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(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->qpd.proc_doorbells) {
 337                 err = kfd_alloc_process_doorbells(dev->kfd, pdd);
 338                 if (err) {
 339                         pr_debug("failed to allocate process doorbells\n");
 340                         goto err_bind_process;
 341                 }
 342         }
 343
 344         /* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work
 345          * on unmapped queues for usermode queue oversubscription (no aggregated doorbell)
 346          */
 347         if (dev->kfd->shared_resources.enable_mes &&
 348                         ((dev->adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK)
 349                         >> AMDGPU_MES_API_VERSION_SHIFT) >= 2) {
 350                 struct amdgpu_bo_va_mapping *wptr_mapping;
 351                 struct amdgpu_vm *wptr_vm;
 352
 353                 wptr_vm = drm_priv_to_vm(pdd->drm_priv);
 354                 err = amdgpu_bo_reserve(wptr_vm->root.bo, false);
 355                 if (err)
 356                         goto err_wptr_map_gart;
 357
 358                 wptr_mapping = amdgpu_vm_bo_lookup_mapping(
 359                                 wptr_vm, args->write_pointer_address >> PAGE_SHIFT);
 360                 amdgpu_bo_unreserve(wptr_vm->root.bo);
 361                 if (!wptr_mapping) {
 362                         pr_err("Failed to lookup wptr bo\n");
 363                         err = -EINVAL;
 364                         goto err_wptr_map_gart;
 365                 }
 366
 367                 wptr_bo = wptr_mapping->bo_va->base.bo;
 368                 if (wptr_bo->tbo.base.size > PAGE_SIZE) {
 369                         pr_err("Requested GART mapping for wptr bo larger than one page\n");
 370                         err = -EINVAL;
 371                         goto err_wptr_map_gart;
 372                 }
 373
 374                 err = amdgpu_amdkfd_map_gtt_bo_to_gart(dev->adev, wptr_bo);
 375                 if (err) {
 376                         pr_err("Failed to map wptr bo to GART\n");
 377                         goto err_wptr_map_gart;
 378                 }
 379         }
 380
 381         pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n",
 382                         p->pasid,
 383                         dev->id);
 384
 385         err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, wptr_bo,
 386                         NULL, NULL, NULL, &doorbell_offset_in_process);
 387         if (err != 0)
 388                 goto err_create_queue;
 389
 390         args->queue_id = queue_id;
 391
 392
 393         /* Return gpu_id as doorbell offset for mmap usage */
 394         args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
 395         args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
 396         if (KFD_IS_SOC15(dev))
 397                 /* On SOC15 ASICs, include the doorbell offset within the
 398                  * process doorbell frame, which is 2 pages.
 399                  */
 400                 args->doorbell_offset |= doorbell_offset_in_process;
 401
 402         mutex_unlock(&p->mutex);
 403
 404         pr_debug("Queue id %d was created successfully\n", args->queue_id);
 405
 406         pr_debug("Ring buffer address == 0x%016llX\n",
 407                         args->ring_base_address);
 408
 409         pr_debug("Read ptr address    == 0x%016llX\n",
 410                         args->read_pointer_address);
 411
 412         pr_debug("Write ptr address   == 0x%016llX\n",
 413                         args->write_pointer_address);
 414
 415         kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0);
 416         return 0;
 417
 418 err_create_queue:
 419         if (wptr_bo)
 420                 amdgpu_amdkfd_free_gtt_mem(dev->adev, wptr_bo);
 421 err_wptr_map_gart:
 422 err_bind_process:
 423 err_pdd:
 424         mutex_unlock(&p->mutex);
 425         return err;
 426 }
 427
 428 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p,
 429                                         void *data)
 430 {
 431         int retval;
 432         struct kfd_ioctl_destroy_queue_args *args = data;
 433
 434         pr_debug("Destroying queue id %d for pasid 0x%x\n",
 435                                 args->queue_id,
 436                                 p->pasid);
 437
 438         mutex_lock(&p->mutex);
 439
 440         retval = pqm_destroy_queue(&p->pqm, args->queue_id);
 441
 442         mutex_unlock(&p->mutex);
 443         return retval;
 444 }
 445
 446 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p,
 447                                         void *data)
 448 {
 449         int retval;
 450         struct kfd_ioctl_update_queue_args *args = data;
 451         struct queue_properties properties;
 452
 453         /*
 454          * Repurpose queue percentage to accommodate new features:
 455          * bit 0-7: queue percentage
 456          * bit 8-15: pm4_target_xcc
 457          */
 458         if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) {
 459                 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");
 460                 return -EINVAL;
 461         }
 462
 463         if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) {
 464                 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n");
 465                 return -EINVAL;
 466         }
 467
 468         if ((args->ring_base_address) &&
 469                 (!access_ok((const void __user *) args->ring_base_address,
 470                         sizeof(uint64_t)))) {
 471                 pr_err("Can't access ring base address\n");
 472                 return -EFAULT;
 473         }
 474
 475         if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) {
 476                 pr_err("Ring size must be a power of 2 or 0\n");
 477                 return -EINVAL;
 478         }
 479
 480         properties.queue_address = args->ring_base_address;
 481         properties.queue_size = args->ring_size;
 482         properties.queue_percent = args->queue_percentage & 0xFF;
 483         /* bit 8-15 are repurposed to be PM4 target XCC */
 484         properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF;
 485         properties.priority = args->queue_priority;
 486
 487         pr_debug("Updating queue id %d for pasid 0x%x\n",
 488                         args->queue_id, p->pasid);
 489
 490         mutex_lock(&p->mutex);
 491
 492         retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties);
 493
 494         mutex_unlock(&p->mutex);
 495
 496         return retval;
 497 }
 498
 499 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p,
 500                                         void *data)
 501 {
 502         int retval;
 503         const int max_num_cus = 1024;
 504         struct kfd_ioctl_set_cu_mask_args *args = data;
 505         struct mqd_update_info minfo = {0};
 506         uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr;
 507         size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32);
 508
 509         if ((args->num_cu_mask % 32) != 0) {
 510                 pr_debug("num_cu_mask 0x%x must be a multiple of 32",
 511                                 args->num_cu_mask);
 512                 return -EINVAL;
 513         }
 514
 515         minfo.cu_mask.count = args->num_cu_mask;
 516         if (minfo.cu_mask.count == 0) {
 517                 pr_debug("CU mask cannot be 0");
 518                 return -EINVAL;
 519         }
 520
 521         /* To prevent an unreasonably large CU mask size, set an arbitrary
 522          * limit of max_num_cus bits.  We can then just drop any CU mask bits
 523          * past max_num_cus bits and just use the first max_num_cus bits.
 524          */
 525         if (minfo.cu_mask.count > max_num_cus) {
 526                 pr_debug("CU mask cannot be greater than 1024 bits");
 527                 minfo.cu_mask.count = max_num_cus;
 528                 cu_mask_size = sizeof(uint32_t) * (max_num_cus/32);
 529         }
 530
 531         minfo.cu_mask.ptr = kzalloc(cu_mask_size, GFP_KERNEL);
 532         if (!minfo.cu_mask.ptr)
 533                 return -ENOMEM;
 534
 535         retval = copy_from_user(minfo.cu_mask.ptr, cu_mask_ptr, cu_mask_size);
 536         if (retval) {
 537                 pr_debug("Could not copy CU mask from userspace");
 538                 retval = -EFAULT;
 539                 goto out;
 540         }
 541
 542         mutex_lock(&p->mutex);
 543
 544         retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo);
 545
 546         mutex_unlock(&p->mutex);
 547
 548 out:
 549         kfree(minfo.cu_mask.ptr);
 550         return retval;
 551 }
 552
 553 static int kfd_ioctl_get_queue_wave_state(struct file *filep,
 554                                           struct kfd_process *p, void *data)
 555 {
 556         struct kfd_ioctl_get_queue_wave_state_args *args = data;
 557         int r;
 558
 559         mutex_lock(&p->mutex);
 560
 561         r = pqm_get_wave_state(&p->pqm, args->queue_id,
 562                                (void __user *)args->ctl_stack_address,
 563                                &args->ctl_stack_used_size,
 564                                &args->save_area_used_size);
 565
 566         mutex_unlock(&p->mutex);
 567
 568         return r;
 569 }
 570
 571 static int kfd_ioctl_set_memory_policy(struct file *filep,
 572                                         struct kfd_process *p, void *data)
 573 {
 574         struct kfd_ioctl_set_memory_policy_args *args = data;
 575         int err = 0;
 576         struct kfd_process_device *pdd;
 577         enum cache_policy default_policy, alternate_policy;
 578
 579         if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT
 580             && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 581                 return -EINVAL;
 582         }
 583
 584         if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT
 585             && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) {
 586                 return -EINVAL;
 587         }
 588
 589         mutex_lock(&p->mutex);
 590         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 591         if (!pdd) {
 592                 pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
 593                 err = -EINVAL;
 594                 goto err_pdd;
 595         }
 596
 597         pdd = kfd_bind_process_to_device(pdd->dev, p);
 598         if (IS_ERR(pdd)) {
 599                 err = -ESRCH;
 600                 goto out;
 601         }
 602
 603         default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 604                          ? cache_policy_coherent : cache_policy_noncoherent;
 605
 606         alternate_policy =
 607                 (args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)
 608                    ? cache_policy_coherent : cache_policy_noncoherent;
 609
 610         if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm,
 611                                 &pdd->qpd,
 612                                 default_policy,
 613                                 alternate_policy,
 614                                 (void __user *)args->alternate_aperture_base,
 615                                 args->alternate_aperture_size))
 616                 err = -EINVAL;
 617
 618 out:
 619 err_pdd:
 620         mutex_unlock(&p->mutex);
 621
 622         return err;
 623 }
 624
 625 static int kfd_ioctl_set_trap_handler(struct file *filep,
 626                                         struct kfd_process *p, void *data)
 627 {
 628         struct kfd_ioctl_set_trap_handler_args *args = data;
 629         int err = 0;
 630         struct kfd_process_device *pdd;
 631
 632         mutex_lock(&p->mutex);
 633
 634         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 635         if (!pdd) {
 636                 err = -EINVAL;
 637                 goto err_pdd;
 638         }
 639
 640         pdd = kfd_bind_process_to_device(pdd->dev, p);
 641         if (IS_ERR(pdd)) {
 642                 err = -ESRCH;
 643                 goto out;
 644         }
 645
 646         kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr);
 647
 648 out:
 649 err_pdd:
 650         mutex_unlock(&p->mutex);
 651
 652         return err;
 653 }
 654
 655 static int kfd_ioctl_dbg_register(struct file *filep,
 656                                 struct kfd_process *p, void *data)
 657 {
 658         return -EPERM;
 659 }
 660
 661 static int kfd_ioctl_dbg_unregister(struct file *filep,
 662                                 struct kfd_process *p, void *data)
 663 {
 664         return -EPERM;
 665 }
 666
 667 static int kfd_ioctl_dbg_address_watch(struct file *filep,
 668                                         struct kfd_process *p, void *data)
 669 {
 670         return -EPERM;
 671 }
 672
 673 /* Parse and generate fixed size data structure for wave control */
 674 static int kfd_ioctl_dbg_wave_control(struct file *filep,
 675                                         struct kfd_process *p, void *data)
 676 {
 677         return -EPERM;
 678 }
 679
 680 static int kfd_ioctl_get_clock_counters(struct file *filep,
 681                                 struct kfd_process *p, void *data)
 682 {
 683         struct kfd_ioctl_get_clock_counters_args *args = data;
 684         struct kfd_process_device *pdd;
 685
 686         mutex_lock(&p->mutex);
 687         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 688         mutex_unlock(&p->mutex);
 689         if (pdd)
 690                 /* Reading GPU clock counter from KGD */
 691                 args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev);
 692         else
 693                 /* Node without GPU resource */
 694                 args->gpu_clock_counter = 0;
 695
 696         /* No access to rdtsc. Using raw monotonic time */
 697         args->cpu_clock_counter = ktime_get_raw_ns();
 698         args->system_clock_counter = ktime_get_boottime_ns();
 699
 700         /* Since the counter is in nano-seconds we use 1GHz frequency */
 701         args->system_clock_freq = 1000000000;
 702
