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/file.h>
  28 #include <linux/sched.h>
  29 #include <linux/slab.h>
  30 #include <linux/uaccess.h>
  31 #include <linux/compat.h>
  32 #include <uapi/linux/kfd_ioctl.h>
  33 #include <linux/time.h>
  34 #include <linux/mm.h>
  35 #include <linux/mman.h>
  36 #include <asm/processor.h>
  37 #include "kfd_priv.h"
  38 #include "kfd_device_queue_manager.h"
  39 #include "kfd_dbgmgr.h"
  40
  41 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
  42 static int kfd_open(struct inode *, struct file *);
  43 static int kfd_mmap(struct file *, struct vm_area_struct *);
  44
  45 static const char kfd_dev_name[] = "kfd";
  46
  47 static const struct file_operations kfd_fops = {
  48         .owner = THIS_MODULE,
  49         .unlocked_ioctl = kfd_ioctl,
  50         .compat_ioctl = kfd_ioctl,
  51         .open = kfd_open,
  52         .mmap = kfd_mmap,
  53 };
  54
  55 static int kfd_char_dev_major = -1;
  56 static struct class *kfd_class;
  57 struct device *kfd_device;
  58
  59 int kfd_chardev_init(void)
  60 {
  61         int err = 0;
  62
  63         kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
  64         err = kfd_char_dev_major;
  65         if (err < 0)
  66                 goto err_register_chrdev;
  67
  68         kfd_class = class_create(THIS_MODULE, kfd_dev_name);
  69         err = PTR_ERR(kfd_class);
  70         if (IS_ERR(kfd_class))
  71                 goto err_class_create;
  72
  73         kfd_device = device_create(kfd_class, NULL,
  74                                         MKDEV(kfd_char_dev_major, 0),
  75                                         NULL, kfd_dev_name);
  76         err = PTR_ERR(kfd_device);
  77         if (IS_ERR(kfd_device))
  78                 goto err_device_create;
  79
  80         return 0;
  81
  82 err_device_create:
  83         class_destroy(kfd_class);
  84 err_class_create:
  85         unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
  86 err_register_chrdev:
  87         return err;
  88 }
  89
  90 void kfd_chardev_exit(void)
  91 {
  92         device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
  93         class_destroy(kfd_class);
  94         unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
  95 }
  96
  97 struct device *kfd_chardev(void)
  98 {
  99         return kfd_device;
 100 }
 101
 102
 103 static int kfd_open(struct inode *inode, struct file *filep)
 104 {
 105         struct kfd_process *process;
 106         bool is_32bit_user_mode;
 107
 108         if (iminor(inode) != 0)
 109                 return -ENODEV;
 110
 111         is_32bit_user_mode = in_compat_syscall();
 112
 113         if (is_32bit_user_mode) {
 114                 dev_warn(kfd_device,
 115                         "Process %d (32-bit) failed to open /dev/kfd\n"
 116                         "32-bit processes are not supported by amdkfd\n",
 117                         current->pid);
 118                 return -EPERM;
 119         }
 120
 121         process = kfd_create_process(filep);
 122         if (IS_ERR(process))
 123                 return PTR_ERR(process);
 124
 125         dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
 126                 process->pasid, process->is_32bit_user_mode);
 127
 128         return 0;
 129 }
 130
 131 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
 132                                         void *data)
 133 {
 134         struct kfd_ioctl_get_version_args *args = data;
 135
 136         args->major_version = KFD_IOCTL_MAJOR_VERSION;
 137         args->minor_version = KFD_IOCTL_MINOR_VERSION;
 138
 139         return 0;
 140 }
 141
 142 static int set_queue_properties_from_user(struct queue_properties *q_properties,
 143                                 struct kfd_ioctl_create_queue_args *args)
 144 {
 145         if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
 146                 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
 147                 return -EINVAL;
 148         }
 149
 150         if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
 151                 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
 152                 return -EINVAL;
 153         }
 154
 155         if ((args->ring_base_address) &&
 156                 (!access_ok(VERIFY_WRITE,
 157                         (const void __user *) args->ring_base_address,
 158                         sizeof(uint64_t)))) {
 159                 pr_err("Can't access ring base address\n");
 160                 return -EFAULT;
 161         }
 162
 163         if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
 164                 pr_err("Ring size must be a power of 2 or 0\n");
 165                 return -EINVAL;
 166         }
 167
 168         if (!access_ok(VERIFY_WRITE,
 169                         (const void __user *) args->read_pointer_address,
 170                         sizeof(uint32_t))) {
 171                 pr_err("Can't access read pointer\n");
 172                 return -EFAULT;
 173         }
 174
 175         if (!access_ok(VERIFY_WRITE,
 176                         (const void __user *) args->write_pointer_address,
 177                         sizeof(uint32_t))) {
 178                 pr_err("Can't access write pointer\n");
 179                 return -EFAULT;
 180         }
 181
 182         if (args->eop_buffer_address &&
 183                 !access_ok(VERIFY_WRITE,
 184                         (const void __user *) args->eop_buffer_address,
 185                         sizeof(uint32_t))) {
 186                 pr_debug("Can't access eop buffer");
 187                 return -EFAULT;
 188         }
 189
 190         if (args->ctx_save_restore_address &&
 191                 !access_ok(VERIFY_WRITE,
 192                         (const void __user *) args->ctx_save_restore_address,
 193                         sizeof(uint32_t))) {
 194                 pr_debug("Can't access ctx save restore buffer");
 195                 return -EFAULT;
 196         }
 197
 198         q_properties->is_interop = false;
 199         q_properties->queue_percent = args->queue_percentage;
 200         q_properties->priority = args->queue_priority;
 201         q_properties->queue_address = args->ring_base_address;
 202         q_properties->queue_size = args->ring_size;
 203         q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
 204         q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
 205         q_properties->eop_ring_buffer_address = args->eop_buffer_address;
 206         q_properties->eop_ring_buffer_size = args->eop_buffer_size;
 207         q_properties->ctx_save_restore_area_address =
 208                         args->ctx_save_restore_address;
 209         q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
 210         q_properties->ctl_stack_size = args->ctl_stack_size;
 211         if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
 212                 args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
 213                 q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
 214         else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
 215                 q_properties->type = KFD_QUEUE_TYPE_SDMA;
 216         else
 217                 return -ENOTSUPP;
 218
 219         if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
