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

   1 // SPDX-License-Identifier: GPL-2.0 OR MIT
   2 /*
   3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be included in
  13  * all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  21  * OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include <linux/device.h>
  25 #include <linux/export.h>
  26 #include <linux/err.h>
  27 #include <linux/fs.h>
  28 #include <linux/file.h>
  29 #include <linux/sched.h>
  30 #include <linux/slab.h>
  31 #include <linux/uaccess.h>
  32 #include <linux/compat.h>
  33 #include <uapi/linux/kfd_ioctl.h>
  34 #include <linux/time.h>
  35 #include <linux/mm.h>
  36 #include <linux/mman.h>
  37 #include <linux/ptrace.h>
  38 #include <linux/dma-buf.h>
  39 #include <linux/fdtable.h>
  40 #include <linux/processor.h>
  41 #include "kfd_priv.h"
  42 #include "kfd_device_queue_manager.h"
  43 #include "kfd_svm.h"
  44 #include "amdgpu_amdkfd.h"
  45 #include "kfd_smi_events.h"
  46 #include "amdgpu_dma_buf.h"
  47 #include "kfd_debug.h"
  48
  49 static long kfd_ioctl(struct file *, unsigned int, unsigned long);
  50 static int kfd_open(struct inode *, struct file *);
  51 static int kfd_release(struct inode *, struct file *);
  52 static int kfd_mmap(struct file *, struct vm_area_struct *);
  53
  54 static const char kfd_dev_name[] = "kfd";
  55
  56 static const struct file_operations kfd_fops = {
  57         .owner = THIS_MODULE,
  58         .unlocked_ioctl = kfd_ioctl,
  59         .compat_ioctl = compat_ptr_ioctl,
  60         .open = kfd_open,
  61         .release = kfd_release,
  62         .mmap = kfd_mmap,
  63 };
  64
  65 static int kfd_char_dev_major = -1;
  66 static struct class *kfd_class;
  67 struct device *kfd_device;
  68
  69 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id)
  70 {
  71         struct kfd_process_device *pdd;
  72
  73         mutex_lock(&p->mutex);
  74         pdd = kfd_process_device_data_by_id(p, gpu_id);
  75
  76         if (pdd)
  77                 return pdd;
  78
  79         mutex_unlock(&p->mutex);
  80         return NULL;
  81 }
  82
  83 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd)
  84 {
  85         mutex_unlock(&pdd->process->mutex);
  86 }
  87
  88 int kfd_chardev_init(void)
  89 {
  90         int err = 0;
  91
  92         kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
  93         err = kfd_char_dev_major;
  94         if (err < 0)
  95                 goto err_register_chrdev;
  96
  97         kfd_class = class_create(THIS_MODULE, kfd_dev_name);
  98         err = PTR_ERR(kfd_class);
  99         if (IS_ERR(kfd_class))
 100                 goto err_class_create;
 101
 102         kfd_device = device_create(kfd_class, NULL,
 103                                         MKDEV(kfd_char_dev_major, 0),
 104                                         NULL, kfd_dev_name);
 105         err = PTR_ERR(kfd_device);
 106         if (IS_ERR(kfd_device))
 107                 goto err_device_create;
 108
 109         return 0;
 110
 111 err_device_create:
 112         class_destroy(kfd_class);
 113 err_class_create:
 114         unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
 115 err_register_chrdev:
 116         return err;
 117 }
 118
 119 void kfd_chardev_exit(void)
 120 {
 121         device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
 122         class_destroy(kfd_class);
 123         unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
 124         kfd_device = NULL;
 125 }
 126
 127
 128 static int kfd_open(struct inode *inode, struct file *filep)
 129 {
 130         struct kfd_process *process;
 131         bool is_32bit_user_mode;
 132
 133         if (iminor(inode) != 0)
 134                 return -ENODEV;
 135
 136         is_32bit_user_mode = in_compat_syscall();
 137
 138         if (is_32bit_user_mode) {
 139                 dev_warn(kfd_device,
 140                         "Process %d (32-bit) failed to open /dev/kfd\n"
 141                         "32-bit processes are not supported by amdkfd\n",
 142                         current->pid);
 143                 return -EPERM;
 144         }
 145
 146         process = kfd_create_process(current);
 147         if (IS_ERR(process))
 148                 return PTR_ERR(process);
 149
 150         if (kfd_process_init_cwsr_apu(process, filep)) {
 151                 kfd_unref_process(process);
 152                 return -EFAULT;
 153         }
 154
 155         /* filep now owns the reference returned by kfd_create_process */
 156         filep->private_data = process;
 157
 158         dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
 159                 process->pasid, process->is_32bit_user_mode);
 160
 161         return 0;
 162 }
 163
 164 static int kfd_release(struct inode *inode, struct file *filep)
 165 {
 166         struct kfd_process *process = filep->private_data;
 167
 168         if (process)
 169                 kfd_unref_process(process);
 170
 171         return 0;
 172 }
 173
 174 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p,
 175                                         void *data)
 176 {
 177         struct kfd_ioctl_get_version_args *args = data;
 178
 179         args->major_version = KFD_IOCTL_MAJOR_VERSION;
 180         args->minor_version = KFD_IOCTL_MINOR_VERSION;
 181
 182         return 0;
 183 }
 184
 185 static int set_queue_properties_from_user(struct queue_properties *q_properties,
 186                                 struct kfd_ioctl_create_queue_args *args)
 187 {
 188         /*
 189          * Repurpose queue percentage to accommodate new features:
 190          * bit 0-7: queue percentage
 191          * bit 8-15: pm4_target_xcc
 192          */
 193         if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
 194                 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
 195                 return -EINVAL;
 196         }
 197
 198         if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
 199                 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
 200                 return -EINVAL;
 201         }
 202
 203         if ((args->ring_base_address) &&
 204                 (!access_ok((const void __user *) args->ring_base_address,
 205                         sizeof(uint64_t)))) {
 206                 pr_err("Can't access ring base address\n");
 207                 return -EFAULT;
 208         }
 209
 210         if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
 211                 pr_err("Ring size must be a power of 2 or 0\n");
 212                 return -EINVAL;
 213         }
 214
 215         if (!access_ok((const void __user *) args->read_pointer_address,
 216                         sizeof(uint32_t))) {
 217                 pr_err("Can't access read pointer\n");
 218                 return -EFAULT;
 219         }
 220
 221         if (!access_ok((const void __user *) args->write_pointer_address,
 222                         sizeof(uint32_t))) {
 223                 pr_err("Can't access write pointer\n");
 224                 return -EFAULT;
 225         }
 226
 227         if (args->eop_buffer_address &&
 228                 !access_ok((const void __user *) args->eop_buffer_address,
 229                         sizeof(uint32_t))) {
 230                 pr_debug("Can't access eop buffer");
 231                 return -EFAULT;
 232         }
 233
 234         if (args->ctx_save_restore_address &&
 235                 !access_ok((const void __user *) args->ctx_save_restore_address,
 236                         sizeof(uint32_t))) {
 237                 pr_debug("Can't access ctx save restore buffer");
 238                 return -EFAULT;
 239         }
 240
 241         q_properties->is_interop = false;
 242         q_properties->is_gws = false;
 243         q_properties->queue_percent = args->queue_percentage & 0xFF;
 244         /* bit 8-15 are repurposed to be PM4 target XCC */
 245         q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
 246         q_properties->priority = args->queue_priority;
 247         q_properties->queue_address = args->ring_base_address;
 248         q_properties->queue_size = args->ring_size;
 249         q_properties->read_ptr = (uint32_t *) args->read_pointer_address;
 250         q_properties->write_ptr = (uint32_t *) args->write_pointer_address;
 251         q_properties->eop_ring_buffer_address = args->eop_buffer_address;
 252         q_properties->eop_ring_buffer_size = args->eop_buffer_size;
 253         q_properties->ctx_save_restore_area_address =
 254                         args->ctx_save_restore_address;
 255         q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
 256         q_properties->ctl_stack_size = args->ctl_stack_size;
 257         if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
 258                 args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
 259                 q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
 260         else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA)
 261                 q_properties->type = KFD_QUEUE_TYPE_SDMA;
 262         else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI)
 263                 q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI;
 264         else
 265                 return -ENOTSUPP;
 266
 267         if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
 268                 q_properties->format = KFD_QUEUE_FORMAT_AQL;
 269         else
 270                 q_properties->format = KFD_QUEUE_FORMAT_PM4;
 271
 272         pr_debug("Queue Percentage: %d, %d\n",
 273                         q_properties->queue_percent, args->queue_percentage);
 274
 275         pr_debug("Queue Priority: %d, %d\n",
 276                         q_properties->priority, args->queue_priority);
 277
 278         pr_debug("Queue Address: 0x%llX, 0x%llX\n",
 279                         q_properties->queue_address, args->ring_base_address);
 280
 281         pr_debug("Queue Size: 0x%llX, %u\n",
 282                         q_properties->queue_size, args->ring_size);
 283
 284         pr_debug("Queue r/w Pointers: %px, %px\n",
 285                         q_properties->read_ptr,
 286                         q_properties->write_ptr);
 287
 288         pr_debug("Queue Format: %d\n", q_properties->format);
 289
 290         pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address);
 291
 292         pr_debug("Queue CTX save area: 0x%llX\n",
 293                         q_properties->ctx_save_restore_area_address);
 294
 295         return 0;
 296 }
 297
 298 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 299                                         void *data)
 300 {
 301         struct kfd_ioctl_create_queue_args *args = data;
 302         struct kfd_node *dev;
 303         int err = 0;
 304         unsigned int queue_id;
 305         struct kfd_process_device *pdd;
 306         struct queue_properties q_properties;
 307         uint32_t doorbell_offset_in_process = 0;
 308         struct amdgpu_bo *wptr_bo = NULL;
 309
 310         memset(&q_properties, 0, sizeof(struct queue_properties));
 311
 312         pr_debug("Creating queue ioctl\n");
 313
 314         err = set_queue_properties_from_user(&q_properties, args);
 315         if (err)
 316                 return err;
 317
 318         pr_debug("Looking for gpu id 0x%x\n", args->gpu_id);
 319
 320         mutex_lock(&p->mutex);
 321
 322         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 323         if (!pdd) {
 324                 pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
 325                 err = -EINVAL;
 326                 goto err_pdd;
 327         }
 328         dev = pdd->dev;
 329
 330         pdd = kfd_bind_process_to_device(dev, p);
 331         if (IS_ERR(pdd)) {
 332                 err = -ESRCH;
 333                 goto err_bind_process;
 334         }
 335
 336         if (!pdd->doorbell_index &&
 337             kfd_alloc_process_doorbells(dev->kfd, &pdd->doorbell_index) < 0) {
 338                 err = -ENOMEM;
 339                 goto err_alloc_doorbells;
 340         }
 341
 342         /* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
 343          * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
 344          */
 345         if (dev->kfd->shared_resources.enable_mes &&
 346                         ((dev->adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK)
 347                         >> AMDGPU_MES_API_VERSION_SHIFT) >= 2) {
 348                 struct amdgpu_bo_va_mapping *wptr_mapping;
 349                 struct amdgpu_vm *wptr_vm;
 350
 351                 wptr_vm = drm_priv_to_vm(pdd->drm_priv);
 352                 err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
 353                 if (err)
 354                         goto err_wptr_map_gart;
 355
 356                 wptr_mapping = amdgpu_vm_bo_lookup_mapping(
 357                                 wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
 358                 amdgpu_bo_unreserve(wptr_vm->root.bo);
 359                 if (!wptr_mapping) {
 360                         pr_err("Failed to lookup wptr bo\n");
 361                         err = -EINVAL;
 362                         goto err_wptr_map_gart;
 363                 }
 364
 365                 wptr_bo = wptr_mapping->bo_va->base.bo;
 366                 if (wptr_bo->tbo.base.size > PAGE_SIZE) {
 367                         pr_err("Requested GART mapping for wptr bo larger than one page\n");
 368                         err = -EINVAL;
 369                         goto err_wptr_map_gart;
 370                 }
 371
 372                 err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
 373                 if (err) {
 374                         pr_err("Failed to map wptr bo to GART\n");
 375                         goto err_wptr_map_gart;
 376                 }
 377         }
 378
 379         pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
 380                         p->pasid,
 381                         dev->id);
 382
 383         err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
 384                         NULL, NULL, NULL, &doorbell_offset_in_process);
 385         if (err != 0)
 386                 goto err_create_queue;
 387
 388         args->queue_id = queue_id;
 389
 390
 391         /* Return gpu_id as doorbell offset for mmap usage */
 392         args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
 393         args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
 394         if (KFD_IS_SOC15(dev))
 395                 /* On SOC15 ASICs, include the doorbell offset within the
 396                  * process doorbell frame, which is 2 pages.
