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 "kfd_svm.h" 42 #include "amdgpu_amdkfd.h" 43 #include "kfd_smi_events.h" 44 45 static long kfd_ioctl(struct file *, unsigned int, unsigned long); 46 static int kfd_open(struct inode *, struct file *); 47 static int kfd_release(struct inode *, struct file *); 48 static int kfd_mmap(struct file *, struct vm_area_struct *); 49 50 static const char kfd_dev_name[] = "kfd"; 51 52 static const struct file_operations kfd_fops = { 53 .owner = THIS_MODULE, 54 .unlocked_ioctl = kfd_ioctl, 55 .compat_ioctl = compat_ptr_ioctl, 56 .open = kfd_open, 57 .release = kfd_release, 58 .mmap = kfd_mmap, 59 }; 60 61 static int kfd_char_dev_major = -1; 62 static struct class *kfd_class; 63 struct device *kfd_device; 64 65 int kfd_chardev_init(void) 66 { 67 int err = 0; 68 69 kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops); 70 err = kfd_char_dev_major; 71 if (err < 0) 72 goto err_register_chrdev; 73 74 kfd_class = class_create(THIS_MODULE, kfd_dev_name); 75 err = PTR_ERR(kfd_class); 76 if (IS_ERR(kfd_class)) 77 goto err_class_create; 78 79 kfd_device = device_create(kfd_class, NULL, 80 MKDEV(kfd_char_dev_major, 0), 81 NULL, kfd_dev_name); 82 err = PTR_ERR(kfd_device); 83 if (IS_ERR(kfd_device)) 84 goto err_device_create; 85 86 return 0; 87 88 err_device_create: 89 class_destroy(kfd_class); 90 err_class_create: 91 unregister_chrdev(kfd_char_dev_major, kfd_dev_name); 92 err_register_chrdev: 93 return err; 94 } 95 96 void kfd_chardev_exit(void) 97 { 98 device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0)); 99 class_destroy(kfd_class); 100 unregister_chrdev(kfd_char_dev_major, kfd_dev_name); 101 kfd_device = NULL; 102 } 103 104 struct device *kfd_chardev(void) 105 { 106 return kfd_device; 107 } 108 109 110 static int kfd_open(struct inode *inode, struct file *filep) 111 { 112 struct kfd_process *process; 113 bool is_32bit_user_mode; 114 115 if (iminor(inode) != 0) 116 return -ENODEV; 117 118 is_32bit_user_mode = in_compat_syscall(); 119 120 if (is_32bit_user_mode) { 121 dev_warn(kfd_device, 122 "Process %d (32-bit) failed to open /dev/kfd\n" 123 "32-bit processes are not supported by amdkfd\n", 124 current->pid); 125 return -EPERM; 126 } 127 128 process = kfd_create_process(filep); 129 if (IS_ERR(process)) 130 return PTR_ERR(process); 131 132 if (kfd_is_locked()) { 133 dev_dbg(kfd_device, "kfd is locked!\n" 134 "process %d unreferenced", process->pasid); 135 kfd_unref_process(process); 136 return -EAGAIN; 137 } 138 139 /* filep now owns the reference returned by kfd_create_process */ 140 filep->private_data = process; 141 142 dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n", 143 process->pasid, process->is_32bit_user_mode); 144 145 return 0; 146 } 147 148 static int kfd_release(struct inode *inode, struct file *filep) 149 { 150 struct kfd_process *process = filep->private_data; 151 152 if (process) 153 kfd_unref_process(process); 154 155 return 0; 156 } 157 158 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p, 159 void *data) 160 { 161 struct kfd_ioctl_get_version_args *args = data; 162 163 args->major_version = KFD_IOCTL_MAJOR_VERSION; 164 args->minor_version = KFD_IOCTL_MINOR_VERSION; 165 166 return 0; 167 } 168 169 static int set_queue_properties_from_user(struct queue_properties *q_properties, 170 struct kfd_ioctl_create_queue_args *args) 171 { 172 if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) { 173 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n"); 174 return -EINVAL; 175 } 176 177 if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) { 178 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n"); 179 return -EINVAL; 180 } 181 182 if ((args->ring_base_address) && 183 (!access_ok((const void __user *) args->ring_base_address, 184 sizeof(uint64_t)))) { 