1 /* 2 * Copyright 2014 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 */ 22 23 #include <linux/device.h> 24 #include <linux/export.h> 25 #include <linux/err.h> 26 #include <linux/fs.h> 27 #include <linux/file.h> 28 #include <linux/sched.h> 29 #include <linux/slab.h> 30 #include <linux/uaccess.h> 31 #include <linux/compat.h> 32 #include <uapi/linux/kfd_ioctl.h> 33 #include <linux/time.h> 34 #include <linux/mm.h> 35 #include <linux/mman.h> 36 #include <linux/dma-buf.h> 37 #include <asm/processor.h> 38 #include "kfd_priv.h" 39 #include "kfd_device_queue_manager.h" 40 #include "kfd_dbgmgr.h" 41 #include "amdgpu_amdkfd.h" 42 #include "kfd_smi_events.h" 43 44 static long kfd_ioctl(struct file *, unsigned int, unsigned long); 45 static int kfd_open(struct inode *, struct file *); 46 static int kfd_release(struct inode *, struct file *); 47 static int kfd_mmap(struct file *, struct vm_area_struct *); 48 49 static const char kfd_dev_name[] = "kfd"; 50 51 static const struct file_operations kfd_fops = { 52 .owner = THIS_MODULE, 53 .unlocked_ioctl = kfd_ioctl, 54 .compat_ioctl = compat_ptr_ioctl, 55 .open = kfd_open, 56 .release = kfd_release, 57 .mmap = kfd_mmap, 58 }; 59 60 static int kfd_char_dev_major = -1; 61 static struct class *kfd_class; 62 struct device *kfd_device; 63 64 int kfd_chardev_init(void) 65 { 66 int err = 0; 67 68 kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops); 69 err = kfd_char_dev_major; 70 if (err < 0) 71 goto err_register_chrdev; 72 73 kfd_class = class_create(THIS_MODULE, kfd_dev_name); 74 err = PTR_ERR(kfd_class); 75 if (IS_ERR(kfd_class)) 76 goto err_class_create; 77 78 kfd_device = device_create(kfd_class, NULL, 79 MKDEV(kfd_char_dev_major, 0), 80 NULL, kfd_dev_name); 81 err = PTR_ERR(kfd_device); 82 if (IS_ERR(kfd_device)) 83 goto err_device_create; 84 85 return 0; 86 87 err_device_create: 88 class_destroy(kfd_class); 89 err_class_create: 90 unregister_chrdev(kfd_char_dev_major, kfd_dev_name); 91 err_register_chrdev: 92 return err; 93 } 94 95 void kfd_chardev_exit(void) 96 { 97 device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0)); 98 class_destroy(kfd_class); 99 unregister_chrdev(kfd_char_dev_major, kfd_dev_name); 100 kfd_device = NULL; 101 } 102 103 struct device *kfd_chardev(void) 104 { 105 return kfd_device; 106 } 107 108 109 static int kfd_open(struct inode *inode, struct file *filep) 110 { 111 struct kfd_process *process; 112 bool is_32bit_user_mode; 113 114 if (iminor(inode) != 0) 115 return -ENODEV; 116 117 is_32bit_user_mode = in_compat_syscall(); 118 119 if (is_32bit_user_mode) { 120 dev_warn(kfd_device, 121 "Process %d (32-bit) failed to open /dev/kfd\n" 122 "32-bit processes are not supported by amdkfd\n", 123 current->pid); 124 return -EPERM; 125 } 126 127 process = kfd_create_process(filep); 128 if (IS_ERR(process)) 129 return PTR_ERR(process); 130 131 if (kfd_is_locked()) { 132 dev_dbg(kfd_device, "kfd is locked!\n" 133 "process %d unreferenced", process->pasid); 134 kfd_unref_process(process); 135 return -EAGAIN; 136 } 137 138 /* filep now owns the reference returned by kfd_create_process */ 139 filep->private_data = process; 140 141 dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n", 142 process->pasid, process->is_32bit_user_mode); 143 144 return 0; 145 } 146 147 static int kfd_release(struct inode *inode, struct file *filep) 148 { 149 struct kfd_process *process = filep->private_data; 150 151 if (process) 152 kfd_unref_process(process); 153 154 return 0; 155 } 156 157 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p, 158 void *data) 159 { 160 struct kfd_ioctl_get_version_args *args = data; 161 162 args->major_version = KFD_IOCTL_MAJOR_VERSION; 163 args->minor_version = KFD_IOCTL_MINOR_VERSION; 164 165 return 0; 166 } 167 168 static int set_queue_properties_from_user(struct queue_properties *q_properties, 169 struct kfd_ioctl_create_queue_args *args) 170 { 171 if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) { 172 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n"); 173 return -EINVAL; 174 } 175 176 if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) { 177 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n"); 178 return -EINVAL; 179 } 180 181 if ((args->ring_base_address) && 182 (!access_ok((const void __user *) args->ring_base_address, 183 sizeof(uint64_t)))) { 184 pr_err("Can't access ring base address\n"); 185 return -EFAULT; 186 } 187 188 if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) { 189 pr_err("Ring size must be a power of 2 or 0\n"); 190 return -EINVAL; 191 } 192 193 if (!access_ok((const void __user *) args->read_pointer_address, 194 sizeof(uint32_t))) { 195 pr_err("Can't access read pointer\n"); 196 return -EFAULT; 197 } 198 199 if (!access_ok((const void __user *) args->write_pointer_address, 200 sizeof(uint32_t))) { 201 pr_err("Can't access write pointer\n"); 202 return -EFAULT; 203 } 204 205 if (args->eop_buffer_address && 206 !access_ok((const void __user *) args->eop_buffer_address, 207 sizeof(uint32_t))) { 208 pr_debug("Can't access eop buffer"); 209 return -EFAULT; 210 } 211 212 if (args->ctx_save_restore_address && 213 !access_ok((const void __user *) args->ctx_save_restore_address, 214 sizeof(uint32_t))) { 215 pr_debug("Can't access ctx save restore buffer"); 216 return -EFAULT; 217 } 218 219 q_properties->is_interop = false; 220 q_properties->is_gws = false; 221 q_properties->queue_percent = args->queue_percentage; 222 q_properties->priority = args->queue_priority; 223 q_properties->queue_address = args->ring_base_address; 224 q_properties->queue_size = args->ring_size; 225 q_properties->read_ptr = (uint32_t *) args->read_pointer_address; 226 q_properties->write_ptr = (uint32_t *) args->write_pointer_address; 227 q_properties->eop_ring_buffer_address = args->eop_buffer_address; 228 q_properties->eop_ring_buffer_size = args->eop_buffer_size; 229 q_properties->ctx_save_restore_area_address = 230 args->ctx_save_restore_address; 231 q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size; 232 q_properties->ctl_stack_size = args->ctl_stack_size; 233 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE || 234 args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL) 235 q_properties->type = KFD_QUEUE_TYPE_COMPUTE; 236 else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA) 237 q_properties->type = KFD_QUEUE_TYPE_SDMA; 238 else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI) 239 q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI; 240 else 241 return -ENOTSUPP; 242 243 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL) 244 q_properties->format = KFD_QUEUE_FORMAT_AQL; 245 else 246 q_properties->format = KFD_QUEUE_FORMAT_PM4; 247 248 pr_debug("Queue Percentage: %d, %d\n", 249 q_properties->queue_percent, args->queue_percentage); 250 251 pr_debug("Queue Priority: %d, %d\n", 252 q_properties->priority, args->queue_priority); 253 254 pr_debug("Queue Address: 0x%llX, 0x%llX\n", 255 q_properties->queue_address, args->ring_base_address); 