 397                  */
 398                 args->doorbell_offset |= doorbell_offset_in_process;
 399
 400         mutex_unlock(&p->mutex);
 401
 402         pr_debug("Queue id %d was created successfully\n", args->queue_id);
 403
 404         pr_debug("Ring buffer address == 0x%016llX\n",
 405                         args->ring_base_address);
 406
 407         pr_debug("Read ptr address    == 0x%016llX\n",
 408                         args->read_pointer_address);
 409
 410         pr_debug("Write ptr address   == 0x%016llX\n",
 411                         args->write_pointer_address);
 412
 413         return 0;
 414
 415 err_create_queue:
 416         if (wptr_bo)
 417                 amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
 418 err_wptr_map_gart:
 419 err_alloc_doorbells:
 420 err_bind_process:
 421 err_pdd:
 422         mutex_unlock(&p->mutex);
 423         return err;
 424 }
 425
 426 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
 427                                         void *data)
 428 {
 429         int retval;
 430         struct kfd_ioctl_destroy_queue_args *args = data;
 431
 432         pr_debug("Destroying queue id %d for pasid 0x%x\n",
 433                                 args->queue_id,
 434                                 p->pasid);
 435
 436         mutex_lock(&p->mutex);
 437
 438         retval = pqm_destroy_queue(&p->pqm, args->queue_id);
 439
 440         mutex_unlock(&p->mutex);
 441         return retval;
 442 }
 443
 444 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
 445                                         void *data)
 446 {
 447         int retval;
 448         struct kfd_ioctl_update_queue_args *args = data;
 449         struct queue_properties properties;
 450
 451         /*
 452          * Repurpose queue percentage to accommodate new features:
 453          * bit 0-7: queue percentage
 454          * bit 8-15: pm4_target_xcc
 455          */
 456         if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
 457                 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
 458                 return -EINVAL;
 459         }
 460
 461         if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
 462                 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
 463                 return -EINVAL;
 464         }
 465
 466         if ((args->ring_base_address) &&
 467                 (!access_ok((const void __user *) args->ring_base_address,
 468                         sizeof(uint64_t)))) {
 469                 pr_err("Can't access ring base address\n");
 470                 return -EFAULT;
 471         }
 472
 473         if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
 474                 pr_err("Ring size must be a power of 2 or 0\n");
 475                 return -EINVAL;
 476         }
 477
 478         properties.queue_address = args->ring_base_address;
 479         properties.queue_size = args->ring_size;
 480         properties.queue_percent = args->queue_percentage & 0xFF;
 481         /* bit 8-15 are repurposed to be PM4 target XCC */
 482         properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
 483         properties.priority = args->queue_priority;
 484
 485         pr_debug("Updating queue id %d for pasid 0x%x\n",
 486                         args->queue_id, p->pasid);
 487
 488         mutex_lock(&p->mutex);
 489
 490         retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
 491
 492         mutex_unlock(&p->mutex);
 493
 494         return retval;
 495 }
 496
 497 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
 498                                         void *data)
 499 {
 500         int retval;
 501         const int max_num_cus = 1024;
 502         struct kfd_ioctl_set_cu_mask_args *args = data;
 503         struct mqd_update_info minfo = {0};
 504         uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
 505         size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
 506
 507         if ((args->num_cu_mask % 32) != 0) {
 508                 pr_debug("num_cu_mask 0x%x must be a multiple of 32",
 509                                 args->num_cu_mask);
 510                 return -EINVAL;
 511         }
 512
 513         minfo.cu_mask.count = args->num_cu_mask;
 514         if (minfo.cu_mask.count == 0) {
 515                 pr_debug("CU mask cannot be 0");
 516                 return -EINVAL;
 517         }
 518
 519         /* To prevent an unreasonably large CU mask size, set an arbitrary
 520          * limit of max_num_cus bits.  We can then just drop any CU mask bits
 521          * past max_num_cus bits and just use the first max_num_cus bits.
 522          */
 523         if (minfo.cu_mask.count > max_num_cus) {
 524                 pr_debug("CU mask cannot be greater than 1024 bits");
 525                 minfo.cu_mask.count = max_num_cus;
 526                 cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
 527         }
 528
 529         minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
 530         if (!minfo.cu_mask.ptr)
 531                 return -ENOMEM;
 532
 533         retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
 534         if (retval) {
 535                 pr_debug("Could not copy CU mask from userspace");
 536                 retval = -EFAULT;
 537                 goto out;
 538         }
 539
 540         mutex_lock(&p->mutex);
 541
 542         retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
 543
 544         mutex_unlock(&p->mutex);
 545
 546 out:
 547         kfree(minfo.cu_mask.ptr);
 548         return retval;
 549 }
 550
 551 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
 552                                           struct kfd_process *p, void *data)
 553 {
 554         struct kfd_ioctl_get_queue_wave_state_args *args = data;
 555         int r;
 556
 557         mutex_lock(&p->mutex);
 558
 559         r = pqm_get_wave_state(&p->pqm, args->queue_id,
 560                                (void __user *)args->ctl_stack_address,
 561                                &args->ctl_stack_used_size,
 562                                &args->save_area_used_size);
 563
 564         mutex_unlock(&p->mutex);
 565
 566         return r;
 567 }
 568
 569 static int kfd_ioctl_set_memory_policy(struct file *filep,
 570                                         struct kfd_process *p, void *data)
 571 {
 572         struct kfd_ioctl_set_memory_policy_args *args = data;
 573         int err = 0;
 574         struct kfd_process_device *pdd;
 575         enum cache_policy default_policy, alternate_policy;
 576
 577         if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
 578             && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 579                 return -EINVAL;
 580         }
 581
 582         if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
 583             && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 584                 return -EINVAL;
 585         }
 586
 587         mutex_lock(&p->mutex);
 588         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 589         if (!pdd) {
 590                 pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
 591                 err = -EINVAL;
 592                 goto err_pdd;
 593         }
 594
 595         pdd = kfd_bind_process_to_device(pdd->dev, p);
 596         if (IS_ERR(pdd)) {
 597                 err = -ESRCH;
 598                 goto out;
 599         }
 600
 601         default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 602                          ? cache_policy_coherent : cache_policy_noncoherent;
 603
 604         alternate_policy =
 605                 (args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 606                    ? cache_policy_coherent : cache_policy_noncoherent;
 607
 608         if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
 609                                 &pdd->qpd,
 610                                 default_policy,
 611                                 alternate_policy,
 612                                 (void __user *)args->alternate_aperture_base,
 613                                 args->alternate_aperture_size))
 614                 err = -EINVAL;
 615
 616 out:
 617 err_pdd:
 618         mutex_unlock(&p->mutex);
 619
 620         return err;
 621 }
 622
 623 static int kfd_ioctl_set_trap_handler(struct file *filep,
 624                                         struct kfd_process *p, void *data)
 625 {
 626         struct kfd_ioctl_set_trap_handler_args *args = data;
 627         int err = 0;
 628         struct kfd_process_device *pdd;
 629
 630         mutex_lock(&p->mutex);
 631
 632         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 633         if (!pdd) {
 634                 err = -EINVAL;
 635                 goto err_pdd;
 636         }
 637
 638         pdd = kfd_bind_process_to_device(pdd->dev, p);
 639         if (IS_ERR(pdd)) {
 640                 err = -ESRCH;
 641                 goto out;
 642         }
 643
 644         kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
 645
 646 out:
 647 err_pdd:
 648         mutex_unlock(&p->mutex);
 649
 650         return err;
 651 }
 652
 653 static int kfd_ioctl_dbg_register(struct file *filep,
 654                                 struct kfd_process *p, void *data)
 655 {
 656         return -EPERM;
 657 }
 658
 659 static int kfd_ioctl_dbg_unregister(struct file *filep,
 660                                 struct kfd_process *p, void *data)
 661 {
 662         return -EPERM;
 663 }
 664
 665 static int kfd_ioctl_dbg_address_watch(struct file *filep,
 666                                         struct kfd_process *p, void *data)
 667 {
 668         return -EPERM;
 669 }
 670
 671 /* Parse and generate fixed size data structure for wave control */
 672 static int kfd_ioctl_dbg_wave_control(struct file *filep,
 673                                         struct kfd_process *p, void *data)
 674 {
 675         return -EPERM;
 676 }
 677
 678 static int kfd_ioctl_get_clock_counters(struct file *filep,
 679                                 struct kfd_process *p, void *data)
 680 {
 681         struct kfd_ioctl_get_clock_counters_args *args = data;
 682         struct kfd_process_device *pdd;
 683
 684         mutex_lock(&p->mutex);
 685         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 686         mutex_unlock(&p->mutex);
 687         if (pdd)
 688                 /* Reading GPU clock counter from KGD */
 689                 args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
 690         else
 691                 /* Node without GPU resource */
 692                 args->gpu_clock_counter = 0;
 693
 694         /* No access to rdtsc. Using raw monotonic time */
 695         args->cpu_clock_counter = ktime_get_raw_ns();
 696         args->system_clock_counter = ktime_get_boottime_ns();
 697
 698         /* Since the counter is in nano-seconds we use 1GHz frequency */
 699         args->system_clock_freq = 1000000000;
 700
 701         return 0;
 702 }
 703
 704
 705 static int kfd_ioctl_get_process_apertures(struct file *filp,
 706                                 struct kfd_process *p, void *data)
 707 {
 708         struct kfd_ioctl_get_process_apertures_args *args = data;
 709         struct kfd_process_device_apertures *pAperture;
 710         int i;
 711
 712         dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
 713
 714         args->num_of_nodes = 0;
 715
 716         mutex_lock(&p->mutex);
 717         /* Run over all pdd of the process */
 718         for (i = 0; i < p->n_pdds; i++) {
 719                 struct kfd_process_device *pdd = p->pdds[i];
 720
 721                 pAperture =
 722                         &args->process_apertures[args->num_of_nodes];
 723                 pAperture->gpu_id = pdd->dev->id;
 724                 pAperture->lds_base = pdd->lds_base;
 725                 pAperture->lds_limit = pdd->lds_limit;
 726                 pAperture->gpuvm_base = pdd->gpuvm_base;
 727                 pAperture->gpuvm_limit = pdd->gpuvm_limit;
 728                 pAperture->scratch_base = pdd->scratch_base;
 729                 pAperture->scratch_limit = pdd->scratch_limit;
 730
 731                 dev_dbg(kfd_device,
 732                         "node id %u\n", args->num_of_nodes);
 733                 dev_dbg(kfd_device,
 734                         "gpu id %u\n", pdd->dev->id);
 735                 dev_dbg(kfd_device,
 736                         "lds_base %llX\n", pdd->lds_base);
 737                 dev_dbg(kfd_device,
 738                         "lds_limit %llX\n", pdd->lds_limit);
 739                 dev_dbg(kfd_device,
 740                         "gpuvm_base %llX\n", pdd->gpuvm_base);
 741                 dev_dbg(kfd_device,
 742                         "gpuvm_limit %llX\n", pdd->gpuvm_limit);
 743                 dev_dbg(kfd_device,
 744                         "scratch_base %llX\n", pdd->scratch_base);
 745                 dev_dbg(kfd_device,
 746                         "scratch_limit %llX\n", pdd->scratch_limit);
 747
 748                 if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
 749                         break;
 750         }
 751         mutex_unlock(&p->mutex);
 752
 753         return 0;
 754 }
 755
 756 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
 757                                 struct kfd_process *p, void *data)
 758 {
 759         struct kfd_ioctl_get_process_apertures_new_args *args = data;
 760         struct kfd_process_device_apertures *pa;
 761         int ret;
 762         int i;
 763
 764         dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
 765
 766         if (args->num_of_nodes == 0) {
 767                 /* Return number of nodes, so that user space can alloacate
 768                  * sufficient memory
 769                  */
 770                 mutex_lock(&p->mutex);
 771                 args->num_of_nodes = p->n_pdds;
 772                 goto out_unlock;
 773         }
 774
 775         /* Fill in process-aperture information for all available
 776          * nodes, but not more than args->num_of_nodes as that is
 777          * the amount of memory allocated by user
 778          */
 779         pa = kzalloc((sizeof(struct kfd_process_device_apertures) *
 780                                 args->num_of_nodes), GFP_KERNEL);
 781         if (!pa)
 782                 return -ENOMEM;
 783
 784         mutex_lock(&p->mutex);
 785
 786         if (!p->n_pdds) {
 787                 args->num_of_nodes = 0;
 788                 kfree(pa);
 789                 goto out_unlock;
 790         }
 791
 792         /* Run over all pdd of the process */
 793         for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
 794                 struct kfd_process_device *pdd = p->pdds[i];
 795
 796                 pa[i].gpu_id = pdd->dev->id;
 797                 pa[i].lds_base = pdd->lds_base;
 798                 pa[i].lds_limit = pdd->lds_limit;
 799                 pa[i].gpuvm_base = pdd->gpuvm_base;
 800                 pa[i].gpuvm_limit = pdd->gpuvm_limit;
 801                 pa[i].scratch_base = pdd->scratch_base;
 802                 pa[i].scratch_limit = pdd->scratch_limit;
 803
 804                 dev_dbg(kfd_device,
 805                         "gpu id %u\n", pdd->dev->id);
 806                 dev_dbg(kfd_device,
 807                         "lds_base %llX\n", pdd->lds_base);
 808                 dev_dbg(kfd_device,
 809                         "lds_limit %llX\n", pdd->lds_limit);
 810                 dev_dbg(kfd_device,
 811                         "gpuvm_base %llX\n", pdd->gpuvm_base);
 812                 dev_dbg(kfd_device,
 813                         "gpuvm_limit %llX\n", pdd->gpuvm_limit);
 814                 dev_dbg(kfd_device,
 815                         "scratch_base %llX\n", pdd->scratch_base);
 816                 dev_dbg(kfd_device,
 817                         "scratch_limit %llX\n", pdd->scratch_limit);
 818         }
 819         mutex_unlock(&p->mutex);
 820
 821         args->num_of_nodes = i;
 822         ret = copy_to_user(
 823                         (void __user *)args->kfd_process_device_apertures_ptr,
 824                         pa,
 825                         (i * sizeof(struct kfd_process_device_apertures)));
 826         kfree(pa);
 827         return ret ? -EFAULT : 0;
 828
 829 out_unlock:
 830         mutex_unlock(&p->mutex);
 831         return 0;
 832 }
 833
 834 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
 835                                         void *data)
 836 {
 837         struct kfd_ioctl_create_event_args *args = data;
 838         int err;
 839
 840         /* For dGPUs the event page is allocated in user mode. The
 841          * handle is passed to KFD with the first call to this IOCTL
 842          * through the event_page_offset field.