185 pr_err("Can't access ring base address\n"); 186 return -EFAULT; 187 } 188 189 if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) { 190 pr_err("Ring size must be a power of 2 or 0\n"); 191 return -EINVAL; 192 } 193 194 if (!access_ok((const void __user *) args->read_pointer_address, 195 sizeof(uint32_t))) { 196 pr_err("Can't access read pointer\n"); 197 return -EFAULT; 198 } 199 200 if (!access_ok((const void __user *) args->write_pointer_address, 201 sizeof(uint32_t))) { 202 pr_err("Can't access write pointer\n"); 203 return -EFAULT; 204 } 205 206 if (args->eop_buffer_address && 207 !access_ok((const void __user *) args->eop_buffer_address, 208 sizeof(uint32_t))) { 209 pr_debug("Can't access eop buffer"); 210 return -EFAULT; 211 } 212 213 if (args->ctx_save_restore_address && 214 !access_ok((const void __user *) args->ctx_save_restore_address, 215 sizeof(uint32_t))) { 216 pr_debug("Can't access ctx save restore buffer"); 217 return -EFAULT; 218 } 219 220 q_properties->is_interop = false; 221 q_properties->is_gws = false; 222 q_properties->queue_percent = args->queue_percentage; 223 q_properties->priority = args->queue_priority; 224 q_properties->queue_address = args->ring_base_address; 225 q_properties->queue_size = args->ring_size; 226 q_properties->read_ptr = (uint32_t *) args->read_pointer_address; 227 q_properties->write_ptr = (uint32_t *) args->write_pointer_address; 228 q_properties->eop_ring_buffer_address = args->eop_buffer_address; 229 q_properties->eop_ring_buffer_size = args->eop_buffer_size; 230 q_properties->ctx_save_restore_area_address = 231 args->ctx_save_restore_address; 232 q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size; 233 q_properties->ctl_stack_size = args->ctl_stack_size; 234 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE || 235 args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL) 236 q_properties->type = KFD_QUEUE_TYPE_COMPUTE; 237 else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA) 238 q_properties->type = KFD_QUEUE_TYPE_SDMA; 239 else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI) 240 q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI; 241 else 242 return -ENOTSUPP; 243 244 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL) 245 q_properties->format = KFD_QUEUE_FORMAT_AQL; 246 else 247 q_properties->format = KFD_QUEUE_FORMAT_PM4; 248 249 pr_debug("Queue Percentage: %d, %d\n", 250 q_properties->queue_percent, args->queue_percentage); 251 252 pr_debug("Queue Priority: %d, %d\n", 253 q_properties->priority, args->queue_priority); 254 255 pr_debug("Queue Address: 0x%llX, 0x%llX\n", 256 q_properties->queue_address, args->ring_base_address); 257 258 pr_debug("Queue Size: 0x%llX, %u\n", 259 q_properties->queue_size, args->ring_size); 260 261 pr_debug("Queue r/w Pointers: %px, %px\n", 262 q_properties->read_ptr, 263 q_properties->write_ptr); 264 265 pr_debug("Queue Format: %d\n", q_properties->format); 266 267 pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address); 268 269 pr_debug("Queue CTX save area: 0x%llX\n", 270 q_properties->ctx_save_restore_area_address); 271 272 return 0; 273 } 274 275 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, 276 void *data) 277 { 278 struct kfd_ioctl_create_queue_args *args = data; 279 struct kfd_dev *dev; 280 int err = 0; 281 unsigned int queue_id; 282 struct kfd_process_device *pdd; 283 struct queue_properties q_properties; 284 uint32_t doorbell_offset_in_process = 0; 285 286 memset(&q_properties, 0, sizeof(struct queue_properties)); 287 288 pr_debug("Creating queue ioctl\n"); 289 290 err = set_queue_properties_from_user(&q_properties, args); 291 if (err) 292 return err; 293 294 pr_debug("Looking for gpu id 0x%x\n", args->gpu_id); 295 dev = kfd_device_by_id(args->gpu_id); 296 if (!dev) { 297 