256 257 pr_debug("Queue Size: 0x%llX, %u\n", 258 q_properties->queue_size, args->ring_size); 259 260 pr_debug("Queue r/w Pointers: %px, %px\n", 261 q_properties->read_ptr, 262 q_properties->write_ptr); 263 264 pr_debug("Queue Format: %d\n", q_properties->format); 265 266 pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address); 267 268 pr_debug("Queue CTX save area: 0x%llX\n", 269 q_properties->ctx_save_restore_area_address); 270 271 return 0; 272 } 273 274 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, 275 void *data) 276 { 277 struct kfd_ioctl_create_queue_args *args = data; 278 struct kfd_dev *dev; 279 int err = 0; 280 unsigned int queue_id; 281 struct kfd_process_device *pdd; 282 struct queue_properties q_properties; 283 uint32_t doorbell_offset_in_process = 0; 284 285 memset(&q_properties, 0, sizeof(struct queue_properties)); 286 287 pr_debug("Creating queue ioctl\n"); 288 289 err = set_queue_properties_from_user(&q_properties, args); 290 if (err) 291 return err; 292 293 pr_debug("Looking for gpu id 0x%x\n", args->gpu_id); 294 dev = kfd_device_by_id(args->gpu_id); 295 if (!dev) { 296 pr_debug("Could not find gpu id 0x%x\n", args->gpu_id); 297 return -EINVAL; 298 } 299 300 mutex_lock(&p->mutex); 301 302 pdd = kfd_bind_process_to_device(dev, p); 303 if (IS_ERR(pdd)) { 304 err = -ESRCH; 305 goto err_bind_process; 306 } 307 308 pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n", 309 p->pasid, 310 dev->id); 311 312 err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, 313 &doorbell_offset_in_process); 314 if (err != 0) 315 goto err_create_queue; 316 317 args->queue_id = queue_id; 318 319 320 /* Return gpu_id as doorbell offset for mmap usage */ 321 args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL; 322 args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id); 323 if (KFD_IS_SOC15(dev->device_info->asic_family)) 324 /* On SOC15 ASICs, include the doorbell offset within the 325 * process doorbell frame, which is 2 pages. 326 */ 327 args->doorbell_offset |= doorbell_offset_in_process; 328 329 mutex_unlock(&p->mutex); 330 331 pr_debug("Queue id %d was created successfully\n", args->queue_id); 332 333 pr_debug("Ring buffer address == 0x%016llX\n", 334 args->ring_base_address); 335 336 pr_debug("Read ptr address == 0x%016llX\n", 337 args->read_pointer_address); 338 339 pr_debug("Write ptr address == 0x%016llX\n", 340 args->write_pointer_address); 341 342 return 0; 343 344 err_create_queue: 345 err_bind_process: 346 mutex_unlock(&p->mutex); 347 return err; 348 } 349 350 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p, 351 void *data) 352 { 353 int retval; 354 struct kfd_ioctl_destroy_queue_args *args = data; 355 356 pr_debug("Destroying queue id %d for pasid 0x%x\n", 357 args->queue_id, 358 p->pasid); 359 360 mutex_lock(&p->mutex); 361 362 retval = pqm_destroy_queue(&p->pqm, args->queue_id); 363 364 mutex_unlock(&p->mutex); 365 return retval; 366 } 367 368 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p, 369 void *data) 370 { 371 int retval; 372 struct kfd_ioctl_update_queue_args *args = data; 373 struct queue_properties properties; 374 375 if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) { 376 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n"); 377 return -EINVAL; 378 } 379 380 if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) { 381 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n"); 382 return -EINVAL; 383 } 384 385 if ((args->ring_base_address) && 386 (!access_ok((const void __user *) args->ring_base_address, 387 sizeof(uint64_t)))) { 388 pr_err("Can't access ring base address\n"); 389 return -EFAULT; 390 } 391 392 if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) { 393 pr_err("Ring size must be a power of 2 or 0\n"); 394 return -EINVAL; 395 } 396 397 properties.queue_address = args->ring_base_address; 398 properties.queue_size = args->ring_size; 399 properties.queue_percent = args->queue_percentage; 400 properties.priority = args->queue_priority; 401 402 pr_debug("Updating queue id %d for pasid 0x%x\n", 403 args->queue_id, p->pasid); 404 405 mutex_lock(&p->mutex); 406 407 retval = pqm_update_queue(&p->pqm, args->queue_id, &properties); 408 409 mutex_unlock(&p->mutex); 410 411 return retval; 412 } 413 414 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p, 415 void *data) 416 { 417 int retval; 418 const int max_num_cus = 1024; 419 struct kfd_ioctl_set_cu_mask_args *args = data; 420 struct queue_properties properties; 421 uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr; 422 size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32); 423 424 if ((args->num_cu_mask % 32) != 0) { 425 pr_debug("num_cu_mask 0x%x must be a multiple of 32", 426 args->num_cu_mask); 427 return -EINVAL; 428 } 429 430 properties.cu_mask_count = args->num_cu_mask; 431 if (properties.cu_mask_count == 0) { 432 pr_debug("CU mask cannot be 0"); 433 return -EINVAL; 434 } 435 436 /* To prevent an unreasonably large CU mask size, set an arbitrary 437 * limit of max_num_cus bits. We can then just drop any CU mask bits 438 * past max_num_cus bits and just use the first max_num_cus bits. 439 */ 440 if (properties.cu_mask_count > max_num_cus) { 441 pr_debug("CU mask cannot be greater than 1024 bits"); 442 properties.cu_mask_count = max_num_cus; 443 cu_mask_size = sizeof(uint32_t) * (max_num_cus/32); 444 } 445 446 properties.cu_mask = kzalloc(cu_mask_size, GFP_KERNEL); 447 if (!properties.cu_mask) 448 return -ENOMEM; 449 450 retval = copy_from_user(properties.cu_mask, cu_mask_ptr, cu_mask_size); 451 if (retval) { 452 pr_debug("Could not copy CU mask from userspace"); 453 kfree(properties.cu_mask); 454 return -EFAULT; 455 } 456 457 mutex_lock(&p->mutex); 458 459 retval = pqm_set_cu_mask(&p->pqm, args->queue_id, &properties); 460 461 mutex_unlock(&p->mutex); 462 463 if (retval) 464 kfree(properties.cu_mask); 465 466 return retval; 467 } 468 469 static int kfd_ioctl_get_queue_wave_state(struct file *filep, 470 struct kfd_process *p, void *data) 471 { 472 struct kfd_ioctl_get_queue_wave_state_args *args = data; 473 int r; 474 475 mutex_lock(&p->mutex); 476 477 r = pqm_get_wave_state(&p->pqm, args->queue_id, 478 (void __user *)args->ctl_stack_address, 479 &args->ctl_stack_used_size, 480 &args->save_area_used_size); 481 482 mutex_unlock(&p->mutex); 483 484 return r; 485 } 486 487 static int kfd_ioctl_set_memory_policy(struct file *filep, 488 struct kfd_process *p, void *data) 489 { 490 struct kfd_ioctl_set_memory_policy_args *args = data; 491 struct kfd_dev *dev; 492 int err = 0; 493 struct kfd_process_device *pdd; 494 enum cache_policy default_policy, alternate_policy; 495 496 if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT 497 && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) { 498 return -EINVAL; 499 } 500 501 if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT 502 && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) { 503 return -EINVAL; 504 } 505 506 dev = kfd_device_by_id(args->gpu_id); 507 if (!dev) 508 return -EINVAL; 509 510 mutex_lock(&p->mutex); 511 512 pdd = kfd_bind_process_to_device(dev, p); 513 if (IS_ERR(pdd)) { 514 err = -ESRCH; 515 goto out; 516 } 517 518 default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT) 519 ? cache_policy_coherent : cache_policy_noncoherent; 520 521 alternate_policy = 522 (args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT) 523 ? cache_policy_coherent : cache_policy_noncoherent; 524 525 if (!dev->dqm->ops.set_cache_memory_policy(dev->dqm, 526 &pdd->qpd, 527 default_policy, 528 alternate_policy, 529 (void __user *)args->alternate_aperture_base, 530 args->alternate_aperture_size)) 531 err = -EINVAL; 532 533 out: 534 mutex_unlock(&p->mutex); 535 536 return err; 537 } 538 539 static int kfd_ioctl_set_trap_handler(struct file *filep, 540 struct kfd_process *p, void *data) 541 { 542 struct kfd_ioctl_set_trap_handler_args *args = data; 543 struct kfd_dev *dev; 544 int err = 0; 545 struct kfd_process_device *pdd; 546 547 dev = kfd_device_by_id(args->gpu_id); 548 if (!dev) 549 return -EINVAL; 550 551 mutex_lock(&p->mutex); 552 553 pdd = kfd_bind_process_to_device(dev, p); 554 if (IS_ERR(pdd)) { 555 err = -ESRCH; 556 goto out; 557 } 558 559 kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr); 560 561 out: 562 mutex_unlock(&p->mutex); 563 564 return err; 565 } 566 567 static int kfd_ioctl_dbg_register(struct file *filep, 568 struct kfd_process *p, void *data) 569 { 570 struct kfd_ioctl_dbg_register_args *args = data; 571 struct kfd_dev *dev; 572 struct kfd_dbgmgr *dbgmgr_ptr; 573 struct kfd_process_device *pdd; 574 bool create_ok; 575 long status = 0; 576 577 dev = kfd_device_by_id(args->gpu_id); 578 if (!dev) 579 return -EINVAL; 580 581 if (dev->device_info->asic_family == CHIP_CARRIZO) { 582 pr_debug("kfd_ioctl_dbg_register not supported on CZ\n"); 583 return -EINVAL; 584 } 585 586 mutex_lock(&p->mutex); 587 mutex_lock(kfd_get_dbgmgr_mutex()); 588 589 /* 590 * make sure that we have pdd, if this the first queue created for 591 * this process 592 */ 593 pdd = kfd_bind_process_to_device(dev, p); 594 if (IS_ERR(pdd)) { 595 status = PTR_ERR(pdd); 596 goto out; 597 } 598 599 if (!dev->dbgmgr) { 600 /* In case of a legal call, we have no dbgmgr yet */ 601 create_ok = kfd_dbgmgr_create(&dbgmgr_ptr, dev); 602 if (create_ok) { 603 status = kfd_dbgmgr_register(dbgmgr_ptr, p); 604 if (status != 0) 605 kfd_dbgmgr_destroy(dbgmgr_ptr); 606 else 607 dev->dbgmgr = dbgmgr_ptr; 608 } 609 } else { 610 pr_debug("debugger already registered\n"); 611 status = -EINVAL; 612 } 613 614 out: 615 mutex_unlock(kfd_get_dbgmgr_mutex()); 616 mutex_unlock(&p->mutex); 617 618 return status; 619 } 620 621 static int kfd_ioctl_dbg_unregister(struct file *filep, 622 struct kfd_process *p, void *data) 623 { 624 struct kfd_ioctl_dbg_unregister_args *args = data; 625 struct kfd_dev *dev; 626 long status; 627 628 dev = kfd_device_by_id(args->gpu_id); 629 if (!dev || !dev->dbgmgr) 630 return -EINVAL; 631 632 if (dev->device_info->asic_family == CHIP_CARRIZO) { 633 pr_debug("kfd_ioctl_dbg_unregister not supported on CZ\n"); 634 return -EINVAL; 635 } 636 637 mutex_lock(kfd_get_dbgmgr_mutex()); 638 639 status = kfd_dbgmgr_unregister(dev->dbgmgr, p); 640 if (!status) { 641 kfd_dbgmgr_destroy(dev->dbgmgr); 642 dev->dbgmgr = NULL; 643 } 644 645 mutex_unlock(kfd_get_dbgmgr_mutex()); 646 647 return status; 648 } 649 650 /* 651 * Parse and generate variable size data structure for address watch. 652 * Total size of the buffer and # watch points is limited in order 653 * to prevent kernel abuse. (no bearing to the much smaller HW limitation 654 * which is enforced by dbgdev module) 655 * please also note that the watch address itself are not "copied from user", 656 * since it be set into the HW in user mode values. 657 * 658 */ 659 static int kfd_ioctl_dbg_address_watch(struct file *filep, 660 struct kfd_process *p, void *data) 661 { 662 struct kfd_ioctl_dbg_address_watch_args *args = data; 663 struct kfd_dev *dev; 664 struct dbg_address_watch_info aw_info; 665 unsigned char *args_buff; 666 long status; 667 void __user *cmd_from_user; 668 uint64_t watch_mask_value = 0; 669 unsigned int args_idx = 0; 670 671 memset((void *) &aw_info, 0, sizeof(struct dbg_address_watch_info)); 672 673 dev = kfd_device_by_id(args->gpu_id); 674 if (!dev) 675 return -EINVAL; 676 677 if (dev->device_info->asic_family == CHIP_CARRIZO) { 678 pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n"); 679 return -EINVAL; 680 } 681 682 cmd_from_user = (void __user *) args->content_ptr; 683 684 /* Validate arguments */ 685 686 if ((args->buf_size_in_bytes > MAX_ALLOWED_AW_BUFF_SIZE) || 687 (args->buf_size_in_bytes <= sizeof(*args) + sizeof(int) * 2) || 688 (cmd_from_user == NULL)) 689 return -EINVAL; 690 691 /* this is the actual buffer to work with */ 692 args_buff = memdup_user(cmd_from_user, 693 args->buf_size_in_bytes - sizeof(*args)); 694 if (IS_ERR(args_buff)) 695 return PTR_ERR(args_buff); 696 697 aw_info.process = p; 698 699 aw_info.num_watch_points = *((uint32_t *)(&args_buff[args_idx])); 700 args_idx += sizeof(aw_info.num_watch_points); 701 702 aw_info.watch_mode = (enum HSA_DBG_WATCH_MODE *) &args_buff[args_idx]; 703 args_idx += sizeof(enum HSA_DBG_WATCH_MODE) * aw_info.num_watch_points; 704 705 /* 706 * set watch address base pointer to point on the array base 707 * within args_buff 708 */ 709 aw_info.watch_address = (uint64_t *) &args_buff[args_idx]; 710 711 /* skip over the addresses buffer */ 712 args_idx += sizeof(aw_info.watch_address) * aw_info.num_watch_points; 713 714 if (args_idx >= args->buf_size_in_bytes - sizeof(*args)) { 715 status = -EINVAL; 716 goto out; 717 } 718 719 watch_mask_value = (uint64_t) args_buff[args_idx]; 720 721 if (watch_mask_value > 0) { 722 /* 723 * There is an array of masks. 724 * set watch mask base pointer to point on the array base 725 * within args_buff 726 */ 727 aw_info.watch_mask = (uint64_t *) &args_buff[args_idx]; 728 729 /* skip over the masks buffer */ 730 args_idx += sizeof(aw_info.watch_mask) * 731 aw_info.num_watch_points; 732 } else { 733 /* just the NULL mask, set to NULL and skip over it */ 734 aw_info.watch_mask = NULL; 735 args_idx += sizeof(aw_info.watch_mask); 736 } 737 738 if (args_idx >= args->buf_size_in_bytes - sizeof(args)) { 739 status = -EINVAL; 740 goto out; 741 } 742 743 /* Currently HSA Event is not supported for DBG */ 744 aw_info.watch_event = NULL; 745 746 mutex_lock(kfd_get_dbgmgr_mutex()); 747 748 status = kfd_dbgmgr_address_watch(dev->dbgmgr, &aw_info); 749 750 mutex_unlock(kfd_get_dbgmgr_mutex()); 751 752 out: 753 kfree(args_buff); 754 755 return status; 756 } 757 758 /* Parse and generate fixed size data structure for wave control */ 759 static int kfd_ioctl_dbg_wave_control(struct file *filep, 760 struct kfd_process *p, void *data) 761 { 762 struct kfd_ioctl_dbg_wave_control_args *args = data; 763 struct kfd_dev *dev; 764 struct dbg_wave_control_info wac_info; 765 