 843          */
 844         if (args->event_page_offset) {
 845                 mutex_lock(&p->mutex);
 846                 err = kfd_kmap_event_page(p, args->event_page_offset);
 847                 mutex_unlock(&p->mutex);
 848                 if (err)
 849                         return err;
 850         }
 851
 852         err = kfd_event_create(filp, p, args->event_type,
 853                                 args->auto_reset != 0, args->node_id,
 854                                 &args->event_id, &args->event_trigger_data,
 855                                 &args->event_page_offset,
 856                                 &args->event_slot_index);
 857
 858         pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
 859         return err;
 860 }
 861
 862 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
 863                                         void *data)
 864 {
 865         struct kfd_ioctl_destroy_event_args *args = data;
 866
 867         return kfd_event_destroy(p, args->event_id);
 868 }
 869
 870 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
 871                                 void *data)
 872 {
 873         struct kfd_ioctl_set_event_args *args = data;
 874
 875         return kfd_set_event(p, args->event_id);
 876 }
 877
 878 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
 879                                 void *data)
 880 {
 881         struct kfd_ioctl_reset_event_args *args = data;
 882
 883         return kfd_reset_event(p, args->event_id);
 884 }
 885
 886 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
 887                                 void *data)
 888 {
 889         struct kfd_ioctl_wait_events_args *args = data;
 890
 891         return kfd_wait_on_events(p, args->num_events,
 892                         (void __user *)args->events_ptr,
 893                         (args->wait_for_all != 0),
 894                         &args->timeout, &args->wait_result);
 895 }
 896 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
 897                                         struct kfd_process *p, void *data)
 898 {
 899         struct kfd_ioctl_set_scratch_backing_va_args *args = data;
 900         struct kfd_process_device *pdd;
 901         struct kfd_node *dev;
 902         long err;
 903
 904         mutex_lock(&p->mutex);
 905         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 906         if (!pdd) {
 907                 err = -EINVAL;
 908                 goto err_pdd;
 909         }
 910         dev = pdd->dev;
 911
 912         pdd = kfd_bind_process_to_device(dev, p);
 913         if (IS_ERR(pdd)) {
 914                 err = PTR_ERR(pdd);
 915                 goto bind_process_to_device_fail;
 916         }
 917
 918         pdd->qpd.sh_hidden_private_base = args->va_addr;
 919
 920         mutex_unlock(&p->mutex);
 921
 922         if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
 923             pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
 924                 dev->kfd2kgd->set_scratch_backing_va(
 925                         dev->adev, args->va_addr, pdd->qpd.vmid);
 926
 927         return 0;
 928
 929 bind_process_to_device_fail:
 930 err_pdd:
 931         mutex_unlock(&p->mutex);
 932         return err;
 933 }
 934
 935 static int kfd_ioctl_get_tile_config(struct file *filep,
 936                 struct kfd_process *p, void *data)
 937 {
 938         struct kfd_ioctl_get_tile_config_args *args = data;
 939         struct kfd_process_device *pdd;
 940         struct tile_config config;
 941         int err = 0;
 942
 943         mutex_lock(&p->mutex);
 944         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 945         mutex_unlock(&p->mutex);
 946         if (!pdd)
 947                 return -EINVAL;
 948
 949         amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
 950
 951         args->gb_addr_config = config.gb_addr_config;
 952         args->num_banks = config.num_banks;
 953         args->num_ranks = config.num_ranks;
 954
 955         if (args->num_tile_configs > config.num_tile_configs)
 956                 args->num_tile_configs = config.num_tile_configs;
 957         err = copy_to_user((void __user *)args->tile_config_ptr,
 958                         config.tile_config_ptr,
 959                         args->num_tile_configs * sizeof(uint32_t));
 960         if (err) {
 961                 args->num_tile_configs = 0;
 962                 return -EFAULT;
 963         }
 964
 965         if (args->num_macro_tile_configs > config.num_macro_tile_configs)
 966                 args->num_macro_tile_configs =
 967                                 config.num_macro_tile_configs;
 968         err = copy_to_user((void __user *)args->macro_tile_config_ptr,
 969                         config.macro_tile_config_ptr,
 970                         args->num_macro_tile_configs * sizeof(uint32_t));
 971         if (err) {
 972                 args->num_macro_tile_configs = 0;
 973                 return -EFAULT;
 974         }
 975
 976         return 0;
 977 }
 978
 979 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
 980                                 void *data)
 981 {
 982         struct kfd_ioctl_acquire_vm_args *args = data;
 983         struct kfd_process_device *pdd;
 984         struct file *drm_file;
 985         int ret;
 986
 987         drm_file = fget(args->drm_fd);
 988         if (!drm_file)
 989                 return -EINVAL;
 990
 991         mutex_lock(&p->mutex);
 992         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 993         if (!pdd) {
 994                 ret = -EINVAL;
 995                 goto err_pdd;
 996         }
 997
 998         if (pdd->drm_file) {
 999                 ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
1000                 goto err_drm_file;
1001         }
1002
1003         ret = kfd_process_device_init_vm(pdd, drm_file);
1004         if (ret)
1005                 goto err_unlock;
1006
1007         /* On success, the PDD keeps the drm_file reference */
1008         mutex_unlock(&p->mutex);
1009
1010         return 0;
1011
1012 err_unlock:
1013 err_pdd:
1014 err_drm_file:
1015         mutex_unlock(&p->mutex);
1016         fput(drm_file);
1017         return ret;
1018 }
1019
1020 bool kfd_dev_is_large_bar(struct kfd_node *dev)
1021 {
1022         if (debug_largebar) {
1023                 pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1024                 return true;
1025         }
1026
1027         if (dev->kfd->use_iommu_v2)
1028                 return false;
1029
1030         if (dev->local_mem_info.local_mem_size_private == 0 &&
1031             dev->local_mem_info.local_mem_size_public > 0)
1032                 return true;
1033
1034         if (dev->local_mem_info.local_mem_size_public == 0 &&
1035             dev->kfd->adev->gmc.is_app_apu) {
1036                 pr_debug("APP APU, Consider like a large bar system\n");
1037                 return true;
1038         }
1039
1040         return false;
1041 }
1042
1043 static int kfd_ioctl_get_available_memory(struct file *filep,
1044                                           struct kfd_process *p, void *data)
1045 {
1046         struct kfd_ioctl_get_available_memory_args *args = data;
1047         struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1048
1049         if (!pdd)
1050                 return -EINVAL;
1051         args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1052                                                         pdd->dev->node_id);
1053         kfd_unlock_pdd(pdd);
1054         return 0;
1055 }
1056
1057 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1058                                         struct kfd_process *p, void *data)
1059 {
1060         struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1061         struct kfd_process_device *pdd;
1062         void *mem;
1063         struct kfd_node *dev;
1064         int idr_handle;
1065         long err;
1066         uint64_t offset = args->mmap_offset;
1067         uint32_t flags = args->flags;
1068
1069         if (args->size == 0)
1070                 return -EINVAL;
1071
1072 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1073         /* Flush pending deferred work to avoid racing with deferred actions
1074          * from previous memory map changes (e.g. munmap).
1075          */
1076         svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1077         mutex_lock(&p->svms.lock);
1078         mmap_write_unlock(current->mm);
1079         if (interval_tree_iter_first(&p->svms.objects,
1080                                      args->va_addr >> PAGE_SHIFT,
1081                                      (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1082                 pr_err("Address: 0x%llx already allocated by SVM\n",
1083                         args->va_addr);
1084                 mutex_unlock(&p->svms.lock);
1085                 return -EADDRINUSE;
1086         }
1087
1088         /* When register user buffer check if it has been registered by svm by
1089          * buffer cpu virtual address.
1090          */
1091         if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1092             interval_tree_iter_first(&p->svms.objects,
1093                                      args->mmap_offset >> PAGE_SHIFT,
1094                                      (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1095                 pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1096                         args->mmap_offset);
1097                 mutex_unlock(&p->svms.lock);
1098                 return -EADDRINUSE;
1099         }
1100
1101         mutex_unlock(&p->svms.lock);
1102 #endif
1103         mutex_lock(&p->mutex);
1104         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1105         if (!pdd) {
1106                 err = -EINVAL;
1107                 goto err_pdd;
1108         }
1109
1110         dev = pdd->dev;
1111
1112         if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1113                 (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1114                 !kfd_dev_is_large_bar(dev)) {
1115                 pr_err("Alloc host visible vram on small bar is not allowed\n");
1116                 err = -EINVAL;
1117                 goto err_large_bar;
1118         }
1119
1120         pdd = kfd_bind_process_to_device(dev, p);
1121         if (IS_ERR(pdd)) {
1122                 err = PTR_ERR(pdd);
1123                 goto err_unlock;
1124         }
1125
1126         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1127                 if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1128                         err = -EINVAL;
1129                         goto err_unlock;
1130                 }
1131                 offset = kfd_get_process_doorbells(pdd);
1132                 if (!offset) {
1133                         err = -ENOMEM;
1134                         goto err_unlock;
1135                 }
1136         } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1137                 if (args->size != PAGE_SIZE) {
1138                         err = -EINVAL;
1139                         goto err_unlock;
1140                 }
1141                 offset = dev->adev->rmmio_remap.bus_addr;
1142                 if (!offset) {
1143                         err = -ENOMEM;
1144                         goto err_unlock;
1145                 }
1146         }
1147
1148         err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1149                 dev->adev, args->va_addr, args->size,
1150                 pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1151                 flags, false);
1152
1153         if (err)
1154                 goto err_unlock;
1155
1156         idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1157         if (idr_handle < 0) {
1158                 err = -EFAULT;
1159                 goto err_free;
1160         }
1161
1162         /* Update the VRAM usage count */
1163         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1164                 uint64_t size = args->size;
1165
1166                 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1167                         size >>= 1;
1168                 WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1169         }
1170
1171         mutex_unlock(&p->mutex);
1172
1173         args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1174         args->mmap_offset = offset;
1175
1176         /* MMIO is mapped through kfd device
1177          * Generate a kfd mmap offset
1178          */
1179         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1180                 args->mmap_offset = KFD_MMAP_TYPE_MMIO
1181                                         | KFD_MMAP_GPU_ID(args->gpu_id);
1182
1183         return 0;
1184
1185 err_free:
1186         amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1187                                                pdd->drm_priv, NULL);
1188 err_unlock:
1189 err_pdd:
1190 err_large_bar:
1191         mutex_unlock(&p->mutex);
1192         return err;
1193 }
1194
1195 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1196                                         struct kfd_process *p, void *data)
1197 {
1198         struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1199         struct kfd_process_device *pdd;
1200         void *mem;
1201         int ret;
1202         uint64_t size = 0;
1203
1204         mutex_lock(&p->mutex);
1205         /*
1206          * Safeguard to prevent user space from freeing signal BO.