 703         return 0;
 704 }
 705
 706
 707 static int kfd_ioctl_get_process_apertures(struct file *filp,
 708                                 struct kfd_process *p, void *data)
 709 {
 710         struct kfd_ioctl_get_process_apertures_args *args = data;
 711         struct kfd_process_device_apertures *pAperture;
 712         int i;
 713
 714         dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
 715
 716         args->num_of_nodes = 0;
 717
 718         mutex_lock(&p->mutex);
 719         /* Run over all pdd of the process */
 720         for (i = 0; i < p->n_pdds; i++) {
 721                 struct kfd_process_device *pdd = p->pdds[i];
 722
 723                 pAperture =
 724                         &args->process_apertures[args->num_of_nodes];
 725                 pAperture->gpu_id = pdd->dev->id;
 726                 pAperture->lds_base = pdd->lds_base;
 727                 pAperture->lds_limit = pdd->lds_limit;
 728                 pAperture->gpuvm_base = pdd->gpuvm_base;
 729                 pAperture->gpuvm_limit = pdd->gpuvm_limit;
 730                 pAperture->scratch_base = pdd->scratch_base;
 731                 pAperture->scratch_limit = pdd->scratch_limit;
 732
 733                 dev_dbg(kfd_device,
 734                         "node id %u\n", args->num_of_nodes);
 735                 dev_dbg(kfd_device,
 736                         "gpu id %u\n", pdd->dev->id);
 737                 dev_dbg(kfd_device,
 738                         "lds_base %llX\n", pdd->lds_base);
 739                 dev_dbg(kfd_device,
 740                         "lds_limit %llX\n", pdd->lds_limit);
 741                 dev_dbg(kfd_device,
 742                         "gpuvm_base %llX\n", pdd->gpuvm_base);
 743                 dev_dbg(kfd_device,
 744                         "gpuvm_limit %llX\n", pdd->gpuvm_limit);
 745                 dev_dbg(kfd_device,
 746                         "scratch_base %llX\n", pdd->scratch_base);
 747                 dev_dbg(kfd_device,
 748                         "scratch_limit %llX\n", pdd->scratch_limit);
 749
 750                 if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS)
 751                         break;
 752         }
 753         mutex_unlock(&p->mutex);
 754
 755         return 0;
 756 }
 757
 758 static int kfd_ioctl_get_process_apertures_new(struct file *filp,
 759                                 struct kfd_process *p, void *data)
 760 {
 761         struct kfd_ioctl_get_process_apertures_new_args *args = data;
 762         struct kfd_process_device_apertures *pa;
 763         int ret;
 764         int i;
 765
 766         dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid);
 767
 768         if (args->num_of_nodes == 0) {
 769                 /* Return number of nodes, so that user space can alloacate
 770                  * sufficient memory
 771                  */
 772                 mutex_lock(&p->mutex);
 773                 args->num_of_nodes = p->n_pdds;
 774                 goto out_unlock;
 775         }
 776
 777         /* Fill in process-aperture information for all available
 778          * nodes, but not more than args->num_of_nodes as that is
 779          * the amount of memory allocated by user
 780          */
 781         pa = kzalloc((sizeof(struct kfd_process_device_apertures) *
 782                                 args->num_of_nodes), GFP_KERNEL);
 783         if (!pa)
 784                 return -ENOMEM;
 785
 786         mutex_lock(&p->mutex);
 787
 788         if (!p->n_pdds) {
 789                 args->num_of_nodes = 0;
 790                 kfree(pa);
 791                 goto out_unlock;
 792         }
 793
 794         /* Run over all pdd of the process */
 795         for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) {
 796                 struct kfd_process_device *pdd = p->pdds[i];
 797
 798                 pa[i].gpu_id = pdd->dev->id;
 799                 pa[i].lds_base = pdd->lds_base;
 800                 pa[i].lds_limit = pdd->lds_limit;
 801                 pa[i].gpuvm_base = pdd->gpuvm_base;
 802                 pa[i].gpuvm_limit = pdd->gpuvm_limit;
 803                 pa[i].scratch_base = pdd->scratch_base;
 804                 pa[i].scratch_limit = pdd->scratch_limit;
 805
 806                 dev_dbg(kfd_device,
 807                         "gpu id %u\n", pdd->dev->id);
 808                 dev_dbg(kfd_device,
 809                         "lds_base %llX\n", pdd->lds_base);
 810                 dev_dbg(kfd_device,
 811                         "lds_limit %llX\n", pdd->lds_limit);
 812                 dev_dbg(kfd_device,
 813                         "gpuvm_base %llX\n", pdd->gpuvm_base);
 814                 dev_dbg(kfd_device,
 815                         "gpuvm_limit %llX\n", pdd->gpuvm_limit);
 816                 dev_dbg(kfd_device,
 817                         "scratch_base %llX\n", pdd->scratch_base);
 818                 dev_dbg(kfd_device,
 819                         "scratch_limit %llX\n", pdd->scratch_limit);
 820         }
 821         mutex_unlock(&p->mutex);
 822
 823         args->num_of_nodes = i;
 824         ret = copy_to_user(
 825                         (void __user *)args->kfd_process_device_apertures_ptr,
 826                         pa,
 827                         (i * sizeof(struct kfd_process_device_apertures)));
 828         kfree(pa);
 829         return ret ? -EFAULT : 0;
 830
 831 out_unlock:
 832         mutex_unlock(&p->mutex);
 833         return 0;
 834 }
 835
 836 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p,
 837                                         void *data)
 838 {
 839         struct kfd_ioctl_create_event_args *args = data;
 840         int err;
 841
 842         /* For dGPUs the event page is allocated in user mode. The
 843          * handle is passed to KFD with the first call to this IOCTL
 844          * through the event_page_offset field.
 845          */
 846         if (args->event_page_offset) {
 847                 mutex_lock(&p->mutex);
 848                 err = kfd_kmap_event_page(p, args->event_page_offset);
 849                 mutex_unlock(&p->mutex);
 850                 if (err)
 851                         return err;
 852         }
 853
 854         err = kfd_event_create(filp, p, args->event_type,
 855                                 args->auto_reset != 0, args->node_id,
 856                                 &args->event_id, &args->event_trigger_data,
 857                                 &args->event_page_offset,
 858                                 &args->event_slot_index);
 859
 860         pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__);
 861         return err;
 862 }
 863
 864 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p,
 865                                         void *data)
 866 {
 867         struct kfd_ioctl_destroy_event_args *args = data;
 868
 869         return kfd_event_destroy(p, args->event_id);
 870 }
 871
 872 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p,
 873                                 void *data)
 874 {
 875         struct kfd_ioctl_set_event_args *args = data;
 876
 877         return kfd_set_event(p, args->event_id);
 878 }
 879
 880 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p,
 881                                 void *data)
 882 {
 883         struct kfd_ioctl_reset_event_args *args = data;
 884
 885         return kfd_reset_event(p, args->event_id);
 886 }
 887
 888 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p,
 889                                 void *data)
 890 {
 891         struct kfd_ioctl_wait_events_args *args = data;
 892
 893         return kfd_wait_on_events(p, args->num_events,
 894                         (void __user *)args->events_ptr,
 895                         (args->wait_for_all != 0),
 896                         &args->timeout, &args->wait_result);
 897 }
 898 static int kfd_ioctl_set_scratch_backing_va(struct file *filep,
 899                                         struct kfd_process *p, void *data)
 900 {
 901         struct kfd_ioctl_set_scratch_backing_va_args *args = data;
 902         struct kfd_process_device *pdd;
 903         struct kfd_node *dev;
 904         long err;
 905
 906         mutex_lock(&p->mutex);
 907         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 908         if (!pdd) {
 909                 err = -EINVAL;
 910                 goto err_pdd;
 911         }
 912         dev = pdd->dev;
 913
 914         pdd = kfd_bind_process_to_device(dev, p);
 915         if (IS_ERR(pdd)) {
 916                 err = PTR_ERR(pdd);
 917                 goto bind_process_to_device_fail;
 918         }
 919
 920         pdd->qpd.sh_hidden_private_base = args->va_addr;
 921
 922         mutex_unlock(&p->mutex);
 923
 924         if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS &&
 925             pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va)
 926                 dev->kfd2kgd->set_scratch_backing_va(
 927                         dev->adev, args->va_addr, pdd->qpd.vmid);
 928
 929         return 0;
 930
 931 bind_process_to_device_fail:
 932 err_pdd:
 933         mutex_unlock(&p->mutex);
 934         return err;
 935 }
 936
 937 static int kfd_ioctl_get_tile_config(struct file *filep,
 938                 struct kfd_process *p, void *data)
 939 {
 940         struct kfd_ioctl_get_tile_config_args *args = data;
 941         struct kfd_process_device *pdd;
 942         struct tile_config config;
 943         int err = 0;
 944
 945         mutex_lock(&p->mutex);
 946         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 947         mutex_unlock(&p->mutex);
 948         if (!pdd)
 949                 return -EINVAL;
 950
 951         amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config);
 952
 953         args->gb_addr_config = config.gb_addr_config;
 954         args->num_banks = config.num_banks;
 955         args->num_ranks = config.num_ranks;
 956
 957         if (args->num_tile_configs > config.num_tile_configs)
 958                 args->num_tile_configs = config.num_tile_configs;
 959         err = copy_to_user((void __user *)args->tile_config_ptr,
 960                         config.tile_config_ptr,
 961                         args->num_tile_configs * sizeof(uint32_t));
 962         if (err) {
 963                 args->num_tile_configs = 0;
 964                 return -EFAULT;
 965         }
 966
 967         if (args->num_macro_tile_configs > config.num_macro_tile_configs)
 968                 args->num_macro_tile_configs =
 969                                 config.num_macro_tile_configs;
 970         err = copy_to_user((void __user *)args->macro_tile_config_ptr,
 971                         config.macro_tile_config_ptr,
 972                         args->num_macro_tile_configs * sizeof(uint32_t));
 973         if (err) {
 974                 args->num_macro_tile_configs = 0;
 975                 return -EFAULT;
 976         }
 977
 978         return 0;
 979 }
 980
 981 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p,
 982                                 void *data)
 983 {
 984         struct kfd_ioctl_acquire_vm_args *args = data;
 985         struct kfd_process_device *pdd;
 986         struct file *drm_file;
 987         int ret;
 988
 989         drm_file = fget(args->drm_fd);
 990         if (!drm_file)
 991                 return -EINVAL;
 992
 993         mutex_lock(&p->mutex);
 994         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
 995         if (!pdd) {
 996                 ret = -EINVAL;
 997                 goto err_pdd;
 998         }
 999
1000         if (pdd->drm_file) {
1001                 ret = pdd->drm_file == drm_file ? 0 : -EBUSY;
1002                 goto err_drm_file;
1003         }
1004
1005         ret = kfd_process_device_init_vm(pdd, drm_file);
1006         if (ret)
1007                 goto err_unlock;
1008
1009         /* On success, the PDD keeps the drm_file reference */
1010         mutex_unlock(&p->mutex);
1011
1012         return 0;
1013
1014 err_unlock:
1015 err_pdd:
1016 err_drm_file:
1017         mutex_unlock(&p->mutex);
1018         fput(drm_file);
1019         return ret;
1020 }
1021
1022 bool kfd_dev_is_large_bar(struct kfd_node *dev)
1023 {
1024         if (debug_largebar) {
1025                 pr_debug("Simulate large-bar allocation on non large-bar machine\n");
1026                 return true;
1027         }
1028
1029         if (dev->local_mem_info.local_mem_size_private == 0 &&
1030             dev->local_mem_info.local_mem_size_public > 0)
1031                 return true;
1032
1033         if (dev->local_mem_info.local_mem_size_public == 0 &&
1034             dev->kfd->adev->gmc.is_app_apu) {
1035                 pr_debug("APP APU, Consider like a large bar system\n");
1036                 return true;
1037         }
1038
1039         return false;
1040 }
1041
1042 static int kfd_ioctl_get_available_memory(struct file *filep,
1043                                           struct kfd_process *p, void *data)
1044 {
1045         struct kfd_ioctl_get_available_memory_args *args = data;
1046         struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
1047
1048         if (!pdd)
1049                 return -EINVAL;
1050         args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
1051                                                         pdd->dev->node_id);
1052         kfd_unlock_pdd(pdd);
1053         return 0;
1054 }
1055
1056 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep,
1057                                         struct kfd_process *p, void *data)
1058 {
1059         struct kfd_ioctl_alloc_memory_of_gpu_args *args = data;
1060         struct kfd_process_device *pdd;
1061         void *mem;
1062         struct kfd_node *dev;
1063         int idr_handle;
1064         long err;
1065         uint64_t offset = args->mmap_offset;
1066         uint32_t flags = args->flags;
1067
1068         if (args->size == 0)
1069                 return -EINVAL;
1070
1071 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
1072         /* Flush pending deferred work to avoid racing with deferred actions
1073          * from previous memory map changes (e.g. munmap).