unsigned char *args_buff; 766 uint32_t computed_buff_size; 767 long status; 768 void __user *cmd_from_user; 769 unsigned int args_idx = 0; 770 771 memset((void *) &wac_info, 0, sizeof(struct dbg_wave_control_info)); 772 773 /* we use compact form, independent of the packing attribute value */ 774 computed_buff_size = sizeof(*args) + 775 sizeof(wac_info.mode) + 776 sizeof(wac_info.operand) + 777 sizeof(wac_info.dbgWave_msg.DbgWaveMsg) + 778 sizeof(wac_info.dbgWave_msg.MemoryVA) + 779 sizeof(wac_info.trapId); 780 781 dev = kfd_device_by_id(args->gpu_id); 782 if (!dev) 783 return -EINVAL; 784 785 if (dev->device_info->asic_family == CHIP_CARRIZO) { 786 pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n"); 787 return -EINVAL; 788 } 789 790 /* input size must match the computed "compact" size */ 791 if (args->buf_size_in_bytes != computed_buff_size) { 792 pr_debug("size mismatch, computed : actual %u : %u\n", 793 args->buf_size_in_bytes, computed_buff_size); 794 return -EINVAL; 795 } 796 797 cmd_from_user = (void __user *) args->content_ptr; 798 799 if (cmd_from_user == NULL) 800 return -EINVAL; 801 802 /* copy the entire buffer from user */ 803 804 args_buff = memdup_user(cmd_from_user, 805 args->buf_size_in_bytes - sizeof(*args)); 806 if (IS_ERR(args_buff)) 807 return PTR_ERR(args_buff); 808 809 /* move ptr to the start of the "pay-load" area */ 810 wac_info.process = p; 811 812 wac_info.operand = *((enum HSA_DBG_WAVEOP *)(&args_buff[args_idx])); 813 args_idx += sizeof(wac_info.operand); 814 815 wac_info.mode = *((enum HSA_DBG_WAVEMODE *)(&args_buff[args_idx])); 816 args_idx += sizeof(wac_info.mode); 817 818 wac_info.trapId = *((uint32_t *)(&args_buff[args_idx])); 819 args_idx += sizeof(wac_info.trapId); 820 821 wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value = 822 *((uint32_t *)(&args_buff[args_idx])); 823 wac_info.dbgWave_msg.MemoryVA = NULL; 824 825 mutex_lock(kfd_get_dbgmgr_mutex()); 826 827 pr_debug("Calling dbg manager process %p, operand %u, mode %u, trapId %u, message %u\n", 828 wac_info.process, wac_info.operand, 829 wac_info.mode, wac_info.trapId, 830 wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value); 831 832 status = kfd_dbgmgr_wave_control(dev->dbgmgr, &wac_info); 833 834 pr_debug("Returned status of dbg manager is %ld\n", status); 835 836 mutex_unlock(kfd_get_dbgmgr_mutex()); 837 838 kfree(args_buff); 839 840 return status; 841 } 842 843 static int kfd_ioctl_get_clock_counters(struct file *filep, 844 struct kfd_process *p, void *data) 845 { 846 struct kfd_ioctl_get_clock_counters_args *args = data; 847 struct kfd_dev *dev; 848 849 dev = kfd_device_by_id(args->gpu_id); 850 if (dev) 851 /* Reading GPU clock counter from KGD */ 852 args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(dev->kgd); 853 else 854 /* Node without GPU resource */ 855 args->gpu_clock_counter = 0; 856 857 /* No access to rdtsc. Using raw monotonic time */ 858 args->cpu_clock_counter = ktime_get_raw_ns(); 859 args->system_clock_counter = ktime_get_boottime_ns(); 860 861 /* Since the counter is in nano-seconds we use 1GHz frequency */ 862 args->system_clock_freq = 1000000000; 863 864 return 0; 865 } 866 867 868 static int kfd_ioctl_get_process_apertures(struct file *filp, 869 struct kfd_process *p, void *data) 870 { 871 struct kfd_ioctl_get_process_apertures_args *args = data; 872 struct kfd_process_device_apertures *pAperture; 873 struct kfd_process_device *pdd; 874 875 dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid); 876 877 args->num_of_nodes = 0; 878 879 mutex_lock(&p->mutex); 880 881 /*if the process-device list isn't empty*/ 882 if (kfd_has_process_device_data(p)) { 883 /* Run over all pdd of the process */ 884 pdd = kfd_get_first_process_device_data(p); 885 do { 886 pAperture = 887 &args->process_apertures[args->num_of_nodes]; 888 pAperture->gpu_id = pdd->dev->id; 889 pAperture->lds_base = pdd->lds_base; 890 pAperture->lds_limit = pdd->lds_limit; 891 pAperture->gpuvm_base = pdd->gpuvm_base; 892 pAperture->gpuvm_limit = pdd->gpuvm_limit; 893 pAperture->scratch_base = pdd->scratch_base; 894 pAperture->scratch_limit = pdd->scratch_limit; 895 896 dev_dbg(kfd_device, 897 "node id %u\n", args->num_of_nodes); 898 dev_dbg(kfd_device, 899 "gpu id %u\n", pdd->dev->id); 900 dev_dbg(kfd_device, 901 "lds_base %llX\n", pdd->lds_base); 902 dev_dbg(kfd_device, 903 "lds_limit %llX\n", pdd->lds_limit); 904 dev_dbg(kfd_device, 905 "gpuvm_base %llX\n", pdd->gpuvm_base); 906 dev_dbg(kfd_device, 907 "gpuvm_limit %llX\n", pdd->gpuvm_limit); 908 dev_dbg(kfd_device, 909 "scratch_base %llX\n", pdd->scratch_base); 910 dev_dbg(kfd_device, 911 "scratch_limit %llX\n", pdd->scratch_limit); 912 913 args->num_of_nodes++; 914 915 pdd = kfd_get_next_process_device_data(p, pdd); 916 } while (pdd && (args->num_of_nodes < NUM_OF_SUPPORTED_GPUS)); 917 } 918 919 mutex_unlock(&p->mutex); 920 921 return 0; 922 } 923 924 static int kfd_ioctl_get_process_apertures_new(struct file *filp, 925 struct kfd_process *p, void *data) 926 { 927 struct kfd_ioctl_get_process_apertures_new_args *args = data; 928 struct kfd_process_device_apertures *pa; 929 struct kfd_process_device *pdd; 930 uint32_t nodes = 0; 931 int ret; 932 933 dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid); 934 935 if (args->num_of_nodes == 0) { 936 /* Return number of nodes, so that user space can alloacate 937 * sufficient memory 938 */ 939 mutex_lock(&p->mutex); 940 941 if (!kfd_has_process_device_data(p)) 942 goto out_unlock; 943 944 /* Run over all pdd of the process */ 945 pdd = kfd_get_first_process_device_data(p); 946 do { 947 args->num_of_nodes++; 948 pdd = kfd_get_next_process_device_data(p, pdd); 949 } while (pdd); 950 951 goto out_unlock; 952 } 953 954 /* Fill in process-aperture information for all available 955 * nodes, but not more than args->num_of_nodes as that is 956 * the amount of memory allocated by user 957 */ 958 pa = kzalloc((sizeof(struct kfd_process_device_apertures) * 959 args->num_of_nodes), GFP_KERNEL); 960 if (!pa) 961 return -ENOMEM; 962 963 mutex_lock(&p->mutex); 964 965 if (!kfd_has_process_device_data(p)) { 966 args->num_of_nodes = 0; 967 kfree(pa); 968 goto out_unlock; 969 } 970 971 /* Run over all pdd of the process */ 972 pdd = kfd_get_first_process_device_data(p); 973 do { 974 pa[nodes].gpu_id = pdd->dev->id; 975 pa[nodes].lds_base = pdd->lds_base; 976 pa[nodes].lds_limit = pdd->lds_limit; 977 pa[nodes].gpuvm_base = pdd->gpuvm_base; 978 pa[nodes].gpuvm_limit = pdd->gpuvm_limit; 979 pa[nodes].scratch_base = pdd->scratch_base; 980 pa[nodes].scratch_limit = pdd->scratch_limit; 981 982 dev_dbg(kfd_device, 983 "gpu id %u\n", pdd->dev->id); 984 dev_dbg(kfd_device, 985 "lds_base %llX\n", pdd->lds_base); 986 dev_dbg(kfd_device, 987 "lds_limit %llX\n", pdd->lds_limit); 988 dev_dbg(kfd_device, 989 "gpuvm_base %llX\n", pdd->gpuvm_base); 990 dev_dbg(kfd_device, 991 "gpuvm_limit %llX\n", pdd->gpuvm_limit); 992 dev_dbg(kfd_device, 993 "scratch_base %llX\n", pdd->scratch_base); 994 dev_dbg(kfd_device, 995 "scratch_limit %llX\n", pdd->scratch_limit); 996 nodes++; 997 998 pdd = kfd_get_next_process_device_data(p, pdd); 999 } while (pdd && (nodes < args->num_of_nodes)); 1000 mutex_unlock(&p->mutex); 1001 1002 args->num_of_nodes = nodes; 1003 ret = copy_to_user( 1004 (void __user *)args->kfd_process_device_apertures_ptr, 1005 pa, 1006 (nodes * sizeof(struct kfd_process_device_apertures))); 1007 kfree(pa); 1008 return ret ? -EFAULT : 0; 1009 1010 out_unlock: 1011 mutex_unlock(&p->mutex); 1012 return 0; 1013 } 1014 1015 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p, 1016 void *data) 1017 { 1018 struct kfd_ioctl_create_event_args *args = data; 1019 int err; 1020 1021 /* For dGPUs the event page is allocated in user mode. The 1022 * handle is passed to KFD with the first call to this IOCTL 1023 * through the event_page_offset field. 