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 <linux/uaccess.h>
  35 #include <uapi/asm-generic/mman-common.h>
  36 #include <asm/processor.h>
  37 #include "kfd_priv.h"
  38 #include "kfd_device_queue_manager.h"
  39
  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 *);
  43
  44 static const char kfd_dev_name[] = "kfd";
  45
  46 static const struct file_operations kfd_fops = {
  47         .owner = THIS_MODULE,
  48         .unlocked_ioctl = kfd_ioctl,
  49         .compat_ioctl = kfd_ioctl,
  50         .open = kfd_open,
  51         .mmap = kfd_mmap,
  52 };
  53
  54 static int kfd_char_dev_major = -1;
  55 static struct class *kfd_class;
  56 struct device *kfd_device;
  57
  58 int kfd_chardev_init(void)
  59 {
  60         int err = 0;
  61
  62         kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
  63         err = kfd_char_dev_major;
  64         if (err < 0)
  65                 goto err_register_chrdev;
  66
  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;
  71
  72         kfd_device = device_create(kfd_class, NULL,
  73                                         MKDEV(kfd_char_dev_major, 0),
  74                                         NULL, kfd_dev_name);
  75         err = PTR_ERR(kfd_device);
  76         if (IS_ERR(kfd_device))
  77                 goto err_device_create;
  78
  79         return 0;
  80
  81 err_device_create:
  82         class_destroy(kfd_class);
  83 err_class_create:
  84         unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
  85 err_register_chrdev:
  86         return err;
  87 }
  88
  89 void kfd_chardev_exit(void)
  90 {
  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);
  94 }
  95
  96 struct device *kfd_chardev(void)
  97 {
  98         return kfd_device;
  99 }
 100
 101
 102 static int kfd_open(struct inode *inode, struct file *filep)
 103 {
 104         struct kfd_process *process;
 105         bool is_32bit_user_mode;
 106
 107         if (iminor(inode) != 0)
 108                 return -ENODEV;
 109
 110         is_32bit_user_mode = is_compat_task();
 111
 112         if (is_32bit_user_mode == true) {
 113                 dev_warn(kfd_device,
 114                         "Process %d (32-bit) failed to open /dev/kfd\n"
 115                         "32-bit processes are not supported by amdkfd\n",
 116                         current->pid);
 117                 return -EPERM;
 118         }
 119
 120         process = kfd_create_process(current);
 121         if (IS_ERR(process))
 122                 return PTR_ERR(process);
 123
 124         dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
 125                 process->pasid, process->is_32bit_user_mode);
 126
 127         return 0;
 128 }
 129
 130 static long kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
 131                                         void __user *arg)
 132 {
 133         struct kfd_ioctl_get_version_args args;
 134         int err = 0;
 135
 136         args.major_version = KFD_IOCTL_MAJOR_VERSION;
 137         args.minor_version = KFD_IOCTL_MINOR_VERSION;
 138
 139         if (copy_to_user(arg, &args, sizeof(args)))
 140                 err = -EFAULT;
 141
 142         return err;
 143 }
 144
 145 static int set_queue_properties_from_user(struct queue_properties *q_properties,
 146                                 struct kfd_ioctl_create_queue_args *args)
 147 {
 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");
 150                 return -EINVAL;
 151         }
 152
 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");
 155                 return -EINVAL;
 156         }
 157
 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");
 163                 return -EFAULT;
 164         }
 165
 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");
 168                 return -EINVAL;
 169         }
 170
 171         if (!access_ok(VERIFY_WRITE,
 172                         (const void __user *) args->read_pointer_address,
 173                         sizeof(uint32_t))) {
 174                 pr_err("kfd: can't access read pointer\n");
 175                 return -EFAULT;
 176         }
 177
 178         if (!access_ok(VERIFY_WRITE,
 179                         (const void __user *) args->write_pointer_address,
 180                         sizeof(uint32_t))) {
 181                 pr_err("kfd: can't access write pointer\n");
 182                 return -EFAULT;
 183         }
 184
 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;
 195         else
