2 * Copyright 2014 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
23 #include <linux/device.h>
24 #include <linux/export.h>
25 #include <linux/err.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>
34 #include <linux/uaccess.h>
35 #include <uapi/asm-generic/mman-common.h>
36 #include <asm/processor.h>
38 #include "kfd_device_queue_manager.h"
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 *);
44 static const char kfd_dev_name[] = "kfd";
46 static const struct file_operations kfd_fops = {
48 .unlocked_ioctl = kfd_ioctl,
49 .compat_ioctl = kfd_ioctl,
54 static int kfd_char_dev_major = -1;
55 static struct class *kfd_class;
56 struct device *kfd_device;
58 int kfd_chardev_init(void)
62 kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
63 err = kfd_char_dev_major;
65 goto err_register_chrdev;
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;
72 kfd_device = device_create(kfd_class, NULL,
73 MKDEV(kfd_char_dev_major, 0),
75 err = PTR_ERR(kfd_device);
76 if (IS_ERR(kfd_device))
77 goto err_device_create;
82 class_destroy(kfd_class);
84 unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
89 void kfd_chardev_exit(void)
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);
96 struct device *kfd_chardev(void)
102 static int kfd_open(struct inode *inode, struct file *filep)
104 struct kfd_process *process;
105 bool is_32bit_user_mode;
107 if (iminor(inode) != 0)
110 is_32bit_user_mode = is_compat_task();
112 if (is_32bit_user_mode == true) {
114 "Process %d (32-bit) failed to open /dev/kfd\n"
115 "32-bit processes are not supported by amdkfd\n",
120 process = kfd_create_process(current);
122 return PTR_ERR(process);
124 dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
125 process->pasid, process->is_32bit_user_mode);
130 static long kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
133 struct kfd_ioctl_get_version_args args;
136 args.major_version = KFD_IOCTL_MAJOR_VERSION;
137 args.minor_version = KFD_IOCTL_MINOR_VERSION;
139 if (copy_to_user(arg, &args, sizeof(args)))
145 static int set_queue_properties_from_user(struct queue_properties *q_properties,
146 struct kfd_ioctl_create_queue_args *args)
148 if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
149 pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
153 if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
154 pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
158 if ((args->ring_base_address) &&
159 (!access_ok(VERIFY_WRITE,
160 (const void __user *) args->ring_base_address,
161 sizeof(uint64_t)))) {
162 pr_err("kfd: can't access ring base address\n");
166 if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
167 pr_err("kfd: ring size must be a power of 2 or 0\n");
171 if (!access_ok(VERIFY_WRITE,
172 (const void __user *) args->read_pointer_address,
174 pr_err("kfd: can't access read pointer\n");
178 if (!access_ok(VERIFY_WRITE,
179 (const void __user *) args->write_pointer_address,
181 pr_err("kfd: can't access write pointer\n");
185 q_properties->is_interop = false;
186 q_properties->queue_percent = args->queue_percentage;
187 q_properties->priority = args->queue_priority;
188 q_properties->queue_address = args->ring_base_address;
189 q_properties->queue_size = args->ring_size;
190 q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
191 q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
192 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
193 args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
194 q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
198 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
199 q_properties->format = KFD_QUEUE_FORMAT_AQL;
201 q_properties->format = KFD_QUEUE_FORMAT_PM4;
203 pr_debug("Queue Percentage (%d, %d)\n",
204 q_properties->queue_percent, args->queue_percentage);
206 pr_debug("Queue Priority (%d, %d)\n",
207 q_properties->priority, args->queue_priority);
209 pr_debug("Queue Address (0x%llX, 0x%llX)\n",