1024 */ 1025 if (args->event_page_offset) { 1026 struct kfd_dev *kfd; 1027 struct kfd_process_device *pdd; 1028 void *mem, *kern_addr; 1029 uint64_t size; 1030 1031 if (p->signal_page) { 1032 pr_err("Event page is already set\n"); 1033 return -EINVAL; 1034 } 1035 1036 kfd = kfd_device_by_id(GET_GPU_ID(args->event_page_offset)); 1037 if (!kfd) { 1038 pr_err("Getting device by id failed in %s\n", __func__); 1039 return -EINVAL; 1040 } 1041 1042 mutex_lock(&p->mutex); 1043 pdd = kfd_bind_process_to_device(kfd, p); 1044 if (IS_ERR(pdd)) { 1045 err = PTR_ERR(pdd); 1046 goto out_unlock; 1047 } 1048 1049 mem = kfd_process_device_translate_handle(pdd, 1050 GET_IDR_HANDLE(args->event_page_offset)); 1051 if (!mem) { 1052 pr_err("Can't find BO, offset is 0x%llx\n", 1053 args->event_page_offset); 1054 err = -EINVAL; 1055 goto out_unlock; 1056 } 1057 mutex_unlock(&p->mutex); 1058 1059 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(kfd->kgd, 1060 mem, &kern_addr, &size); 1061 if (err) { 1062 pr_err("Failed to map event page to kernel\n"); 1063 return err; 1064 } 1065 1066 err = kfd_event_page_set(p, kern_addr, size); 1067 if (err) { 1068 pr_err("Failed to set event page\n"); 1069 return err; 1070 } 1071 } 1072 1073 err = kfd_event_create(filp, p, args->event_type, 1074 args->auto_reset != 0, args->node_id, 1075 &args->event_id, &args->event_trigger_data, 1076 &args->event_page_offset, 1077 &args->event_slot_index); 1078 1079 return err; 1080 1081 out_unlock: 1082 mutex_unlock(&p->mutex); 1083 return err; 1084 } 1085 1086 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p, 1087 void *data) 1088 { 1089 struct kfd_ioctl_destroy_event_args *args = data; 1090 1091 return kfd_event_destroy(p, args->event_id); 1092 } 1093 1094 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p, 1095 void *data) 1096 { 1097 struct kfd_ioctl_set_event_args *args = data; 1098 1099 return kfd_set_event(p, args->event_id); 1100 } 1101 1102 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p, 1103 void *data) 1104 { 1105 struct kfd_ioctl_reset_event_args *args = data; 1106 1107 return kfd_reset_event(p, args->event_id); 1108 } 1109 1110 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p, 1111 void *data) 1112 { 1113 struct kfd_ioctl_wait_events_args *args = data; 1114 int err; 1115 1116 err = kfd_wait_on_events(p, args->num_events, 1117 (void __user *)args->events_ptr, 1118 (args->wait_for_all != 0), 1119 args->timeout, &args->wait_result); 1120 1121 return err; 1122 } 1123 static int kfd_ioctl_set_scratch_backing_va(struct file *filep, 1124 struct kfd_process *p, void *data) 1125 { 1126 struct kfd_ioctl_set_scratch_backing_va_args *args = data; 1127 struct kfd_process_device *pdd; 1128 struct kfd_dev *dev; 1129 long err; 1130 1131 dev = kfd_device_by_id(args->gpu_id); 1132 if (!dev) 1133 return -EINVAL; 1134 1135 mutex_lock(&p->mutex); 1136 1137 pdd = kfd_bind_process_to_device(dev, p); 1138 if (IS_ERR(pdd)) { 1139 err = PTR_ERR(pdd); 1140 goto bind_process_to_device_fail; 1141 } 1142 1143 pdd->qpd.sh_hidden_private_base = args->va_addr; 1144 1145 mutex_unlock(&p->mutex); 1146 1147 if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS && 1148 pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va) 1149 dev->kfd2kgd->set_scratch_backing_va( 1150 dev->kgd, args->va_addr, pdd->qpd.vmid); 1151 1152 return 0; 1153 1154 bind_process_to_device_fail: 1155 mutex_unlock(&p->mutex); 1156 return err; 1157 } 1158 1159 static int kfd_ioctl_get_tile_config(struct file *filep, 1160 struct kfd_process *p, void *data) 1161 { 1162 struct kfd_ioctl_get_tile_config_args *args = data; 1163 struct kfd_dev *dev; 1164 struct tile_config config; 1165 int err = 0; 1166 1167 dev = kfd_device_by_id(args->gpu_id); 1168 if (!dev) 1169 return -EINVAL; 1170 1171 amdgpu_amdkfd_get_tile_config(dev->kgd, &config); 1172 1173 args->gb_addr_config = config.gb_addr_config; 1174 args->num_banks = config.num_banks; 1175 args->num_ranks = config.num_ranks; 1176 1177 if (args->num_tile_configs > config.num_tile_configs) 1178 args->num_tile_configs = config.num_tile_configs; 1179 err = copy_to_user((void __user *)args->tile_config_ptr, 1180 config.tile_config_ptr, 1181 args->num_tile_configs * sizeof(uint32_t)); 1182 if (err) { 1183 args->num_tile_configs = 0; 1184 return -EFAULT; 1185 } 1186 1187 if (args->num_macro_tile_configs > config.num_macro_tile_configs) 1188 args->num_macro_tile_configs = 1189 config.num_macro_tile_configs; 1190 err = copy_to_user((void __user *)args->macro_tile_config_ptr, 1191 config.macro_tile_config_ptr, 1192 args->num_macro_tile_configs * sizeof(uint32_t)); 1193 if (err) { 1194 args->num_macro_tile_configs = 0; 1195 return -EFAULT; 1196 } 1197 1198 return 0; 1199 } 1200 1201 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p, 1202 void *data) 1203 { 1204 struct kfd_ioctl_acquire_vm_args *args = data; 1205 struct kfd_process_device *pdd; 1206 struct kfd_dev *dev; 1207 struct file *drm_file; 1208 int ret; 1209 1210 dev = kfd_device_by_id(args->gpu_id); 1211 if (!dev) 1212 return -EINVAL; 1213 1214 drm_file = fget(args->drm_fd); 1215 if (!drm_file) 1216 return -EINVAL; 1217 1218 mutex_lock(&p->mutex); 1219 1220 pdd = kfd_get_process_device_data(dev, p); 1221 if (!pdd) { 1222 ret = -EINVAL; 1223 goto err_unlock; 1224 } 1225 1226 if (pdd->drm_file) { 1227 ret = pdd->drm_file == drm_file ? 