1074          */
1075         svm_range_list_lock_and_flush_work(&p->svms, current->mm);
1076         mutex_lock(&p->svms.lock);
1077         mmap_write_unlock(current->mm);
1078         if (interval_tree_iter_first(&p->svms.objects,
1079                                      args->va_addr >> PAGE_SHIFT,
1080                                      (args->va_addr + args->size - 1) >> PAGE_SHIFT)) {
1081                 pr_err("Address: 0x%llx already allocated by SVM\n",
1082                         args->va_addr);
1083                 mutex_unlock(&p->svms.lock);
1084                 return -EADDRINUSE;
1085         }
1086
1087         /* When register user buffer check if it has been registered by svm by
1088          * buffer cpu virtual address.
1089          */
1090         if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) &&
1091             interval_tree_iter_first(&p->svms.objects,
1092                                      args->mmap_offset >> PAGE_SHIFT,
1093                                      (args->mmap_offset  + args->size - 1) >> PAGE_SHIFT)) {
1094                 pr_err("User Buffer Address: 0x%llx already allocated by SVM\n",
1095                         args->mmap_offset);
1096                 mutex_unlock(&p->svms.lock);
1097                 return -EADDRINUSE;
1098         }
1099
1100         mutex_unlock(&p->svms.lock);
1101 #endif
1102         mutex_lock(&p->mutex);
1103         pdd = kfd_process_device_data_by_id(p, args->gpu_id);
1104         if (!pdd) {
1105                 err = -EINVAL;
1106                 goto err_pdd;
1107         }
1108
1109         dev = pdd->dev;
1110
1111         if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) &&
1112                 (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) &&
1113                 !kfd_dev_is_large_bar(dev)) {
1114                 pr_err("Alloc host visible vram on small bar is not allowed\n");
1115                 err = -EINVAL;
1116                 goto err_large_bar;
1117         }
1118
1119         pdd = kfd_bind_process_to_device(dev, p);
1120         if (IS_ERR(pdd)) {
1121                 err = PTR_ERR(pdd);
1122                 goto err_unlock;
1123         }
1124
1125         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
1126                 if (args->size != kfd_doorbell_process_slice(dev->kfd)) {
1127                         err = -EINVAL;
1128                         goto err_unlock;
1129                 }
1130                 offset = kfd_get_process_doorbells(pdd);
1131                 if (!offset) {
1132                         err = -ENOMEM;
1133                         goto err_unlock;
1134                 }
1135         } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
1136                 if (args->size != PAGE_SIZE) {
1137                         err = -EINVAL;
1138                         goto err_unlock;
1139                 }
1140                 offset = dev->adev->rmmio_remap.bus_addr;
1141                 if (!offset) {
1142                         err = -ENOMEM;
1143                         goto err_unlock;
1144                 }
1145         }
1146
1147         err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1148                 dev->adev, args->va_addr, args->size,
1149                 pdd->drm_priv, (struct kgd_mem **) &mem, &offset,
1150                 flags, false);
1151
1152         if (err)
1153                 goto err_unlock;
1154
1155         idr_handle = kfd_process_device_create_obj_handle(pdd, mem);
1156         if (idr_handle < 0) {
1157                 err = -EFAULT;
1158                 goto err_free;
1159         }
1160
1161         /* Update the VRAM usage count */
1162         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1163                 uint64_t size = args->size;
1164
1165                 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM)
1166                         size >>= 1;
1167                 WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + PAGE_ALIGN(size));
1168         }
1169
1170         mutex_unlock(&p->mutex);
1171
1172         args->handle = MAKE_HANDLE(args->gpu_id, idr_handle);
1173         args->mmap_offset = offset;
1174
1175         /* MMIO is mapped through kfd device
1176          * Generate a kfd mmap offset
1177          */
1178         if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1179                 args->mmap_offset = KFD_MMAP_TYPE_MMIO
1180                                         | KFD_MMAP_GPU_ID(args->gpu_id);
1181
1182         return 0;
1183
1184 err_free:
1185         amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem,
1186                                                pdd->drm_priv, NULL);
1187 err_unlock:
1188 err_pdd:
1189 err_large_bar:
1190         mutex_unlock(&p->mutex);
1191         return err;
1192 }
1193
1194 static int kfd_ioctl_free_memory_of_gpu(struct file *filep,
1195                                         struct kfd_process *p, void *data)
1196 {
1197         struct kfd_ioctl_free_memory_of_gpu_args *args = data;
1198         struct kfd_process_device *pdd;
1199         void *mem;
1200         int ret;
1201         uint64_t size = 0;
1202
1203         mutex_lock(&p->mutex);
1204         /*
1205          * Safeguard to prevent user space from freeing signal BO.
1206          * It will be freed at process termination.
1207          */
1208         if (p->signal_handle && (p->signal_handle == args->handle)) {
1209                 pr_err("Free signal BO is not allowed\n");
1210                 ret = -EPERM;
1211                 goto err_unlock;
1212         }
1213
1214         pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1215         if (!pdd) {
1216                 pr_err("Process device data doesn't exist\n");
1217                 ret = -EINVAL;
1218                 goto err_pdd;
1219         }
1220
1221         mem = kfd_process_device_translate_handle(
1222                 pdd, GET_IDR_HANDLE(args->handle));
1223         if (!mem) {
1224                 ret = -EINVAL;
1225                 goto err_unlock;
1226         }
1227
1228         ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev,
1229                                 (struct kgd_mem *)mem, pdd->drm_priv, &size);
1230
1231         /* If freeing the buffer failed, leave the handle in place for
1232          * clean-up during process tear-down.