210 q_properties->queue_address, args->ring_base_address);
212 pr_debug("Queue Size (0x%llX, %u)\n",
213 q_properties->queue_size, args->ring_size);
215 pr_debug("Queue r/w Pointers (0x%llX, 0x%llX)\n",
216 (uint64_t) q_properties->read_ptr,
217 (uint64_t) q_properties->write_ptr);
219 pr_debug("Queue Format (%d)\n", q_properties->format);
224 static long kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
227 struct kfd_ioctl_create_queue_args args;
230 unsigned int queue_id;
231 struct kfd_process_device *pdd;
232 struct queue_properties q_properties;
234 memset(&q_properties, 0, sizeof(struct queue_properties));
236 if (copy_from_user(&args, arg, sizeof(args)))
239 pr_debug("kfd: creating queue ioctl\n");
241 err = set_queue_properties_from_user(&q_properties, &args);
245 dev = kfd_device_by_id(args.gpu_id);
249 mutex_lock(&p->mutex);
251 pdd = kfd_bind_process_to_device(dev, p);
254 goto err_bind_process;
257 pr_debug("kfd: creating queue for PASID %d on GPU 0x%x\n",
261 err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, 0,
262 KFD_QUEUE_TYPE_COMPUTE, &queue_id);
264 goto err_create_queue;
266 args.queue_id = queue_id;
268 /* Return gpu_id as doorbell offset for mmap usage */
269 args.doorbell_offset = args.gpu_id << PAGE_SHIFT;
271 if (copy_to_user(arg, &args, sizeof(args))) {
273 goto err_copy_args_out;
276 mutex_unlock(&p->mutex);
278 pr_debug("kfd: queue id %d was created successfully\n", args.queue_id);
280 pr_debug("ring buffer address == 0x%016llX\n",
281 args.ring_base_address);
283 pr_debug("read ptr address == 0x%016llX\n",
284 args.read_pointer_address);
286 pr_debug("write ptr address == 0x%016llX\n",
287 args.write_pointer_address);
292 pqm_destroy_queue(&p->pqm, queue_id);
295 mutex_unlock(&p->mutex);
299 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
303 struct kfd_ioctl_destroy_queue_args args;
305 if (copy_from_user(&args, arg, sizeof(args)))
308 pr_debug("kfd: destroying queue id %d for PASID %d\n",
312 mutex_lock(&p->mutex);
314 retval = pqm_destroy_queue(&p->pqm, args.queue_id);
316 mutex_unlock(&p->mutex);
320 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
324 struct kfd_ioctl_update_queue_args args;
325 struct queue_properties properties;
327 if (copy_from_user(&args, arg, sizeof(args)))
330 if (args.queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
331 pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
335 if (args.queue_priority > KFD_MAX_QUEUE_PRIORITY) {
336 pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
340 if ((args.ring_base_address) &&
341 (!access_ok(VERIFY_WRITE,
342 (const void __user *) args.ring_base_address,
343 sizeof(uint64_t)))) {
344 pr_err("kfd: can't access ring base address\n");
348 if (!is_power_of_2(args.ring_size) && (args.ring_size != 0)) {
349 pr_err("kfd: ring size must be a power of 2 or 0\n");
353 properties.queue_address = args.ring_base_address;
354 properties.queue_size = args.ring_size;
355 properties.queue_percent = args.queue_percentage;
356 properties.priority = args.queue_priority;
358 pr_debug("kfd: updating queue id %d for PASID %d\n",
359 args.queue_id, p->pasid);
361 mutex_lock(&p->mutex);
363 retval = pqm_update_queue(&p->pqm, args.queue_id, &properties);
365 mutex_unlock(&p->mutex);
370 static long kfd_ioctl_set_memory_policy(struct file *filep,
371 struct kfd_process *p, void __user *arg)
373 struct kfd_ioctl_set_memory_policy_args args;
376 struct kfd_process_device *pdd;
377 enum cache_policy default_policy, alternate_policy;
379 if (copy_from_user(&args, arg, sizeof(args)))
382 if (args.default_policy != KFD_IOC_CACHE_POLICY_COHERENT
383 && args.default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
387 if (args.alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
388 && args.alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
392 dev = kfd_device_by_id(args.gpu_id);
396 mutex_lock(&p->mutex);
398 pdd = kfd_bind_process_to_device(dev, p);
