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)
150 if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
151 pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
155 if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
156 pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
160 if ((args->ring_base_address) &&
161 (!access_ok(VERIFY_WRITE,
162 (const void __user *) args->ring_base_address,
163 sizeof(uint64_t)))) {
164 pr_err("kfd: can't access ring base address\n");
168 if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
169 pr_err("kfd: ring size must be a power of 2 or 0\n");
173 if (!access_ok(VERIFY_WRITE,
174 (const void __user *) args->read_pointer_address,
176 pr_err("kfd: can't access read pointer\n");
180 if (!access_ok(VERIFY_WRITE,
181 (const void __user *) args->write_pointer_address,
183 pr_err("kfd: can't access write pointer\n");
187 tmp = (void *)(uintptr_t)args->eop_buffer_address;
189 !access_ok(VERIFY_WRITE, tmp, sizeof(uint32_t))) {
190 pr_debug("kfd: can't access eop buffer");
194 tmp = (void *)(uintptr_t)args->ctx_save_restore_address;
196 !access_ok(VERIFY_WRITE, tmp, sizeof(uint32_t))) {
197 pr_debug("kfd: can't access ctx save restore buffer");
201 q_properties->is_interop = false;
202 q_properties->queue_percent = args->queue_percentage;
203 q_properties->priority = args->queue_priority;
204 q_properties->queue_address = args->ring_base_address;
205 q_properties->queue_size = args->ring_size;
206 q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
207 q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
208 q_properties->eop_ring_buffer_address = args->eop_buffer_address;
209 q_properties->eop_ring_buffer_size = args->eop_buffer_size;
210 q_properties->ctx_save_restore_area_address =
211 args->ctx_save_restore_address;
212 q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
213 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
214 args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
215 q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
216 else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
217 q_properties->type = KFD_QUEUE_TYPE_SDMA;
221 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
222 q_properties->format = KFD_QUEUE_FORMAT_AQL;
224 q_properties->format = KFD_QUEUE_FORMAT_PM4;
226 pr_debug("Queue Percentage (%d, %d)\n",
227 q_properties->queue_percent, args->queue_percentage);
229 pr_debug("Queue Priority (%d, %d)\n",
230 q_properties->priority, args->queue_priority);
232 pr_debug("Queue Address (0x%llX, 0x%llX)\n",
233 q_properties->queue_address, args->ring_base_address);
235 pr_debug("Queue Size (0x%llX, %u)\n",
236 q_properties->queue_size, args->ring_size);
238 pr_debug("Queue r/w Pointers (0x%llX, 0x%llX)\n",
239 (uint64_t) q_properties->read_ptr,
240 (uint64_t) q_properties->write_ptr);
242 pr_debug("Queue Format (%d)\n", q_properties->format);
244 pr_debug("Queue EOP (0x%llX)\n", q_properties->eop_ring_buffer_address);
246 pr_debug("Queue CTX save arex (0x%llX)\n",
247 q_properties->ctx_save_restore_area_address);
252 static long kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
255 struct kfd_ioctl_create_queue_args args;
258 unsigned int queue_id;
259 struct kfd_process_device *pdd;
260 struct queue_properties q_properties;
262 memset(&q_properties, 0, sizeof(struct queue_properties));
264 if (copy_from_user(&args, arg, sizeof(args)))
267 pr_debug("kfd: creating queue ioctl\n");
269 err = set_queue_properties_from_user(&q_properties, &args);
273 pr_debug("kfd: looking for gpu id 0x%x\n", args.gpu_id);