 220                 q_properties->format = KFD_QUEUE_FORMAT_AQL;
 221         else
 222                 q_properties->format = KFD_QUEUE_FORMAT_PM4;
 223
 224         pr_debug("Queue Percentage: %d, %d\n",
 225                         q_properties->queue_percent, args->queue_percentage);
 226
 227         pr_debug("Queue Priority: %d, %d\n",
 228                         q_properties->priority, args->queue_priority);
 229
 230         pr_debug("Queue Address: 0x%llX, 0x%llX\n",
 231                         q_properties->queue_address, args->ring_base_address);
 232
 233         pr_debug("Queue Size: 0x%llX, %u\n",
 234                         q_properties->queue_size, args->ring_size);
 235
 236         pr_debug("Queue r/w Pointers: %p, %p\n",
 237                         q_properties->read_ptr,
 238                         q_properties->write_ptr);
 239
 240         pr_debug("Queue Format: %d\n", q_properties->format);
 241
 242         pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
 243
 244         pr_debug("Queue CTX save area: 0x%llX\n",
 245                         q_properties->ctx_save_restore_area_address);
 246
 247         return 0;
 248 }
 249
 250 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 251                                         void *data)
 252 {
 253         struct kfd_ioctl_create_queue_args *args = data;
 254         struct kfd_dev *dev;
 255         int err = 0;
 256         unsigned int queue_id;
 257         struct kfd_process_device *pdd;
 258         struct queue_properties q_properties;
 259
 260         memset(&q_properties, 0, sizeof(struct queue_properties));
 261
 262         pr_debug("Creating queue ioctl\n");
 263
 264         err = set_queue_properties_from_user(&q_properties, args);
 265         if (err)
 266                 return err;
 267
 268         pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
 269         dev = kfd_device_by_id(args->gpu_id);
 270         if (!dev) {
 271                 pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
 272                 return -EINVAL;
 273         }
 274
 275         mutex_lock(&p->mutex);
 276
 277         pdd = kfd_bind_process_to_device(dev, p);
 278         if (IS_ERR(pdd)) {
 279                 err = -ESRCH;
 280                 goto err_bind_process;
 281         }
 282
 283         pr_debug("Creating queue for PASID %d on gpu 0x%x\n",
 284                         p->pasid,
 285                         dev->id);
 286
 287         err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id);
 288         if (err != 0)
 289                 goto err_create_queue;
 290
 291         args->queue_id = queue_id;
 292
 293
 294         /* Return gpu_id as doorbell offset for mmap usage */
 295         args->doorbell_offset = (KFD_MMAP_DOORBELL_MASK | args->gpu_id);
 296         args->doorbell_offset <<= PAGE_SHIFT;
 297
 298         mutex_unlock(&p->mutex);
 299
 300         pr_debug("Queue id %d was created successfully\n", args->queue_id);
 301
 302         pr_debug("Ring buffer address == 0x%016llX\n",
 303                         args->ring_base_address);
 304
 305         pr_debug("Read ptr address    == 0x%016llX\n",
 306                         args->read_pointer_address);
 307
 308         pr_debug("Write ptr address   == 0x%016llX\n",
 309                         args->write_pointer_address);
 310
 311         return 0;
 312
 313 err_create_queue:
 314 err_bind_process:
 315         mutex_unlock(&p->mutex);
 316         return err;
 317 }
 318
 319 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
 320                                         void *data)
 321 {
 322         int retval;
 323         struct kfd_ioctl_destroy_queue_args *args = data;
 324
 325         pr_debug("Destroying queue id %d for pasid %d\n",
 326                                 args->queue_id,
 327                                 p->pasid);
 328
 329         mutex_lock(&p->mutex);
 330
 331         retval = pqm_destroy_queue(&p->pqm, args->queue_id);
 332
 333         mutex_unlock(&p->mutex);
 334         return retval;
 335 }
 336
 337 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
 338                                         void *data)
 339 {
 340         int retval;
 341         struct kfd_ioctl_update_queue_args *args = data;
 342         struct queue_properties properties;
 343
 344         if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {
 345                 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
 346                 return -EINVAL;
 347         }
 348
 349         if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
 350                 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
 351                 return -EINVAL;
 352         }
 353
 354         if ((args->ring_base_address) &&
 355                 (!access_ok(VERIFY_WRITE,
 356                         (const void __user *) args->ring_base_address,
 357                         sizeof(uint64_t)))) {
 358                 pr_err("Can't access ring base address\n");
 359                 return -EFAULT;
 360         }
 361
 362         if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
 363                 pr_err("Ring size must be a power of 2 or 0\n");
 364                 return -EINVAL;
 365         }
 366
 367         properties.queue_address = args->ring_base_address;
 368         properties.queue_size = args->ring_size;
 369         properties.queue_percent = args->queue_percentage;
 370         properties.priority = args->queue_priority;
 371
 372         pr_debug("Updating queue id %d for pasid %d\n",
 373                         args->queue_id, p->pasid);
 374
 375         mutex_lock(&p->mutex);
 376
 377         retval = pqm_update_queue(&p->pqm, args->queue_id, &properties);
 378
 379         mutex_unlock(&p->mutex);
 380
 381         return retval;
 382 }
 383
 384 static int kfd_ioctl_set_memory_policy(struct file *filep,
 385                                         struct kfd_process *p, void *data)
 386 {
 387         struct kfd_ioctl_set_memory_policy_args *args = data;
 388         struct kfd_dev *dev;
 389         int err = 0;
 390         struct kfd_process_device *pdd;
 391         enum cache_policy default_policy, alternate_policy;
 392
 393         if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
 394             && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 395                 return -EINVAL;
 396         }
 397
 398         if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
 399             && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 400                 return -EINVAL;
 401         }
 402
 403         dev = kfd_device_by_id(args->gpu_id);
 404         if (!dev)
 405                 return -EINVAL;
 406
 407         mutex_lock(&p->mutex);
 408
 409         pdd = kfd_bind_process_to_device(dev, p);
 410         if (IS_ERR(pdd)) {
 411                 err = -ESRCH;
 412                 goto out;
 413         }
 414
 415         default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 416                          ? cache_policy_coherent : cache_policy_noncoherent;
 417
 418         alternate_policy =
 419                 (args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 420                    ? cache_policy_coherent : cache_policy_noncoherent;
 421
 422         if (!dev->dqm->ops.set_cache_memory_policy(dev->dqm,