404 default_policy = (args.default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
405 ? cache_policy_coherent : cache_policy_noncoherent;
408 (args.alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
409 ? cache_policy_coherent : cache_policy_noncoherent;
411 if (!dev->dqm->set_cache_memory_policy(dev->dqm,
415 (void __user *)args.alternate_aperture_base,
416 args.alternate_aperture_size))
420 mutex_unlock(&p->mutex);
425 static long kfd_ioctl_get_clock_counters(struct file *filep,
426 struct kfd_process *p, void __user *arg)
428 struct kfd_ioctl_get_clock_counters_args args;
430 struct timespec time;
432 if (copy_from_user(&args, arg, sizeof(args)))
435 dev = kfd_device_by_id(args.gpu_id);
439 /* Reading GPU clock counter from KGD */
440 args.gpu_clock_counter = kfd2kgd->get_gpu_clock_counter(dev->kgd);
442 /* No access to rdtsc. Using raw monotonic time */
443 getrawmonotonic(&time);
444 args.cpu_clock_counter = (uint64_t)timespec_to_ns(&time);
446 get_monotonic_boottime(&time);
447 args.system_clock_counter = (uint64_t)timespec_to_ns(&time);
449 /* Since the counter is in nano-seconds we use 1GHz frequency */
450 args.system_clock_freq = 1000000000;
452 if (copy_to_user(arg, &args, sizeof(args)))
459 static int kfd_ioctl_get_process_apertures(struct file *filp,
460 struct kfd_process *p, void __user *arg)
462 struct kfd_ioctl_get_process_apertures_args args;
463 struct kfd_process_device_apertures *pAperture;
464 struct kfd_process_device *pdd;
466 dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
468 if (copy_from_user(&args, arg, sizeof(args)))
471 args.num_of_nodes = 0;
473 mutex_lock(&p->mutex);
475 /*if the process-device list isn't empty*/
476 if (kfd_has_process_device_data(p)) {
477 /* Run over all pdd of the process */
478 pdd = kfd_get_first_process_device_data(p);
480 pAperture = &args.process_apertures[args.num_of_nodes];
481 pAperture->gpu_id = pdd->dev->id;
482 pAperture->lds_base = pdd->lds_base;
483 pAperture->lds_limit = pdd->lds_limit;
484 pAperture->gpuvm_base = pdd->gpuvm_base;
485 pAperture->gpuvm_limit = pdd->gpuvm_limit;
486 pAperture->scratch_base = pdd->scratch_base;
487 pAperture->scratch_limit = pdd->scratch_limit;
490 "node id %u\n", args.num_of_nodes);
492 "gpu id %u\n", pdd->dev->id);
494 "lds_base %llX\n", pdd->lds_base);
496 "lds_limit %llX\n", pdd->lds_limit);
498 "gpuvm_base %llX\n", pdd->gpuvm_base);
500 "gpuvm_limit %llX\n", pdd->gpuvm_limit);
502 "scratch_base %llX\n", pdd->scratch_base);
504 "scratch_limit %llX\n", pdd->scratch_limit);
507 } while ((pdd = kfd_get_next_process_device_data(p, pdd)) != NULL &&
508 (args.num_of_nodes < NUM_OF_SUPPORTED_GPUS));
511 mutex_unlock(&p->mutex);
513 if (copy_to_user(arg, &args, sizeof(args)))
519 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
521 struct kfd_process *process;
525 "ioctl cmd 0x%x (#%d), arg 0x%lx\n",
526 cmd, _IOC_NR(cmd), arg);
528 process = kfd_get_process(current);
530 return PTR_ERR(process);
533 case KFD_IOC_GET_VERSION:
534 err = kfd_ioctl_get_version(filep, process, (void __user *)arg);
536 case KFD_IOC_CREATE_QUEUE:
537 err = kfd_ioctl_create_queue(filep, process,
541 case KFD_IOC_DESTROY_QUEUE:
542 err = kfd_ioctl_destroy_queue(filep, process,
546 case KFD_IOC_SET_MEMORY_POLICY:
547 err = kfd_ioctl_set_memory_policy(filep, process,
551 case KFD_IOC_GET_CLOCK_COUNTERS:
552 err = kfd_ioctl_get_clock_counters(filep, process,
556 case KFD_IOC_GET_PROCESS_APERTURES:
557 err = kfd_ioctl_get_process_apertures(filep, process,
561 case KFD_IOC_UPDATE_QUEUE:
562 err = kfd_ioctl_update_queue(filep, process,
568 "unknown ioctl cmd 0x%x, arg 0x%lx)\n",
576 "ioctl error %ld for ioctl cmd 0x%x (#%d)\n",
577 err, cmd, _IOC_NR(cmd));
582 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
584 struct kfd_process *process;
586 process = kfd_get_process(current);
588 return PTR_ERR(process);
590 return kfd_doorbell_mmap(process, vma);