1233          */
1234         if (!ret)
1235                 kfd_process_device_remove_obj_handle(
1236                         pdd, GET_IDR_HANDLE(args->handle));
1237
1238         WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size);
1239
1240 err_unlock:
1241 err_pdd:
1242         mutex_unlock(&p->mutex);
1243         return ret;
1244 }
1245
1246 static int kfd_ioctl_map_memory_to_gpu(struct file *filep,
1247                                         struct kfd_process *p, void *data)
1248 {
1249         struct kfd_ioctl_map_memory_to_gpu_args *args = data;
1250         struct kfd_process_device *pdd, *peer_pdd;
1251         void *mem;
1252         struct kfd_node *dev;
1253         long err = 0;
1254         int i;
1255         uint32_t *devices_arr = NULL;
1256
1257         if (!args->n_devices) {
1258                 pr_debug("Device IDs array empty\n");
1259                 return -EINVAL;
1260         }
1261         if (args->n_success > args->n_devices) {
1262                 pr_debug("n_success exceeds n_devices\n");
1263                 return -EINVAL;
1264         }
1265
1266         devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1267                                     GFP_KERNEL);
1268         if (!devices_arr)
1269                 return -ENOMEM;
1270
1271         err = copy_from_user(devices_arr,
1272                              (void __user *)args->device_ids_array_ptr,
1273                              args->n_devices * sizeof(*devices_arr));
1274         if (err != 0) {
1275                 err = -EFAULT;
1276                 goto copy_from_user_failed;
1277         }
1278
1279         mutex_lock(&p->mutex);
1280         pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1281         if (!pdd) {
1282                 err = -EINVAL;
1283                 goto get_process_device_data_failed;
1284         }
1285         dev = pdd->dev;
1286
1287         pdd = kfd_bind_process_to_device(dev, p);
1288         if (IS_ERR(pdd)) {
1289                 err = PTR_ERR(pdd);
1290                 goto bind_process_to_device_failed;
1291         }
1292
1293         mem = kfd_process_device_translate_handle(pdd,
1294                                                 GET_IDR_HANDLE(args->handle));
1295         if (!mem) {
1296                 err = -ENOMEM;
1297                 goto get_mem_obj_from_handle_failed;
1298         }
1299
1300         for (i = args->n_success; i < args->n_devices; i++) {
1301                 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1302                 if (!peer_pdd) {
1303                         pr_debug("Getting device by id failed for 0x%x\n",
1304                                  devices_arr[i]);
1305                         err = -EINVAL;
1306                         goto get_mem_obj_from_handle_failed;
1307                 }
1308
1309                 peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p);
1310                 if (IS_ERR(peer_pdd)) {
1311                         err = PTR_ERR(peer_pdd);
1312                         goto get_mem_obj_from_handle_failed;
1313                 }
1314
1315                 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1316                         peer_pdd->dev->adev, (struct kgd_mem *)mem,
1317                         peer_pdd->drm_priv);
1318                 if (err) {
1319                         struct pci_dev *pdev = peer_pdd->dev->adev->pdev;
1320
1321                         dev_err(dev->adev->dev,
1322                                "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n",
1323                                pci_domain_nr(pdev->bus),
1324                                pdev->bus->number,
1325                                PCI_SLOT(pdev->devfn),
1326                                PCI_FUNC(pdev->devfn),
1327                                ((struct kgd_mem *)mem)->domain);
1328                         goto map_memory_to_gpu_failed;
1329                 }
1330                 args->n_success = i+1;
1331         }
1332
1333         err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true);
1334         if (err) {
1335                 pr_debug("Sync memory failed, wait interrupted by user signal\n");
1336                 goto sync_memory_failed;
1337         }
1338
1339         mutex_unlock(&p->mutex);
1340
1341         /* Flush TLBs after waiting for the page table updates to complete */
1342         for (i = 0; i < args->n_devices; i++) {
1343                 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1344                 if (WARN_ON_ONCE(!peer_pdd))
1345                         continue;
1346                 kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY);
1347         }
1348         kfree(devices_arr);
1349
1350         return err;
1351
1352 get_process_device_data_failed:
1353 bind_process_to_device_failed:
1354 get_mem_obj_from_handle_failed:
1355 map_memory_to_gpu_failed:
1356 sync_memory_failed:
1357         mutex_unlock(&p->mutex);
1358 copy_from_user_failed:
1359         kfree(devices_arr);
1360
1361         return err;
1362 }
1363
1364 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
1365                                         struct kfd_process *p, void *data)
1366 {
1367         struct kfd_ioctl_unmap_memory_from_gpu_args *args = data;
1368         struct kfd_process_device *pdd, *peer_pdd;
1369         void *mem;
1370         long err = 0;
1371         uint32_t *devices_arr = NULL, i;
1372         bool flush_tlb;
1373
1374         if (!args->n_devices) {
1375                 pr_debug("Device IDs array empty\n");
1376                 return -EINVAL;
1377         }
1378         if (args->n_success > args->n_devices) {
1379                 pr_debug("n_success exceeds n_devices\n");
1380                 return -EINVAL;
1381         }
1382
1383         devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr),
1384                                     GFP_KERNEL);
1385         if (!devices_arr)
1386                 return -ENOMEM;
1387
1388         err = copy_from_user(devices_arr,
1389                              (void __user *)args->device_ids_array_ptr,
1390                              args->n_devices * sizeof(*devices_arr));
1391         if (err != 0) {
1392                 err = -EFAULT;
1393                 goto copy_from_user_failed;
1394         }
1395
1396         mutex_lock(&p->mutex);
1397         pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle));
1398         if (!pdd) {
1399                 err = -EINVAL;
1400                 goto bind_process_to_device_failed;
1401         }
1402
1403         mem = kfd_process_device_translate_handle(pdd,
1404                                                 GET_IDR_HANDLE(args->handle));
1405         if (!mem) {
1406                 err = -ENOMEM;
1407                 goto get_mem_obj_from_handle_failed;
1408         }
1409
1410         for (i = args->n_success; i < args->n_devices; i++) {
1411                 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1412                 if (!peer_pdd) {
1413                         err = -EINVAL;
1414                         goto get_mem_obj_from_handle_failed;
1415                 }
1416                 err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1417                         peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);
1418                 if (err) {
1419                         pr_err("Failed to unmap from gpu %d/%d\n",
1420                                i, args->n_devices);
1421                         goto unmap_memory_from_gpu_failed;
1422                 }
1423                 args->n_success = i+1;
1424         }
1425
1426         flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd);
1427         if (flush_tlb) {
1428                 err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev,
1429                                 (struct kgd_mem *) mem, true);
1430                 if (err) {
1431                         pr_debug("Sync memory failed, wait interrupted by user signal\n");
1432                         goto sync_memory_failed;
1433                 }
1434         }
1435         mutex_unlock(&p->mutex);
1436
1437         if (flush_tlb) {
1438                 /* Flush TLBs after waiting for the page table updates to complete */
1439                 for (i = 0; i < args->n_devices; i++) {
1440                         peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]);
1441                         if (WARN_ON_ONCE(!peer_pdd))
1442                                 continue;
1443                         kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
1444                 }
1445         }
1446         kfree(devices_arr);
1447
1448         return 0;
1449
1450 bind_process_to_device_failed:
1451 get_mem_obj_from_handle_failed:
1452 unmap_memory_from_gpu_failed:
1453 sync_memory_failed:
1454         mutex_unlock(&p->mutex);
1455 copy_from_user_failed:
1456         kfree(devices_arr);
1457         return err;
1458 }
1459
1460 static int kfd_ioctl_alloc_queue_gws(struct file *filep,
1461                 struct kfd_process *p, void *data)
1462 {
1463         int retval;
1464         struct kfd_ioctl_alloc_queue_gws_args *args = data;
1465         struct queue *q;
1466         struct kfd_node *dev;
1467
1468         mutex_lock(&p->mutex);
1469         q = pqm_get_user_queue(&p->pqm, args->queue_id);
1470
1471         if (q) {
1472                 dev = q->device;
1473         } else {
1474                 retval = -EINVAL;
1475                 goto out_unlock;
1476         }
1477
1478         if (!dev->gws) {
1479                 retval = -ENODEV;
1480                 goto out_unlock;
1481         }
1482
1483         if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1484                 retval = -ENODEV;
1485                 goto out_unlock;
1486         }
1487
1488         if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
1489                                       kfd_dbg_has_cwsr_workaround(dev))) {
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 (!kgd_mem->va || 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 kgd_mem *mem, int flags,
1855                                       u32 *shared_fd)
1856 {
1857         struct dma_buf *dmabuf;
1858         int ret;
1859
1860         ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
1861         if (ret) {
1862                 pr_err("dmabuf export failed for the BO\n");
1863                 return ret;
1864         }
1865
1866         ret = dma_buf_fd(dmabuf, flags);
1867         if (ret < 0) {
1868                 pr_err("dmabuf create fd failed, ret:%d\n", ret);
1869                 goto out_free_dmabuf;
1870         }
1871
1872         *shared_fd = ret;
1873         return 0;
1874
1875 out_free_dmabuf:
1876         dma_buf_put(dmabuf);
1877         return ret;
1878 }
1879
1880 static int criu_checkpoint_bos(struct kfd_process *p,
1881                                uint32_t num_bos,
1882                                uint8_t __user *user_bos,
1883                                uint8_t __user *user_priv_data,
1884                                uint64_t *priv_offset)
1885 {
1886         struct kfd_criu_bo_bucket *bo_buckets;
1887         struct kfd_criu_bo_priv_data *bo_privs;
1888         int ret = 0, pdd_index, bo_index = 0, id;
1889         void *mem;
1890
1891         bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL);
1892         if (!bo_buckets)
1893                 return -ENOMEM;
1894
1895         bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL);
1896         if (!bo_privs) {
1897                 ret = -ENOMEM;
1898                 goto exit;
1899         }
1900
1901         for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
1902                 struct kfd_process_device *pdd = p->pdds[pdd_index];
1903                 struct amdgpu_bo *dumper_bo;
1904                 struct kgd_mem *kgd_mem;
1905
1906                 idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1907                         struct kfd_criu_bo_bucket *bo_bucket;
1908                         struct kfd_criu_bo_priv_data *bo_priv;
1909                         int i, dev_idx = 0;
1910
1911                         if (!mem) {
1912                                 ret = -ENOMEM;
1913                                 goto exit;
1914                         }
1915
1916                         kgd_mem = (struct kgd_mem *)mem;
1917                         dumper_bo = kgd_mem->bo;
1918
1919                         /* Skip checkpointing BOs that are used for Trap handler
1920                          * code and state. Currently, these BOs have a VA that
1921                          * is less GPUVM Base
1922                          */
1923                         if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base)
1924                                 continue;
1925
1926                         bo_bucket = &bo_buckets[bo_index];
1927                         bo_priv = &bo_privs[bo_index];
1928
1929                         bo_bucket->gpu_id = pdd->user_gpu_id;
1930                         bo_bucket->addr = (uint64_t)kgd_mem->va;
1931                         bo_bucket->size = amdgpu_bo_size(dumper_bo);
1932                         bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags;
1933                         bo_priv->idr_handle = id;
1934
1935                         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1936                                 ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo,
1937                                                                 &bo_priv->user_addr);
1938                                 if (ret) {
1939                                         pr_err("Failed to obtain user address for user-pointer bo\n");
1940                                         goto exit;
1941                                 }
1942                         }
1943                         if (bo_bucket->alloc_flags
1944                             & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
1945                                 ret = criu_get_prime_handle(kgd_mem,
1946                                                 bo_bucket->alloc_flags &
1947                                                 KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0,
1948                                                 &bo_bucket->dmabuf_fd);
1949                                 if (ret)
1950                                         goto exit;
1951                         } else {
1952                                 bo_bucket->dmabuf_fd = KFD_INVALID_FD;
1953                         }
1954
1955                         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
1956                                 bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL |
1957                                         KFD_MMAP_GPU_ID(pdd->dev->id);
1958                         else if (bo_bucket->alloc_flags &
1959                                 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)
1960                                 bo_bucket->offset = KFD_MMAP_TYPE_MMIO |
1961                                         KFD_MMAP_GPU_ID(pdd->dev->id);
1962                         else
1963                                 bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo);
1964
1965                         for (i = 0; i < p->n_pdds; i++) {
1966                                 if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->dev->adev, kgd_mem))
1967                                         bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id;
1968                         }
1969
1970                         pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n"
1971                                         "gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x",
1972                                         bo_bucket->size,
1973                                         bo_bucket->addr,
1974                                         bo_bucket->offset,
1975                                         bo_bucket->gpu_id,
1976                                         bo_bucket->alloc_flags,
1977                                         bo_priv->idr_handle);
1978                         bo_index++;
1979                 }
1980         }
1981
1982         ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets));
1983         if (ret) {
1984                 pr_err("Failed to copy BO information to user\n");
1985                 ret = -EFAULT;
1986                 goto exit;
1987         }
1988
1989         ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs));
1990         if (ret) {
1991                 pr_err("Failed to copy BO priv information to user\n");
1992                 ret = -EFAULT;
1993                 goto exit;
1994         }
1995
1996         *priv_offset += num_bos * sizeof(*bo_privs);
1997
1998 exit:
1999         while (ret && bo_index--) {
2000                 if (bo_buckets[bo_index].alloc_flags
2001                     & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2002                         close_fd(bo_buckets[bo_index].dmabuf_fd);
2003         }
2004
2005         kvfree(bo_buckets);
2006         kvfree(bo_privs);
2007         return ret;
2008 }
2009
2010 static int criu_get_process_object_info(struct kfd_process *p,
2011                                         uint32_t *num_devices,
2012                                         uint32_t *num_bos,
2013                                         uint32_t *num_objects,
2014                                         uint64_t *objs_priv_size)
2015 {
2016         uint64_t queues_priv_data_size, svm_priv_data_size, priv_size;
2017         uint32_t num_queues, num_events, num_svm_ranges;
2018         int ret;
2019
2020         *num_devices = p->n_pdds;
2021         *num_bos = get_process_num_bos(p);
2022
2023         ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size);
2024         if (ret)
2025                 return ret;
2026
2027         num_events = kfd_get_num_events(p);
2028
2029         ret = svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size);
2030         if (ret)
2031                 return ret;
2032
2033         *num_objects = num_queues + num_events + num_svm_ranges;
2034
2035         if (objs_priv_size) {
2036                 priv_size = sizeof(struct kfd_criu_process_priv_data);
2037                 priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data);
2038                 priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data);
2039                 priv_size += queues_priv_data_size;
2040                 priv_size += num_events * sizeof(struct kfd_criu_event_priv_data);
2041                 priv_size += svm_priv_data_size;
2042                 *objs_priv_size = priv_size;
2043         }
2044         return 0;
2045 }
2046
2047 static int criu_checkpoint(struct file *filep,
2048                            struct kfd_process *p,
2049                            struct kfd_ioctl_criu_args *args)
2050 {
2051         int ret;
2052         uint32_t num_devices, num_bos, num_objects;
2053         uint64_t priv_size, priv_offset = 0, bo_priv_offset;
2054
2055         if (!args->devices || !args->bos || !args->priv_data)
2056                 return -EINVAL;
2057
2058         mutex_lock(&p->mutex);
2059
2060         if (!p->n_pdds) {
2061                 pr_err("No pdd for given process\n");
2062                 ret = -ENODEV;
2063                 goto exit_unlock;
2064         }
2065
2066         /* Confirm all process queues are evicted */
2067         if (!p->queues_paused) {
2068                 pr_err("Cannot dump process when queues are not in evicted state\n");
2069                 /* CRIU plugin did not call op PROCESS_INFO before checkpointing */
2070                 ret = -EINVAL;
2071                 goto exit_unlock;
2072         }
2073
2074         ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size);
2075         if (ret)
2076                 goto exit_unlock;
2077
2078         if (num_devices != args->num_devices ||
2079             num_bos != args->num_bos ||
2080             num_objects != args->num_objects ||
2081             priv_size != args->priv_data_size) {
2082
2083                 ret = -EINVAL;
2084                 goto exit_unlock;
2085         }
2086
2087         /* each function will store private data inside priv_data and adjust priv_offset */
2088         ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset);
2089         if (ret)
2090                 goto exit_unlock;
2091
2092         ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices,
2093                                 (uint8_t __user *)args->priv_data, &priv_offset);
2094         if (ret)
2095                 goto exit_unlock;
2096
2097         /* Leave room for BOs in the private data. They need to be restored
2098          * before events, but we checkpoint them last to simplify the error
2099          * handling.