 423                                 &pdd->qpd,
 424                                 default_policy,
 425                                 alternate_policy,
 426                                 (void __user *)args->alternate_aperture_base,
 427                                 args->alternate_aperture_size))
 428                 err = -EINVAL;
 429
 430 out:
 431         mutex_unlock(&p->mutex);
 432
 433         return err;
 434 }
 435
 436 static int kfd_ioctl_set_trap_handler(struct file *filep,
 437                                         struct kfd_process *p, void *data)
 438 {
 439         struct kfd_ioctl_set_trap_handler_args *args = data;
 440         struct kfd_dev *dev;
 441         int err = 0;
 442         struct kfd_process_device *pdd;
 443
 444         dev = kfd_device_by_id(args->gpu_id);
 445         if (dev == NULL)
 446                 return -EINVAL;
 447
 448         mutex_lock(&p->mutex);
 449
 450         pdd = kfd_bind_process_to_device(dev, p);
 451         if (IS_ERR(pdd)) {
 452                 err = -ESRCH;
 453                 goto out;
 454         }
 455
 456         if (dev->dqm->ops.set_trap_handler(dev->dqm,
 457                                         &pdd->qpd,
 458                                         args->tba_addr,
 459                                         args->tma_addr))
 460                 err = -EINVAL;
 461
 462 out:
 463         mutex_unlock(&p->mutex);
 464
 465         return err;
 466 }
 467
 468 static int kfd_ioctl_dbg_register(struct file *filep,
 469                                 struct kfd_process *p, void *data)
 470 {
 471         struct kfd_ioctl_dbg_register_args *args = data;
 472         struct kfd_dev *dev;
 473         struct kfd_dbgmgr *dbgmgr_ptr;
 474         struct kfd_process_device *pdd;
 475         bool create_ok;
 476         long status = 0;
 477
 478         dev = kfd_device_by_id(args->gpu_id);
 479         if (!dev)
 480                 return -EINVAL;
 481
 482         if (dev->device_info->asic_family == CHIP_CARRIZO) {
 483                 pr_debug("kfd_ioctl_dbg_register not supported on CZ\n");
 484                 return -EINVAL;
 485         }
 486
 487         mutex_lock(&p->mutex);
 488         mutex_lock(kfd_get_dbgmgr_mutex());
 489
 490         /*
 491          * make sure that we have pdd, if this the first queue created for
 492          * this process
 493          */
 494         pdd = kfd_bind_process_to_device(dev, p);
 495         if (IS_ERR(pdd)) {
 496                 status = PTR_ERR(pdd);
 497                 goto out;
 498         }
 499
 500         if (!dev->dbgmgr) {
 501                 /* In case of a legal call, we have no dbgmgr yet */
 502                 create_ok = kfd_dbgmgr_create(&dbgmgr_ptr, dev);
 503                 if (create_ok) {
 504                         status = kfd_dbgmgr_register(dbgmgr_ptr, p);
 505                         if (status != 0)
 506                                 kfd_dbgmgr_destroy(dbgmgr_ptr);
 507                         else
 508                                 dev->dbgmgr = dbgmgr_ptr;
 509                 }
 510         } else {
 511                 pr_debug("debugger already registered\n");
 512                 status = -EINVAL;
 513         }
 514
 515 out:
 516         mutex_unlock(kfd_get_dbgmgr_mutex());
 517         mutex_unlock(&p->mutex);
 518
 519         return status;
 520 }
 521
 522 static int kfd_ioctl_dbg_unregister(struct file *filep,
 523                                 struct kfd_process *p, void *data)
 524 {
 525         struct kfd_ioctl_dbg_unregister_args *args = data;
 526         struct kfd_dev *dev;
 527         long status;
 528
 529         dev = kfd_device_by_id(args->gpu_id);
 530         if (!dev || !dev->dbgmgr)
 531                 return -EINVAL;
 532
 533         if (dev->device_info->asic_family == CHIP_CARRIZO) {
 534                 pr_debug("kfd_ioctl_dbg_unregister not supported on CZ\n");
 535                 return -EINVAL;
 536         }
 537
 538         mutex_lock(kfd_get_dbgmgr_mutex());
 539
 540         status = kfd_dbgmgr_unregister(dev->dbgmgr, p);
 541         if (!status) {
 542                 kfd_dbgmgr_destroy(dev->dbgmgr);
 543                 dev->dbgmgr = NULL;
 544         }
 545
 546         mutex_unlock(kfd_get_dbgmgr_mutex());
 547
 548         return status;
 549 }
 550
 551 /*
 552  * Parse and generate variable size data structure for address watch.
 553  * Total size of the buffer and # watch points is limited in order
 554  * to prevent kernel abuse. (no bearing to the much smaller HW limitation
 555  * which is enforced by dbgdev module)
 556  * please also note that the watch address itself are not "copied from user",
 557  * since it be set into the HW in user mode values.
 558  *
 559  */
 560 static int kfd_ioctl_dbg_address_watch(struct file *filep,
 561                                         struct kfd_process *p, void *data)
 562 {
 563         struct kfd_ioctl_dbg_address_watch_args *args = data;
 564         struct kfd_dev *dev;
 565         struct dbg_address_watch_info aw_info;
 566         unsigned char *args_buff;
 567         long status;
 568         void __user *cmd_from_user;
 569         uint64_t watch_mask_value = 0;
 570         unsigned int args_idx = 0;
 571
 572         memset((void *) &aw_info, 0, sizeof(struct dbg_address_watch_info));
 573
 574         dev = kfd_device_by_id(args->gpu_id);
 575         if (!dev)
 576                 return -EINVAL;
 577
 578         if (dev->device_info->asic_family == CHIP_CARRIZO) {
 579                 pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
 580                 return -EINVAL;
 581         }
 582
 583         cmd_from_user = (void __user *) args->content_ptr;
 584
 585         /* Validate arguments */
 586
 587         if ((args->buf_size_in_bytes > MAX_ALLOWED_AW_BUFF_SIZE) ||
 588                 (args->buf_size_in_bytes <= sizeof(*args) + sizeof(int) * 2) ||
 589                 (cmd_from_user == NULL))
 590                 return -EINVAL;
 591
 592         /* this is the actual buffer to work with */
 593         args_buff = memdup_user(cmd_from_user,
 594                                 args->buf_size_in_bytes - sizeof(*args));
 595         if (IS_ERR(args_buff))
 596                 return PTR_ERR(args_buff);
 597
 598         aw_info.process = p;
 599
 600         aw_info.num_watch_points = *((uint32_t *)(&args_buff[args_idx]));
 601         args_idx += sizeof(aw_info.num_watch_points);
 602
 603         aw_info.watch_mode = (enum HSA_DBG_WATCH_MODE *) &args_buff[args_idx];
 604         args_idx += sizeof(enum HSA_DBG_WATCH_MODE) * aw_info.num_watch_points;
 605
 606         /*
 607          * set watch address base pointer to point on the array base
 608          * within args_buff
 609          */
 610         aw_info.watch_address = (uint64_t *) &args_buff[args_idx];
 611
 612         /* skip over the addresses buffer */
 613         args_idx += sizeof(aw_info.watch_address) * aw_info.num_watch_points;
 614
 615         if (args_idx >= args->buf_size_in_bytes - sizeof(*args)) {
 616                 status = -EINVAL;
 617                 goto out;
 618         }
 619
 620         watch_mask_value = (uint64_t) args_buff[args_idx];
 621
 622         if (watch_mask_value > 0) {
 623                 /*
 624                  * There is an array of masks.