1207          * It will be freed at process termination.
1208          */
1209         if (p->signal_handle && (p->signal_handle == args->handle)) {
1210                 pr_err("Free signal BO is not allowed\n");
1211                 ret = -EPERM;
1212                 goto err_unlock;
1213         }
1214
1215         pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1216         if (!pdd) {
1217                 pr_err("Process device data doesn't exist\n");
1218                 ret = -EINVAL;
1219                 goto err_pdd;
1220         }
1221
1222         mem = kfd_process_device_translate_handle(
1223                 pdd, GET_IDR_HANDLE(args->handle));
1224         if (!mem) {
1225                 ret = -EINVAL;
1226                 goto err_unlock;
1227         }
1228
1229         ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1230                                 (struct kgd_mem *)mem, pdd->drm_priv, &size);
1231
1232         /* If freeing the buffer failed, leave the handle in place for
1233          * clean-up during process tear-down.
1234          */
1235         if (!ret)
1236                 kfd_process_device_remove_obj_handle(
1237                         pdd, GET_IDR_HANDLE(args->handle));
1238
1239         WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1240
1241 err_unlock:
1242 err_pdd:
1243         mutex_unlock(&p->mutex);
1244         return ret;
1245 }
1246
1247 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1248                                         struct kfd_process *p, void *data)
1249 {
1250         struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1251         struct kfd_process_device *pdd, *peer_pdd;
1252         void *mem;
1253         struct kfd_node *dev;
1254         long err = 0;
1255         int i;
1256         uint32_t *devices_arr = NULL;
1257
1258         if (!args->n_devices) {
1259                 pr_debug("Device IDs array empty\n");
1260                 return -EINVAL;
1261         }
1262         if (args->n_success > args->n_devices) {
1263                 pr_debug("n_success exceeds n_devices\n");
1264                 return -EINVAL;
1265         }
1266
1267         devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1268                                     GFP_KERNEL);
1269         if (!devices_arr)
1270                 return -ENOMEM;
1271
1272         err = copy_from_user(devices_arr,
1273                              (void __user *)args->device_ids_array_ptr,
1274                              args->n_devices * sizeof(*devices_arr));
1275         if (err != 0) {
1276                 err = -EFAULT;
1277                 goto copy_from_user_failed;
1278         }
1279
1280         mutex_lock(&p->mutex);
1281         pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1282         if (!pdd) {
1283                 err = -EINVAL;
1284                 goto get_process_device_data_failed;
1285         }
1286         dev = pdd->dev;
1287
1288         pdd = kfd_bind_process_to_device(dev, p);
1289         if (IS_ERR(pdd)) {
1290                 err = PTR_ERR(pdd);
1291                 goto bind_process_to_device_failed;
1292         }
1293
1294         mem = kfd_process_device_translate_handle(pdd,
1295                                                 GET_IDR_HANDLE(args->handle));
1296         if (!mem) {
1297                 err = -ENOMEM;
1298                 goto get_mem_obj_from_handle_failed;
1299         }
1300
1301         for (i = args->n_success; i < args->n_devices; i++) {
1302                 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1303                 if (!peer_pdd) {
1304                         pr_debug("Getting device by id failed for 0x%x\n",
1305                                  devices_arr[i]);
1306                         err = -EINVAL;
1307                         goto get_mem_obj_from_handle_failed;
1308                 }
1309
1310                 peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1311                 if (IS_ERR(peer_pdd)) {
1312                         err = PTR_ERR(peer_pdd);
1313                         goto get_mem_obj_from_handle_failed;
1314                 }
1315
1316                 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1317                         peer_pdd->dev->adev, (struct kgd_mem *)mem,
1318                         peer_pdd->drm_priv);
1319                 if (err) {
1320                         struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1321
1322                         dev_err(dev->adev->dev,
1323                                "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1324                                pci_domain_nr(pdev->bus),
1325                                pdev->bus->number,
1326                                PCI_SLOT(pdev->devfn),
1327                                PCI_FUNC(pdev->devfn),
1328                                ((struct kgd_mem *)mem)->domain);
1329                         goto map_memory_to_gpu_failed;
1330                 }
1331                 args->n_success = i+1;
1332         }
1333
1334         err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1335         if (err) {
1336                 pr_debug("Sync memory failed, wait interrupted by user signal\n");
1337                 goto sync_memory_failed;
1338         }
1339
1340         mutex_unlock(&p->mutex);
1341
1342         /* Flush TLBs after waiting for the page table updates to complete */
1343         for (i = 0; i < args->n_devices; i++) {
1344                 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1345                 if (WARN_ON_ONCE(!peer_pdd))
1346                         continue;
1347                 kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1348         }
1349         kfree(devices_arr);
1350
1351         return err;
1352
1353 get_process_device_data_failed:
1354 bind_process_to_device_failed:
1355 get_mem_obj_from_handle_failed:
1356 map_memory_to_gpu_failed:
1357 sync_memory_failed:
1358         mutex_unlock(&p->mutex);
1359 copy_from_user_failed:
1360         kfree(devices_arr);
1361
1362         return err;
1363 }
1364
1365 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1366                                         struct kfd_process *p, void *data)
1367 {
1368         struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1369         struct kfd_process_device *pdd, *peer_pdd;
1370         void *mem;
1371         long err = 0;
1372         uint32_t *devices_arr = NULL, i;
1373         bool flush_tlb;
1374
1375         if (!args->n_devices) {
1376                 pr_debug("Device IDs array empty\n");
1377                 return -EINVAL;
1378         }
1379         if (args->n_success > args->n_devices) {
1380                 pr_debug("n_success exceeds n_devices\n");
1381                 return -EINVAL;
1382         }
1383
1384         devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1385                                     GFP_KERNEL);
1386         if (!devices_arr)
1387                 return -ENOMEM;
1388
1389         err = copy_from_user(devices_arr,
1390                              (void __user *)args->device_ids_array_ptr,
1391                              args->n_devices * sizeof(*devices_arr));
1392         if (err != 0) {
1393                 err = -EFAULT;
1394                 goto copy_from_user_failed;
1395         }
1396
1397         mutex_lock(&p->mutex);
1398         pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1399         if (!pdd) {
1400                 err = -EINVAL;
1401                 goto bind_process_to_device_failed;
1402         }
1403
1404         mem = kfd_process_device_translate_handle(pdd,
1405                                                 GET_IDR_HANDLE(args->handle));
1406         if (!mem) {
1407                 err = -ENOMEM;
1408                 goto get_mem_obj_from_handle_failed;
1409         }
1410
1411         for (i = args->n_success; i < args->n_devices; i++) {
1412                 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1413                 if (!peer_pdd) {
1414                         err = -EINVAL;
1415                         goto get_mem_obj_from_handle_failed;
1416                 }
1417                 err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1418                         peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1419                 if (err) {
1420                         pr_err("Failed to unmap from gpu %d/%d\n",
1421                                i, args->n_devices);
1422                         goto unmap_memory_from_gpu_failed;
1423                 }
1424                 args->n_success = i+1;
1425         }
1426
1427         flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1428         if (flush_tlb) {
1429                 err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1430                                 (struct kgd_mem *) mem, true);
1431                 if (err) {
1432                         pr_debug("Sync memory failed, wait interrupted by user signal\n");
1433                         goto sync_memory_failed;
1434                 }
1435         }
1436         mutex_unlock(&p->mutex);
1437
1438         if (flush_tlb) {
1439                 /* Flush TLBs after waiting for the page table updates to complete */
1440                 for (i = 0; i < args->n_devices; i++) {
1441                         peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1442                         if (WARN_ON_ONCE(!peer_pdd))
1443                                 continue;
1444                         kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1445                 }
1446         }
1447         kfree(devices_arr);
1448
1449         return 0;
1450
1451 bind_process_to_device_failed:
1452 get_mem_obj_from_handle_failed:
1453 unmap_memory_from_gpu_failed:
1454 sync_memory_failed:
1455         mutex_unlock(&p->mutex);
1456 copy_from_user_failed:
1457         kfree(devices_arr);
1458         return err;
1459 }
1460
1461 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1462                 struct kfd_process *p, void *data)
1463 {
1464         int retval;
1465         struct kfd_ioctl_alloc_queue_gws_args *args = data;
1466         struct queue *q;
1467         struct kfd_node *dev;
1468
1469         mutex_lock(&p->mutex);
1470         q = pqm_get_user_queue(&p->pqm, args->queue_id);
1471
1472         if (q) {
1473                 dev = q->device;
1474         } else {
1475                 retval = -EINVAL;
1476                 goto out_unlock;
1477         }
1478
1479         if (!dev->gws) {
1480                 retval = -ENODEV;
1481                 goto out_unlock;
1482         }
1483
1484         if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1485                 retval = -ENODEV;
1486                 goto out_unlock;
1487         }
1488
1489         if (!kfd_dbg_has_gws_support(dev) && p->debug_trap_enabled) {
1490                 retval = -EBUSY;
1491                 goto out_unlock;
1492         }
1493
1494         retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
1495         mutex_unlock(&p->mutex);
1496
1497         args->first_gws = 0;
1498         return retval;
1499
1500 out_unlock:
1501         mutex_unlock(&p->mutex);
1502         return retval;
1503 }
1504
1505 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
1506                 struct kfd_process *p, void *data)
1507 {
1508         struct kfd_ioctl_get_dmabuf_info_args *args = data;
1509         struct kfd_node *dev = NULL;
1510         struct amdgpu_device *dmabuf_adev;
1511         void *metadata_buffer = NULL;
1512         uint32_t flags;
1513         int8_t xcp_id;
1514         unsigned int i;
1515         int r;
1516
1517         /* Find a KFD GPU device that supports the get_dmabuf_info query */
1518         for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++)
1519                 if (dev)
1520                         break;
1521         if (!dev)
1522                 return -EINVAL;
1523
1524         if (args->metadata_ptr) {
1525                 metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL);
1526                 if (!metadata_buffer)
1527                         return -ENOMEM;
1528         }
1529
1530         /* Get dmabuf info from KGD */
1531         r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd,
1532                                           &dmabuf_adev, &args->size,
1533                                           metadata_buffer, args->metadata_size,
1534                                           &args->metadata_size, &flags, &xcp_id);
1535         if (r)
1536                 goto exit;
1537
1538         if (xcp_id >= 0)
1539                 args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id;
1540         else
1541                 args->gpu_id = dmabuf_adev->kfd.dev->nodes[0]->id;
1542         args->flags = flags;
1543
1544         /* Copy metadata buffer to user mode */
1545         if (metadata_buffer) {
1546                 r = copy_to_user((void __user *)args->metadata_ptr,
1547                                  metadata_buffer, args->metadata_size);
1548                 if (r != 0)
1549                         r = -EFAULT;
1550         }
1551
1552 exit:
1553         kfree(metadata_buffer);
1554
1555         return r;
1556 }
1557
1558 static int kfd_ioctl_import_dmabuf(struct file *filep,
1559                                    struct kfd_process *p, void *data)
1560 {
1561         struct kfd_ioctl_import_dmabuf_args *args = data;
1562         struct kfd_process_device *pdd;
1563         struct dma_buf *dmabuf;
1564         int idr_handle;
1565         uint64_t size;
1566         void *mem;
1567         int r;
1568
1569         dmabuf = dma_buf_get(args->dmabuf_fd);
1570         if (IS_ERR(dmabuf))
1571                 return PTR_ERR(dmabuf);
1572
1573         mutex_lock(&p->mutex);
1574         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1575         if (!pdd) {
1576                 r = -EINVAL;
1577                 goto err_unlock;
1578         }
1579
1580         pdd = kfd_bind_process_to_device(pdd->dev, p);
1581         if (IS_ERR(pdd)) {
1582                 r = PTR_ERR(pdd);
1583                 goto err_unlock;
1584         }
1585
1586         r = amdgpu_amdkfd_gpuvm_import_dmabuf(pdd->dev->adev, dmabuf,
1587                                               args->va_addr, pdd->drm_priv,
1588                                               (struct kgd_mem **)&mem, &size,
1589                                               NULL);
1590         if (r)
1591                 goto err_unlock;
1592
1593         idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1594         if (idr_handle < 0) {
1595                 r = -EFAULT;
1596                 goto err_free;
1597         }
1598
1599         mutex_unlock(&p->mutex);
1600         dma_buf_put(dmabuf);
1601
1602         args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1603
1604         return 0;
1605
1606 err_free:
1607         amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem,
1608                                                pdd->drm_priv, NULL);
1609 err_unlock:
1610         mutex_unlock(&p->mutex);
1611         dma_buf_put(dmabuf);
1612         return r;
1613 }
1614
1615 static int kfd_ioctl_export_dmabuf(struct file *filep,
1616                                    struct kfd_process *p, void *data)
1617 {
1618         struct kfd_ioctl_export_dmabuf_args *args = data;
1619         struct kfd_process_device *pdd;
1620         struct dma_buf *dmabuf;
1621         struct kfd_node *dev;
1622         void *mem;
1623         int ret = 0;
1624
1625         dev = kfd_device_by_id(GET_GPU_ID(args->handle));
1626         if (!dev)
1627                 return -EINVAL;
1628
1629         mutex_lock(&p->mutex);
1630
1631         pdd = kfd_get_process_device_data(dev, p);
1632         if (!pdd) {
1633                 ret = -EINVAL;
1634                 goto err_unlock;
1635         }
1636
1637         mem = kfd_process_device_translate_handle(pdd,
1638                                                 GET_IDR_HANDLE(args->handle));
1639         if (!mem) {
1640                 ret = -EINVAL;
1641                 goto err_unlock;
1642         }
1643
1644         ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1645         mutex_unlock(&p->mutex);
1646         if (ret)
1647                 goto err_out;
1648
1649         ret = dma_buf_fd(dmabuf, args->flags);
1650         if (ret < 0) {
1651                 dma_buf_put(dmabuf);
1652                 goto err_out;
1653         }
1654         /* dma_buf_fd assigns the reference count to the fd, no need to
1655          * put the reference here.