2100          */
2101         bo_priv_offset = priv_offset;
2102         priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data);
2103
2104         if (num_objects) {
2105                 ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data,
2106                                                  &priv_offset);
2107                 if (ret)
2108                         goto exit_unlock;
2109
2110                 ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data,
2111                                                  &priv_offset);
2112                 if (ret)
2113                         goto exit_unlock;
2114
2115                 ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset);
2116                 if (ret)
2117                         goto exit_unlock;
2118         }
2119
2120         /* This must be the last thing in this function that can fail.
2121          * Otherwise we leak dmabuf file descriptors.
2122          */
2123         ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos,
2124                            (uint8_t __user *)args->priv_data, &bo_priv_offset);
2125
2126 exit_unlock:
2127         mutex_unlock(&p->mutex);
2128         if (ret)
2129                 pr_err("Failed to dump CRIU ret:%d\n", ret);
2130         else
2131                 pr_debug("CRIU dump ret:%d\n", ret);
2132
2133         return ret;
2134 }
2135
2136 static int criu_restore_process(struct kfd_process *p,
2137                                 struct kfd_ioctl_criu_args *args,
2138                                 uint64_t *priv_offset,
2139                                 uint64_t max_priv_data_size)
2140 {
2141         int ret = 0;
2142         struct kfd_criu_process_priv_data process_priv;
2143
2144         if (*priv_offset + sizeof(process_priv) > max_priv_data_size)
2145                 return -EINVAL;
2146
2147         ret = copy_from_user(&process_priv,
2148                                 (void __user *)(args->priv_data + *priv_offset),
2149                                 sizeof(process_priv));
2150         if (ret) {
2151                 pr_err("Failed to copy process private information from user\n");
2152                 ret = -EFAULT;
2153                 goto exit;
2154         }
2155         *priv_offset += sizeof(process_priv);
2156
2157         if (process_priv.version != KFD_CRIU_PRIV_VERSION) {
2158                 pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n",
2159                         process_priv.version, KFD_CRIU_PRIV_VERSION);
2160                 return -EINVAL;
2161         }
2162
2163         pr_debug("Setting XNACK mode\n");
2164         if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) {
2165                 pr_err("xnack mode cannot be set\n");
2166                 ret = -EPERM;
2167                 goto exit;
2168         } else {
2169                 pr_debug("set xnack mode: %d\n", process_priv.xnack_mode);
2170                 p->xnack_enabled = process_priv.xnack_mode;
2171         }
2172
2173 exit:
2174         return ret;
2175 }
2176
2177 static int criu_restore_devices(struct kfd_process *p,
2178                                 struct kfd_ioctl_criu_args *args,
2179                                 uint64_t *priv_offset,
2180                                 uint64_t max_priv_data_size)
2181 {
2182         struct kfd_criu_device_bucket *device_buckets;
2183         struct kfd_criu_device_priv_data *device_privs;
2184         int ret = 0;
2185         uint32_t i;
2186
2187         if (args->num_devices != p->n_pdds)
2188                 return -EINVAL;
2189
2190         if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size)
2191                 return -EINVAL;
2192
2193         device_buckets = kmalloc_array(args->num_devices, sizeof(*device_buckets), GFP_KERNEL);
2194         if (!device_buckets)
2195                 return -ENOMEM;
2196
2197         ret = copy_from_user(device_buckets, (void __user *)args->devices,
2198                                 args->num_devices * sizeof(*device_buckets));
2199         if (ret) {
2200                 pr_err("Failed to copy devices buckets from user\n");
2201                 ret = -EFAULT;
2202                 goto exit;
2203         }
2204
2205         for (i = 0; i < args->num_devices; i++) {
2206                 struct kfd_node *dev;
2207                 struct kfd_process_device *pdd;
2208                 struct file *drm_file;
2209
2210                 /* device private data is not currently used */
2211
2212                 if (!device_buckets[i].user_gpu_id) {
2213                         pr_err("Invalid user gpu_id\n");
2214                         ret = -EINVAL;
2215                         goto exit;
2216                 }
2217
2218                 dev = kfd_device_by_id(device_buckets[i].actual_gpu_id);
2219                 if (!dev) {
2220                         pr_err("Failed to find device with gpu_id = %x\n",
2221                                 device_buckets[i].actual_gpu_id);
2222                         ret = -EINVAL;
2223                         goto exit;
2224                 }
2225
2226                 pdd = kfd_get_process_device_data(dev, p);
2227                 if (!pdd) {
2228                         pr_err("Failed to get pdd for gpu_id = %x\n",
2229                                         device_buckets[i].actual_gpu_id);
2230                         ret = -EINVAL;
2231                         goto exit;
2232                 }
2233                 pdd->user_gpu_id = device_buckets[i].user_gpu_id;
2234
2235                 drm_file = fget(device_buckets[i].drm_fd);
2236                 if (!drm_file) {
2237                         pr_err("Invalid render node file descriptor sent from plugin (%d)\n",
2238                                 device_buckets[i].drm_fd);
2239                         ret = -EINVAL;
2240                         goto exit;
2241                 }
2242
2243                 if (pdd->drm_file) {
2244                         ret = -EINVAL;
2245                         goto exit;
2246                 }
2247
2248                 /* create the vm using render nodes for kfd pdd */
2249                 if (kfd_process_device_init_vm(pdd, drm_file)) {
2250                         pr_err("could not init vm for given pdd\n");
2251                         /* On success, the PDD keeps the drm_file reference */
2252                         fput(drm_file);
2253                         ret = -EINVAL;
2254                         goto exit;
2255                 }
2256                 /*
2257                  * pdd now already has the vm bound to render node so below api won't create a new
2258                  * exclusive kfd mapping but use existing one with renderDXXX but is still needed
2259                  * for iommu v2 binding  and runtime pm.
2260                  */
2261                 pdd = kfd_bind_process_to_device(dev, p);
2262                 if (IS_ERR(pdd)) {
2263                         ret = PTR_ERR(pdd);
2264                         goto exit;
2265                 }
2266
2267                 if (!pdd->qpd.proc_doorbells) {
2268                         ret = kfd_alloc_process_doorbells(dev->kfd, pdd);
2269                         if (ret)
2270                                 goto exit;
2271                 }
2272         }
2273
2274         /*
2275          * We are not copying device private data from user as we are not using the data for now,
2276          * but we still adjust for its private data.