 196                 return -ENOTSUPP;
 197
 198         if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
 199                 q_properties->format = KFD_QUEUE_FORMAT_AQL;
 200         else
 201                 q_properties->format = KFD_QUEUE_FORMAT_PM4;
 202
 203         pr_debug("Queue Percentage (%d, %d)\n",
 204                         q_properties->queue_percent, args->queue_percentage);
 205
 206         pr_debug("Queue Priority (%d, %d)\n",
 207                         q_properties->priority, args->queue_priority);
 208
 209         pr_debug("Queue Address (0x%llX, 0x%llX)\n",
 210                         q_properties->queue_address, args->ring_base_address);
 211
 212         pr_debug("Queue Size (0x%llX, %u)\n",
 213                         q_properties->queue_size, args->ring_size);
 214
 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);
 218
 219         pr_debug("Queue Format (%d)\n", q_properties->format);
 220
 221         return 0;
 222 }
 223
 224 static long kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 225                                         void __user *arg)
 226 {
 227         struct kfd_ioctl_create_queue_args args;
 228         struct kfd_dev *dev;
 229         int err = 0;
 230         unsigned int queue_id;
 231         struct kfd_process_device *pdd;
 232         struct queue_properties q_properties;
 233
 234         memset(&q_properties, 0, sizeof(struct queue_properties));
 235
 236         if (copy_from_user(&args, arg, sizeof(args)))
 237                 return -EFAULT;
 238
 239         pr_debug("kfd: creating queue ioctl\n");
 240
 241         err = set_queue_properties_from_user(&q_properties, &args);
 242         if (err)
 243                 return err;
 244
 245         dev = kfd_device_by_id(args.gpu_id);
 246         if (dev == NULL)
 247                 return -EINVAL;
 248
 249         mutex_lock(&p->mutex);
 250
 251         pdd = kfd_bind_process_to_device(dev, p);
 252         if (IS_ERR(pdd)) {
 253                 err = PTR_ERR(pdd);
 254                 goto err_bind_process;
 255         }
 256
 257         pr_debug("kfd: creating queue for PASID %d on GPU 0x%x\n",
 258                         p->pasid,
 259                         dev->id);
 260
 261         err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, 0,
 262                                 KFD_QUEUE_TYPE_COMPUTE, &queue_id);
 263         if (err != 0)
 264                 goto err_create_queue;
 265
 266         args.queue_id = queue_id;
 267
 268         /* Return gpu_id as doorbell offset for mmap usage */
 269         args.doorbell_offset = args.gpu_id << PAGE_SHIFT;
 270
 271         if (copy_to_user(arg, &args, sizeof(args))) {
 272                 err = -EFAULT;
 273                 goto err_copy_args_out;
 274         }
 275
 276         mutex_unlock(&p->mutex);
 277
 278         pr_debug("kfd: queue id %d was created successfully\n", args.queue_id);
 279
 280         pr_debug("ring buffer address == 0x%016llX\n",
 281                         args.ring_base_address);
 282
 283         pr_debug("read ptr address    == 0x%016llX\n",
 284                         args.read_pointer_address);
 285
 286         pr_debug("write ptr address   == 0x%016llX\n",
 287                         args.write_pointer_address);
 288
 289         return 0;
 290
 291 err_copy_args_out:
 292         pqm_destroy_queue(&p->pqm, queue_id);
 293 err_create_queue:
 294 err_bind_process:
 295         mutex_unlock(&p->mutex);
 296         return err;
 297 }
 298
 299 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
 300                                         void __user *arg)
 301 {
 302         int retval;
 303         struct kfd_ioctl_destroy_queue_args args;
 304
 305         if (copy_from_user(&args, arg, sizeof(args)))
 306                 return -EFAULT;
 307
 308         pr_debug("kfd: destroying queue id %d for PASID %d\n",
 309                                 args.queue_id,
 310                                 p->pasid);
 311
 312         mutex_lock(&p->mutex);
 313
 314         retval = pqm_destroy_queue(&p->pqm, args.queue_id);
 315
 316         mutex_unlock(&p->mutex);
 317         return retval;
 318 }
 319
 320 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
 321                                         void __user *arg)
 322 {
 323         int retval;
 324         struct kfd_ioctl_update_queue_args args;
 325         struct queue_properties properties;
 326
 327         if (copy_from_user(&args, arg, sizeof(args)))
 328                 return -EFAULT;