1656          */
1657         args->dmabuf_fd = ret;
1658
1659         return 0;
1660
1661 err_unlock:
1662         mutex_unlock(&p->mutex);
1663 err_out:
1664         return ret;
1665 }
1666
1667 /* Handle requests for watching SMI events */
1668 static int kfd_ioctl_smi_events(struct file *filep,
1669                                 struct kfd_process *p, void *data)
1670 {
1671         struct kfd_ioctl_smi_events_args *args = data;
1672         struct kfd_process_device *pdd;
1673
1674         mutex_lock(&p->mutex);
1675
1676         pdd = kfd_process_device_data_by_id(p, args->gpuid);
1677         mutex_unlock(&p->mutex);
1678         if (!pdd)
1679                 return -EINVAL;
1680
1681         return kfd_smi_event_open(pdd->dev, &args->anon_fd);
1682 }
1683
1684 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1685
1686 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1687                                     struct kfd_process *p, void *data)
1688 {
1689         struct kfd_ioctl_set_xnack_mode_args *args = data;
1690         int r = 0;
1691
1692         mutex_lock(&p->mutex);
1693         if (args->xnack_enabled >= 0) {
1694                 if (!list_empty(&p->pqm.queues)) {
1695                         pr_debug("Process has user queues running\n");
1696                         r = -EBUSY;
1697                         goto out_unlock;
1698                 }
1699
1700                 if (p->xnack_enabled == args->xnack_enabled)
1701                         goto out_unlock;
1702
1703                 if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) {
1704                         r = -EPERM;
1705                         goto out_unlock;
1706                 }
1707
1708                 r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled);
1709         } else {
1710                 args->xnack_enabled = p->xnack_enabled;
1711         }
1712
1713 out_unlock:
1714         mutex_unlock(&p->mutex);
1715
1716         return r;
1717 }
1718
1719 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1720 {
1721         struct kfd_ioctl_svm_args *args = data;
1722         int r = 0;
1723
1724         pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n",
1725                  args->start_addr, args->size, args->op, args->nattr);
1726
1727         if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK))
1728                 return -EINVAL;
1729         if (!args->start_addr || !args->size)
1730                 return -EINVAL;
1731
1732         r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr,
1733                       args->attrs);
1734
1735         return r;
1736 }
1737 #else
1738 static int kfd_ioctl_set_xnack_mode(struct file *filep,
1739                                     struct kfd_process *p, void *data)
1740 {
1741         return -EPERM;
1742 }
1743 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
1744 {
1745         return -EPERM;
1746 }
1747 #endif
1748
1749 static int criu_checkpoint_process(struct kfd_process *p,
1750                              uint8_t __user *user_priv_data,
1751                              uint64_t *priv_offset)
1752 {
1753         struct kfd_criu_process_priv_data process_priv;
1754         int ret;
1755
1756         memset(&process_priv, 0, sizeof(process_priv));
1757
1758         process_priv.version = KFD_CRIU_PRIV_VERSION;
1759         /* For CR, we don't consider negative xnack mode which is used for
1760          * querying without changing it, here 0 simply means disabled and 1
1761          * means enabled so retry for finding a valid PTE.
1762          */
1763         process_priv.xnack_mode = p->xnack_enabled ? 1 : 0;
1764
1765         ret = copy_to_user(user_priv_data + *priv_offset,
1766                                 &process_priv, sizeof(process_priv));
1767
1768         if (ret) {
1769                 pr_err("Failed to copy process information to user\n");
1770                 ret = -EFAULT;
1771         }
1772
1773         *priv_offset += sizeof(process_priv);
1774         return ret;
1775 }
1776
1777 static int criu_checkpoint_devices(struct kfd_process *p,
1778                              uint32_t num_devices,
1779                              uint8_t __user *user_addr,
1780                              uint8_t __user *user_priv_data,
1781                              uint64_t *priv_offset)
1782 {
1783         struct kfd_criu_device_priv_data *device_priv = NULL;
1784         struct kfd_criu_device_bucket *device_buckets = NULL;
1785         int ret = 0, i;
1786
1787         device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL);
1788         if (!device_buckets) {
1789                 ret = -ENOMEM;
1790                 goto exit;
1791         }
1792
1793         device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL);
1794         if (!device_priv) {
1795                 ret = -ENOMEM;
1796                 goto exit;
1797         }
1798
1799         for (i = 0; i < num_devices; i++) {
1800                 struct kfd_process_device *pdd = p->pdds[i];
1801
1802                 device_buckets[i].user_gpu_id = pdd->user_gpu_id;
1803                 device_buckets[i].actual_gpu_id = pdd->dev->id;
1804
1805                 /*
1806                  * priv_data does not contain useful information for now and is reserved for
1807                  * future use, so we do not set its contents.
1808                  */
1809         }
1810
1811         ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets));
1812         if (ret) {
1813                 pr_err("Failed to copy device information to user\n");
1814                 ret = -EFAULT;
1815                 goto exit;
1816         }
1817
1818         ret = copy_to_user(user_priv_data + *priv_offset,
1819                            device_priv,
1820                            num_devices * sizeof(*device_priv));
1821         if (ret) {
1822                 pr_err("Failed to copy device information to user\n");
1823                 ret = -EFAULT;
1824         }
1825         *priv_offset += num_devices * sizeof(*device_priv);
1826
1827 exit:
1828         kvfree(device_buckets);
1829         kvfree(device_priv);
1830         return ret;
1831 }
1832
1833 static uint32_t get_process_num_bos(struct kfd_process *p)
1834 {
1835         uint32_t num_of_bos = 0;
1836         int i;
1837
1838         /* Run over all PDDs of the process */
1839         for (i = 0; i < p->n_pdds; i++) {
1840                 struct kfd_process_device *pdd = p->pdds[i];
1841                 void *mem;
1842                 int id;
1843
1844                 idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1845                         struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
1846
1847                         if ((uint64_t)kgd_mem->va > pdd->gpuvm_base)
1848                                 num_of_bos++;
1849                 }
1850         }
1851         return num_of_bos;
1852 }
1853
1854 static int criu_get_prime_handle(struct drm_gem_object *gobj, int flags,
1855                                       u32 *shared_fd)
1856 {
1857         struct dma_buf *dmabuf;
1858         int ret;
1859
1860         dmabuf = amdgpu_gem_prime_export(gobj, flags);
1861         if (IS_ERR(dmabuf)) {
1862                 ret = PTR_ERR(dmabuf);
1863                 pr_err("dmabuf export failed for the BO\n");
1864                 return ret;
1865         }
1866
1867         ret = dma_buf_fd(dmabuf, flags);
1868         if (ret < 0) {
1869                 pr_err("dmabuf create fd failed, ret:%d\n", ret);
1870                 goto out_free_dmabuf;
1871         }
1872
1873         *shared_fd = ret;
1874         return 0;
1875
1876 out_free_dmabuf:
1877         dma_buf_put(dmabuf);
1878         return ret;
1879 }
1880
1881 static int criu_checkpoint_bos(struct kfd_process *p,
1882                                uint32_t num_bos,
1883                                uint8_t __user *user_bos,
1884                                uint8_t __user *user_priv_data,
1885                                uint64_t *priv_offset)
1886 {
1887         struct kfd_criu_bo_bucket *bo_buckets;
1888         struct kfd_criu_bo_priv_data *bo_privs;
1889         int ret = 0, pdd_index, bo_index = 0, id;
1890         void *mem;
1891
1892         bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1893         if (!bo_buckets)
1894                 return -ENOMEM;
1895
1896         bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1897         if (!bo_privs) {
1898                 ret = -ENOMEM;
1899                 goto exit;
1900         }
1901
1902         for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1903                 struct kfd_process_device *pdd = p->pdds[pdd_index];
1904                 struct amdgpu_bo *dumper_bo;
1905                 struct kgd_mem *kgd_mem;
1906
1907                 idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1908                         struct kfd_criu_bo_bucket *bo_bucket;
1909                         struct kfd_criu_bo_priv_data *bo_priv;
1910                         int i, dev_idx = 0;
1911
1912                         if (!mem) {
1913                                 ret = -ENOMEM;
1914                                 goto exit;
1915                         }
1916
1917                         kgd_mem = (struct kgd_mem *)mem;
1918                         dumper_bo = kgd_mem->bo;
1919
1920                         if ((uint64_t)kgd_mem->va <= pdd->gpuvm_base)
1921                                 continue;
1922
1923                         bo_bucket = &bo_buckets[bo_index];
1924                         bo_priv = &bo_privs[bo_index];
1925
1926                         bo_bucket->gpu_id = pdd->user_gpu_id;
1927                         bo_bucket->addr = (uint64_t)kgd_mem->va;
1928                         bo_bucket->size = amdgpu_bo_size(dumper_bo);
1929                         bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1930                         bo_priv->idr_handle = id;
1931
1932                         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1933                                 ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1934                                                                 &bo_priv->user_addr);
1935                                 if (ret) {
1936                                         pr_err("Failed to obtain user address for user-pointer bo\n");
1937                                         goto exit;
1938                                 }
1939                         }
1940                         if (bo_bucket->alloc_flags
1941                             & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1942                                 ret = criu_get_prime_handle(&dumper_bo->tbo.base,
1943                                                 bo_bucket->alloc_flags &
1944                                                 KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1945                                                 &bo_bucket->dmabuf_fd);
1946                                 if (ret)
1947                                         goto exit;
1948                         } else {
1949                                 bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1950                         }
1951
1952                         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1953                                 bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1954                                         KFD_MMAP_GPU_ID(pdd->dev->id);
1955                         else if (bo_bucket->alloc_flags &
1956                                 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1957                                 bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1958                                         KFD_MMAP_GPU_ID(pdd->dev->id);
1959                         else
1960                                 bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1961
1962                         for (i = 0; i < p->n_pdds; i++) {
1963                                 if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->dev->adev, kgd_mem))
1964                                         bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1965                         }
1966
1967                         pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1968                                         "gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1969                                         bo_bucket->size,
1970                                         bo_bucket->addr,
1971                                         bo_bucket->offset,
1972                                         bo_bucket->gpu_id,
1973                                         bo_bucket->alloc_flags,
1974                                         bo_priv->idr_handle);
1975                         bo_index++;
1976                 }
1977         }
1978
1979         ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1980         if (ret) {
1981                 pr_err("Failed to copy BO information to user\n");
1982                 ret = -EFAULT;
1983                 goto exit;
1984         }
1985
1986         ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1987         if (ret) {
1988                 pr_err("Failed to copy BO priv information to user\n");
1989                 ret = -EFAULT;
1990                 goto exit;
1991         }
1992
1993         *priv_offset += num_bos * sizeof(*bo_privs);
1994
1995 exit:
1996         while (ret && bo_index--) {
1997                 if (bo_buckets[bo_index].alloc_flags
1998                     & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
1999                         close_fd(bo_buckets[bo_index].dmabuf_fd);
2000         }
2001
2002         kvfree(bo_buckets);
2003         kvfree(bo_privs);
2004         return ret;
2005 }
2006
2007 static int criu_get_process_object_info(struct kfd_process *p,
2008                                         uint32_t *num_devices,
2009                                         uint32_t *num_bos,
2010                                         uint32_t *num_objects,
2011                                         uint64_t *objs_priv_size)
2012 {
2013         uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2014         uint32_t num_queues, num_events, num_svm_ranges;
2015         int ret;
2016
2017         *num_devices = p->n_pdds;
2018         *num_bos = get_process_num_bos(p);
2019
2020         ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2021         if (ret)
2022                 return ret;
2023
2024         num_events = kfd_get_num_events(p);
2025
2026         ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2027         if (ret)
2028                 return ret;
2029
2030         *num_objects = num_queues + num_events + num_svm_ranges;
2031
2032         if (objs_priv_size) {
2033                 priv_size = sizeof(struct kfd_criu_process_priv_data);
2034                 priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2035                 priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2036                 priv_size += queues_priv_data_size;
2037                 priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2038                 priv_size += svm_priv_data_size;
2039                 *objs_priv_size = priv_size;
2040         }
2041         return 0;
2042 }
2043
2044 static int criu_checkpoint(struct file *filep,
2045                            struct kfd_process *p,
2046                            struct kfd_ioctl_criu_args *args)
2047 {
2048         int ret;
2049         uint32_t num_devices, num_bos, num_objects;
2050         uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2051
2052         if (!args->devices || !args->bos || !args->priv_data)
2053                 return -EINVAL;
2054
2055         mutex_lock(&p->mutex);
2056
2057         if (!p->n_pdds) {
2058                 pr_err("No pdd for given process\n");
2059                 ret = -ENODEV;
2060                 goto exit_unlock;
2061         }
2062
2063         /* Confirm all process queues are evicted */
2064         if (!p->queues_paused) {
2065                 pr_err("Cannot dump process when queues are not in evicted state\n");
2066                 /* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2067                 ret = -EINVAL;
2068                 goto exit_unlock;
2069         }
2070
2071         ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2072         if (ret)
2073                 goto exit_unlock;
2074
2075         if (num_devices != args->num_devices ||
2076             num_bos != args->num_bos ||
2077             num_objects != args->num_objects ||
2078             priv_size != args->priv_data_size) {
2079
2080                 ret = -EINVAL;
2081                 goto exit_unlock;
2082         }
2083
2084         /* each function will store private data inside priv_data and adjust priv_offset */
2085         ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2086         if (ret)
2087                 goto exit_unlock;
2088
2089         ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2090                                 (uint8_t __user *)args->priv_data, &priv_offset);
2091         if (ret)
2092                 goto exit_unlock;
2093
2094         /* Leave room for BOs in the private data. They need to be restored
2095          * before events, but we checkpoint them last to simplify the error
2096          * handling.