274 dev = kfd_device_by_id(args.gpu_id);
276 pr_debug("kfd: gpu id 0x%x was not found\n", args.gpu_id);
280 mutex_lock(&p->mutex);
282 pdd = kfd_bind_process_to_device(dev, p);
285 goto err_bind_process;
288 pr_debug("kfd: creating queue for PASID %d on GPU 0x%x\n",
292 err = pqm_create_queue(&p->pqm, dev, filep, &q_properties,
293 0, q_properties.type, &queue_id);
295 goto err_create_queue;
297 args.queue_id = queue_id;
299 /* Return gpu_id as doorbell offset for mmap usage */
300 args.doorbell_offset = args.gpu_id << PAGE_SHIFT;
302 if (copy_to_user(arg, &args, sizeof(args))) {
304 goto err_copy_args_out;
307 mutex_unlock(&p->mutex);
309 pr_debug("kfd: queue id %d was created successfully\n", args.queue_id);
311 pr_debug("ring buffer address == 0x%016llX\n",
312 args.ring_base_address);
314 pr_debug("read ptr address == 0x%016llX\n",
315 args.read_pointer_address);
317 pr_debug("write ptr address == 0x%016llX\n",
318 args.write_pointer_address);
323 pqm_destroy_queue(&p->pqm, queue_id);
326 mutex_unlock(&p->mutex);
330 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
334 struct kfd_ioctl_destroy_queue_args args;
336 if (copy_from_user(&args, arg, sizeof(args)))
339 pr_debug("kfd: destroying queue id %d for PASID %d\n",
343 mutex_lock(&p->mutex);
345 retval = pqm_destroy_queue(&p->pqm, args.queue_id);
347 mutex_unlock(&p->mutex);
351 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
355 struct kfd_ioctl_update_queue_args args;
356 struct queue_properties properties;
358 if (copy_from_user(&args, arg, sizeof(args)))
361 if (args.queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
362 pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
366 if (args.queue_priority > KFD_MAX_QUEUE_PRIORITY) {
367 pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
371 if ((args.ring_base_address) &&
372 (!access_ok(VERIFY_WRITE,
373 (const void __user *) args.ring_base_address,
374 sizeof(uint64_t)))) {
375 pr_err("kfd: can't access ring base address\n");
379 if (!is_power_of_2(args.ring_size) && (args.ring_size != 0)) {
380 pr_err("kfd: ring size must be a power of 2 or 0\n");
384 properties.queue_address = args.ring_base_address;
385 properties.queue_size = args.ring_size;
386 properties.queue_percent = args.queue_percentage;
387 properties.priority = args.queue_priority;
389 pr_debug("kfd: updating queue id %d for PASID %d\n",
390 args.queue_id, p->pasid);
392 mutex_lock(&p->mutex);
394 retval = pqm_update_queue(&p->pqm, args.queue_id, &properties);
396 mutex_unlock(&p->mutex);
401 static long kfd_ioctl_set_memory_policy(struct file *filep,
402 struct kfd_process *p, void __user *arg)
404 struct kfd_ioctl_set_memory_policy_args args;
407 struct kfd_process_device *pdd;
408 enum cache_policy default_policy, alternate_policy;
410 if (copy_from_user(&args, arg, sizeof(args)))
413 if (args.default_policy != KFD_IOC_CACHE_POLICY_COHERENT
414 && args.default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
418 if (args.alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
419 && args.alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
423 dev = kfd_device_by_id(args.gpu_id);
427 mutex_lock(&p->mutex);
429 pdd = kfd_bind_process_to_device(dev, p);
435 default_policy = (args.default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
436 ? cache_policy_coherent : cache_policy_noncoherent;
439 (args.alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
440 ? cache_policy_coherent : cache_policy_noncoherent;
442 if (!dev->dqm->ops.set_cache_memory_policy(dev->dqm,
446 (void __user *)args.alternate_aperture_base,
