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
 106         if (iminor(inode) != 0)
 107                 return -ENODEV;
 108
 109         process = kfd_create_process(current);
 110         if (IS_ERR(process))
 111                 return PTR_ERR(process);
 112
 113         process->is_32bit_user_mode = is_compat_task();
 114
 115         dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
 116                 process->pasid, process->is_32bit_user_mode);
 117
 118         kfd_init_apertures(process);
 119
 120         return 0;
 121 }
 122
 123 static long kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
 124                                         void __user *arg)
 125 {
 126         struct kfd_ioctl_get_version_args args;
 127         int err = 0;
 128
 129         args.major_version = KFD_IOCTL_MAJOR_VERSION;
 130         args.minor_version = KFD_IOCTL_MINOR_VERSION;
 131
 132         if (copy_to_user(arg, &args, sizeof(args)))
 133                 err = -EFAULT;
 134
 135         return err;
 136 }
 137
 138 static int set_queue_properties_from_user(struct queue_properties *q_properties,
 139                                 struct kfd_ioctl_create_queue_args *args)
 140 {
 141         if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
 142                 pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
 143                 return -EINVAL;
 144         }
 145
 146         if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
 147                 pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
 148                 return -EINVAL;
 149         }
 150
 151         if ((args->ring_base_address) &&
 152                 (!access_ok(VERIFY_WRITE, args->ring_base_address, sizeof(uint64_t)))) {
 153                 pr_err("kfd: can't access ring base address\n");
 154                 return -EFAULT;
 155         }
 156
 157         if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
 158                 pr_err("kfd: ring size must be a power of 2 or 0\n");
 159                 return -EINVAL;
 160         }
 161
 162         if (!access_ok(VERIFY_WRITE, args->read_pointer_address, sizeof(uint32_t))) {
 163                 pr_err("kfd: can't access read pointer\n");
 164                 return -EFAULT;
 165         }
 166
 167         if (!access_ok(VERIFY_WRITE, args->write_pointer_address, sizeof(uint32_t))) {
 168                 pr_err("kfd: can't access write pointer\n");
 169                 return -EFAULT;
 170         }
 171
 172         q_properties->is_interop = false;
 173         q_properties->queue_percent = args->queue_percentage;
 174         q_properties->priority = args->queue_priority;
 175         q_properties->queue_address = args->ring_base_address;
 176         q_properties->queue_size = args->ring_size;
 177         q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
 178         q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
 179         if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
 180                 args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
 181                 q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
 182         else
 183                 return -ENOTSUPP;
 184
 185         if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
 186                 q_properties->format = KFD_QUEUE_FORMAT_AQL;
 187         else
 188                 q_properties->format = KFD_QUEUE_FORMAT_PM4;
 189
 190         pr_debug("Queue Percentage (%d, %d)\n",
 191                         q_properties->queue_percent, args->queue_percentage);
 192
 193         pr_debug("Queue Priority (%d, %d)\n",
 194                         q_properties->priority, args->queue_priority);
 195
 196         pr_debug("Queue Address (0x%llX, 0x%llX)\n",
 197                         q_properties->queue_address, args->ring_base_address);
 198
 199         pr_debug("Queue Size (0x%llX, %u)\n",
 200                         q_properties->queue_size, args->ring_size);
 201
 202         pr_debug("Queue r/w Pointers (0x%llX, 0x%llX)\n",
 203                         (uint64_t) q_properties->read_ptr,
 204                         (uint64_t) q_properties->write_ptr);
 205
 206         pr_debug("Queue Format (%d)\n", q_properties->format);
 207
 208         return 0;
 209 }
 210
 211 static long kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 212                                         void __user *arg)
 213 {
 214         struct kfd_ioctl_create_queue_args args;
 215         struct kfd_dev *dev;
 216         int err = 0;
