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 process->is_32bit_user_mode = is_32bit_user_mode;
126 dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
127 process->pasid, process->is_32bit_user_mode);
129 kfd_init_apertures(process);
134 static long kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
137 struct kfd_ioctl_get_version_args args;
140 args.major_version = KFD_IOCTL_MAJOR_VERSION;
141 args.minor_version = KFD_IOCTL_MINOR_VERSION;
143 if (copy_to_user(arg, &args, sizeof(args)))
149 static int set_queue_properties_from_user(struct queue_properties *q_properties,
150 struct kfd_ioctl_create_queue_args *args)
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");
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");
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");
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");
175 if (!access_ok(VERIFY_WRITE,
176 (const void __user *) args->read_pointer_address,
178 pr_err("kfd: can't access read pointer\n");
182 if (!access_ok(VERIFY_WRITE,
183 (const void __user *) args->write_pointer_address,
185 pr_err("kfd: can't access write pointer\n");
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;
202 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
203 q_properties->format = KFD_QUEUE_FORMAT_AQL;
205 q_properties->format = KFD_QUEUE_FORMAT_PM4;
207 pr_debug("Queue Percentage (%d, %d)\n",
208 q_properties->queue_percent, args->queue_percentage);
210 pr_debug("Queue Priority (%d, %d)\n",
211 q_properties->priority, args->queue_priority);
213 pr_debug("Queue Address (0x%llX, 0x%llX)\n",
214 q_properties->queue_address, args->ring_base_address);
216 pr_debug("Queue Size (0x%llX, %u)\n",
217 q_properties->queue_size, args->ring_size);
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);
223 pr_debug("Queue Format (%d)\n", q_properties->format);
228 static long kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
231 struct kfd_ioctl_create_queue_args args;
234 unsigned int queue_id;
235 struct kfd_process_device *pdd;
236 struct queue_properties q_properties;
238 memset(&q_properties, 0, sizeof(struct queue_properties));
240 if (copy_from_user(&args, arg, sizeof(args)))
243 pr_debug("kfd: creating queue ioctl\n");
245 err = set_queue_properties_from_user(&q_properties, &args);
249 dev = kfd_device_by_id(args.gpu_id);
253 mutex_lock(&p->mutex);
255 pdd = kfd_bind_process_to_device(dev, p);
258 goto err_bind_process;
261 pr_debug("kfd: creating queue for PASID %d on GPU 0x%x\n",
265 err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, 0,
266 KFD_QUEUE_TYPE_COMPUTE, &queue_id);
268 goto err_create_queue;
270 args.queue_id = queue_id;
272 /* Return gpu_id as doorbell offset for mmap usage */
273 args.doorbell_offset = args.gpu_id << PAGE_SHIFT;
275 if (copy_to_user(arg, &args, sizeof(args))) {
277 goto err_copy_args_out;
280 mutex_unlock(&p->mutex);
282 pr_debug("kfd: queue id %d was created successfully\n", args.queue_id);
284 pr_debug("ring buffer address == 0x%016llX\n",
285 args.ring_base_address);
287 pr_debug("read ptr address == 0x%016llX\n",
288 args.read_pointer_address);
290 pr_debug("write ptr address == 0x%016llX\n",
291 args.write_pointer_address);
296 pqm_destroy_queue(&p->pqm, queue_id);
299 mutex_unlock(&p->mutex);
303 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
307 struct kfd_ioctl_destroy_queue_args args;
309 if (copy_from_user(&args, arg, sizeof(args)))
312 pr_debug("kfd: destroying queue id %d for PASID %d\n",
316 mutex_lock(&p->mutex);
318 retval = pqm_destroy_queue(&p->pqm, args.queue_id);
320 mutex_unlock(&p->mutex);
324 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
328 struct kfd_ioctl_update_queue_args args;
329 struct queue_properties properties;
331 if (copy_from_user(&args, arg, sizeof(args)))
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");
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");
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");
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");
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;
362 pr_debug("kfd: updating queue id %d for PASID %d\n",
363 args.queue_id, p->pasid);
365 mutex_lock(&p->mutex);