2097          */
2098         bo_priv_offset = priv_offset;
2099         priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2100
2101         if (num_objects) {
2102                 ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2103                                                  &priv_offset);
2104                 if (ret)
2105                         goto exit_unlock;
2106
2107                 ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2108                                                  &priv_offset);
2109                 if (ret)
2110                         goto exit_unlock;
2111
2112                 ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2113                 if (ret)
2114                         goto exit_unlock;
2115         }
2116
2117         /* This must be the last thing in this function that can fail.
2118          * Otherwise we leak dmabuf file descriptors.
2119          */
2120         ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2121                            (uint8_t __user *)args->priv_data, &bo_priv_offset);
2122
2123 exit_unlock:
2124         mutex_unlock(&p->mutex);
2125         if (ret)
2126                 pr_err("Failed to dump CRIU ret:%d\n", ret);
2127         else
2128                 pr_debug("CRIU dump ret:%d\n", ret);
2129
2130         return ret;
2131 }
2132
2133 static int criu_restore_process(struct kfd_process *p,
2134                                 struct kfd_ioctl_criu_args *args,
2135                                 uint64_t *priv_offset,
2136                                 uint64_t max_priv_data_size)
2137 {
2138         int ret = 0;
2139         struct kfd_criu_process_priv_data process_priv;
2140
2141         if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2142                 return -EINVAL;
2143
2144         ret = copy_from_user(&process_priv,
2145                                 (void __user *)(args->priv_data + *priv_offset),
2146                                 sizeof(process_priv));
2147         if (ret) {
2148                 pr_err("Failed to copy process private information from user\n");
2149                 ret = -EFAULT;
2150                 goto exit;
2151         }
2152         *priv_offset += sizeof(process_priv);
2153
2154         if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2155                 pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2156                         process_priv.version, KFD_CRIU_PRIV_VERSION);
2157                 return -EINVAL;
2158         }
2159
2160         pr_debug("Setting XNACK mode\n");
2161         if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2162                 pr_err("xnack mode cannot be set\n");
2163                 ret = -EPERM;
2164                 goto exit;
2165         } else {
2166                 pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2167                 p->xnack_enabled = process_priv.xnack_mode;
2168         }
2169
2170 exit:
2171         return ret;
2172 }
2173
2174 static int criu_restore_devices(struct kfd_process *p,
2175                                 struct kfd_ioctl_criu_args *args,
2176                                 uint64_t *priv_offset,
2177                                 uint64_t max_priv_data_size)
2178 {
2179         struct kfd_criu_device_bucket *device_buckets;
2180         struct kfd_criu_device_priv_data *device_privs;
2181         int ret = 0;
2182         uint32_t i;
2183
2184         if (args->num_devices != p->n_pdds)
2185                 return -EINVAL;
2186
2187         if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2188                 return -EINVAL;
2189
2190         device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2191         if (!device_buckets)
2192                 return -ENOMEM;
2193
2194         ret = copy_from_user(device_buckets, (void __user *)args->devices,
2195                                 args->num_devices * sizeof(*device_buckets));
2196         if (ret) {
2197                 pr_err("Failed to copy devices buckets from user\n");
2198                 ret = -EFAULT;
2199                 goto exit;
2200         }
2201
2202         for (i = 0; i < args->num_devices; i++) {
2203                 struct kfd_node *dev;
2204                 struct kfd_process_device *pdd;
2205                 struct file *drm_file;
2206
2207                 /* device private data is not currently used */
2208
2209                 if (!device_buckets[i].user_gpu_id) {
2210                         pr_err("Invalid user gpu_id\n");
2211                         ret = -EINVAL;
2212                         goto exit;
2213                 }
2214
2215                 dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2216                 if (!dev) {
2217                         pr_err("Failed to find device with gpu_id = %x\n",
2218                                 device_buckets[i].actual_gpu_id);
2219                         ret = -EINVAL;
2220                         goto exit;
2221                 }
2222
2223                 pdd = kfd_get_process_device_data(dev, p);
2224                 if (!pdd) {
2225                         pr_err("Failed to get pdd for gpu_id = %x\n",
2226                                         device_buckets[i].actual_gpu_id);
2227                         ret = -EINVAL;
2228                         goto exit;
2229                 }
2230                 pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2231
2232                 drm_file = fget(device_buckets[i].drm_fd);
2233                 if (!drm_file) {
2234                         pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2235                                 device_buckets[i].drm_fd);
2236                         ret = -EINVAL;
2237                         goto exit;
2238                 }
2239
2240                 if (pdd->drm_file) {
2241                         ret = -EINVAL;
2242                         goto exit;
2243                 }
2244
2245                 /* create the vm using render nodes for kfd pdd */
2246                 if (kfd_process_device_init_vm(pdd, drm_file)) {
2247                         pr_err("could not init vm for given pdd\n");
2248                         /* On success, the PDD keeps the drm_file reference */
2249                         fput(drm_file);
2250                         ret = -EINVAL;
2251                         goto exit;
2252                 }
2253                 /*
2254                  * pdd now already has the vm bound to render node so below api won't create a new
2255                  * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2256                  * for iommu v2 binding  and runtime pm.
2257                  */
2258                 pdd = kfd_bind_process_to_device(dev, p);
2259                 if (IS_ERR(pdd)) {
2260                         ret = PTR_ERR(pdd);
2261                         goto exit;
2262                 }
2263
2264                 if (!pdd->doorbell_index &&
2265                     kfd_alloc_process_doorbells(pdd->dev->kfd, &pdd->doorbell_index) < 0) {
2266                         ret = -ENOMEM;
2267                         goto exit;
2268                 }
2269         }
2270
2271         /*
2272          * We are not copying device private data from user as we are not using the data for now,
2273          * but we still adjust for its private data.
2274          */
2275         *priv_offset += args->num_devices * sizeof(*device_privs);
2276
2277 exit:
2278         kfree(device_buckets);
2279         return ret;
2280 }
2281
2282 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2283                                       struct kfd_criu_bo_bucket *bo_bucket,
2284                                       struct kfd_criu_bo_priv_data *bo_priv,
2285                                       struct kgd_mem **kgd_mem)
2286 {
2287         int idr_handle;
2288         int ret;
2289         const bool criu_resume = true;
2290         u64 offset;
2291
2292         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2293                 if (bo_bucket->size !=
2294                                 kfd_doorbell_process_slice(pdd->dev->kfd))
2295                         return -EINVAL;
2296
2297                 offset = kfd_get_process_doorbells(pdd);
2298                 if (!offset)
2299                         return -ENOMEM;
2300         } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2301                 /* MMIO BOs need remapped bus address */
2302                 if (bo_bucket->size != PAGE_SIZE) {
2303                         pr_err("Invalid page size\n");
2304                         return -EINVAL;
2305                 }
2306                 offset = pdd->dev->adev->rmmio_remap.bus_addr;
2307                 if (!offset) {
2308                         pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2309                         return -ENOMEM;
2310                 }
2311         } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2312                 offset = bo_priv->user_addr;
2313         }
2314         /* Create the BO */
2315         ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2316                                                       bo_bucket->size, pdd->drm_priv, kgd_mem,
2317                                                       &offset, bo_bucket->alloc_flags, criu_resume);
2318         if (ret) {
2319                 pr_err("Could not create the BO\n");
2320                 return ret;
2321         }
2322         pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2323                  bo_bucket->size, bo_bucket->addr, offset);
2324
2325         /* Restore previous IDR handle */
2326         pr_debug("Restoring old IDR handle for the BO");
2327         idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2328                                bo_priv->idr_handle + 1, GFP_KERNEL);
2329
2330         if (idr_handle < 0) {
2331                 pr_err("Could not allocate idr\n");
2332                 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2333                                                        NULL);
2334                 return -ENOMEM;
2335         }
2336
2337         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2338                 bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2339         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2340                 bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2341         } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2342                 bo_bucket->restored_offset = offset;
2343         } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2344                 bo_bucket->restored_offset = offset;
2345                 /* Update the VRAM usage count */
2346                 WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2347         }
2348         return 0;
2349 }
2350
2351 static int criu_restore_bo(struct kfd_process *p,
2352                            struct kfd_criu_bo_bucket *bo_bucket,
2353                            struct kfd_criu_bo_priv_data *bo_priv)
2354 {
2355         struct kfd_process_device *pdd;
2356         struct kgd_mem *kgd_mem;
2357         int ret;
2358         int j;
2359
2360         pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2361                  bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2362                  bo_priv->idr_handle);
2363
2364         pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2365         if (!pdd) {
2366                 pr_err("Failed to get pdd\n");
2367                 return -ENODEV;
2368         }
2369
2370         ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2371         if (ret)
2372                 return ret;
2373
2374         /* now map these BOs to GPU/s */
2375         for (j = 0; j < p->n_pdds; j++) {
2376                 struct kfd_node *peer;
2377                 struct kfd_process_device *peer_pdd;
2378
2379                 if (!bo_priv->mapped_gpuids[j])
2380                         break;
2381
2382                 peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2383                 if (!peer_pdd)
2384                         return -EINVAL;
2385
2386                 peer = peer_pdd->dev;
2387
2388                 peer_pdd = kfd_bind_process_to_device(peer, p);
2389                 if (IS_ERR(peer_pdd))
2390                         return PTR_ERR(peer_pdd);
2391
2392                 ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2393                                                             peer_pdd->drm_priv);
2394                 if (ret) {
2395                         pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2396                         return ret;
2397                 }
2398         }
2399
2400         pr_debug("map memory was successful for the BO\n");
2401         /* create the dmabuf object and export the bo */
2402         if (bo_bucket->alloc_flags
2403             & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2404                 ret = criu_get_prime_handle(&kgd_mem->bo->tbo.base, DRM_RDWR,
2405                                             &bo_bucket->dmabuf_fd);
2406                 if (ret)
2407                         return ret;
2408         } else {
2409                 bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2410         }
2411
2412         return 0;
2413 }
2414
2415 static int criu_restore_bos(struct kfd_process *p,
2416                             struct kfd_ioctl_criu_args *args,
2417                             uint64_t *priv_offset,
2418                             uint64_t max_priv_data_size)
2419 {
2420         struct kfd_criu_bo_bucket *bo_buckets = NULL;
2421         struct kfd_criu_bo_priv_data *bo_privs = NULL;
2422         int ret = 0;
2423         uint32_t i = 0;
2424
2425         if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2426                 return -EINVAL;
2427
2428         /* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2429         amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2430
2431         bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2432         if (!bo_buckets)
2433                 return -ENOMEM;
2434
2435         ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2436                              args->num_bos * sizeof(*bo_buckets));
2437         if (ret) {
2438                 pr_err("Failed to copy BOs information from user\n");
2439                 ret = -EFAULT;
2440                 goto exit;
2441         }
2442
2443         bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2444         if (!bo_privs) {
2445                 ret = -ENOMEM;
2446                 goto exit;
2447         }
2448
2449         ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2450                              args->num_bos * sizeof(*bo_privs));
2451         if (ret) {
2452                 pr_err("Failed to copy BOs information from user\n");
2453                 ret = -EFAULT;
2454                 goto exit;
2455         }
2456         *priv_offset += args->num_bos * sizeof(*bo_privs);
2457
2458         /* Create and map new BOs */
2459         for (; i < args->num_bos; i++) {
2460                 ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2461                 if (ret) {
2462                         pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2463                         goto exit;
2464                 }
2465         } /* done */
2466
2467         /* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2468         ret = copy_to_user((void __user *)args->bos,
2469                                 bo_buckets,
2470                                 (args->num_bos * sizeof(*bo_buckets)));
2471         if (ret)
2472                 ret = -EFAULT;
2473
2474 exit:
2475         while (ret && i--) {
2476                 if (bo_buckets[i].alloc_flags
2477                    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2478                         close_fd(bo_buckets[i].dmabuf_fd);
2479         }
2480         kvfree(bo_buckets);
2481         kvfree(bo_privs);
2482         return ret;
2483 }
2484
2485 static int criu_restore_objects(struct file *filep,
2486                                 struct kfd_process *p,
2487                                 struct kfd_ioctl_criu_args *args,
2488                                 uint64_t *priv_offset,
2489                                 uint64_t max_priv_data_size)
2490 {
2491         int ret = 0;
2492         uint32_t i;
2493
2494         BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2495         BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2496         BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2497
2498         for (i = 0; i < args->num_objects; i++) {
2499                 uint32_t object_type;
2500
2501                 if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2502                         pr_err("Invalid private data size\n");
2503                         return -EINVAL;
2504                 }
2505
2506                 ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2507                 if (ret) {
2508                         pr_err("Failed to copy private information from user\n");
2509                         goto exit;
2510                 }
2511
2512                 switch (object_type) {
2513                 case KFD_CRIU_OBJECT_TYPE_QUEUE:
2514                         ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2515                                                      priv_offset, max_priv_data_size);
2516                         if (ret)
2517                                 goto exit;
2518                         break;
2519                 case KFD_CRIU_OBJECT_TYPE_EVENT:
2520                         ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2521                                                      priv_offset, max_priv_data_size);
2522                         if (ret)
2523                                 goto exit;
2524                         break;
2525                 case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2526                         ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2527                                                      priv_offset, max_priv_data_size);
2528                         if (ret)
2529                                 goto exit;
2530                         break;
2531                 default:
2532                         pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2533                         ret = -EINVAL;
2534                         goto exit;
2535                 }
2536         }
2537 exit:
2538         return ret;
2539 }
2540
2541 static int criu_restore(struct file *filep,
2542                         struct kfd_process *p,
2543                         struct kfd_ioctl_criu_args *args)
2544 {
2545         uint64_t priv_offset = 0;
2546         int ret = 0;
2547
2548         pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2549                  args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2550
2551         if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2552             !args->num_devices || !args->num_bos)
2553                 return -EINVAL;
2554
2555         mutex_lock(&p->mutex);
2556
2557         /*
2558          * Set the process to evicted state to avoid running any new queues before all the memory
2559          * mappings are ready.