 625                  * set watch mask base pointer to point on the array base
 626                  * within args_buff
 627                  */
 628                 aw_info.watch_mask = (uint64_t *) &args_buff[args_idx];
 629
 630                 /* skip over the masks buffer */
 631                 args_idx += sizeof(aw_info.watch_mask) *
 632                                 aw_info.num_watch_points;
 633         } else {
 634                 /* just the NULL mask, set to NULL and skip over it */
 635                 aw_info.watch_mask = NULL;
 636                 args_idx += sizeof(aw_info.watch_mask);
 637         }
 638
 639         if (args_idx >= args->buf_size_in_bytes - sizeof(args)) {
 640                 status = -EINVAL;
 641                 goto out;
 642         }
 643
 644         /* Currently HSA Event is not supported for DBG */
 645         aw_info.watch_event = NULL;
 646
 647         mutex_lock(kfd_get_dbgmgr_mutex());
 648
 649         status = kfd_dbgmgr_address_watch(dev->dbgmgr, &aw_info);
 650
 651         mutex_unlock(kfd_get_dbgmgr_mutex());
 652
 653 out:
 654         kfree(args_buff);
 655
 656         return status;
 657 }
 658
 659 /* Parse and generate fixed size data structure for wave control */
 660 static int kfd_ioctl_dbg_wave_control(struct file *filep,
 661                                         struct kfd_process *p, void *data)
 662 {
 663         struct kfd_ioctl_dbg_wave_control_args *args = data;
 664         struct kfd_dev *dev;
 665         struct dbg_wave_control_info wac_info;
 666         unsigned char *args_buff;
 667         uint32_t computed_buff_size;
 668         long status;
 669         void __user *cmd_from_user;
 670         unsigned int args_idx = 0;
 671
 672         memset((void *) &wac_info, 0, sizeof(struct dbg_wave_control_info));
 673
 674         /* we use compact form, independent of the packing attribute value */
 675         computed_buff_size = sizeof(*args) +
 676                                 sizeof(wac_info.mode) +
 677                                 sizeof(wac_info.operand) +
 678                                 sizeof(wac_info.dbgWave_msg.DbgWaveMsg) +
 679                                 sizeof(wac_info.dbgWave_msg.MemoryVA) +
 680                                 sizeof(wac_info.trapId);
 681
 682         dev = kfd_device_by_id(args->gpu_id);
 683         if (!dev)
 684                 return -EINVAL;
 685
 686         if (dev->device_info->asic_family == CHIP_CARRIZO) {
 687                 pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n");
 688                 return -EINVAL;
 689         }
 690
 691         /* input size must match the computed "compact" size */
 692         if (args->buf_size_in_bytes != computed_buff_size) {
 693                 pr_debug("size mismatch, computed : actual %u : %u\n",
 694                                 args->buf_size_in_bytes, computed_buff_size);
 695                 return -EINVAL;
 696         }
 697
 698         cmd_from_user = (void __user *) args->content_ptr;
 699
 700         if (cmd_from_user == NULL)
 701                 return -EINVAL;
 702
 703         /* copy the entire buffer from user */
 704
 705         args_buff = memdup_user(cmd_from_user,
 706                                 args->buf_size_in_bytes - sizeof(*args));
 707         if (IS_ERR(args_buff))
 708                 return PTR_ERR(args_buff);
 709
 710         /* move ptr to the start of the "pay-load" area */
 711         wac_info.process = p;
 712
 713         wac_info.operand = *((enum HSA_DBG_WAVEOP *)(&args_buff[args_idx]));
 714         args_idx += sizeof(wac_info.operand);
 715
 716         wac_info.mode = *((enum HSA_DBG_WAVEMODE *)(&args_buff[args_idx]));
 717         args_idx += sizeof(wac_info.mode);
 718
 719         wac_info.trapId = *((uint32_t *)(&args_buff[args_idx]));
 720         args_idx += sizeof(wac_info.trapId);
 721
 722         wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value =
 723                                         *((uint32_t *)(&args_buff[args_idx]));
 724         wac_info.dbgWave_msg.MemoryVA = NULL;
 725
 726         mutex_lock(kfd_get_dbgmgr_mutex());
 727
 728         pr_debug("Calling dbg manager process %p, operand %u, mode %u, trapId %u, message %u\n",
 729                         wac_info.process, wac_info.operand,
 730                         wac_info.mode, wac_info.trapId,
 731                         wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value);
 732
 733         status = kfd_dbgmgr_wave_control(dev->dbgmgr, &wac_info);
 734
 735         pr_debug("Returned status of dbg manager is %ld\n", status);
 736
 737         mutex_unlock(kfd_get_dbgmgr_mutex());
 738
 739         kfree(args_buff);
 740
 741         return status;
 742 }
 743
 744 static int kfd_ioctl_get_clock_counters(struct file *filep,
 745                                 struct kfd_process *p, void *data)
 746 {
 747         struct kfd_ioctl_get_clock_counters_args *args = data;
 748         struct kfd_dev *dev;
 749         struct timespec64 time;
 750
 751         dev = kfd_device_by_id(args->gpu_id);
 752         if (dev == NULL)
 753                 return -EINVAL;
 754
 755         /* Reading GPU clock counter from KGD */
 756         args->gpu_clock_counter =
 757                 dev->kfd2kgd->get_gpu_clock_counter(dev->kgd);
 758
 759         /* No access to rdtsc. Using raw monotonic time */
 760         getrawmonotonic64(&time);
 761         args->cpu_clock_counter = (uint64_t)timespec64_to_ns(&time);
 762
 763         get_monotonic_boottime64(&time);
 764         args->system_clock_counter = (uint64_t)timespec64_to_ns(&time);
 765
 766         /* Since the counter is in nano-seconds we use 1GHz frequency */
 767         args->system_clock_freq = 1000000000;
 768
 769         return 0;
 770 }
 771
 772
 773 static int kfd_ioctl_get_process_apertures(struct file *filp,
 774                                 struct kfd_process *p, void *data)
 775 {
 776         struct kfd_ioctl_get_process_apertures_args *args = data;
 777         struct kfd_process_device_apertures *pAperture;
 778         struct kfd_process_device *pdd;
 779
 780         dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
 781
 782         args->num_of_nodes = 0;
 783
 784         mutex_lock(&p->mutex);
 785
 786         /*if the process-device list isn't empty*/
 787         if (kfd_has_process_device_data(p)) {
 788                 /* Run over all pdd of the process */
 789                 pdd = kfd_get_first_process_device_data(p);
 790                 do {
 791                         pAperture =
 792                                 &args->process_apertures[args->num_of_nodes];
 793                         pAperture->gpu_id = pdd->dev->id;
 794                         pAperture->lds_base = pdd->lds_base;
 795                         pAperture->lds_limit = pdd->lds_limit;
 796                         pAperture->gpuvm_base = pdd->gpuvm_base;
 797                         pAperture->gpuvm_limit = pdd->gpuvm_limit;
 798                         pAperture->scratch_base = pdd->scratch_base;
 799                         pAperture->scratch_limit = pdd->scratch_limit;
 800
 801                         dev_dbg(kfd_device,
 802                                 "node id %u\n", args->num_of_nodes);
 803                         dev_dbg(kfd_device,
 804                                 "gpu id %u\n", pdd->dev->id);
 805                         dev_dbg(kfd_device,
 806                                 "lds_base %llX\n", pdd->lds_base);
 807                         dev_dbg(kfd_device,
 808                                 "lds_limit %llX\n", pdd->lds_limit);
 809                         dev_dbg(kfd_device,
 810                                 "gpuvm_base %llX\n", pdd->gpuvm_base);
 811                         dev_dbg(kfd_device,
 812                                 "gpuvm_limit %llX\n", pdd->gpuvm_limit);
 813                         dev_dbg(kfd_device,
 814                                 "scratch_base %llX\n", pdd->scratch_base);
 815                         dev_dbg(kfd_device,
 816                                 "scratch_limit %llX\n", pdd->scratch_limit);
