(&p->mutex);
442 
443 	/*if the process-device list isn't empty*/
444 	if (kfd_has_process_device_data(p)) {
445 		/* Run over all pdd of the process */
446 		pdd = kfd_get_first_process_device_data(p);
447 		do {
448 			pAperture =
449 				&args->process_apertures[args->num_of_nodes];
450 			pAperture->gpu_id = pdd->dev->id;
451 			pAperture->lds_base = pdd->lds_base;
452 			pAperture->lds_limit = pdd->lds_limit;
453 			pAperture->gpuvm_base = pdd->gpuvm_base;
454 			pAperture->gpuvm_limit = pdd->gpuvm_limit;
455 			pAperture->scratch_base = pdd->scratch_base;
456 			pAperture->scratch_limit = pdd->scratch_limit;
457 
458 			dev_dbg(kfd_device,
459 				"node id %u\n", args->num_of_nodes);
460 			dev_dbg(kfd_device,
461 				"gpu id %u\n", pdd->dev->id);
462 			dev_dbg(kfd_device,
463 				"lds_base %llX\n", pdd->lds_base);
464 			dev_dbg(kfd_device,
465 				"lds_limit %llX\n", pdd->lds_limit);
466 			dev_dbg(kfd_device,
467 				"gpuvm_base %llX\n", pdd->gpuvm_base);
468 			dev_dbg(kfd_device,
469 				"gpuvm_limit %llX\n", pdd->gpuvm_limit);
470 			dev_dbg(kfd_device,
471 				"scratch_base %llX\n", pdd->scratch_base);
472 			dev_dbg(kfd_device,
473 				"scratch_limit %llX\n", pdd->scratch_limit);
474 
475 			args->num_of_nodes++;
476 		} while ((pdd = kfd_get_next_process_device_data(p, pdd)) != NULL &&
477 				(args->num_of_nodes < NUM_OF_SUPPORTED_GPUS));
478 	}
479 
480 	mutex_unlock(&p->mutex);
481 
482 	return 0;
483 }
484 
485 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
486 	[_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, .cmd_drv = 0, .name = #ioctl}
487 
488 /** Ioctl table */
489 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
490 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
491 			kfd_ioctl_get_version, 0),
492 
493 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
494 			kfd_ioctl_create_queue, 0),
495 
496 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
497 			kfd_ioctl_destroy_queue, 0),
498 
499 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
500 			kfd_ioctl_set_memory_policy, 0),
501 
502 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
503 			kfd_ioctl_get_clock_counters, 0),
504 
505 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
506 			kfd_ioctl_get_process_apertures, 0),
507 
508 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
509 			kfd_ioctl_update_queue, 0),
510 };
511 
512 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
513 
514 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
515 {
516 	struct kfd_process *process;
517 	amdkfd_ioctl_t *func;
518 	const struct amdkfd_ioctl_desc *ioctl = NULL;
519 	unsigned int nr = _IOC_NR(cmd);
520 	char stack_kdata[128];
521 	char *kdata = NULL;
522 	unsigned int usize, asize;
523 	int retcode = -EINVAL;
524 
525 	if (nr >= AMDKFD_CORE_IOCTL_COUNT)
526 		goto err_i1;
527 
528 	if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
529 		u32 amdkfd_size;
530 
531 		ioctl = &amdkfd_ioctls[nr];
532 
533 		amdkfd_size = _IOC_SIZE(ioctl->cmd);
534 		usize = asize = _IOC_SIZE(cmd);
535 		if (amdkfd_size > asize)
536 			asize = amdkfd_size;
537 
538 		cmd = ioctl->cmd;
539 	} else
540 		goto err_i1;
541 
542 	dev_dbg(kfd_device, "ioctl cmd 0x%x (#%d), arg 0x%lx\n", cmd, nr, arg);
543 
544 	process = kfd_get_process(current);
545 	if (IS_ERR(process)) {
546 		dev_dbg(kfd_device, "no process\n");
547 		goto err_i1;
548 	}
549 
550 	/* Do not trust userspace, use our own definition */
551 	func = ioctl->func;
552 
553 	if (unlikely(!func)) {
554 		dev_dbg(kfd_device, "no function\n");
555 		retcode = -EINVAL;
556 		goto err_i1;
557 	}
558 
559 	if (cmd & (IOC_IN | IOC_OUT)) {
560 		if (asize <= sizeof(stack_kdata)) {
561 			kdata = stack_kdata;
562 		} else {
563 			kdata = kmalloc(asize, GFP_KERNEL);
564 			if (!kdata) {
565 				retcode = -ENOMEM;
566 				goto err_i1;
567 			}
568 		}
569 		if (asize > usize)
570 			memset(kdata + usize, 0, asize - usize);
571 	}
572 
573 	if (cmd & IOC_IN) {
574 		if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
575 			retcode = -EFAULT;
576 			goto err_i1;
577 		}
578 	} else if (cmd & IOC_OUT) {
579 		memset(kdata, 0, usize);
580 	}
581 
582 	retcode = func(filep, process, kdata);
583 
584 	if (cmd & IOC_OUT)
585 		if (copy_to_user((void __user *)arg, kdata, usize) != 0)
586 			retcode = -EFAULT;
587 
588 err_i1:
589 	if (!ioctl)
590 		dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
591 			  task_pid_nr(current), cmd, nr);
592 
593 	if (kdata != stack_kdata)
594 		kfree(kdata);
595 
596 	if (retcode)
597 		dev_dbg(kfd_device, "ret = %d\n", retcode);
598 
599 	return retcode;
600 }
601 
602 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
603 {
604 	struct kfd_process *process;
605 
606 	process = kfd_get_process(current);
607 	if (IS_ERR(process))
608 		return PTR_ERR(process);
609 
610 	return kfd_doorbell_mmap(process, vma);
611 }
612