2277          */
2278         *priv_offset += args->num_devices * sizeof(*device_privs);
2279
2280 exit:
2281         kfree(device_buckets);
2282         return ret;
2283 }
2284
2285 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd,
2286                                       struct kfd_criu_bo_bucket *bo_bucket,
2287                                       struct kfd_criu_bo_priv_data *bo_priv,
2288                                       struct kgd_mem **kgd_mem)
2289 {
2290         int idr_handle;
2291         int ret;
2292         const bool criu_resume = true;
2293         u64 offset;
2294
2295         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
2296                 if (bo_bucket->size !=
2297                                 kfd_doorbell_process_slice(pdd->dev->kfd))
2298                         return -EINVAL;
2299
2300                 offset = kfd_get_process_doorbells(pdd);
2301                 if (!offset)
2302                         return -ENOMEM;
2303         } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2304                 /* MMIO BOs need remapped bus address */
2305                 if (bo_bucket->size != PAGE_SIZE) {
2306                         pr_err("Invalid page size\n");
2307                         return -EINVAL;
2308                 }
2309                 offset = pdd->dev->adev->rmmio_remap.bus_addr;
2310                 if (!offset) {
2311                         pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n");
2312                         return -ENOMEM;
2313                 }
2314         } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
2315                 offset = bo_priv->user_addr;
2316         }
2317         /* Create the BO */
2318         ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr,
2319                                                       bo_bucket->size, pdd->drm_priv, kgd_mem,
2320                                                       &offset, bo_bucket->alloc_flags, criu_resume);
2321         if (ret) {
2322                 pr_err("Could not create the BO\n");
2323                 return ret;
2324         }
2325         pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n",
2326                  bo_bucket->size, bo_bucket->addr, offset);
2327
2328         /* Restore previous IDR handle */
2329         pr_debug("Restoring old IDR handle for the BO");
2330         idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle,
2331                                bo_priv->idr_handle + 1, GFP_KERNEL);
2332
2333         if (idr_handle < 0) {
2334                 pr_err("Could not allocate idr\n");
2335                 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv,
2336                                                        NULL);
2337                 return -ENOMEM;
2338         }
2339
2340         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL)
2341                 bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id);
2342         if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) {
2343                 bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id);
2344         } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
2345                 bo_bucket->restored_offset = offset;
2346         } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
2347                 bo_bucket->restored_offset = offset;
2348                 /* Update the VRAM usage count */
2349                 WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + bo_bucket->size);
2350         }
2351         return 0;
2352 }
2353
2354 static int criu_restore_bo(struct kfd_process *p,
2355                            struct kfd_criu_bo_bucket *bo_bucket,
2356                            struct kfd_criu_bo_priv_data *bo_priv)
2357 {
2358         struct kfd_process_device *pdd;
2359         struct kgd_mem *kgd_mem;
2360         int ret;
2361         int j;
2362
2363         pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n",
2364                  bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags,
2365                  bo_priv->idr_handle);
2366
2367         pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id);
2368         if (!pdd) {
2369                 pr_err("Failed to get pdd\n");
2370                 return -ENODEV;
2371         }
2372
2373         ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem);
2374         if (ret)
2375                 return ret;
2376
2377         /* now map these BOs to GPU/s */
2378         for (j = 0; j < p->n_pdds; j++) {
2379                 struct kfd_node *peer;
2380                 struct kfd_process_device *peer_pdd;
2381
2382                 if (!bo_priv->mapped_gpuids[j])
2383                         break;
2384
2385                 peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]);
2386                 if (!peer_pdd)
2387                         return -EINVAL;
2388
2389                 peer = peer_pdd->dev;
2390
2391                 peer_pdd = kfd_bind_process_to_device(peer, p);
2392                 if (IS_ERR(peer_pdd))
2393                         return PTR_ERR(peer_pdd);
2394
2395                 ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem,
2396                                                             peer_pdd->drm_priv);
2397                 if (ret) {
2398                         pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds);
2399                         return ret;
2400                 }
2401         }
2402
2403         pr_debug("map memory was successful for the BO\n");
2404         /* create the dmabuf object and export the bo */
2405         if (bo_bucket->alloc_flags
2406             & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) {
2407                 ret = criu_get_prime_handle(kgd_mem, DRM_RDWR,
2408                                             &bo_bucket->dmabuf_fd);
2409                 if (ret)
2410                         return ret;
2411         } else {
2412                 bo_bucket->dmabuf_fd = KFD_INVALID_FD;
2413         }
2414
2415         return 0;
2416 }
2417
2418 static int criu_restore_bos(struct kfd_process *p,
2419                             struct kfd_ioctl_criu_args *args,
2420                             uint64_t *priv_offset,
2421                             uint64_t max_priv_data_size)
2422 {
2423         struct kfd_criu_bo_bucket *bo_buckets = NULL;
2424         struct kfd_criu_bo_priv_data *bo_privs = NULL;
2425         int ret = 0;
2426         uint32_t i = 0;
2427
2428         if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size)
2429                 return -EINVAL;
2430
2431         /* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */
2432         amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info);
2433
2434         bo_buckets = kvmalloc_array(args->num_bos, sizeof(*bo_buckets), GFP_KERNEL);
2435         if (!bo_buckets)
2436                 return -ENOMEM;
2437
2438         ret = copy_from_user(bo_buckets, (void __user *)args->bos,
2439                              args->num_bos * sizeof(*bo_buckets));
2440         if (ret) {
2441                 pr_err("Failed to copy BOs information from user\n");
2442                 ret = -EFAULT;
2443                 goto exit;
2444         }
2445
2446         bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), GFP_KERNEL);
2447         if (!bo_privs) {
2448                 ret = -ENOMEM;
2449                 goto exit;
2450         }
2451
2452         ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset,
2453                              args->num_bos * sizeof(*bo_privs));
2454         if (ret) {
2455                 pr_err("Failed to copy BOs information from user\n");
2456                 ret = -EFAULT;
2457                 goto exit;
2458         }
2459         *priv_offset += args->num_bos * sizeof(*bo_privs);
2460
2461         /* Create and map new BOs */
2462         for (; i < args->num_bos; i++) {
2463                 ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i]);
2464                 if (ret) {
2465                         pr_debug("Failed to restore BO[%d] ret%d\n", i, ret);
2466                         goto exit;
2467                 }
2468         } /* done */
2469
2470         /* Copy only the buckets back so user can read bo_buckets[N].restored_offset */
2471         ret = copy_to_user((void __user *)args->bos,
2472                                 bo_buckets,
2473                                 (args->num_bos * sizeof(*bo_buckets)));
2474         if (ret)
2475                 ret = -EFAULT;
2476
2477 exit:
2478         while (ret && i--) {
2479                 if (bo_buckets[i].alloc_flags
2480                    & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT))
2481                         close_fd(bo_buckets[i].dmabuf_fd);
2482         }
2483         kvfree(bo_buckets);
2484         kvfree(bo_privs);
2485         return ret;
2486 }
2487
2488 static int criu_restore_objects(struct file *filep,
2489                                 struct kfd_process *p,
2490                                 struct kfd_ioctl_criu_args *args,
2491                                 uint64_t *priv_offset,
2492                                 uint64_t max_priv_data_size)
2493 {
2494         int ret = 0;
2495         uint32_t i;
2496
2497         BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type));
2498         BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type));
2499         BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type));
2500
2501         for (i = 0; i < args->num_objects; i++) {
2502                 uint32_t object_type;
2503
2504                 if (*priv_offset + sizeof(object_type) > max_priv_data_size) {
2505                         pr_err("Invalid private data size\n");
2506                         return -EINVAL;
2507                 }
2508
2509                 ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset));
2510                 if (ret) {
2511                         pr_err("Failed to copy private information from user\n");
2512                         goto exit;
2513                 }
2514
2515                 switch (object_type) {
2516                 case KFD_CRIU_OBJECT_TYPE_QUEUE:
2517                         ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data,
2518                                                      priv_offset, max_priv_data_size);
2519                         if (ret)
2520                                 goto exit;
2521                         break;
2522                 case KFD_CRIU_OBJECT_TYPE_EVENT:
2523                         ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data,
2524                                                      priv_offset, max_priv_data_size);
2525                         if (ret)
2526                                 goto exit;
2527                         break;
2528                 case KFD_CRIU_OBJECT_TYPE_SVM_RANGE:
2529                         ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data,
2530                                                      priv_offset, max_priv_data_size);
2531                         if (ret)
2532                                 goto exit;
2533                         break;
2534                 default:
2535                         pr_err("Invalid object type:%u at index:%d\n", object_type, i);
2536                         ret = -EINVAL;
2537                         goto exit;
2538                 }
2539         }
2540 exit:
2541         return ret;
2542 }
2543
2544 static int criu_restore(struct file *filep,
2545                         struct kfd_process *p,
2546                         struct kfd_ioctl_criu_args *args)
2547 {
2548         uint64_t priv_offset = 0;
2549         int ret = 0;
2550
2551         pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n",
2552                  args->num_devices, args->num_bos, args->num_objects, args->priv_data_size);
2553
2554         if (!args->bos || !args->devices || !args->priv_data || !args->priv_data_size ||
2555             !args->num_devices || !args->num_bos)
2556                 return -EINVAL;
2557
2558         mutex_lock(&p->mutex);
2559
2560         /*
2561          * Set the process to evicted state to avoid running any new queues before all the memory
2562          * mappings are ready.
2563          */
2564         ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE);
2565         if (ret)
2566                 goto exit_unlock;
2567
2568         /* Each function will adjust priv_offset based on how many bytes they consumed */
2569         ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size);
2570         if (ret)
2571                 goto exit_unlock;
2572
2573         ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size);
2574         if (ret)
2575                 goto exit_unlock;
2576
2577         ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size);
2578         if (ret)
2579                 goto exit_unlock;
2580
2581         ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size);
2582         if (ret)
2583                 goto exit_unlock;
2584
2585         if (priv_offset != args->priv_data_size) {
2586                 pr_err("Invalid private data size\n");
2587                 ret = -EINVAL;
2588         }
2589
2590 exit_unlock:
2591         mutex_unlock(&p->mutex);
2592         if (ret)
2593                 pr_err("Failed to restore CRIU ret:%d\n", ret);
2594         else
2595                 pr_debug("CRIU restore successful\n");
2596
2597         return ret;
2598 }
2599
2600 static int criu_unpause(struct file *filep,
2601                         struct kfd_process *p,
2602                         struct kfd_ioctl_criu_args *args)
2603 {
2604         int ret;
2605
2606         mutex_lock(&p->mutex);
2607
2608         if (!p->queues_paused) {
2609                 mutex_unlock(&p->mutex);
2610                 return -EINVAL;
2611         }
2612
2613         ret = kfd_process_restore_queues(p);
2614         if (ret)
2615                 pr_err("Failed to unpause queues ret:%d\n", ret);
2616         else
2617                 p->queues_paused = false;
2618
2619         mutex_unlock(&p->mutex);
2620
2621         return ret;
2622 }
2623
2624 static int criu_resume(struct file *filep,
2625                         struct kfd_process *p,
2626                         struct kfd_ioctl_criu_args *args)
2627 {
2628         struct kfd_process *target = NULL;
2629         struct pid *pid = NULL;
2630         int ret = 0;
2631
2632         pr_debug("Inside %s, target pid for criu restore: %d\n", __func__,
2633                  args->pid);
2634
2635         pid = find_get_pid(args->pid);
2636         if (!pid) {
2637                 pr_err("Cannot find pid info for %i\n", args->pid);
2638                 return -ESRCH;
2639         }
2640
2641         pr_debug("calling kfd_lookup_process_by_pid\n");
2642         target = kfd_lookup_process_by_pid(pid);
2643
2644         put_pid(pid);
2645
2646         if (!target) {
2647                 pr_debug("Cannot find process info for %i\n", args->pid);
2648                 return -ESRCH;
2649         }
2650
2651         mutex_lock(&target->mutex);
2652         ret = kfd_criu_resume_svm(target);
2653         if (ret) {
2654                 pr_err("kfd_criu_resume_svm failed for %i\n", args->pid);
2655                 goto exit;
2656         }
2657
2658         ret =  amdgpu_amdkfd_criu_resume(target->kgd_process_info);
2659         if (ret)
2660                 pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid);
2661
2662 exit:
2663         mutex_unlock(&target->mutex);
2664
2665         kfd_unref_process(target);
2666         return ret;
2667 }
2668
2669 static int criu_process_info(struct file *filep,
2670                                 struct kfd_process *p,
2671                                 struct kfd_ioctl_criu_args *args)
2672 {
2673         int ret = 0;
2674
2675         mutex_lock(&p->mutex);
2676
2677         if (!p->n_pdds) {
2678                 pr_err("No pdd for given process\n");
2679                 ret = -ENODEV;
2680                 goto err_unlock;
2681         }
2682
2683         ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT);
2684         if (ret)
2685                 goto err_unlock;
2686
2687         p->queues_paused = true;
2688
2689         args->pid = task_pid_nr_ns(p->lead_thread,
2690                                         task_active_pid_ns(p->lead_thread));
2691
2692         ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos,
2693                                            &args->num_objects, &args->priv_data_size);
2694         if (ret)
2695                 goto err_unlock;
2696
2697         dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n",
2698                                 args->num_devices, args->num_bos, args->num_objects,
2699                                 args->priv_data_size);
2700
2701 err_unlock:
2702         if (ret) {
2703                 kfd_process_restore_queues(p);
2704                 p->queues_paused = false;
2705         }
2706         mutex_unlock(&p->mutex);
2707         return ret;
2708 }
2709
2710 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data)
2711 {
2712         struct kfd_ioctl_criu_args *args = data;
2713         int ret;
2714
2715         dev_dbg(kfd_device, "CRIU operation: %d\n", args->op);
2716         switch (args->op) {
2717         case KFD_CRIU_OP_PROCESS_INFO:
2718                 ret = criu_process_info(filep, p, args);
2719                 break;
2720         case KFD_CRIU_OP_CHECKPOINT:
2721                 ret = criu_checkpoint(filep, p, args);
2722                 break;
2723         case KFD_CRIU_OP_UNPAUSE:
2724                 ret = criu_unpause(filep, p, args);
2725                 break;
2726         case KFD_CRIU_OP_RESTORE:
2727                 ret = criu_restore(filep, p, args);
2728                 break;
2729         case KFD_CRIU_OP_RESUME:
2730                 ret = criu_resume(filep, p, args);
2731                 break;
2732         default:
2733                 dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op);
2734                 ret = -EINVAL;
2735                 break;
2736         }
2737
2738         if (ret)
2739                 dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret);
2740
2741         return ret;
2742 }
2743
2744 static int runtime_enable(struct kfd_process *p, uint64_t r_debug,
2745                         bool enable_ttmp_setup)
2746 {
2747         int i = 0, ret = 0;
2748
2749         if (p->is_runtime_retry)
2750                 goto retry;
2751
2752         if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
2753                 return -EBUSY;
2754
2755         for (i = 0; i < p->n_pdds; i++) {
2756                 struct kfd_process_device *pdd = p->pdds[i];
2757
2758                 if (pdd->qpd.queue_count)
2759                         return -EEXIST;
2760
2761                 /*
2762                  * Setup TTMPs by default.