 817
 818                         args->num_of_nodes++;
 819
 820                         pdd = kfd_get_next_process_device_data(p, pdd);
 821                 } while (pdd && (args->num_of_nodes < NUM_OF_SUPPORTED_GPUS));
 822         }
 823
 824         mutex_unlock(&p->mutex);
 825
 826         return 0;
 827 }
 828
 829 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
 830                                 struct kfd_process *p, void *data)
 831 {
 832         struct kfd_ioctl_get_process_apertures_new_args *args = data;
 833         struct kfd_process_device_apertures *pa;
 834         struct kfd_process_device *pdd;
 835         uint32_t nodes = 0;
 836         int ret;
 837
 838         dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid);
 839
 840         if (args->num_of_nodes == 0) {
 841                 /* Return number of nodes, so that user space can alloacate
 842                  * sufficient memory
 843                  */
 844                 mutex_lock(&p->mutex);
 845
 846                 if (!kfd_has_process_device_data(p))
 847                         goto out_unlock;
 848
 849                 /* Run over all pdd of the process */
 850                 pdd = kfd_get_first_process_device_data(p);
 851                 do {
 852                         args->num_of_nodes++;
 853                         pdd = kfd_get_next_process_device_data(p, pdd);
 854                 } while (pdd);
 855
 856                 goto out_unlock;
 857         }
 858
 859         /* Fill in process-aperture information for all available
 860          * nodes, but not more than args->num_of_nodes as that is
 861          * the amount of memory allocated by user
 862          */
 863         pa = kzalloc((sizeof(struct kfd_process_device_apertures) *
 864                                 args->num_of_nodes), GFP_KERNEL);
 865         if (!pa)
 866                 return -ENOMEM;
 867
 868         mutex_lock(&p->mutex);
 869
 870         if (!kfd_has_process_device_data(p)) {
 871                 args->num_of_nodes = 0;
 872                 kfree(pa);
 873                 goto out_unlock;
 874         }
 875
 876         /* Run over all pdd of the process */
 877         pdd = kfd_get_first_process_device_data(p);
 878         do {
 879                 pa[nodes].gpu_id = pdd->dev->id;
 880                 pa[nodes].lds_base = pdd->lds_base;
 881                 pa[nodes].lds_limit = pdd->lds_limit;
 882                 pa[nodes].gpuvm_base = pdd->gpuvm_base;
 883                 pa[nodes].gpuvm_limit = pdd->gpuvm_limit;
 884                 pa[nodes].scratch_base = pdd->scratch_base;
 885                 pa[nodes].scratch_limit = pdd->scratch_limit;
 886
 887                 dev_dbg(kfd_device,
 888                         "gpu id %u\n", pdd->dev->id);
 889                 dev_dbg(kfd_device,
 890                         "lds_base %llX\n", pdd->lds_base);
 891                 dev_dbg(kfd_device,
 892                         "lds_limit %llX\n", pdd->lds_limit);
 893                 dev_dbg(kfd_device,
 894                         "gpuvm_base %llX\n", pdd->gpuvm_base);
 895                 dev_dbg(kfd_device,
 896                         "gpuvm_limit %llX\n", pdd->gpuvm_limit);
 897                 dev_dbg(kfd_device,
 898                         "scratch_base %llX\n", pdd->scratch_base);
 899                 dev_dbg(kfd_device,
 900                         "scratch_limit %llX\n", pdd->scratch_limit);
 901                 nodes++;
 902
 903                 pdd = kfd_get_next_process_device_data(p, pdd);
 904         } while (pdd && (nodes < args->num_of_nodes));
 905         mutex_unlock(&p->mutex);
 906
 907         args->num_of_nodes = nodes;
 908         ret = copy_to_user(
 909                         (void __user *)args->kfd_process_device_apertures_ptr,
 910                         pa,
 911                         (nodes * sizeof(struct kfd_process_device_apertures)));
 912         kfree(pa);
 913         return ret ? -EFAULT : 0;
 914
 915 out_unlock:
 916         mutex_unlock(&p->mutex);
 917         return 0;
 918 }
 919
 920 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
 921                                         void *data)
 922 {
 923         struct kfd_ioctl_create_event_args *args = data;
 924         int err;
 925
 926         /* For dGPUs the event page is allocated in user mode. The
 927          * handle is passed to KFD with the first call to this IOCTL
 928          * through the event_page_offset field.
 929          */
 930         if (args->event_page_offset) {
 931                 struct kfd_dev *kfd;
 932                 struct kfd_process_device *pdd;
 933                 void *mem, *kern_addr;
 934                 uint64_t size;
 935
 936                 if (p->signal_page) {
 937                         pr_err("Event page is already set\n");
 938                         return -EINVAL;
 939                 }
 940
 941                 kfd = kfd_device_by_id(GET_GPU_ID(args->event_page_offset));
 942                 if (!kfd) {
 943                         pr_err("Getting device by id failed in %s\n", __func__);
 944                         return -EINVAL;
 945                 }
 946
 947                 mutex_lock(&p->mutex);
 948                 pdd = kfd_bind_process_to_device(kfd, p);
 949                 if (IS_ERR(pdd)) {
 950                         err = PTR_ERR(pdd);
 951                         goto out_unlock;
 952                 }
 953
 954                 mem = kfd_process_device_translate_handle(pdd,
 955                                 GET_IDR_HANDLE(args->event_page_offset));
 956                 if (!mem) {
 957                         pr_err("Can't find BO, offset is 0x%llx\n",
 958                                args->event_page_offset);
 959                         err = -EINVAL;
 960                         goto out_unlock;
 961                 }
 962                 mutex_unlock(&p->mutex);
 963
 964                 err = kfd->kfd2kgd->map_gtt_bo_to_kernel(kfd->kgd,
 965                                                 mem, &kern_addr, &size);
 966                 if (err) {
 967                         pr_err("Failed to map event page to kernel\n");
 968                         return err;
 969                 }
 970
 971                 err = kfd_event_page_set(p, kern_addr, size);
 972                 if (err) {
 973                         pr_err("Failed to set event page\n");
 974                         return err;
 975                 }
 976         }
 977
 978         err = kfd_event_create(filp, p, args->event_type,
 979                                 args->auto_reset != 0, args->node_id,
 980                                 &args->event_id, &args->event_trigger_data,
 981                                 &args->event_page_offset,
 982                                 &args->event_slot_index);
 983
 984         return err;
 985
 986 out_unlock:
 987         mutex_unlock(&p->mutex);
 988         return err;
 989 }
 990
 991 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
 992                                         void *data)
 993 {
 994         struct kfd_ioctl_destroy_event_args *args = data;
 995
 996         return kfd_event_destroy(p, args->event_id);
 997 }
 998
 999 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
1000                                 void *data)
1001 {
1002         struct kfd_ioctl_set_event_args *args = data;
1003
1004         return kfd_set_event(p, args->event_id);
1005 }
1006
1007 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
1008                                 void *data)
1009 {
1010         struct kfd_ioctl_reset_event_args *args = data;
1011
1012         return kfd_reset_event(p, args->event_id);
1013 }
1014
1015 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
1016                                 void *data)
1017 {
1018         struct kfd_ioctl_wait_events_args *args = data;
1019         int err;
1020
1021         err = kfd_wait_on_events(p, args->num_events,
1022                         (void __user *)args->events_ptr,
1023                         (args->wait_for_all != 0),