 217         unsigned int queue_id;
 218         struct kfd_process_device *pdd;
 219         struct queue_properties q_properties;
 220
 221         memset(&q_properties, 0, sizeof(struct queue_properties));
 222
 223         if (copy_from_user(&args, arg, sizeof(args)))
 224                 return -EFAULT;
 225
 226         pr_debug("kfd: creating queue ioctl\n");
 227
 228         err = set_queue_properties_from_user(&q_properties, &args);
 229         if (err)
 230                 return err;
 231
 232         dev = kfd_device_by_id(args.gpu_id);
 233         if (dev == NULL)
 234                 return -EINVAL;
 235
 236         mutex_lock(&p->mutex);
 237
 238         pdd = kfd_bind_process_to_device(dev, p);
 239         if (IS_ERR(pdd) < 0) {
 240                 err = PTR_ERR(pdd);
 241                 goto err_bind_process;
 242         }
 243
 244         pr_debug("kfd: creating queue for PASID %d on GPU 0x%x\n",
 245                         p->pasid,
 246                         dev->id);
 247
 248         err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, 0,
 249                                 KFD_QUEUE_TYPE_COMPUTE, &queue_id);
 250         if (err != 0)
 251                 goto err_create_queue;
 252
 253         args.queue_id = queue_id;
 254
 255         /* Return gpu_id as doorbell offset for mmap usage */
 256         args.doorbell_offset = args.gpu_id << PAGE_SHIFT;
 257
 258         if (copy_to_user(arg, &args, sizeof(args))) {
 259                 err = -EFAULT;
 260                 goto err_copy_args_out;
 261         }
 262
 263         mutex_unlock(&p->mutex);
 264
 265         pr_debug("kfd: queue id %d was created successfully\n", args.queue_id);
 266
 267         pr_debug("ring buffer address == 0x%016llX\n",
 268                         args.ring_base_address);
 269
 270         pr_debug("read ptr address    == 0x%016llX\n",
 271                         args.read_pointer_address);
 272
 273         pr_debug("write ptr address   == 0x%016llX\n",
 274                         args.write_pointer_address);
 275
 276         return 0;
 277
 278 err_copy_args_out:
 279         pqm_destroy_queue(&p->pqm, queue_id);
 280 err_create_queue:
 281 err_bind_process:
 282         mutex_unlock(&p->mutex);
 283         return err;
 284 }
 285
 286 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
 287                                         void __user *arg)
 288 {
 289         int retval;
 290         struct kfd_ioctl_destroy_queue_args args;
 291
 292         if (copy_from_user(&args, arg, sizeof(args)))
 293                 return -EFAULT;
 294
 295         pr_debug("kfd: destroying queue id %d for PASID %d\n",
 296                                 args.queue_id,
 297                                 p->pasid);
 298
 299         mutex_lock(&p->mutex);
 300
 301         retval = pqm_destroy_queue(&p->pqm, args.queue_id);
 302
 303         mutex_unlock(&p->mutex);
 304         return retval;
 305 }
 306
 307 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
 308                                         void __user *arg)
 309 {
 310         int retval;
 311         struct kfd_ioctl_update_queue_args args;
 312         struct queue_properties properties;
 313
 314         if (copy_from_user(&args, arg, sizeof(args)))
 315                 return -EFAULT;
 316
 317         if (args.queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
 318                 pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
 319                 return -EINVAL;
 320         }
 321
 322         if (args.queue_priority > KFD_MAX_QUEUE_PRIORITY) {
 323                 pr_err("kfd: queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
 324                 return -EINVAL;
 325         }
 326
 327         if ((args.ring_base_address) &&
 328                 (!access_ok(VERIFY_WRITE, args.ring_base_address, sizeof(uint64_t)))) {
 329                 pr_err("kfd: can't access ring base address\n");
 330                 return -EFAULT;
 331         }
 332
 333         if (!is_power_of_2(args.ring_size) && (args.ring_size != 0)) {
 334                 pr_err("kfd: ring size must be a power of 2 or 0\n");
 335                 return -EINVAL;
 336         }
 337
 338         properties.queue_address = args.ring_base_address;
 339         properties.queue_size = args.ring_size;
 340         properties.queue_percent = args.queue_percentage;
 341         properties.priority = args.queue_priority;
 342
 343         pr_debug("kfd: updating queue id %d for PASID %d\n",
 344                         args.queue_id, p->pasid);