 329
 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");
 332                 return -EINVAL;
 333         }
 334
 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");
 337                 return -EINVAL;
 338         }
 339
 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");
 345                 return -EFAULT;
 346         }
 347
 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");
 350                 return -EINVAL;
 351         }
 352
 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;
 357
 358         pr_debug("kfd: updating queue id %d for PASID %d\n",
 359                         args.queue_id, p->pasid);
 360
 361         mutex_lock(&p->mutex);
 362
 363         retval = pqm_update_queue(&p->pqm, args.queue_id, &properties);
 364
 365         mutex_unlock(&p->mutex);
 366
 367         return retval;
 368 }
 369
 370 static long kfd_ioctl_set_memory_policy(struct file *filep,
 371                                 struct kfd_process *p, void __user *arg)
 372 {
 373         struct kfd_ioctl_set_memory_policy_args args;
 374         struct kfd_dev *dev;
 375         int err = 0;
 376         struct kfd_process_device *pdd;
 377         enum cache_policy default_policy, alternate_policy;
 378
 379         if (copy_from_user(&args, arg, sizeof(args)))
 380                 return -EFAULT;
 381
 382         if (args.default_policy != KFD_IOC_CACHE_POLICY_COHERENT
 383             && args.default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 384                 return -EINVAL;
 385         }
 386
 387         if (args.alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
 388             && args.alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 389                 return -EINVAL;
 390         }
 391
 392         dev = kfd_device_by_id(args.gpu_id);
 393         if (dev == NULL)
 394                 return -EINVAL;
 395
 396         mutex_lock(&p->mutex);
 397
 398         pdd = kfd_bind_process_to_device(dev, p);
 399         if (IS_ERR(pdd)) {
 400                 err = PTR_ERR(pdd);
 401                 goto out;
 402         }
 403
 404         default_policy = (args.default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 405                          ? cache_policy_coherent : cache_policy_noncoherent;
 406
 407         alternate_policy =
 408                 (args.alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 409                    ? cache_policy_coherent : cache_policy_noncoherent;
 410
 411         if (!dev->dqm->set_cache_memory_policy(dev->dqm,
 412                                 &pdd->qpd,
 413                                 default_policy,
 414                                 alternate_policy,
 415                                 (void __user *)args.alternate_aperture_base,
 416                                 args.alternate_aperture_size))
 417                 err = -EINVAL;
 418
 419 out:
 420         mutex_unlock(&p->mutex);
 421
 422         return err;
 423 }
 424
 425 static long kfd_ioctl_get_clock_counters(struct file *filep,
 426                                 struct kfd_process *p, void __user *arg)
 427 {
 428         struct kfd_ioctl_get_clock_counters_args args;
 429         struct kfd_dev *dev;
 430         struct timespec time;
 431
 432         if (copy_from_user(&args, arg, sizeof(args)))
 433                 return -EFAULT;
 434
 435         dev = kfd_device_by_id(args.gpu_id);
 436         if (dev == NULL)
 437                 return -EINVAL;
 438
 439         /* Reading GPU clock counter from KGD */
 440         args.gpu_clock_counter = kfd2kgd->get_gpu_clock_counter(dev->kgd);
 441
 442         /* No access to rdtsc. Using raw monotonic time */
 443         getrawmonotonic(&time);
 444         args.cpu_clock_counter = (uint64_t)timespec_to_ns(&time);
 445
 446         get_monotonic_boottime(&time);
 447         args.system_clock_counter = (uint64_t)timespec_to_ns(&time);
 448
 449         /* Since the counter is in nano-seconds we use 1GHz frequency */
 450         args.system_clock_freq = 1000000000;
 451
 452         if (copy_to_user(arg, &args, sizeof(args)))
 453                 return -EFAULT;
 454
 455         return 0;
 456 }
 457
 458
 459 static int kfd_ioctl_get_process_apertures(struct file *filp,
 460                                 struct kfd_process *p, void __user *arg)
 461 {
 462         struct kfd_ioctl_get_process_apertures_args args;
 463         struct kfd_process_device_apertures *pAperture;
 464         struct kfd_process_device *pdd;
