drivers/gpu/drm/amd/amdkfd/kfd_chardev.c

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