 345
 346         mutex_lock(&p->mutex);
 347
 348         retval = pqm_update_queue(&p->pqm, args.queue_id, &properties);
 349
 350         mutex_unlock(&p->mutex);
 351
 352         return retval;
 353 }
 354
 355 static long kfd_ioctl_set_memory_policy(struct file *filep,
 356                                 struct kfd_process *p, void __user *arg)
 357 {
 358         struct kfd_ioctl_set_memory_policy_args args;
 359         struct kfd_dev *dev;
 360         int err = 0;
 361         struct kfd_process_device *pdd;
 362         enum cache_policy default_policy, alternate_policy;
 363
 364         if (copy_from_user(&args, arg, sizeof(args)))
 365                 return -EFAULT;
 366
 367         if (args.default_policy != KFD_IOC_CACHE_POLICY_COHERENT
 368             && args.default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 369                 return -EINVAL;
 370         }
 371
 372         if (args.alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
 373             && args.alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 374                 return -EINVAL;
 375         }
 376
 377         dev = kfd_device_by_id(args.gpu_id);
 378         if (dev == NULL)
 379                 return -EINVAL;
 380
 381         mutex_lock(&p->mutex);
 382
 383         pdd = kfd_bind_process_to_device(dev, p);
 384         if (IS_ERR(pdd) < 0) {
 385                 err = PTR_ERR(pdd);
 386                 goto out;
 387         }
 388
 389         default_policy = (args.default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 390                          ? cache_policy_coherent : cache_policy_noncoherent;
 391
 392         alternate_policy =
 393                 (args.alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 394                    ? cache_policy_coherent : cache_policy_noncoherent;
 395
 396         if (!dev->dqm->set_cache_memory_policy(dev->dqm,
 397                                 &pdd->qpd,
 398                                 default_policy,
 399                                 alternate_policy,
 400                                 (void __user *)args.alternate_aperture_base,
 401                                 args.alternate_aperture_size))
 402                 err = -EINVAL;
 403
 404 out:
 405         mutex_unlock(&p->mutex);
 406
 407         return err;
 408 }
 409
 410 static long kfd_ioctl_get_clock_counters(struct file *filep,
 411                                 struct kfd_process *p, void __user *arg)
 412 {
 413         struct kfd_ioctl_get_clock_counters_args args;
 414         struct kfd_dev *dev;
 415         struct timespec time;
 416
 417         if (copy_from_user(&args, arg, sizeof(args)))
 418                 return -EFAULT;
 419
 420         dev = kfd_device_by_id(args.gpu_id);
 421         if (dev == NULL)
 422                 return -EINVAL;
 423
 424         /* Reading GPU clock counter from KGD */
 425         args.gpu_clock_counter = kfd2kgd->get_gpu_clock_counter(dev->kgd);
 426
 427         /* No access to rdtsc. Using raw monotonic time */
 428         getrawmonotonic(&time);
 429         args.cpu_clock_counter = (uint64_t)timespec_to_ns(&time);
 430
 431         get_monotonic_boottime(&time);
 432         args.system_clock_counter = (uint64_t)timespec_to_ns(&time);
 433
 434         /* Since the counter is in nano-seconds we use 1GHz frequency */
 435         args.system_clock_freq = 1000000000;
 436
 437         if (copy_to_user(arg, &args, sizeof(args)))
 438                 return -EFAULT;
 439
 440         return 0;
 441 }
 442
 443
 444 static int kfd_ioctl_get_process_apertures(struct file *filp,
 445                                 struct kfd_process *p, void __user *arg)
 446 {
 447         struct kfd_ioctl_get_process_apertures_args args;
 448         struct kfd_process_device_apertures *pAperture;
 449         struct kfd_process_device *pdd;
 450
 451         dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
 452
 453         if (copy_from_user(&args, arg, sizeof(args)))
 454                 return -EFAULT;
 455
 456         args.num_of_nodes = 0;
 457
 458         mutex_lock(&p->mutex);
 459
 460         /*if the process-device list isn't empty*/
 461         if (kfd_has_process_device_data(p)) {
 462                 /* Run over all pdd of the process */
 463                 pdd = kfd_get_first_process_device_data(p);
 464                 do {
 465                         pAperture = &args.process_apertures[args.num_of_nodes];
 466                         pAperture->gpu_id = pdd->dev->id;