2763                  * Note that this call must remain here for MES ADD QUEUE to
2764                  * skip_process_ctx_clear unconditionally as the first call to
2765                  * SET_SHADER_DEBUGGER clears any stale process context data
2766                  * saved in MES.
2767                  */
2768                 if (pdd->dev->kfd->shared_resources.enable_mes)
2769                         kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
2770         }
2771
2772         p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
2773         p->runtime_info.r_debug = r_debug;
2774         p->runtime_info.ttmp_setup = enable_ttmp_setup;
2775
2776         if (p->runtime_info.ttmp_setup) {
2777                 for (i = 0; i < p->n_pdds; i++) {
2778                         struct kfd_process_device *pdd = p->pdds[i];
2779
2780                         if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) {
2781                                 amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
2782                                 pdd->dev->kfd2kgd->enable_debug_trap(
2783                                                 pdd->dev->adev,
2784                                                 true,
2785                                                 pdd->dev->vm_info.last_vmid_kfd);
2786                         } else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2787                                 pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
2788                                                 pdd->dev->adev,
2789                                                 false,
2790                                                 0);
2791                         }
2792                 }
2793         }
2794
2795 retry:
2796         if (p->debug_trap_enabled) {
2797                 if (!p->is_runtime_retry) {
2798                         kfd_dbg_trap_activate(p);
2799                         kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2800                                         p, NULL, 0, false, NULL, 0);
2801                 }
2802
2803                 mutex_unlock(&p->mutex);
2804                 ret = down_interruptible(&p->runtime_enable_sema);
2805                 mutex_lock(&p->mutex);
2806
2807                 p->is_runtime_retry = !!ret;
2808         }
2809
2810         return ret;
2811 }
2812
2813 static int runtime_disable(struct kfd_process *p)
2814 {
2815         int i = 0, ret;
2816         bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED;
2817
2818         p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED;
2819         p->runtime_info.r_debug = 0;
2820
2821         if (p->debug_trap_enabled) {
2822                 if (was_enabled)
2823                         kfd_dbg_trap_deactivate(p, false, 0);
2824
2825                 if (!p->is_runtime_retry)
2826                         kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME),
2827                                         p, NULL, 0, false, NULL, 0);
2828
2829                 mutex_unlock(&p->mutex);
2830                 ret = down_interruptible(&p->runtime_enable_sema);
2831                 mutex_lock(&p->mutex);
2832
2833                 p->is_runtime_retry = !!ret;
2834                 if (ret)
2835                         return ret;
2836         }
2837
2838         if (was_enabled && p->runtime_info.ttmp_setup) {
2839                 for (i = 0; i < p->n_pdds; i++) {
2840                         struct kfd_process_device *pdd = p->pdds[i];
2841
2842                         if (!kfd_dbg_is_rlc_restore_supported(pdd->dev))
2843                                 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
2844                 }
2845         }
2846
2847         p->runtime_info.ttmp_setup = false;
2848
2849         /* disable ttmp setup */
2850         for (i = 0; i < p->n_pdds; i++) {
2851                 struct kfd_process_device *pdd = p->pdds[i];
2852
2853                 if (kfd_dbg_is_per_vmid_supported(pdd->dev)) {
2854                         pdd->spi_dbg_override =
2855                                         pdd->dev->kfd2kgd->disable_debug_trap(
2856                                         pdd->dev->adev,
2857                                         false,
2858                                         pdd->dev->vm_info.last_vmid_kfd);
2859
2860                         if (!pdd->dev->kfd->shared_resources.enable_mes)
2861                                 debug_refresh_runlist(pdd->dev->dqm);
2862                         else
2863                                 kfd_dbg_set_mes_debug_mode(pdd,
2864                                                            !kfd_dbg_has_cwsr_workaround(pdd->dev));
2865                 }
2866         }
2867
2868         return 0;
2869 }
2870
2871 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data)
2872 {
2873         struct kfd_ioctl_runtime_enable_args *args = data;
2874         int r;
2875
2876         mutex_lock(&p->mutex);
2877
2878         if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK)
2879                 r = runtime_enable(p, args->r_debug,
2880                                 !!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK));
2881         else
2882                 r = runtime_disable(p);
2883
2884         mutex_unlock(&p->mutex);
2885
2886         return r;
2887 }
2888
2889 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data)
2890 {
2891         struct kfd_ioctl_dbg_trap_args *args = data;
2892         struct task_struct *thread = NULL;
2893         struct mm_struct *mm = NULL;
2894         struct pid *pid = NULL;
2895         struct kfd_process *target = NULL;
2896         struct kfd_process_device *pdd = NULL;
2897         int r = 0;
2898
2899         if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2900                 pr_err("Debugging does not support sched_policy %i", sched_policy);
2901                 return -EINVAL;
2902         }
2903
2904         pid = find_get_pid(args->pid);
2905         if (!pid) {
2906                 pr_debug("Cannot find pid info for %i\n", args->pid);
2907                 r = -ESRCH;
2908                 goto out;
2909         }
2910
2911         thread = get_pid_task(pid, PIDTYPE_PID);
2912         if (!thread) {
2913                 r = -ESRCH;
2914                 goto out;
2915         }
2916
2917         mm = get_task_mm(thread);
2918         if (!mm) {
2919                 r = -ESRCH;
2920                 goto out;
2921         }
2922
2923         if (args->op == KFD_IOC_DBG_TRAP_ENABLE) {
2924                 bool create_process;
2925
2926                 rcu_read_lock();
2927                 create_process = thread && thread != current && ptrace_parent(thread) == current;
2928                 rcu_read_unlock();
2929
2930                 target = create_process ? kfd_create_process(thread) :
2931                                         kfd_lookup_process_by_pid(pid);
2932         } else {
2933                 target = kfd_lookup_process_by_pid(pid);
2934         }
2935
2936         if (IS_ERR_OR_NULL(target)) {
2937                 pr_debug("Cannot find process PID %i to debug\n", args->pid);
2938                 r = target ? PTR_ERR(target) : -ESRCH;
2939                 goto out;
2940         }
2941
2942         /* Check if target is still PTRACED. */
2943         rcu_read_lock();
2944         if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE
2945                                 && ptrace_parent(target->lead_thread) != current) {
2946                 pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid);
2947                 r = -EPERM;
2948         }
2949         rcu_read_unlock();
2950
2951         if (r)
2952                 goto out;
2953
2954         mutex_lock(&target->mutex);
2955
2956         if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) {
2957                 pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op);
2958                 r = -EINVAL;
2959                 goto unlock_out;
2960         }
2961
2962         if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED &&
2963                         (args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE ||
2964                          args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE ||
2965                          args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES ||
2966                          args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES ||
2967                          args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2968                          args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH ||
2969                          args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) {
2970                 r = -EPERM;
2971                 goto unlock_out;
2972         }
2973
2974         if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
2975             args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
2976                 int user_gpu_id = kfd_process_get_user_gpu_id(target,
2977                                 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
2978                                         args->set_node_address_watch.gpu_id :
2979                                         args->clear_node_address_watch.gpu_id);
2980
2981                 pdd = kfd_process_device_data_by_id(target, user_gpu_id);
2982                 if (user_gpu_id == -EINVAL || !pdd) {
2983                         r = -ENODEV;
2984                         goto unlock_out;
2985                 }
2986         }
2987
2988         switch (args->op) {
2989         case KFD_IOC_DBG_TRAP_ENABLE:
2990                 if (target != p)
2991                         target->debugger_process = p;
2992
2993                 r = kfd_dbg_trap_enable(target,
2994                                         args->enable.dbg_fd,
2995                                         (void __user *)args->enable.rinfo_ptr,
2996                                         &args->enable.rinfo_size);
2997                 if (!r)
2998                         target->exception_enable_mask = args->enable.exception_mask;
2999
3000                 break;
3001         case KFD_IOC_DBG_TRAP_DISABLE:
3002                 r = kfd_dbg_trap_disable(target);
3003                 break;
3004         case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT:
3005                 r = kfd_dbg_send_exception_to_runtime(target,
3006                                 args->send_runtime_event.gpu_id,
3007                                 args->send_runtime_event.queue_id,
3008                                 args->send_runtime_event.exception_mask);
3009                 break;
3010         case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED:
3011                 kfd_dbg_set_enabled_debug_exception_mask(target,
3012                                 args->set_exceptions_enabled.exception_mask);
3013                 break;
3014         case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
3015                 r = kfd_dbg_trap_set_wave_launch_override(target,
3016                                 args->launch_override.override_mode,
3017                                 args->launch_override.enable_mask,
3018                                 args->launch_override.support_request_mask,
3019                                 &args->launch_override.enable_mask,
3020                                 &args->launch_override.support_request_mask);
3021                 break;
3022         case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
3023                 r = kfd_dbg_trap_set_wave_launch_mode(target,
3024                                 args->launch_mode.launch_mode);
3025                 break;
3026         case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES:
3027                 r = suspend_queues(target,
3028                                 args->suspend_queues.num_queues,
3029                                 args->suspend_queues.grace_period,
3030                                 args->suspend_queues.exception_mask,
3031                                 (uint32_t *)args->suspend_queues.queue_array_ptr);
3032
3033                 break;
3034         case KFD_IOC_DBG_TRAP_RESUME_QUEUES:
3035                 r = resume_queues(target, args->resume_queues.num_queues,
3036                                 (uint32_t *)args->resume_queues.queue_array_ptr);
3037                 break;
3038         case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
3039                 r = kfd_dbg_trap_set_dev_address_watch(pdd,
3040                                 args->set_node_address_watch.address,
3041                                 args->set_node_address_watch.mask,
3042                                 &args->set_node_address_watch.id,
3043                                 args->set_node_address_watch.mode);
3044                 break;
3045         case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
3046                 r = kfd_dbg_trap_clear_dev_address_watch(pdd,
3047                                 args->clear_node_address_watch.id);
3048                 break;
3049         case KFD_IOC_DBG_TRAP_SET_FLAGS:
3050                 r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags);
3051                 break;
3052         case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
3053                 r = kfd_dbg_ev_query_debug_event(target,
3054                                 &args->query_debug_event.queue_id,
3055                                 &args->query_debug_event.gpu_id,
3056                                 args->query_debug_event.exception_mask,
3057                                 &args->query_debug_event.exception_mask);
3058                 break;
3059         case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:
3060                 r = kfd_dbg_trap_query_exception_info(target,
3061                                 args->query_exception_info.source_id,
3062                                 args->query_exception_info.exception_code,
3063                                 args->query_exception_info.clear_exception,