1024                         args->timeout, &args->wait_result);
1025
1026         return err;
1027 }
1028 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
1029                                         struct kfd_process *p, void *data)
1030 {
1031         struct kfd_ioctl_set_scratch_backing_va_args *args = data;
1032         struct kfd_process_device *pdd;
1033         struct kfd_dev *dev;
1034         long err;
1035
1036         dev = kfd_device_by_id(args->gpu_id);
1037         if (!dev)
1038                 return -EINVAL;
1039
1040         mutex_lock(&p->mutex);
1041
1042         pdd = kfd_bind_process_to_device(dev, p);
1043         if (IS_ERR(pdd)) {
1044                 err = PTR_ERR(pdd);
1045                 goto bind_process_to_device_fail;
1046         }
1047
1048         pdd->qpd.sh_hidden_private_base = args->va_addr;
1049
1050         mutex_unlock(&p->mutex);
1051
1052         if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
1053             pdd->qpd.vmid != 0)
1054                 dev->kfd2kgd->set_scratch_backing_va(
1055                         dev->kgd, args->va_addr, pdd->qpd.vmid);
1056
1057         return 0;
1058
1059 bind_process_to_device_fail:
1060         mutex_unlock(&p->mutex);
1061         return err;
1062 }
1063
1064 static int kfd_ioctl_get_tile_config(struct file *filep,
1065                 struct kfd_process *p, void *data)
1066 {
1067         struct kfd_ioctl_get_tile_config_args *args = data;
1068         struct kfd_dev *dev;
1069         struct tile_config config;
1070         int err = 0;
1071
1072         dev = kfd_device_by_id(args->gpu_id);
1073         if (!dev)
1074                 return -EINVAL;
1075
1076         dev->kfd2kgd->get_tile_config(dev->kgd, &config);
1077
1078         args->gb_addr_config = config.gb_addr_config;
1079         args->num_banks = config.num_banks;
1080         args->num_ranks = config.num_ranks;
1081
1082         if (args->num_tile_configs > config.num_tile_configs)
1083                 args->num_tile_configs = config.num_tile_configs;
1084         err = copy_to_user((void __user *)args->tile_config_ptr,
1085                         config.tile_config_ptr,
1086                         args->num_tile_configs * sizeof(uint32_t));
1087         if (err) {
1088                 args->num_tile_configs = 0;
1089                 return -EFAULT;
1090         }
1091
1092         if (args->num_macro_tile_configs > config.num_macro_tile_configs)
1093                 args->num_macro_tile_configs =
1094                                 config.num_macro_tile_configs;
1095         err = copy_to_user((void __user *)args->macro_tile_config_ptr,
1096                         config.macro_tile_config_ptr,
1097                         args->num_macro_tile_configs * sizeof(uint32_t));
1098         if (err) {
1099                 args->num_macro_tile_configs = 0;
1100                 return -EFAULT;
1101         }
1102
1103         return 0;
1104 }
1105
1106 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
1107                                 void *data)
1108 {
1109         struct kfd_ioctl_acquire_vm_args *args = data;
1110         struct kfd_process_device *pdd;
1111         struct kfd_dev *dev;
1112         struct file *drm_file;
1113         int ret;
1114
1115         dev = kfd_device_by_id(args->gpu_id);
1116         if (!dev)
1117                 return -EINVAL;
1118
1119         drm_file = fget(args->drm_fd);
1120         if (!drm_file)
1121                 return -EINVAL;
1122
1123         mutex_lock(&p->mutex);
1124
1125         pdd = kfd_get_process_device_data(dev, p);
1126         if (!pdd) {
1127                 ret = -EINVAL;
1128                 goto err_unlock;
1129         }
1130
1131         if (pdd->drm_file) {
1132                 ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
1133                 goto err_unlock;
1134         }
1135
1136         ret = kfd_process_device_init_vm(pdd, drm_file);
1137         if (ret)
1138                 goto err_unlock;
1139         /* On success, the PDD keeps the drm_file reference */
1140         mutex_unlock(&p->mutex);
1141
1142         return 0;
1143
1144 err_unlock:
1145         mutex_unlock(&p->mutex);
1146         fput(drm_file);
1147         return ret;
1148 }
1149
1150 static bool kfd_dev_is_large_bar(struct kfd_dev *dev)
1151 {
1152         struct kfd_local_mem_info mem_info;
1153
1154         if (debug_largebar) {
1155                 pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1156                 return true;
1157         }
1158
1159         if (dev->device_info->needs_iommu_device)
1160                 return false;
1161
1162         dev->kfd2kgd->get_local_mem_info(dev->kgd, &mem_info);
1163         if (mem_info.local_mem_size_private == 0 &&
1164                         mem_info.local_mem_size_public > 0)
1165                 return true;
1166         return false;
1167 }
1168
1169 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1170                                         struct kfd_process *p, void *data)
1171 {
1172         struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1173         struct kfd_process_device *pdd;
1174         void *mem;
1175         struct kfd_dev *dev;
1176         int idr_handle;
1177         long err;
1178         uint64_t offset = args->mmap_offset;
1179         uint32_t flags = args->flags;
1180
1181         if (args->size == 0)
1182                 return -EINVAL;
1183
1184         dev = kfd_device_by_id(args->gpu_id);
1185         if (!dev)
1186                 return -EINVAL;
1187
1188         if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1189                 (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1190                 !kfd_dev_is_large_bar(dev)) {
1191                 pr_err("Alloc host visible vram on small bar is not allowed\n");
1192                 return -EINVAL;
1193         }
1194
1195         mutex_lock(&p->mutex);
1196
1197         pdd = kfd_bind_process_to_device(dev, p);
1198         if (IS_ERR(pdd)) {
1199                 err = PTR_ERR(pdd);
1200                 goto err_unlock;
1201         }
1202
1203         err = dev->kfd2kgd->alloc_memory_of_gpu(
1204                 dev->kgd, args->va_addr, args->size,
1205                 pdd->vm, (struct kgd_mem **) &mem, &offset,
1206                 flags);
1207
1208         if (err)
1209                 goto err_unlock;
1210
1211         idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1212         if (idr_handle < 0) {
1213                 err = -EFAULT;
1214                 goto err_free;
1215         }
1216
1217         mutex_unlock(&p->mutex);
1218
1219         args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1220         args->mmap_offset = offset;
1221
1222         return 0;
1223
1224 err_free:
1225         dev->kfd2kgd->free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem);
1226 err_unlock:
1227         mutex_unlock(&p->mutex);
1228         return err;
1229 }
1230
1231 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1232                                         struct kfd_process *p, void *data)
1233 {
1234         struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1235         struct kfd_process_device *pdd;
1236         void *mem;
1237         struct kfd_dev *dev;
1238         int ret;
1239
1240         dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1241         if (!dev)
1242                 return -EINVAL;
1243
1244         mutex_lock(&p->mutex);
1245
1246         pdd = kfd_get_process_device_data(dev, p);
1247         if (!pdd) {
1248                 pr_err("Process device data doesn't exist\n");
1249                 ret = -EINVAL;
1250                 goto err_unlock;
1251         }
1252
1253         mem = kfd_process_device_translate_handle(
1254                 pdd, GET_IDR_HANDLE(args->handle));
1255         if (!mem) {
1256                 ret = -EINVAL;
1257                 goto err_unlock;
1258         }
1259
1260         ret = dev->kfd2kgd->free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem);
1261
1262         /* If freeing the buffer failed, leave the handle in place for
1263          * clean-up during process tear-down.