2560          */
2561         ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2562         if (ret)
2563                 goto exit_unlock;
2564
2565         /* Each function will adjust priv_offset based on how many bytes they consumed */
2566         ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2567         if (ret)
2568                 goto exit_unlock;
2569
2570         ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2571         if (ret)
2572                 goto exit_unlock;
2573
2574         ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2575         if (ret)
2576                 goto exit_unlock;
2577
2578         ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2579         if (ret)
2580                 goto exit_unlock;
2581
2582         if (priv_offset != args->priv_data_size) {
2583                 pr_err("Invalid private data size\n");
2584                 ret = -EINVAL;
2585         }
2586
2587 exit_unlock:
2588         mutex_unlock(&p->mutex);
2589         if (ret)
2590                 pr_err("Failed to restore CRIU ret:%d\n", ret);
2591         else
2592                 pr_debug("CRIU restore successful\n");
2593
2594         return ret;
2595 }
2596
2597 static int criu_unpause(struct file *filep,
2598                         struct kfd_process *p,
2599                         struct kfd_ioctl_criu_args *args)
2600 {
2601         int ret;
2602
2603         mutex_lock(&p->mutex);
2604
2605         if (!p->queues_paused) {
2606                 mutex_unlock(&p->mutex);
2607                 return -EINVAL;
2608         }
2609
2610         ret = kfd_process_restore_queues(p);
2611         if (ret)
2612                 pr_err("Failed to unpause queues ret:%d\n", ret);
2613         else
2614                 p->queues_paused = false;
2615
2616         mutex_unlock(&p->mutex);
2617
2618         return ret;
2619 }
2620
2621 static int criu_resume(struct file *filep,
2622                         struct kfd_process *p,
2623                         struct kfd_ioctl_criu_args *args)
2624 {
2625         struct kfd_process *target = NULL;
2626         struct pid *pid = NULL;
2627         int ret = 0;
2628
2629         pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2630                  args->pid);
2631
2632         pid = find_get_pid(args->pid);
2633         if (!pid) {
2634                 pr_err("Cannot find pid info for %i\n", args->pid);
2635                 return -ESRCH;
2636         }
2637
2638         pr_debug("calling kfd_lookup_process_by_pid\n");
2639         target = kfd_lookup_process_by_pid(pid);
2640
2641         put_pid(pid);
2642
2643         if (!target) {
2644                 pr_debug("Cannot find process info for %i\n", args->pid);
2645                 return -ESRCH;
2646         }
2647
2648         mutex_lock(&target->mutex);
2649         ret = kfd_criu_resume_svm(target);
2650         if (ret) {
2651                 pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2652                 goto exit;
2653         }
2654
2655         ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2656         if (ret)
2657                 pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2658
2659 exit:
2660         mutex_unlock(&target->mutex);
2661
2662         kfd_unref_process(target);
2663         return ret;
2664 }
2665
2666 static int criu_process_info(struct file *filep,
2667                                 struct kfd_process *p,
2668                                 struct kfd_ioctl_criu_args *args)
2669 {
2670         int ret = 0;
2671
2672         mutex_lock(&p->mutex);
2673
2674         if (!p->n_pdds) {
2675                 pr_err("No pdd for given process\n");
2676                 ret = -ENODEV;
2677                 goto err_unlock;
2678         }
2679
2680         ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2681         if (ret)
2682                 goto err_unlock;
2683
2684         p->queues_paused = true;
2685
2686         args->pid = task_pid_nr_ns(p->lead_thread,
2687                                         task_active_pid_ns(p->lead_thread));
2688
2689         ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2690                                            &args->num_objects, &args->priv_data_size);
2691         if (ret)
2692                 goto err_unlock;
2693
2694         dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2695                                 args->num_devices, args->num_bos, args->num_objects,
2696                                 args->priv_data_size);
2697
2698 err_unlock:
2699         if (ret) {
2700                 kfd_process_restore_queues(p);
2701                 p->queues_paused = false;
2702         }
2703         mutex_unlock(&p->mutex);
2704         return ret;
2705 }
2706
2707 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2708 {
2709         struct kfd_ioctl_criu_args *args = data;
2710         int ret;
2711
2712         dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2713         switch (args->op) {
2714         case KFD_CRIU_OP_PROCESS_INFO:
2715                 ret = criu_process_info(filep, p, args);
2716                 break;
2717         case KFD_CRIU_OP_CHECKPOINT:
2718                 ret = criu_checkpoint(filep, p, args);
2719                 break;
2720         case KFD_CRIU_OP_UNPAUSE:
2721                 ret = criu_unpause(filep, p, args);
2722                 break;
2723         case KFD_CRIU_OP_RESTORE:
2724                 ret = criu_restore(filep, p, args);
2725                 break;
2726         case KFD_CRIU_OP_RESUME:
2727                 ret = criu_resume(filep, p, args);
2728                 break;
2729         default:
2730                 dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2731                 ret = -EINVAL;
2732                 break;
2733         }
2734
2735         if (ret)
2736                 dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2737
2738         return ret;
2739 }
2740
2741 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2742                         bool enable_ttmp_setup)
2743 {
2744         int i = 0, ret = 0;
2745
2746         if (p->is_runtime_retry)
2747                 goto retry;
2748
2749         if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2750                 return -EBUSY;
2751
2752         for (i = 0; i < p->n_pdds; i++) {
2753                 struct kfd_process_device *pdd = p->pdds[i];
2754
2755                 if (pdd->qpd.queue_count)
2756                         return -EEXIST;
2757         }
2758
2759         p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2760         p->runtime_info.r_debug = r_debug;
2761         p->runtime_info.ttmp_setup = enable_ttmp_setup;
2762
2763         if (p->runtime_info.ttmp_setup) {
2764                 for (i = 0; i < p->n_pdds; i++) {
2765                         struct kfd_process_device *pdd = p->pdds[i];
2766
2767                         if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2768                                 amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2769                                 pdd->dev->kfd2kgd->enable_debug_trap(
2770                                                 pdd->dev->adev,
2771                                                 true,
2772                                                 pdd->dev->vm_info.last_vmid_kfd);
2773                         } else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2774                                 pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2775                                                 pdd->dev->adev,
2776                                                 false,
2777                                                 0);
2778                         }
2779                 }
2780         }
2781
2782 retry:
2783         if (p->debug_trap_enabled) {
2784                 if (!p->is_runtime_retry) {
2785                         kfd_dbg_trap_activate(p);
2786                         kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2787                                         p, NULL, 0, false, NULL, 0);
2788                 }
2789
2790                 mutex_unlock(&p->mutex);
2791                 ret = down_interruptible(&p->runtime_enable_sema);
2792                 mutex_lock(&p->mutex);
2793
2794                 p->is_runtime_retry = !!ret;
2795         }
2796
2797         return ret;
2798 }
2799
2800 static int runtime_disable(struct kfd_process *p)
2801 {
2802         int i = 0, ret;
2803         bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2804
2805         p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2806         p->runtime_info.r_debug = 0;
2807
2808         if (p->debug_trap_enabled) {
2809                 if (was_enabled)
2810                         kfd_dbg_trap_deactivate(p, false, 0);
2811
2812                 if (!p->is_runtime_retry)
2813                         kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2814                                         p, NULL, 0, false, NULL, 0);
2815
2816                 mutex_unlock(&p->mutex);
2817                 ret = down_interruptible(&p->runtime_enable_sema);
2818                 mutex_lock(&p->mutex);
2819
2820                 p->is_runtime_retry = !!ret;
2821                 if (ret)
2822                         return ret;
2823         }
2824
2825         if (was_enabled && p->runtime_info.ttmp_setup) {
2826                 for (i = 0; i < p->n_pdds; i++) {
2827                         struct kfd_process_device *pdd = p->pdds[i];
2828
2829                         if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2830                                 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2831                 }
2832         }
2833
2834         p->runtime_info.ttmp_setup = false;
2835
2836         /* disable ttmp setup */
2837         for (i = 0; i < p->n_pdds; i++) {
2838                 struct kfd_process_device *pdd = p->pdds[i];
2839
2840                 if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2841                         pdd->spi_dbg_override =
2842                                         pdd->dev->kfd2kgd->disable_debug_trap(
2843                                         pdd->dev->adev,
2844                                         false,
2845                                         pdd->dev->vm_info.last_vmid_kfd);
2846
2847                         if (!pdd->dev->kfd->shared_resources.enable_mes)
2848                                 debug_refresh_runlist(pdd->dev->dqm);
2849                         else
2850                                 kfd_dbg_set_mes_debug_mode(pdd);
2851                 }
2852         }
2853
2854         return 0;
2855 }
2856
2857 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2858 {
2859         struct kfd_ioctl_runtime_enable_args *args = data;
2860         int r;
2861
2862         mutex_lock(&p->mutex);
2863
2864         if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2865                 r = runtime_enable(p, args->r_debug,
2866                                 !!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2867         else
2868                 r = runtime_disable(p);
2869
2870         mutex_unlock(&p->mutex);
2871
2872         return r;
2873 }
2874
2875 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2876 {
2877         struct kfd_ioctl_dbg_trap_args *args = data;
2878         struct task_struct *thread = NULL;
2879         struct mm_struct *mm = NULL;
2880         struct pid *pid = NULL;
2881         struct kfd_process *target = NULL;
2882         int r = 0;
2883
2884         if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2885                 pr_err("Debugging does not support sched_policy %i", sched_policy);
2886                 return -EINVAL;
2887         }
2888
2889         pid = find_get_pid(args->pid);
2890         if (!pid) {
2891                 pr_debug("Cannot find pid info for %i\n", args->pid);
2892                 r = -ESRCH;
2893                 goto out;
2894         }
2895
2896         thread = get_pid_task(pid, PIDTYPE_PID);
2897         if (!thread) {
2898                 r = -ESRCH;
2899                 goto out;
2900         }
2901
2902         mm = get_task_mm(thread);
2903         if (!mm) {
2904                 r = -ESRCH;
2905                 goto out;
2906         }
2907
2908         if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2909                 bool create_process;
2910
2911                 rcu_read_lock();
2912                 create_process = thread && thread != current && ptrace_parent(thread) == current;
2913                 rcu_read_unlock();
2914
2915                 target = create_process ? kfd_create_process(thread) :
2916                                         kfd_lookup_process_by_pid(pid);
2917         } else {
2918                 target = kfd_lookup_process_by_pid(pid);
2919         }
2920
2921         if (!target) {
2922                 pr_debug("Cannot find process PID %i to debug\n", args->pid);
2923                 r = -ESRCH;
2924                 goto out;
2925         }
2926
2927         /* Check if target is still PTRACED. */
2928         rcu_read_lock();
2929         if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2930                                 && ptrace_parent(target->lead_thread) != current) {
2931                 pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2932                 r = -EPERM;
2933         }
2934         rcu_read_unlock();
2935
2936         if (r)
2937                 goto out;
2938
2939         mutex_lock(&target->mutex);
2940
2941         if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2942                 pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2943                 r = -EINVAL;
2944                 goto unlock_out;
2945         }
2946
2947         if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2948                         (args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2949                          args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2950                          args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2951                          args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2952                          args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2953                          args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2954                          args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2955                 r = -EPERM;
2956                 goto unlock_out;
2957         }
2958
2959         switch (args->op) {
2960         case KFD_IOC_DBG_TRAP_ENABLE:
2961                 if (target != p)
2962                         target->debugger_process = p;
2963
2964                 r = kfd_dbg_trap_enable(target,