367 retval = pqm_update_queue(&p->pqm, args.queue_id, &properties);
369 mutex_unlock(&p->mutex);
374 static long kfd_ioctl_set_memory_policy(struct file *filep,
375 struct kfd_process *p, void __user *arg)
377 struct kfd_ioctl_set_memory_policy_args args;
380 struct kfd_process_device *pdd;
381 enum cache_policy default_policy, alternate_policy;
383 if (copy_from_user(&args, arg, sizeof(args)))
386 if (args.default_policy != KFD_IOC_CACHE_POLICY_COHERENT
387 && args.default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
391 if (args.alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
392 && args.alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
396 dev = kfd_device_by_id(args.gpu_id);
400 mutex_lock(&p->mutex);
402 pdd = kfd_bind_process_to_device(dev, p);
408 default_policy = (args.default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
409 ? cache_policy_coherent : cache_policy_noncoherent;
412 (args.alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
413 ? cache_policy_coherent : cache_policy_noncoherent;
415 if (!dev->dqm->set_cache_memory_policy(dev->dqm,
419 (void __user *)args.alternate_aperture_base,
420 args.alternate_aperture_size))
424 mutex_unlock(&p->mutex);
429 static long kfd_ioctl_get_clock_counters(struct file *filep,
430 struct kfd_process *p, void __user *arg)
432 struct kfd_ioctl_get_clock_counters_args args;
434 struct timespec time;
436 if (copy_from_user(&args, arg, sizeof(args)))
439 dev = kfd_device_by_id(args.gpu_id);
443 /* Reading GPU clock counter from KGD */
444 args.gpu_clock_counter = kfd2kgd->get_gpu_clock_counter(dev->kgd);
446 /* No access to rdtsc. Using raw monotonic time */
447 getrawmonotonic(&time);
448 args.cpu_clock_counter = (uint64_t)timespec_to_ns(&time);
450 get_monotonic_boottime(&time);
451 args.system_clock_counter = (uint64_t)timespec_to_ns(&time);
453 /* Since the counter is in nano-seconds we use 1GHz frequency */
454 args.system_clock_freq = 1000000000;
456 if (copy_to_user(arg, &args, sizeof(args)))
463 static int kfd_ioctl_get_process_apertures(struct file *filp,
464 struct kfd_process *p, void __user *arg)
466 struct kfd_ioctl_get_process_apertures_args args;
467 struct kfd_process_device_apertures *pAperture;
468 struct kfd_process_device *pdd;
470 dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
472 if (copy_from_user(&args, arg, sizeof(args)))
475 args.num_of_nodes = 0;
477 mutex_lock(&p->mutex);
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);
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;
494 "node id %u\n", args.num_of_nodes);
496 "gpu id %u\n", pdd->dev->id);
498 "lds_base %llX\n", pdd->lds_base);
500 "lds_limit %llX\n", pdd->lds_limit);
502 "gpuvm_base %llX\n", pdd->gpuvm_base);
504 "gpuvm_limit %llX\n", pdd->gpuvm_limit);
506 "scratch_base %llX\n", pdd->scratch_base);
508 "scratch_limit %llX\n", pdd->scratch_limit);
511 } while ((pdd = kfd_get_next_process_device_data(p, pdd)) != NULL &&
512 (args.num_of_nodes < NUM_OF_SUPPORTED_GPUS));
515 mutex_unlock(&p->mutex);
517 if (copy_to_user(arg, &args, sizeof(args)))
523 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
525 struct kfd_process *process;
529 "ioctl cmd 0x%x (#%d), arg 0x%lx\n",
530 cmd, _IOC_NR(cmd), arg);
532 process = kfd_get_process(current);
534 return PTR_ERR(process);
537 case KFD_IOC_GET_VERSION:
538 err = kfd_ioctl_get_version(filep, process, (void __user *)arg);
540 case KFD_IOC_CREATE_QUEUE:
541 err = kfd_ioctl_create_queue(filep, process,
545 case KFD_IOC_DESTROY_QUEUE:
546 err = kfd_ioctl_destroy_queue(filep, process,
550 case KFD_IOC_SET_MEMORY_POLICY:
551 err = kfd_ioctl_set_memory_policy(filep, process,
555 case KFD_IOC_GET_CLOCK_COUNTERS:
556 err = kfd_ioctl_get_clock_counters(filep, process,
560 case KFD_IOC_GET_PROCESS_APERTURES:
561 err = kfd_ioctl_get_process_apertures(filep, process,
565 case KFD_IOC_UPDATE_QUEUE:
566 err = kfd_ioctl_update_queue(filep, process,
572 "unknown ioctl cmd 0x%x, arg 0x%lx)\n",
580 "ioctl error %ld for ioctl cmd 0x%x (#%d)\n",
581 err, cmd, _IOC_NR(cmd));
586 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
588 struct kfd_process *process;
590 process = kfd_get_process(current);
592 return PTR_ERR(process);
594 return kfd_doorbell_mmap(process, vma);