3064                                 (void __user *)args->query_exception_info.info_ptr,
3065                                 &args->query_exception_info.info_size);
3066                 break;
3067         case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
3068                 r = pqm_get_queue_snapshot(&target->pqm,
3069                                 args->queue_snapshot.exception_mask,
3070                                 (void __user *)args->queue_snapshot.snapshot_buf_ptr,
3071                                 &args->queue_snapshot.num_queues,
3072                                 &args->queue_snapshot.entry_size);
3073                 break;
3074         case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT:
3075                 r = kfd_dbg_trap_device_snapshot(target,
3076                                 args->device_snapshot.exception_mask,
3077                                 (void __user *)args->device_snapshot.snapshot_buf_ptr,
3078                                 &args->device_snapshot.num_devices,
3079                                 &args->device_snapshot.entry_size);
3080                 break;
3081         default:
3082                 pr_err("Invalid option: %i\n", args->op);
3083                 r = -EINVAL;
3084         }
3085
3086 unlock_out:
3087         mutex_unlock(&target->mutex);
3088
3089 out:
3090         if (thread)
3091                 put_task_struct(thread);
3092
3093         if (mm)
3094                 mmput(mm);
3095
3096         if (pid)
3097                 put_pid(pid);
3098
3099         if (target)
3100                 kfd_unref_process(target);
3101
3102         return r;
3103 }
3104
3105 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \
3106         [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \
3107                             .cmd_drv = 0, .name = #ioctl}
3108
3109 /** Ioctl table */
3110 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
3111         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION,
3112                         kfd_ioctl_get_version, 0),
3113
3114         AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE,
3115                         kfd_ioctl_create_queue, 0),
3116
3117         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE,
3118                         kfd_ioctl_destroy_queue, 0),
3119
3120         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY,
3121                         kfd_ioctl_set_memory_policy, 0),
3122
3123         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS,
3124                         kfd_ioctl_get_clock_counters, 0),
3125
3126         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES,
3127                         kfd_ioctl_get_process_apertures, 0),
3128
3129         AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE,
3130                         kfd_ioctl_update_queue, 0),
3131
3132         AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT,
3133                         kfd_ioctl_create_event, 0),
3134
3135         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT,
3136                         kfd_ioctl_destroy_event, 0),
3137
3138         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT,
3139                         kfd_ioctl_set_event, 0),
3140
3141         AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT,
3142                         kfd_ioctl_reset_event, 0),
3143
3144         AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS,
3145                         kfd_ioctl_wait_events, 0),
3146
3147         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED,
3148                         kfd_ioctl_dbg_register, 0),
3149
3150         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED,
3151                         kfd_ioctl_dbg_unregister, 0),
3152
3153         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED,
3154                         kfd_ioctl_dbg_address_watch, 0),
3155
3156         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED,
3157                         kfd_ioctl_dbg_wave_control, 0),
3158
3159         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA,
3160                         kfd_ioctl_set_scratch_backing_va, 0),
3161
3162         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
3163                         kfd_ioctl_get_tile_config, 0),
3164
3165         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
3166                         kfd_ioctl_set_trap_handler, 0),
3167
3168         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW,
3169                         kfd_ioctl_get_process_apertures_new, 0),
3170
3171         AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM,
3172                         kfd_ioctl_acquire_vm, 0),
3173
3174         AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU,
3175                         kfd_ioctl_alloc_memory_of_gpu, 0),
3176
3177         AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU,
3178                         kfd_ioctl_free_memory_of_gpu, 0),
3179
3180         AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU,
3181                         kfd_ioctl_map_memory_to_gpu, 0),
3182
3183         AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU,
3184                         kfd_ioctl_unmap_memory_from_gpu, 0),
3185
3186         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK,
3187                         kfd_ioctl_set_cu_mask, 0),
3188
3189         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE,
3190                         kfd_ioctl_get_queue_wave_state, 0),
3191
3192         AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO,
3193                                 kfd_ioctl_get_dmabuf_info, 0),
3194
3195         AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
3196                                 kfd_ioctl_import_dmabuf, 0),
3197
3198         AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
3199                         kfd_ioctl_alloc_queue_gws, 0),
3200
3201         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS,
3202                         kfd_ioctl_smi_events, 0),
3203
3204         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
3205
3206         AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
3207                         kfd_ioctl_set_xnack_mode, 0),
3208
3209         AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP,
3210                         kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE),
3211
3212         AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY,
3213                         kfd_ioctl_get_available_memory, 0),
3214
3215         AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
3216                                 kfd_ioctl_export_dmabuf, 0),
3217
3218         AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE,
3219                         kfd_ioctl_runtime_enable, 0),
3220
3221         AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP,
3222                         kfd_ioctl_set_debug_trap, 0),
3223 };
3224
3225 #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)
3226
3227 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
3228 {
3229         struct kfd_process *process;
3230         amdkfd_ioctl_t *func;
3231         const struct amdkfd_ioctl_desc *ioctl = NULL;
3232         unsigned int nr = _IOC_NR(cmd);
3233         char stack_kdata[128];
3234         char *kdata = NULL;
3235         unsigned int usize, asize;
3236         int retcode = -EINVAL;
3237         bool ptrace_attached = false;
3238
3239         if (nr >= AMDKFD_CORE_IOCTL_COUNT)
3240                 goto err_i1;
3241
3242         if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) {
3243                 u32 amdkfd_size;
3244
3245                 ioctl = &amdkfd_ioctls[nr];
3246
3247                 amdkfd_size = _IOC_SIZE(ioctl->cmd);
3248                 usize = asize = _IOC_SIZE(cmd);
3249                 if (amdkfd_size > asize)
3250                         asize = amdkfd_size;
3251
3252                 cmd = ioctl->cmd;
3253         } else
3254                 goto err_i1;
3255
3256         dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg);
3257
3258         /* Get the process struct from the filep. Only the process
3259          * that opened /dev/kfd can use the file descriptor. Child
3260          * processes need to create their own KFD device context.
3261          */
3262         process = filep->private_data;
3263
3264         rcu_read_lock();
3265         if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) &&
3266             ptrace_parent(process->lead_thread) == current)
3267                 ptrace_attached = true;
3268         rcu_read_unlock();
3269
3270         if (process->lead_thread != current->group_leader
3271             && !ptrace_attached) {
3272                 dev_dbg(kfd_device, "Using KFD FD in wrong process\n");
3273                 retcode = -EBADF;
3274                 goto err_i1;
3275         }
3276
3277         /* Do not trust userspace, use our own definition */
3278         func = ioctl->func;
3279
3280         if (unlikely(!func)) {
3281                 dev_dbg(kfd_device, "no function\n");
3282                 retcode = -EINVAL;
3283                 goto err_i1;
3284         }
3285
3286         /*
3287          * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support
3288          * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a
3289          * more priviledged access.
3290          */
3291         if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) {
3292                 if (!capable(CAP_CHECKPOINT_RESTORE) &&
3293                                                 !capable(CAP_SYS_ADMIN)) {
3294                         retcode = -EACCES;
3295                         goto err_i1;
3296                 }
3297         }
3298
3299         if (cmd & (IOC_IN | IOC_OUT)) {
3300                 if (asize <= sizeof(stack_kdata)) {
3301                         kdata = stack_kdata;
3302                 } else {
3303                         kdata = kmalloc(asize, GFP_KERNEL);
3304                         if (!kdata) {
3305                                 retcode = -ENOMEM;
3306                                 goto err_i1;
3307                         }
3308                 }
3309                 if (asize > usize)
3310                         memset(kdata + usize, 0, asize - usize);
3311         }
3312
3313         if (cmd & IOC_IN) {
3314                 if (copy_from_user(kdata, (void __user *)arg, usize) != 0) {
3315                         retcode = -EFAULT;
3316                         goto err_i1;
3317                 }
3318         } else if (cmd & IOC_OUT) {
3319                 memset(kdata, 0, usize);
3320         }
3321
3322         retcode = func(filep, process, kdata);
3323
3324         if (cmd & IOC_OUT)
3325                 if (copy_to_user((void __user *)arg, kdata, usize) != 0)
3326                         retcode = -EFAULT;
3327
3328 err_i1:
3329         if (!ioctl)
3330                 dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n",
3331                           task_pid_nr(current), cmd, nr);
3332
3333         if (kdata != stack_kdata)
3334                 kfree(kdata);
3335
3336         if (retcode)
3337                 dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n",
3338                                 nr, arg, retcode);
3339
3340         return retcode;
3341 }
3342
3343 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process,
3344                       struct vm_area_struct *vma)
3345 {
3346         phys_addr_t address;
3347
3348         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3349                 return -EINVAL;
3350
3351         address = dev->adev->rmmio_remap.bus_addr;
3352
3353         vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
3354                                 VM_DONTDUMP | VM_PFNMAP);
3355
3356         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3357
3358         pr_debug("pasid 0x%x mapping mmio page\n"
3359                  "     target user address == 0x%08llX\n"
3360                  "     physical address    == 0x%08llX\n"
3361                  "     vm_flags            == 0x%04lX\n"
3362                  "     size                == 0x%04lX\n",
3363                  process->pasid, (unsigned long long) vma->vm_start,
3364                  address, vma->vm_flags, PAGE_SIZE);
3365
3366         return io_remap_pfn_range(vma,
3367                                 vma->vm_start,
3368                                 address >> PAGE_SHIFT,
3369                                 PAGE_SIZE,
3370                                 vma->vm_page_prot);
3371 }
3372
3373
3374 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
3375 {
3376         struct kfd_process *process;
3377         struct kfd_node *dev = NULL;
3378         unsigned long mmap_offset;
3379         unsigned int gpu_id;
3380
3381         process = kfd_get_process(current);
3382         if (IS_ERR(process))
3383                 return PTR_ERR(process);
3384
3385         mmap_offset = vma->vm_pgoff << PAGE_SHIFT;
3386         gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset);
3387         if (gpu_id)
3388                 dev = kfd_device_by_id(gpu_id);
3389
3390         switch (mmap_offset & KFD_MMAP_TYPE_MASK) {
3391         case KFD_MMAP_TYPE_DOORBELL:
3392                 if (!dev)
3393                         return -ENODEV;
3394                 return kfd_doorbell_mmap(dev, process, vma);
3395
3396         case KFD_MMAP_TYPE_EVENTS:
3397                 return kfd_event_mmap(process, vma);
3398
3399         case KFD_MMAP_TYPE_RESERVED_MEM:
3400                 if (!dev)
3401                         return -ENODEV;
3402                 return kfd_reserved_mem_mmap(dev, process, vma);
3403         case KFD_MMAP_TYPE_MMIO:
3404                 if (!dev)
3405                         return -ENODEV;
3406                 return kfd_mmio_mmap(dev, process, vma);
3407         }
3408
3409         return -EFAULT;
3410 }