1264          */
1265         if (!ret)
1266                 kfd_process_device_remove_obj_handle(
1267                         pdd, GET_IDR_HANDLE(args->handle));
1268
1269 err_unlock:
1270         mutex_unlock(&p->mutex);
1271         return ret;
1272 }
1273
1274 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1275                                         struct kfd_process *p, void *data)
1276 {
1277         struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1278         struct kfd_process_device *pdd, *peer_pdd;
1279         void *mem;
1280         struct kfd_dev *dev, *peer;
1281         long err = 0;
1282         int i;
1283         uint32_t *devices_arr = NULL;
1284
1285         dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1286         if (!dev)
1287                 return -EINVAL;
1288
1289         if (!args->n_devices) {
1290                 pr_debug("Device IDs array empty\n");
1291                 return -EINVAL;
1292         }
1293         if (args->n_success > args->n_devices) {
1294                 pr_debug("n_success exceeds n_devices\n");
1295                 return -EINVAL;
1296         }
1297
1298         devices_arr = kmalloc(args->n_devices * sizeof(*devices_arr),
1299                               GFP_KERNEL);
1300         if (!devices_arr)
1301                 return -ENOMEM;
1302
1303         err = copy_from_user(devices_arr,
1304                              (void __user *)args->device_ids_array_ptr,
1305                              args->n_devices * sizeof(*devices_arr));
1306         if (err != 0) {
1307                 err = -EFAULT;
1308                 goto copy_from_user_failed;
1309         }
1310
1311         mutex_lock(&p->mutex);
1312
1313         pdd = kfd_bind_process_to_device(dev, p);
1314         if (IS_ERR(pdd)) {
1315                 err = PTR_ERR(pdd);
1316                 goto bind_process_to_device_failed;
1317         }
1318
1319         mem = kfd_process_device_translate_handle(pdd,
1320                                                 GET_IDR_HANDLE(args->handle));
1321         if (!mem) {
1322                 err = -ENOMEM;
1323                 goto get_mem_obj_from_handle_failed;
1324         }
1325
1326         for (i = args->n_success; i < args->n_devices; i++) {
1327                 peer = kfd_device_by_id(devices_arr[i]);
1328                 if (!peer) {
1329                         pr_debug("Getting device by id failed for 0x%x\n",
1330                                  devices_arr[i]);
1331                         err = -EINVAL;
1332                         goto get_mem_obj_from_handle_failed;
1333                 }
1334
1335                 peer_pdd = kfd_bind_process_to_device(peer, p);
1336                 if (IS_ERR(peer_pdd)) {
1337                         err = PTR_ERR(peer_pdd);
1338                         goto get_mem_obj_from_handle_failed;
1339                 }
1340                 err = peer->kfd2kgd->map_memory_to_gpu(
1341                         peer->kgd, (struct kgd_mem *)mem, peer_pdd->vm);
1342                 if (err) {
1343                         pr_err("Failed to map to gpu %d/%d\n",
1344                                i, args->n_devices);
1345                         goto map_memory_to_gpu_failed;
1346                 }
1347                 args->n_success = i+1;
1348         }
1349
1350         mutex_unlock(&p->mutex);
1351
1352         err = dev->kfd2kgd->sync_memory(dev->kgd, (struct kgd_mem *) mem, true);
1353         if (err) {
1354                 pr_debug("Sync memory failed, wait interrupted by user signal\n");
1355                 goto sync_memory_failed;
1356         }
1357
1358         /* Flush TLBs after waiting for the page table updates to complete */
1359         for (i = 0; i < args->n_devices; i++) {
1360                 peer = kfd_device_by_id(devices_arr[i]);
1361                 if (WARN_ON_ONCE(!peer))
1362                         continue;
1363                 peer_pdd = kfd_get_process_device_data(peer, p);
1364                 if (WARN_ON_ONCE(!peer_pdd))
1365                         continue;
1366                 kfd_flush_tlb(peer_pdd);
1367         }
1368
1369         kfree(devices_arr);
1370
1371         return err;
1372
1373 bind_process_to_device_failed:
1374 get_mem_obj_from_handle_failed:
1375 map_memory_to_gpu_failed:
1376         mutex_unlock(&p->mutex);
1377 copy_from_user_failed:
1378 sync_memory_failed:
1379         kfree(devices_arr);
1380
1381         return err;
1382 }
1383
1384 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1385                                         struct kfd_process *p, void *data)
1386 {
1387         struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1388         struct kfd_process_device *pdd, *peer_pdd;
1389         void *mem;
1390         struct kfd_dev *dev, *peer;
1391         long err = 0;
1392         uint32_t *devices_arr = NULL, i;
1393
1394         dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1395         if (!dev)
1396                 return -EINVAL;
1397
1398         if (!args->n_devices) {
1399                 pr_debug("Device IDs array empty\n");
1400                 return -EINVAL;
1401         }
1402         if (args->n_success > args->n_devices) {
1403                 pr_debug("n_success exceeds n_devices\n");
1404                 return -EINVAL;
1405         }
1406
1407         devices_arr = kmalloc(args->n_devices * sizeof(*devices_arr),
1408                               GFP_KERNEL);
1409         if (!devices_arr)
1410                 return -ENOMEM;
1411
1412         err = copy_from_user(devices_arr,
1413                              (void __user *)args->device_ids_array_ptr,
1414                              args->n_devices * sizeof(*devices_arr));
1415         if (err != 0) {
1416                 err = -EFAULT;
1417                 goto copy_from_user_failed;
1418         }
1419
1420         mutex_lock(&p->mutex);
1421
1422         pdd = kfd_get_process_device_data(dev, p);
1423         if (!pdd) {
1424                 err = PTR_ERR(pdd);
1425                 goto bind_process_to_device_failed;
1426         }
1427
1428         mem = kfd_process_device_translate_handle(pdd,
1429                                                 GET_IDR_HANDLE(args->handle));
1430         if (!mem) {
1431                 err = -ENOMEM;
1432                 goto get_mem_obj_from_handle_failed;
1433         }
1434
1435         for (i = args->n_success; i < args->n_devices; i++) {
1436                 peer = kfd_device_by_id(devices_arr[i]);
1437                 if (!peer) {
1438                         err = -EINVAL;
1439                         goto get_mem_obj_from_handle_failed;
1440                 }
1441
1442                 peer_pdd = kfd_get_process_device_data(peer, p);
1443                 if (!peer_pdd) {
1444                         err = -ENODEV;
1445                         goto get_mem_obj_from_handle_failed;
1446                 }
1447                 err = dev->kfd2kgd->unmap_memory_to_gpu(
1448                         peer->kgd, (struct kgd_mem *)mem, peer_pdd->vm);
1449                 if (err) {
1450                         pr_err("Failed to unmap from gpu %d/%d\n",
1451                                i, args->n_devices);
1452                         goto unmap_memory_from_gpu_failed;
1453                 }
1454                 args->n_success = i+1;
1455         }
1456         kfree(devices_arr);
1457
1458         mutex_unlock(&p->mutex);
1459
1460         return 0;
1461
1462 bind_process_to_device_failed:
1463 get_mem_obj_from_handle_failed:
1464 unmap_memory_from_gpu_failed:
1465         mutex_unlock(&p->mutex);
1466 copy_from_user_failed:
1467         kfree(devices_arr);
1468         return err;
1469 }
1470
1471 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