pr_debug("Could not find gpu id 0x%x\n", args->gpu_id); 298 return -EINVAL; 299 } 300 301 mutex_lock(&p->mutex); 302 303 pdd = kfd_bind_process_to_device(dev, p); 304 if (IS_ERR(pdd)) { 305 err = -ESRCH; 306 goto err_bind_process; 307 } 308 309 pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n", 310 p->pasid, 311 dev->id); 312 313 err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, 314 &doorbell_offset_in_process); 315 if (err != 0) 316 goto err_create_queue; 317 318 args->queue_id = queue_id; 319 320 321 /* Return gpu_id as doorbell offset for mmap usage */ 322 args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL; 323 args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id); 324 if (KFD_IS_SOC15(dev->device_info->asic_family)) 325 /* On SOC15 ASICs, include the doorbell offset within the 326 * process doorbell frame, which is 2 pages. 327 */ 328 args->doorbell_offset |= doorbell_offset_in_process; 329 330 mutex_unlock(&p->mutex); 331 332 pr_debug("Queue id %d was created successfully\n", args->queue_id); 333 334 pr_debug("Ring buffer address == 0x%016llX\n", 335 args->ring_base_address); 336 337 pr_debug("Read ptr address == 0x%016llX\n", 338 args->read_pointer_address); 339 340 pr_debug("Write ptr address == 0x%016llX\n", 341 args->write_pointer_address); 342 343 return 0; 344 345 err_create_queue: 346 err_bind_process: 347 mutex_unlock(&p->mutex); 348 return err; 349 } 350 351 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p, 352 void *data) 353 { 354 int retval; 355 struct kfd_ioctl_destroy_queue_args *args = data; 356 357 pr_debug("Destroying queue id %d for pasid 0x%x\n", 358 args->queue_id, 359 p->pasid); 360 361 mutex_lock(&p->mutex); 362 363 retval = pqm_destroy_queue(&p->pqm, args->queue_id); 364 365 mutex_unlock(&p->mutex); 366 return retval; 367 } 368 369 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p, 370 void *data) 371 { 372 int retval; 373 struct kfd_ioctl_update_queue_args *args = data; 374 struct queue_properties properties; 375 376 if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) { 377 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n"); 378 return -EINVAL; 379 } 380 381 if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) { 382 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n"); 383 return -EINVAL; 384 } 385 386 if ((args->ring_base_address) && 387 (!access_ok((const void __user *) args->ring_base_address, 388 sizeof(uint64_t)))) { 389 pr_err("Can't access ring base address\n"); 390 return -EFAULT; 391 } 392 393 if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) { 394 pr_err("Ring size must be a power of 2 or 0\n"); 395 return -EINVAL; 396 } 397 398 properties.queue_address = args->ring_base_address; 399 properties.queue_size = args->ring_size; 400 properties.queue_percent = args->queue_percentage; 401 properties.priority = args->queue_priority; 402 403 pr_debug("Updating queue id %d for pasid 0x%x\n", 404 args->queue_id, p->pasid); 405 406 mutex_lock(&p->mutex); 407 408 retval = pqm_update_queue(&p->pqm, args->queue_id, &properties); 409 410 mutex_unlock(&p->mutex); 411 412 return retval; 413 } 414 415 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p, 416 void *data) 417 { 418 int retval; 419 const int max_num_cus = 1024; 420 struct kfd_ioctl_set_cu_mask_args *args = data; 421 struct queue_properties properties; 422 uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr; 423 size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32); 424 425 if ((args->num_cu_mask % 32) != 0) { 426 pr_debug("num_cu_mask 0x%x must be a multiple of 32", 427 args->num_cu_mask); 428 return -EINVAL; 429 } 430 431 properties.cu_mask_count = args->num_cu_mask; 432 if (properties.cu_mask_count == 0) { 433 pr_debug("CU mask cannot be 0"); 434 return -EINVAL; 435 } 