447 args.alternate_aperture_size))
451 mutex_unlock(&p->mutex);
456 static long kfd_ioctl_get_clock_counters(struct file *filep,
457 struct kfd_process *p, void __user *arg)
459 struct kfd_ioctl_get_clock_counters_args args;
461 struct timespec time;
463 if (copy_from_user(&args, arg, sizeof(args)))
466 dev = kfd_device_by_id(args.gpu_id);
470 /* Reading GPU clock counter from KGD */
471 args.gpu_clock_counter = kfd2kgd->get_gpu_clock_counter(dev->kgd);
473 /* No access to rdtsc. Using raw monotonic time */
474 getrawmonotonic(&time);
475 args.cpu_clock_counter = (uint64_t)timespec_to_ns(&time);
477 get_monotonic_boottime(&time);
478 args.system_clock_counter = (uint64_t)timespec_to_ns(&time);
480 /* Since the counter is in nano-seconds we use 1GHz frequency */
481 args.system_clock_freq = 1000000000;
483 if (copy_to_user(arg, &args, sizeof(args)))
490 static int kfd_ioctl_get_process_apertures(struct file *filp,
491 struct kfd_process *p, void __user *arg)
493 struct kfd_ioctl_get_process_apertures_args args;
494 struct kfd_process_device_apertures *pAperture;
495 struct kfd_process_device *pdd;
497 dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
499 if (copy_from_user(&args, arg, sizeof(args)))
502 args.num_of_nodes = 0;
504 mutex_lock(&p->mutex);
506 /*if the process-device list isn't empty*/
507 if (kfd_has_process_device_data(p)) {
508 /* Run over all pdd of the process */
509 pdd = kfd_get_first_process_device_data(p);
511 pAperture = &args.process_apertures[args.num_of_nodes];
512 pAperture->gpu_id = pdd->dev->id;
513 pAperture->lds_base = pdd->lds_base;
514 pAperture->lds_limit = pdd->lds_limit;
515 pAperture->gpuvm_base = pdd->gpuvm_base;
516 pAperture->gpuvm_limit = pdd->gpuvm_limit;
517 pAperture->scratch_base = pdd->scratch_base;
518 pAperture->scratch_limit = pdd->scratch_limit;
521 "node id %u\n", args.num_of_nodes);
523 "gpu id %u\n", pdd->dev->id);
525 "lds_base %llX\n", pdd->lds_base);
527 "lds_limit %llX\n", pdd->lds_limit);
529 "gpuvm_base %llX\n", pdd->gpuvm_base);
531 "gpuvm_limit %llX\n", pdd->gpuvm_limit);
533 "scratch_base %llX\n", pdd->scratch_base);
535 "scratch_limit %llX\n", pdd->scratch_limit);
538 } while ((pdd = kfd_get_next_process_device_data(p, pdd)) != NULL &&
539 (args.num_of_nodes < NUM_OF_SUPPORTED_GPUS));
542 mutex_unlock(&p->mutex);
544 if (copy_to_user(arg, &args, sizeof(args)))
550 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
552 struct kfd_process *process;
556 "ioctl cmd 0x%x (#%d), arg 0x%lx\n",
557 cmd, _IOC_NR(cmd), arg);
559 process = kfd_get_process(current);
561 return PTR_ERR(process);
564 case KFD_IOC_GET_VERSION:
565 err = kfd_ioctl_get_version(filep, process, (void __user *)arg);
567 case KFD_IOC_CREATE_QUEUE:
568 err = kfd_ioctl_create_queue(filep, process,
572 case KFD_IOC_DESTROY_QUEUE:
573 err = kfd_ioctl_destroy_queue(filep, process,
577 case KFD_IOC_SET_MEMORY_POLICY:
578 err = kfd_ioctl_set_memory_policy(filep, process,
582 case KFD_IOC_GET_CLOCK_COUNTERS:
583 err = kfd_ioctl_get_clock_counters(filep, process,
587 case KFD_IOC_GET_PROCESS_APERTURES:
588 err = kfd_ioctl_get_process_apertures(filep, process,
592 case KFD_IOC_UPDATE_QUEUE:
593 err = kfd_ioctl_update_queue(filep, process,
599 "unknown ioctl cmd 0x%x, arg 0x%lx)\n",
607 "ioctl error %ld for ioctl cmd 0x%x (#%d)\n",
608 err, cmd, _IOC_NR(cmd));
613 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
615 struct kfd_process *process;
617 process = kfd_get_process(current);
619 return PTR_ERR(process);
621 return kfd_doorbell_mmap(process, vma);