 467                         pAperture->lds_base = pdd->lds_base;
 468                         pAperture->lds_limit = pdd->lds_limit;
 469                         pAperture->gpuvm_base = pdd->gpuvm_base;
 470                         pAperture->gpuvm_limit = pdd->gpuvm_limit;
 471                         pAperture->scratch_base = pdd->scratch_base;
 472                         pAperture->scratch_limit = pdd->scratch_limit;
 473
 474                         dev_dbg(kfd_device,
 475                                 "node id %u\n", args.num_of_nodes);
 476                         dev_dbg(kfd_device,
 477                                 "gpu id %u\n", pdd->dev->id);
 478                         dev_dbg(kfd_device,
 479                                 "lds_base %llX\n", pdd->lds_base);
 480                         dev_dbg(kfd_device,
 481                                 "lds_limit %llX\n", pdd->lds_limit);
 482                         dev_dbg(kfd_device,
 483                                 "gpuvm_base %llX\n", pdd->gpuvm_base);
 484                         dev_dbg(kfd_device,
 485                                 "gpuvm_limit %llX\n", pdd->gpuvm_limit);
 486                         dev_dbg(kfd_device,
 487                                 "scratch_base %llX\n", pdd->scratch_base);
 488                         dev_dbg(kfd_device,
 489                                 "scratch_limit %llX\n", pdd->scratch_limit);
 490
 491                         args.num_of_nodes++;
 492                 } while ((pdd = kfd_get_next_process_device_data(p, pdd)) != NULL &&
 493                                 (args.num_of_nodes < NUM_OF_SUPPORTED_GPUS));
 494         }
 495
 496         mutex_unlock(&p->mutex);
 497
 498         if (copy_to_user(arg, &args, sizeof(args)))
 499                 return -EFAULT;
 500
 501         return 0;
 502 }
 503
 504 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
 505 {
 506         struct kfd_process *process;
 507         long err = -EINVAL;
 508
 509         dev_dbg(kfd_device,
 510                 "ioctl cmd 0x%x (#%d), arg 0x%lx\n",
 511                 cmd, _IOC_NR(cmd), arg);
 512
 513         process = kfd_get_process(current);
 514         if (IS_ERR(process))
 515                 return PTR_ERR(process);
 516
 517         switch (cmd) {
 518         case KFD_IOC_GET_VERSION:
 519                 err = kfd_ioctl_get_version(filep, process, (void __user *)arg);
 520                 break;
 521         case KFD_IOC_CREATE_QUEUE:
 522                 err = kfd_ioctl_create_queue(filep, process,
 523                                                 (void __user *)arg);
 524                 break;
 525
 526         case KFD_IOC_DESTROY_QUEUE:
 527                 err = kfd_ioctl_destroy_queue(filep, process,
 528                                                 (void __user *)arg);
 529                 break;
 530
 531         case KFD_IOC_SET_MEMORY_POLICY:
 532                 err = kfd_ioctl_set_memory_policy(filep, process,
 533                                                 (void __user *)arg);
 534                 break;
 535
 536         case KFD_IOC_GET_CLOCK_COUNTERS:
 537                 err = kfd_ioctl_get_clock_counters(filep, process,
 538                                                 (void __user *)arg);
 539                 break;
 540
 541         case KFD_IOC_GET_PROCESS_APERTURES:
 542                 err = kfd_ioctl_get_process_apertures(filep, process,
 543                                                 (void __user *)arg);
 544                 break;
 545
 546         case KFD_IOC_UPDATE_QUEUE:
 547                 err = kfd_ioctl_update_queue(filep, process,
 548                                                 (void __user *)arg);
 549                 break;
 550
 551         default:
 552                 dev_err(kfd_device,
 553                         "unknown ioctl cmd 0x%x, arg 0x%lx)\n",
 554                         cmd, arg);
 555                 err = -EINVAL;
 556                 break;
 557         }
 558
 559         if (err < 0)
 560                 dev_err(kfd_device,
 561                         "ioctl error %ld for ioctl cmd 0x%x (#%d)\n",
 562                         err, cmd, _IOC_NR(cmd));
 563
 564         return err;
 565 }
 566
 567 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
 568 {
 569         struct kfd_process *process;
 570
 571         process = kfd_get_process(current);
 572         if (IS_ERR(process))
 573                 return PTR_ERR(process);
 574
 575         return kfd_doorbell_mmap(process, vma);
 576 }