1472         [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
1473                             .cmd_drv = 0, .name = #ioctl}
1474
1475 /** Ioctl table */
1476 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
1477         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
1478                         kfd_ioctl_get_version, 0),
1479
1480         AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
1481                         kfd_ioctl_create_queue, 0),
1482
1483         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
1484                         kfd_ioctl_destroy_queue, 0),
1485
1486         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
1487                         kfd_ioctl_set_memory_policy, 0),
1488
1489         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
1490                         kfd_ioctl_get_clock_counters, 0),
1491
1492         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
1493                         kfd_ioctl_get_process_apertures, 0),
1494
1495         AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
1496                         kfd_ioctl_update_queue, 0),
1497
1498         AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
1499                         kfd_ioctl_create_event, 0),
1500
1501         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
1502                         kfd_ioctl_destroy_event, 0),
1503
1504         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
1505                         kfd_ioctl_set_event, 0),
1506
1507         AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
1508                         kfd_ioctl_reset_event, 0),
1509
1510         AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
1511                         kfd_ioctl_wait_events, 0),
1512
1513         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER,
1514                         kfd_ioctl_dbg_register, 0),
1515
1516         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER,
1517                         kfd_ioctl_dbg_unregister, 0),
1518
1519         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH,
1520                         kfd_ioctl_dbg_address_watch, 0),
1521
1522         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL,
1523                         kfd_ioctl_dbg_wave_control, 0),
1524
1525         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
1526                         kfd_ioctl_set_scratch_backing_va, 0),
1527
1528         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
1529                         kfd_ioctl_get_tile_config, 0),
1530
1531         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
1532                         kfd_ioctl_set_trap_handler, 0),
1533
1534         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
1535                         kfd_ioctl_get_process_apertures_new, 0),
1536
1537         AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
1538                         kfd_ioctl_acquire_vm, 0),
1539
1540         AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
1541                         kfd_ioctl_alloc_memory_of_gpu, 0),
1542
1543         AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
1544                         kfd_ioctl_free_memory_of_gpu, 0),
1545
1546         AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
1547                         kfd_ioctl_map_memory_to_gpu, 0),
1548
1549         AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
1550                         kfd_ioctl_unmap_memory_from_gpu, 0),
1551
1552 };
1553
1554 #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)
1555
1556 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
1557 {
1558         struct kfd_process *process;
1559         amdkfd_ioctl_t *func;
1560         const struct amdkfd_ioctl_desc *ioctl = NULL;
1561         unsigned int nr = _IOC_NR(cmd);
1562         char stack_kdata[128];
1563         char *kdata = NULL;
1564         unsigned int usize, asize;
1565         int retcode = -EINVAL;
1566
1567         if (nr >= AMDKFD_CORE_IOCTL_COUNT)
1568                 goto err_i1;
1569
1570         if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
1571                 u32 amdkfd_size;
1572
1573                 ioctl = &amdkfd_ioctls[nr];
1574
1575                 amdkfd_size = _IOC_SIZE(ioctl->cmd);
1576                 usize = asize = _IOC_SIZE(cmd);
1577                 if (amdkfd_size > asize)
1578                         asize = amdkfd_size;
1579
1580                 cmd = ioctl->cmd;
1581         } else
1582                 goto err_i1;
1583
1584         dev_dbg(kfd_device, "ioctl cmd 0x%x (#%d), arg 0x%lx\n", cmd, nr, arg);
1585
1586         process = kfd_get_process(current);
1587         if (IS_ERR(process)) {
1588                 dev_dbg(kfd_device, "no process\n");
1589                 goto err_i1;
1590         }
1591
1592         /* Do not trust userspace, use our own definition */
1593         func = ioctl->func;
1594
1595         if (unlikely(!func)) {
1596                 dev_dbg(kfd_device, "no function\n");
1597                 retcode = -EINVAL;
1598                 goto err_i1;
1599         }
1600
1601         if (cmd & (IOC_IN | IOC_OUT)) {
1602                 if (asize <= sizeof(stack_kdata)) {
1603                         kdata = stack_kdata;
1604                 } else {
1605                         kdata = kmalloc(asize, GFP_KERNEL);
1606                         if (!kdata) {
1607                                 retcode = -ENOMEM;
1608                                 goto err_i1;
1609                         }
1610                 }
1611                 if (asize > usize)
1612                         memset(kdata + usize, 0, asize - usize);
1613         }
1614
1615         if (cmd & IOC_IN) {
1616                 if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
1617                         retcode = -EFAULT;
1618                         goto err_i1;
1619                 }
1620         } else if (cmd & IOC_OUT) {
1621                 memset(kdata, 0, usize);
1622         }
1623
1624         retcode = func(filep, process, kdata);
1625
1626         if (cmd & IOC_OUT)
1627                 if (copy_to_user((void __user *)arg, kdata, usize) != 0)
1628                         retcode = -EFAULT;
1629
1630 err_i1:
1631         if (!ioctl)
1632                 dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
1633                           task_pid_nr(current), cmd, nr);
1634
1635         if (kdata != stack_kdata)
1636                 kfree(kdata);
1637
1638         if (retcode)
1639                 dev_dbg(kfd_device, "ret = %d\n", retcode);
1640
1641         return retcode;
1642 }
1643
1644 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
1645 {
1646         struct kfd_process *process;
1647
1648         process = kfd_get_process(current);
1649         if (IS_ERR(process))
1650                 return PTR_ERR(process);
1651
1652         if ((vma->vm_pgoff & KFD_MMAP_DOORBELL_MASK) ==
1653                         KFD_MMAP_DOORBELL_MASK) {
1654                 vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_DOORBELL_MASK;
1655                 return kfd_doorbell_mmap(process, vma);
1656         } else if ((vma->vm_pgoff & KFD_MMAP_EVENTS_MASK) ==
1657                         KFD_MMAP_EVENTS_MASK) {
1658                 vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_EVENTS_MASK;
1659                 return kfd_event_mmap(process, vma);
1660         } else if ((vma->vm_pgoff & KFD_MMAP_RESERVED_MEM_MASK) ==
1661                         KFD_MMAP_RESERVED_MEM_MASK) {
1662                 vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_RESERVED_MEM_MASK;
1663                 return kfd_reserved_mem_mmap(process, vma);
1664         }
1665
1666         return -EFAULT;
1667 }