0 : -EBUSY; 1228 goto err_unlock; 1229 } 1230 1231 ret = kfd_process_device_init_vm(pdd, drm_file); 1232 if (ret) 1233 goto err_unlock; 1234 /* On success, the PDD keeps the drm_file reference */ 1235 mutex_unlock(&p->mutex); 1236 1237 return 0; 1238 1239 err_unlock: 1240 mutex_unlock(&p->mutex); 1241 fput(drm_file); 1242 return ret; 1243 } 1244 1245 bool kfd_dev_is_large_bar(struct kfd_dev *dev) 1246 { 1247 struct kfd_local_mem_info mem_info; 1248 1249 if (debug_largebar) { 1250 pr_debug("Simulate large-bar allocation on non large-bar machine\n"); 1251 return true; 1252 } 1253 1254 if (dev->use_iommu_v2) 1255 return false; 1256 1257 amdgpu_amdkfd_get_local_mem_info(dev->kgd, &mem_info); 1258 if (mem_info.local_mem_size_private == 0 && 1259 mem_info.local_mem_size_public > 0) 1260 return true; 1261 return false; 1262 } 1263 1264 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, 1265 struct kfd_process *p, void *data) 1266 { 1267 struct kfd_ioctl_alloc_memory_of_gpu_args *args = data; 1268 struct kfd_process_device *pdd; 1269 void *mem; 1270 struct kfd_dev *dev; 1271 int idr_handle; 1272 long err; 1273 uint64_t offset = args->mmap_offset; 1274 uint32_t flags = args->flags; 1275 1276 if (args->size == 0) 1277 return -EINVAL; 1278 1279 dev = kfd_device_by_id(args->gpu_id); 1280 if (!dev) 1281 return -EINVAL; 1282 1283 if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) && 1284 (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) && 1285 !kfd_dev_is_large_bar(dev)) { 1286 pr_err("Alloc host visible vram on small bar is not allowed\n"); 1287 return -EINVAL; 1288 } 1289 1290 mutex_lock(&p->mutex); 1291 1292 pdd = kfd_bind_process_to_device(dev, p); 1293 if (IS_ERR(pdd)) { 1294 err = PTR_ERR(pdd); 1295 goto err_unlock; 1296 } 1297 1298 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) { 1299 if (args->size != kfd_doorbell_process_slice(dev)) { 1300 err = -EINVAL; 1301 goto err_unlock; 1302 } 1303 offset = kfd_get_process_doorbells(pdd); 1304 } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) { 1305 if (args->size != PAGE_SIZE) { 1306 err = -EINVAL; 1307 goto err_unlock; 1308 } 1309 offset = amdgpu_amdkfd_get_mmio_remap_phys_addr(dev->kgd); 1310 if (!offset) { 1311 err = -ENOMEM; 1312 goto err_unlock; 1313 } 1314 } 1315 1316 err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( 1317 dev->kgd, args->va_addr, args->size, 1318 pdd->vm, (struct kgd_mem **) &mem, &offset, 1319 flags); 1320 1321 if (err) 1322 goto err_unlock; 1323 1324 idr_handle = kfd_process_device_create_obj_handle(pdd, mem); 1325 if (idr_handle < 0) { 1326 err = -EFAULT; 1327 goto err_free; 1328 } 1329 1330 /* Update the VRAM usage count */ 1331 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) 1332 WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size); 1333 1334 mutex_unlock(&p->mutex); 1335 1336 args->handle = MAKE_HANDLE(args->gpu_id, idr_handle); 1337 args->mmap_offset = offset; 1338 1339 /* MMIO is mapped through kfd device 1340 * Generate a kfd mmap offset 1341 */ 1342 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) 1343 args->mmap_offset = KFD_MMAP_TYPE_MMIO 1344 | KFD_MMAP_GPU_ID(args->gpu_id); 1345 1346 return 0; 1347 1348 err_free: 1349 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem, NULL); 1350 err_unlock: 1351 mutex_unlock(&p->mutex); 1352 return err; 1353 } 1354 1355 static int kfd_ioctl_free_memory_of_gpu(struct file *filep, 1356 struct kfd_process *p, void *data) 1357 { 1358 struct kfd_ioctl_free_memory_of_gpu_args *args = data; 1359 struct kfd_process_device *pdd; 1360 void *mem; 1361 struct kfd_dev *dev; 1362 int ret; 1363 uint64_t size = 0; 1364 1365 dev = kfd_device_by_id(GET_GPU_ID(args->handle)); 1366 if (!dev) 1367 return -EINVAL; 1368 1369 mutex_lock(&p->mutex); 1370 1371 pdd = kfd_get_process_device_data(dev, p); 1372 if (!pdd) { 1373 pr_err("Process device data doesn't exist\n"); 1374 ret = -EINVAL; 1375 goto err_unlock; 1376 } 1377 1378 mem = kfd_process_device_translate_handle( 1379 pdd, GET_IDR_HANDLE(args->handle)); 1380 if (!mem) { 1381 ret = -EINVAL; 1382 goto err_unlock; 1383 } 1384 1385 ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, 1386 (struct kgd_mem *)mem, &size); 1387 1388 /* If freeing the buffer failed, leave the handle in place for 1389 * clean-up during process tear-down. 1390 */ 1391 if (!ret) 1392 kfd_process_device_remove_obj_handle( 1393 pdd, GET_IDR_HANDLE(args->handle)); 1394 1395 WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size); 1396 1397 err_unlock: 1398 mutex_unlock(&p->mutex); 1399 return ret; 1400 } 1401 1402 static int kfd_ioctl_map_memory_to_gpu(struct file *filep, 1403 struct kfd_process *p, void *data) 1404 { 1405 struct kfd_ioctl_map_memory_to_gpu_args *args = data; 1406 struct kfd_process_device *pdd, *peer_pdd; 1407 void *mem; 1408 struct kfd_dev *dev, *peer; 1409 long err = 0; 1410 int i; 1411 uint32_t *devices_arr = NULL; 1412 1413 dev = kfd_device_by_id(GET_GPU_ID(args->handle)); 1414 if (!dev) 1415 return -EINVAL; 1416 1417 if (!args->n_devices) { 1418 pr_debug("Device IDs array empty\n"); 1419 return -EINVAL; 1420 } 1421 if (args->n_success > args->n_devices) { 1422 pr_debug("n_success exceeds n_devices\n"); 1423 return -EINVAL; 1424 } 1425 1426 devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr), 1427 GFP_KERNEL); 1428 if (!devices_arr) 1429 return -ENOMEM; 1430 1431 err = copy_from_user(devices_arr, 1432 (void __user *)args->device_ids_array_ptr, 1433 args->n_devices * sizeof(*devices_arr)); 1434 if (err != 0) { 1435 err = -EFAULT; 1436 goto copy_from_user_failed; 1437 } 1438 1439 mutex_lock(&p->mutex); 1440 1441 pdd = kfd_bind_process_to_device(dev, p); 1442 if (IS_ERR(pdd)) { 1443 err = PTR_ERR(pdd); 1444 goto bind_process_to_device_failed; 1445 } 1446 1447 mem = kfd_process_device_translate_handle(pdd, 1448 GET_IDR_HANDLE(args->handle)); 1449 if (!mem) { 1450 err = -ENOMEM; 1451 goto get_mem_obj_from_handle_failed; 1452 } 1453 1454 for (i = args->n_success; i < args->n_devices; i++) { 1455 peer = kfd_device_by_id(devices_arr[i]); 1456 if (!peer) { 1457 pr_debug("Getting device by id failed for 0x%x\n", 1458 devices_arr[i]); 1459 err = -EINVAL; 1460 goto get_mem_obj_from_handle_failed; 1461 } 1462 1463 peer_pdd = kfd_bind_process_to_device(peer, p); 1464 if (IS_ERR(peer_pdd)) { 1465 err = PTR_ERR(peer_pdd); 1466 goto get_mem_obj_from_handle_failed; 1467 } 1468 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu( 1469 peer->kgd, (struct kgd_mem *)mem, peer_pdd->vm); 1470 if (err) { 1471 pr_err("Failed to map to gpu %d/%d\n", 1472 i, args->n_devices); 1473 goto map_memory_to_gpu_failed; 1474 } 1475 args->n_success = i+1; 1476 } 1477 1478 mutex_unlock(&p->mutex); 1479 1480 err = amdgpu_amdkfd_gpuvm_sync_memory(dev->kgd, (struct kgd_mem *) mem, true); 1481 if (err) { 1482 pr_debug("Sync memory failed, wait interrupted by user signal\n"); 1483 goto sync_memory_failed; 1484 } 1485 1486 /* Flush TLBs after waiting for the page table updates to complete */ 1487 for (i = 0; i < args->n_devices; i++) { 1488 peer = kfd_device_by_id(devices_arr[i]); 1489 if (WARN_ON_ONCE(!peer)) 1490 continue; 1491 peer_pdd = kfd_get_process_device_data(peer, p); 1492 if (WARN_ON_ONCE(!peer_pdd)) 1493 continue; 1494 kfd_flush_tlb(peer_pdd); 1495 } 1496 1497 kfree(devices_arr); 1498 1499 return err; 1500 1501 bind_process_to_device_failed: 1502 get_mem_obj_from_handle_failed: 1503 map_memory_to_gpu_failed: 1504 mutex_unlock(&p->mutex); 1505 copy_from_user_failed: 1506 sync_memory_failed: 1507 kfree(devices_arr); 1508 1509 return err; 1510 } 1511 1512 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, 1513 struct kfd_process *p, void *data) 1514 { 1515 struct kfd_ioctl_unmap_memory_from_gpu_args *args = data; 1516 struct kfd_process_device *pdd, *peer_pdd; 1517 void *mem; 1518 struct kfd_dev *dev, *peer; 1519 long err = 0; 1520 uint32_t *devices_arr = NULL, i; 1521 1522 dev = kfd_device_by_id(GET_GPU_ID(args->handle)); 1523 if (!dev) 1524 return -EINVAL; 1525 1526 if (!args->n_devices) { 1527 pr_debug("Device IDs array empty\n"); 1528 return -EINVAL; 1529 } 1530 if (args->n_success > args->n_devices) { 1531 pr_debug("n_success exceeds n_devices\n"); 1532 return -EINVAL; 1533 } 1534 1535 devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr), 1536 GFP_KERNEL); 1537 if (!devices_arr) 1538 return -ENOMEM; 1539 1540 err = copy_from_user(devices_arr, 1541 (void __user *)args->device_ids_array_ptr, 1542 args->n_devices * sizeof(*devices_arr)); 1543 if (err != 0) { 1544 err = -EFAULT; 1545 goto copy_from_user_failed; 1546 } 1547 1548 mutex_lock(&p->mutex); 1549 1550 pdd = kfd_get_process_device_data(dev, p); 1551 if (!pdd) { 1552 err = -EINVAL; 1553 goto bind_process_to_device_failed; 1554 } 1555 1556 mem = kfd_process_device_translate_handle(pdd, 1557 GET_IDR_HANDLE(args->handle)); 1558 if (!mem) { 1559 err = -ENOMEM; 1560 goto get_mem_obj_from_handle_failed; 1561 } 1562 1563 for (i = args->n_success; i < args->n_devices; i++) { 1564 peer = kfd_device_by_id(devices_arr[i]); 1565 if (!peer) { 1566 err = -EINVAL; 1567 goto get_mem_obj_from_handle_failed; 1568 } 1569 1570 peer_pdd = kfd_get_process_device_data(peer, p); 1571 if (!peer_pdd) { 1572 err = -ENODEV; 1573 goto get_mem_obj_from_handle_failed; 1574 } 1575 err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( 1576 peer->kgd, (struct kgd_mem *)mem, peer_pdd->vm); 1577 if (err) { 1578 pr_err("Failed to unmap from gpu %d/%d\n", 1579 i, args->n_devices); 1580 goto unmap_memory_from_gpu_failed; 1581 } 1582 args->n_success = i+1; 1583 } 1584 kfree(devices_arr); 1585 1586 mutex_unlock(&p->mutex); 1587 1588 return 0; 1589 1590 bind_process_to_device_failed: 1591 get_mem_obj_from_handle_failed: 1592 unmap_memory_from_gpu_failed: 1593 mutex_unlock(&p->mutex); 1594 copy_from_user_failed: 1595 kfree(devices_arr); 1596 return err; 1597 } 1598 1599 static int kfd_ioctl_alloc_queue_gws(struct file *filep, 1600 struct kfd_process *p, void *data) 1601 { 1602 int retval; 1603 struct kfd_ioctl_alloc_queue_gws_args *args = data; 1604 struct queue *q; 1605 struct kfd_dev *dev; 1606 1607 mutex_lock(&p->mutex); 1608 q = pqm_get_user_queue(&p->pqm, args->queue_id); 1609 1610 if (q) { 1611 dev = q->device; 1612 } else { 1613 retval = -EINVAL; 1614 goto out_unlock; 1615 } 1616 1617 if (!dev->gws) { 1618 retval = -ENODEV; 1619 goto out_unlock; 1620 } 1621 1622 if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { 1623 retval = -ENODEV; 1624 goto out_unlock; 1625 } 1626 1627 retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL); 1628 mutex_unlock(&p->mutex); 1629 1630 args->first_gws = 0; 1631 return retval; 1632 1633 out_unlock: 1634 mutex_unlock(&p->mutex); 1635 return retval; 1636 } 1637 1638 static int kfd_ioctl_get_dmabuf_info(struct file *filep, 1639 struct kfd_process *p, void *data) 1640 { 1641 struct kfd_ioctl_get_dmabuf_info_args *args = data; 1642 struct kfd_dev *dev = NULL; 1643 struct kgd_dev *dma_buf_kgd; 1644 void *metadata_buffer = NULL; 1645 uint32_t flags; 1646 unsigned int i; 1647 int r; 1648 1649 /* Find a KFD GPU device that supports the get_dmabuf_info query */ 1650 for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++) 1651 if (dev) 1652 break; 1653 if (!dev) 1654 return -EINVAL; 1655 1656 if (args->metadata_ptr) { 1657 metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL); 1658 if (!metadata_buffer) 1659 return -ENOMEM; 1660 } 1661 1662 /* Get dmabuf info from KGD */ 1663 r = amdgpu_amdkfd_get_dmabuf_info(dev->kgd, args->dmabuf_fd, 1664 &dma_buf_kgd, &args->size, 1665 metadata_buffer, args->metadata_size, 1666 &args->metadata_size, &flags); 1667 if (r) 1668 goto exit; 1669 1670 /* Reverse-lookup gpu_id from kgd pointer */ 1671 dev = kfd_device_by_kgd(dma_buf_kgd); 1672 if (!dev) { 1673 r = -EINVAL; 1674 goto exit; 1675 } 1676 args->gpu_id = dev->id; 1677 args->flags = flags; 1678 1679 /* Copy metadata buffer to user mode */ 1680 if (metadata_buffer) { 1681 r = copy_to_user((void __user *)args->metadata_ptr, 1682 metadata_buffer, args->metadata_size); 1683 if (r != 0) 1684 r = -EFAULT; 1685 } 1686 1687 exit: 1688 kfree(metadata_buffer); 1689 1690 return r; 1691 } 1692 1693 static int kfd_ioctl_import_dmabuf(struct file *filep, 1694 struct kfd_process *p, void *data) 1695 { 1696 struct kfd_ioctl_import_dmabuf_args *args = data; 1697 struct kfd_process_device *pdd; 1698 struct dma_buf *dmabuf; 1699 struct kfd_dev *dev; 1700 int idr_handle; 1701 uint64_t size; 1702 void *mem; 1703 int r; 1704 1705 dev = kfd_device_by_id(args->gpu_id); 1706 if (!dev) 1707 return -EINVAL; 1708 1709 dmabuf = dma_buf_get(args->dmabuf_fd); 1710 if (IS_ERR(dmabuf)) 1711 return PTR_ERR(dmabuf); 1712 1713 mutex_lock(&p->mutex); 1714 1715 pdd = kfd_bind_process_to_device(dev, p); 1716 if (IS_ERR(pdd)) { 1717 r = PTR_ERR(pdd); 1718 goto err_unlock; 1719 } 1720 1721 r = amdgpu_amdkfd_gpuvm_import_dmabuf(dev->kgd, dmabuf, 1722 args->va_addr, pdd->vm, 1723 (struct kgd_mem **)&mem, &size, 1724 NULL); 1725 if (r) 1726 goto err_unlock; 1727 1728 idr_handle = kfd_process_device_create_obj_handle(pdd, mem); 1729 if (idr_handle < 0) { 1730 r = -EFAULT; 1731 goto err_free; 1732 } 1733 1734 mutex_unlock(&p->mutex); 1735 dma_buf_put(dmabuf); 1736 1737 args->handle = MAKE_HANDLE(args->gpu_id, idr_handle); 1738 1739 return 0; 1740 1741 err_free: 1742 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem, NULL); 1743 err_unlock: 1744 mutex_unlock(&p->mutex); 1745 dma_buf_put(dmabuf); 1746 return r; 1747 } 1748 1749 /* Handle requests for watching SMI events */ 1750 static int kfd_ioctl_smi_events(struct file *filep, 1751 struct kfd_process *p, void *data) 1752 { 1753 struct kfd_ioctl_smi_events_args *args = data; 1754 struct kfd_dev *dev; 1755 1756 dev = kfd_device_by_id(args->gpuid); 1757 if (!dev) 1758 return -EINVAL; 1759 1760 return kfd_smi_event_open(dev, &args->anon_fd); 1761 } 1762 1763 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \ 1764 [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \ 1765 .cmd_drv = 0, .name = #ioctl} 1766 1767 /** Ioctl table */ 1768 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = { 1769 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION, 1770 kfd_ioctl_get_version, 0), 1771 1772 AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE, 1773 kfd_ioctl_create_queue, 0), 1774 1775 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE, 1776 kfd_ioctl_destroy_queue, 0), 1777 1778 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY, 1779 kfd_ioctl_set_memory_policy, 0), 1780 1781 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS, 1782 kfd_ioctl_get_clock_counters, 0), 1783 1784 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES, 1785 kfd_ioctl_get_process_apertures, 0), 1786 1787 AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE, 1788 kfd_ioctl_update_queue, 0), 1789 1790 AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT, 1791 kfd_ioctl_create_event, 0), 1792 1793 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT, 1794 kfd_ioctl_destroy_event, 0), 1795 1796 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT, 1797 kfd_ioctl_set_event, 0), 1798 1799 AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT, 1800 kfd_ioctl_reset_event, 0), 1801 1802 AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS, 1803 kfd_ioctl_wait_events, 0), 1804 1805 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER, 1806 kfd_ioctl_dbg_register, 0), 1807 1808 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER, 1809 kfd_ioctl_dbg_unregister, 0), 1810 1811 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH, 1812 kfd_ioctl_dbg_address_watch, 0), 1813 1814 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL, 1815 kfd_ioctl_dbg_wave_control, 0), 1816 1817 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA, 1818 kfd_ioctl_set_scratch_backing_va, 0), 1819 1820 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG, 1821 kfd_ioctl_get_tile_config, 0), 1822 1823 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER, 1824 kfd_ioctl_set_trap_handler, 0), 1825 1826 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW, 1827 kfd_ioctl_get_process_apertures_new, 0), 1828 1829 AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM, 1830 kfd_ioctl_acquire_vm, 0), 1831 1832 AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU, 1833 kfd_ioctl_alloc_memory_of_gpu, 0), 1834 1835 AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU, 1836 kfd_ioctl_free_memory_of_gpu, 0), 1837 1838 AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU, 1839 kfd_ioctl_map_memory_to_gpu, 0), 1840 1841 AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU, 1842 kfd_ioctl_unmap_memory_from_gpu, 0), 1843 1844 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK, 1845 kfd_ioctl_set_cu_mask, 0), 1846 1847 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE, 1848 kfd_ioctl_get_queue_wave_state, 0), 1849 1850 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO, 1851 kfd_ioctl_get_dmabuf_info, 0), 1852 1853 AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF, 1854 kfd_ioctl_import_dmabuf, 0), 1855 1856 AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS, 1857 kfd_ioctl_alloc_queue_gws, 0), 1858 1859 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS, 1860 kfd_ioctl_smi_events, 0), 1861 }; 1862 1863 #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls) 1864 1865 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg) 1866 { 1867 struct kfd_process *process; 1868 amdkfd_ioctl_t *func; 1869 const struct amdkfd_ioctl_desc *ioctl = NULL; 1870 unsigned int nr = _IOC_NR(cmd); 1871 char stack_kdata[128]; 1872 char *kdata = NULL; 1873 unsigned int usize, asize; 1874 int retcode = -EINVAL; 1875 1876 if (nr >= AMDKFD_CORE_IOCTL_COUNT) 1877 goto err_i1; 1878 1879 if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) { 1880 u32 amdkfd_size; 1881 1882 ioctl = &amdkfd_ioctls[nr]; 1883 1884 amdkfd_size = _IOC_SIZE(ioctl->cmd); 1885 usize = asize = _IOC_SIZE(cmd); 1886 if (amdkfd_size > asize) 1887 asize = amdkfd_size; 1888 1889 cmd = ioctl->cmd; 1890 } else 1891 goto err_i1; 1892 1893 dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg); 1894 1895 /* Get the process struct from the filep. Only the process 1896 * that opened /dev/kfd can use the file descriptor. Child 1897 * processes need to create their own KFD device context. 1898 */ 1899 process = filep->private_data; 1900 if (process->lead_thread != current->group_leader) { 1901 dev_dbg(kfd_device, "Using KFD FD in wrong process\n"); 1902 retcode = -EBADF; 1903 goto err_i1; 1904 } 1905 1906 /* Do not trust userspace, use our own definition */ 1907 func = ioctl->func; 1908 1909 if (unlikely(!func)) { 1910 dev_dbg(kfd_device, "no function\n"); 1911 retcode = -EINVAL; 1912 goto err_i1; 1913 } 1914 1915 if (cmd & (IOC_IN | IOC_OUT)) { 1916 if (asize <= sizeof(stack_kdata)) { 1917 kdata = stack_kdata; 1918 } else { 1919 kdata = kmalloc(asize, GFP_KERNEL); 1920 if (!kdata) { 1921 retcode = -ENOMEM; 1922 goto err_i1; 1923 } 1924 } 1925 if (asize > usize) 1926 memset(kdata + usize, 0, asize - usize); 1927 } 1928 1929 if (cmd & IOC_IN) { 1930 if (copy_from_user(kdata, (void __user *)arg, usize) != 0) { 1931 retcode = -EFAULT; 1932 goto err_i1; 1933 } 1934 } else if (cmd & IOC_OUT) { 1935 memset(kdata, 0, usize); 1936 } 1937 1938 retcode = func(filep, process, kdata); 1939 1940 if (cmd & IOC_OUT) 1941 if (copy_to_user((void __user *)arg, kdata, usize) != 0) 1942 retcode = -EFAULT; 1943 1944 err_i1: 1945 if (!ioctl) 1946 dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n", 1947 task_pid_nr(current), cmd, nr); 1948 1949 if (kdata != stack_kdata) 1950 kfree(kdata); 1951 1952 if (retcode) 1953 dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n", 1954 nr, arg, retcode); 1955 1956 return retcode; 1957 } 1958 1959 static int kfd_mmio_mmap(struct kfd_dev *dev, struct kfd_process *process, 1960 struct vm_area_struct *vma) 1961 { 1962 phys_addr_t address; 1963 int ret; 1964 1965 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 1966 return -EINVAL; 1967 1968 address = amdgpu_amdkfd_get_mmio_remap_phys_addr(dev->kgd); 1969 1970 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE | 1971 VM_DONTDUMP | VM_PFNMAP; 1972 1973 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 1974 1975 pr_debug("pasid 0x%x mapping mmio page\n" 1976 " target user address == 0x%08llX\n" 1977 " physical address == 0x%08llX\n" 1978 " vm_flags == 0x%04lX\n" 1979 " size == 0x%04lX\n", 1980 process->pasid, (unsigned long long) vma->vm_start, 1981 address, vma->vm_flags, PAGE_SIZE); 1982 1983 ret = io_remap_pfn_range(vma, 1984 vma->vm_start, 1985 address >> PAGE_SHIFT, 1986 PAGE_SIZE, 1987 vma->vm_page_prot); 1988 return ret; 1989 } 1990 1991 1992 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma) 1993 { 1994 struct kfd_process *process; 1995 struct kfd_dev *dev = NULL; 1996 unsigned long mmap_offset; 1997 unsigned int gpu_id; 1998 1999 process = kfd_get_process(current); 2000 if (IS_ERR(process)) 2001 return PTR_ERR(process); 2002 2003 mmap_offset = vma->vm_pgoff << PAGE_SHIFT; 2004 gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset); 2005 if (gpu_id) 2006 dev = kfd_device_by_id(gpu_id); 2007 2008 switch (mmap_offset & KFD_MMAP_TYPE_MASK) { 2009 case KFD_MMAP_TYPE_DOORBELL: 2010 if (!dev) 2011 return -ENODEV; 2012 return kfd_doorbell_mmap(dev, process, vma); 2013 2014 case KFD_MMAP_TYPE_EVENTS: 2015 return kfd_event_mmap(process, vma); 2016 2017 case KFD_MMAP_TYPE_RESERVED_MEM: 2018 if (!dev) 2019 return -ENODEV; 2020 return kfd_reserved_mem_mmap(dev, process, vma); 2021 case KFD_MMAP_TYPE_MMIO: 2022 if (!dev) 2023 return -ENODEV; 2024 return kfd_mmio_mmap(dev, process, vma); 2025 } 2026 2027 return -EFAULT; 2028 } 2029