2965                                         args->enable.dbg_fd,
2966                                         (void __user *)args->enable.rinfo_ptr,
2967                                         &args->enable.rinfo_size);
2968                 if (!r)
2969                         target->exception_enable_mask = args->enable.exception_mask;
2970
2971                 pr_warn("Debug functions limited\n");
2972                 break;
2973         case KFD_IOC_DBG_TRAP_DISABLE:
2974                 r = kfd_dbg_trap_disable(target);
2975                 break;
2976         case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
2977                 r = kfd_dbg_send_exception_to_runtime(target,
2978                                 args->send_runtime_event.gpu_id,
2979                                 args->send_runtime_event.queue_id,
2980                                 args->send_runtime_event.exception_mask);
2981                 break;
2982         case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
2983                 kfd_dbg_set_enabled_debug_exception_mask(target,
2984                                 args->set_exceptions_enabled.exception_mask);
2985                 break;
2986         case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
2987         case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
2988         case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
2989         case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
2990         case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
2991         case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
2992         case KFD_IOC_DBG_TRAP_SET_FLAGS:
2993         case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
2994         case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
2995         case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
2996         case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
2997                 pr_warn("Debug op %i not supported yet\n", args->op);
2998                 r = -EACCES;
2999                 break;
3000         default:
3001                 pr_err("Invalid option: %i\n", args->op);
3002                 r = -EINVAL;
3003         }
3004
3005 unlock_out:
3006         mutex_unlock(&target->mutex);
3007
3008 out:
3009         if (thread)
3010                 put_task_struct(thread);
3011
3012         if (mm)
3013                 mmput(mm);
3014
3015         if (pid)
3016                 put_pid(pid);
3017
3018         if (target)
3019                 kfd_unref_process(target);
3020
3021         return r;
3022 }
3023
3024 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3025         [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3026                             .cmd_drv = 0, .name = #ioctl}
3027
3028 /** Ioctl table */
3029 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3030         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3031                         kfd_ioctl_get_version, 0),
3032
3033         AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3034                         kfd_ioctl_create_queue, 0),
3035
3036         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3037                         kfd_ioctl_destroy_queue, 0),
3038
3039         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3040                         kfd_ioctl_set_memory_policy, 0),
3041
3042         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3043                         kfd_ioctl_get_clock_counters, 0),
3044
3045         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3046                         kfd_ioctl_get_process_apertures, 0),
3047
3048         AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3049                         kfd_ioctl_update_queue, 0),
3050
3051         AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3052                         kfd_ioctl_create_event, 0),
3053
3054         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3055                         kfd_ioctl_destroy_event, 0),
3056
3057         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3058                         kfd_ioctl_set_event, 0),
3059
3060         AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3061                         kfd_ioctl_reset_event, 0),
3062
3063         AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3064                         kfd_ioctl_wait_events, 0),
3065
3066         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3067                         kfd_ioctl_dbg_register, 0),
3068
3069         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3070                         kfd_ioctl_dbg_unregister, 0),
3071
3072         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3073                         kfd_ioctl_dbg_address_watch, 0),
3074
3075         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3076                         kfd_ioctl_dbg_wave_control, 0),
3077
3078         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3079                         kfd_ioctl_set_scratch_backing_va, 0),
3080
3081         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3082                         kfd_ioctl_get_tile_config, 0),
3083
3084         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3085                         kfd_ioctl_set_trap_handler, 0),
3086
3087         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3088                         kfd_ioctl_get_process_apertures_new, 0),
3089
3090         AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3091                         kfd_ioctl_acquire_vm, 0),
3092
3093         AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3094                         kfd_ioctl_alloc_memory_of_gpu, 0),
3095
3096         AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3097                         kfd_ioctl_free_memory_of_gpu, 0),
3098
3099         AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3100                         kfd_ioctl_map_memory_to_gpu, 0),
3101
3102         AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3103                         kfd_ioctl_unmap_memory_from_gpu, 0),
3104
3105         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3106                         kfd_ioctl_set_cu_mask, 0),
3107
3108         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3109                         kfd_ioctl_get_queue_wave_state, 0),
3110
3111         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3112                                 kfd_ioctl_get_dmabuf_info, 0),
3113
3114         AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3115                                 kfd_ioctl_import_dmabuf, 0),
3116
3117         AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3118                         kfd_ioctl_alloc_queue_gws, 0),
3119
3120         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3121                         kfd_ioctl_smi_events, 0),
3122
3123         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3124
3125         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3126                         kfd_ioctl_set_xnack_mode, 0),
3127
3128         AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3129                         kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3130
3131         AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3132                         kfd_ioctl_get_available_memory, 0),
3133
3134         AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3135                                 kfd_ioctl_export_dmabuf, 0),
3136
3137         AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3138                         kfd_ioctl_runtime_enable, 0),
3139
3140         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3141                         kfd_ioctl_set_debug_trap, 0),
3142 };
3143
3144 #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)
3145
3146 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3147 {
3148         struct kfd_process *process;
3149         amdkfd_ioctl_t *func;
3150         const struct amdkfd_ioctl_desc *ioctl = NULL;
3151         unsigned int nr = _IOC_NR(cmd);
3152         char stack_kdata[128];
3153         char *kdata = NULL;
3154         unsigned int usize, asize;
3155         int retcode = -EINVAL;
3156         bool ptrace_attached = false;
3157
3158         if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3159                 goto err_i1;
3160
3161         if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3162                 u32 amdkfd_size;
3163
3164                 ioctl = &amdkfd_ioctls[nr];
3165
3166                 amdkfd_size = _IOC_SIZE(ioctl->cmd);
3167                 usize = asize = _IOC_SIZE(cmd);
3168                 if (amdkfd_size > asize)
3169                         asize = amdkfd_size;
3170
3171                 cmd = ioctl->cmd;
3172         } else
3173                 goto err_i1;
3174
3175         dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3176
3177         /* Get the process struct from the filep. Only the process
3178          * that opened /dev/kfd can use the file descriptor. Child
3179          * processes need to create their own KFD device context.
3180          */
3181         process = filep->private_data;
3182
3183         rcu_read_lock();
3184         if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3185             ptrace_parent(process->lead_thread) == current)
3186                 ptrace_attached = true;
3187         rcu_read_unlock();
3188
3189         if (process->lead_thread != current->group_leader
3190             && !ptrace_attached) {
3191                 dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3192                 retcode = -EBADF;
3193                 goto err_i1;
3194         }
3195
3196         /* Do not trust userspace, use our own definition */
3197         func = ioctl->func;
3198
3199         if (unlikely(!func)) {
3200                 dev_dbg(kfd_device, "no function\n");
3201                 retcode = -EINVAL;
3202                 goto err_i1;
3203         }
3204
3205         /*
3206          * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3207          * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3208          * more priviledged access.
3209          */
3210         if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3211                 if (!capable(CAP_CHECKPOINT_RESTORE) &&
3212                                                 !capable(CAP_SYS_ADMIN)) {
3213                         retcode = -EACCES;
3214                         goto err_i1;
3215                 }
3216         }
3217
3218         if (cmd & (IOC_IN | IOC_OUT)) {
3219                 if (asize <= sizeof(stack_kdata)) {
3220                         kdata = stack_kdata;
3221                 } else {
3222                         kdata = kmalloc(asize, GFP_KERNEL);
3223                         if (!kdata) {
3224                                 retcode = -ENOMEM;
3225                                 goto err_i1;
3226                         }
3227                 }
3228                 if (asize > usize)
3229                         memset(kdata + usize, 0, asize - usize);
3230         }
3231
3232         if (cmd & IOC_IN) {
3233                 if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3234                         retcode = -EFAULT;
3235                         goto err_i1;
3236                 }
3237         } else if (cmd & IOC_OUT) {
3238                 memset(kdata, 0, usize);
3239         }
3240
3241         retcode = func(filep, process, kdata);
3242
3243         if (cmd & IOC_OUT)
3244                 if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3245                         retcode = -EFAULT;
3246
3247 err_i1:
3248         if (!ioctl)
3249                 dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3250                           task_pid_nr(current), cmd, nr);
3251
3252         if (kdata != stack_kdata)
3253                 kfree(kdata);
3254
3255         if (retcode)
3256                 dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3257                                 nr, arg, retcode);
3258
3259         return retcode;
3260 }
3261
3262 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3263                       struct vm_area_struct *vma)
3264 {
3265         phys_addr_t address;
3266
3267         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3268                 return -EINVAL;
3269
3270         address = dev->adev->rmmio_remap.bus_addr;
3271
3272         vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3273                                 VM_DONTDUMP | VM_PFNMAP);
3274
3275         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3276
3277         pr_debug("pasid 0x%x mapping mmio page\n"
3278                  "     target user address == 0x%08llX\n"
3279                  "     physical address    == 0x%08llX\n"
3280                  "     vm_flags            == 0x%04lX\n"
3281                  "     size                == 0x%04lX\n",
3282                  process->pasid, (unsigned long long) vma->vm_start,
3283                  address, vma->vm_flags, PAGE_SIZE);
3284
3285         return io_remap_pfn_range(vma,
3286                                 vma->vm_start,
3287                                 address >> PAGE_SHIFT,
3288                                 PAGE_SIZE,
3289                                 vma->vm_page_prot);
3290 }
3291
3292
3293 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3294 {
3295         struct kfd_process *process;
3296         struct kfd_node *dev = NULL;
3297         unsigned long mmap_offset;
3298         unsigned int gpu_id;
3299
3300         process = kfd_get_process(current);
3301         if (IS_ERR(process))
3302                 return PTR_ERR(process);
3303
3304         mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3305         gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3306         if (gpu_id)
3307                 dev = kfd_device_by_id(gpu_id);
3308
3309         switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3310         case KFD_MMAP_TYPE_DOORBELL:
3311                 if (!dev)
3312                         return -ENODEV;
3313                 return kfd_doorbell_mmap(dev, process, vma);
3314
3315         case KFD_MMAP_TYPE_EVENTS:
3316                 return kfd_event_mmap(process, vma);
3317
3318         case KFD_MMAP_TYPE_RESERVED_MEM:
3319                 if (!dev)
3320                         return -ENODEV;
3321                 return kfd_reserved_mem_mmap(dev, process, vma);
3322         case KFD_MMAP_TYPE_MMIO:
3323                 if (!dev)
3324                         return -ENODEV;
3325                 return kfd_mmio_mmap(dev, process, vma);
3326         }
3327
3328         return -EFAULT;
3329 }