 465
 466         dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
 467
 468         if (copy_from_user(&args, arg, sizeof(args)))
 469                 return -EFAULT;
 470
 471         args.num_of_nodes = 0;
 472
 473         mutex_lock(&p->mutex);
 474
 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);
 479                 do {
 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;
 488
 489                         dev_dbg(kfd_device,
 490                                 "node id %u\n", args.num_of_nodes);
 491                         dev_dbg(kfd_device,
 492                                 "gpu id %u\n", pdd->dev->id);
 493                         dev_dbg(kfd_device,
 494                                 "lds_base %llX\n", pdd->lds_base);
 495                         dev_dbg(kfd_device,
 496                                 "lds_limit %llX\n", pdd->lds_limit);
 497                         dev_dbg(kfd_device,
 498                                 "gpuvm_base %llX\n", pdd->gpuvm_base);
 499                         dev_dbg(kfd_device,
 500                                 "gpuvm_limit %llX\n", pdd->gpuvm_limit);
 501                         dev_dbg(kfd_device,
 502                                 "scratch_base %llX\n", pdd->scratch_base);
 503                         dev_dbg(kfd_device,
 504                                 "scratch_limit %llX\n", pdd->scratch_limit);
 505
 506                         args.num_of_nodes++;
 507                 } while ((pdd = kfd_get_next_process_device_data(p, pdd)) != NULL &&
 508                                 (args.num_of_nodes < NUM_OF_SUPPORTED_GPUS));
 509         }
 510
 511         mutex_unlock(&p->mutex);
 512
 513         if (copy_to_user(arg, &args, sizeof(args)))
 514                 return -EFAULT;
 515
 516         return 0;
 517 }
 518
 519 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
 520 {
 521         struct kfd_process *process;
 522         long err = -EINVAL;
 523
 524         dev_dbg(kfd_device,
 525                 "ioctl cmd 0x%x (#%d), arg 0x%lx\n",
 526                 cmd, _IOC_NR(cmd), arg);
 527
 528         process = kfd_get_process(current);
 529         if (IS_ERR(process))
 530                 return PTR_ERR(process);
 531
 532         switch (cmd) {
 533         case KFD_IOC_GET_VERSION:
 534                 err = kfd_ioctl_get_version(filep, process, (void __user *)arg);
 535                 break;
 536         case KFD_IOC_CREATE_QUEUE:
 537                 err = kfd_ioctl_create_queue(filep, process,
 538                                                 (void __user *)arg);
 539                 break;
 540
 541         case KFD_IOC_DESTROY_QUEUE:
 542                 err = kfd_ioctl_destroy_queue(filep, process,
 543                                                 (void __user *)arg);
 544                 break;
 545
 546         case KFD_IOC_SET_MEMORY_POLICY:
 547                 err = kfd_ioctl_set_memory_policy(filep, process,
 548                                                 (void __user *)arg);
 549                 break;
 550
 551         case KFD_IOC_GET_CLOCK_COUNTERS:
 552                 err = kfd_ioctl_get_clock_counters(filep, process,
 553                                                 (void __user *)arg);
 554                 break;
 555
 556         case KFD_IOC_GET_PROCESS_APERTURES:
 557                 err = kfd_ioctl_get_process_apertures(filep, process,
 558                                                 (void __user *)arg);
 559                 break;
 560
 561         case KFD_IOC_UPDATE_QUEUE:
 562                 err = kfd_ioctl_update_queue(filep, process,
 563                                                 (void __user *)arg);
 564                 break;
 565
 566         default:
 567                 dev_err(kfd_device,
 568                         "unknown ioctl cmd 0x%x, arg 0x%lx)\n",
 569                         cmd, arg);
 570                 err = -EINVAL;
 571                 break;
 572         }
 573
 574         if (err < 0)
 575                 dev_err(kfd_device,
 576                         "ioctl error %ld for ioctl cmd 0x%x (#%d)\n",
 577                         err, cmd, _IOC_NR(cmd));
 578
 579         return err;
 580 }
 581
 582 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
 583 {
 584         struct kfd_process *process;
 585
 586         process = kfd_get_process(current);
 587         if (IS_ERR(process))
 588                 return PTR_ERR(process);
 589
 590         return kfd_doorbell_mmap(process, vma);
 591 }