436 437 /* To prevent an unreasonably large CU mask size, set an arbitrary 438 * limit of max_num_cus bits. We can then just drop any CU mask bits 439 * past max_num_cus bits and just use the first max_num_cus bits. 440 */ 441 if (properties.cu_mask_count > max_num_cus) { 442 pr_debug("CU mask cannot be greater than 1024 bits"); 443 properties.cu_mask_count = max_num_cus; 444 cu_mask_size = sizeof(uint32_t) * (max_num_cus/32); 445 } 446 447 properties.cu_mask = kzalloc(cu_mask_size, GFP_KERNEL); 448 if (!properties.cu_mask) 449 return -ENOMEM; 450 451 retval = copy_from_user(properties.cu_mask, cu_mask_ptr, cu_mask_size); 452 if (retval) { 453 pr_debug("Could not copy CU mask from userspace"); 454 kfree(properties.cu_mask); 455 return -EFAULT; 456 } 457 458 mutex_lock(&p->mutex); 459 460 retval = pqm_set_cu_mask(&p->pqm, args->queue_id, &properties); 461 462 mutex_unlock(&p->mutex); 463 464 if (retval) 465 kfree(properties.cu_mask); 466 467 return retval; 468 } 469 470 static int kfd_ioctl_get_queue_wave_state(struct file *filep, 471 struct kfd_process *p, void *data) 472 { 473 struct kfd_ioctl_get_queue_wave_state_args *args = data; 474 int r; 475 476 mutex_lock(&p->mutex); 477 478 r = pqm_get_wave_state(&p->pqm, args->queue_id, 479 (void __user *)args->ctl_stack_address, 480 &args->ctl_stack_used_size, 481 &args->save_area_used_size); 482 483 mutex_unlock(&p->mutex); 484 485 return r; 486 } 487 488 static int kfd_ioctl_set_memory_policy(struct file *filep, 489 struct kfd_process *p, void *data) 490 { 491 struct kfd_ioctl_set_memory_policy_args *args = data; 492 struct kfd_dev *dev; 493 int err = 0; 494 struct kfd_process_device *pdd; 495 enum cache_policy default_policy, alternate_policy; 496 497 if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT 498 && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) { 499 return -EINVAL; 500 } 501 502 if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT 503 && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) { 504 return -EINVAL; 505 } 506 507 dev = kfd_device_by_id(args->gpu_id); 508 if (!dev) 509 return -EINVAL; 510 511 mutex_lock(&p->mutex); 512 513 pdd = kfd_bind_process_to_device(dev, p); 514 if (IS_ERR(pdd)) { 515 err = -ESRCH; 516 goto out; 517 } 518 519 default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT) 520 ? cache_policy_coherent : cache_policy_noncoherent; 521 522 alternate_policy = 523 (args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT) 524 ? cache_policy_coherent : cache_policy_noncoherent; 525 526 if (!dev->dqm->ops.set_cache_memory_policy(dev->dqm, 527 &pdd->qpd, 528 default_policy, 529 alternate_policy, 530 (void __user *)args->alternate_aperture_base, 531 args->alternate_aperture_size)) 532 err = -EINVAL; 533 534 out: 535 mutex_unlock(&p->mutex); 536 537 return err; 538 } 539 540 static int kfd_ioctl_set_trap_handler(struct file *filep, 541 struct kfd_process *p, void *data) 542 { 543 struct kfd_ioctl_set_trap_handler_args *args = data; 544 struct kfd_dev *dev; 545 int err = 0; 546 struct kfd_process_device *pdd; 547 548 dev = kfd_device_by_id(args->gpu_id); 549 if (!dev) 550 return -EINVAL; 551 552 mutex_lock(&p->mutex); 553 554 pdd = kfd_bind_process_to_device(dev, p); 555 if (IS_ERR(pdd)) { 556 err = -ESRCH; 557 goto out; 558 } 559 560 kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr); 561 562 out: 563 mutex_unlock(&p->mutex); 564 565 return err; 566 } 567 568 static int kfd_ioctl_dbg_register(struct file *filep, 569 struct kfd_process *p, void *data) 570 { 571 struct kfd_ioctl_dbg_register_args *args = data; 572 struct kfd_dev *dev; 573 struct kfd_dbgmgr *dbgmgr_ptr; 574 struct kfd_process_device *pdd; 575 bool create_ok; 576 long status = 0; 577 578 dev = kfd_device_by_id(args->gpu_id); 579 if (!dev) 580 return -EINVAL; 581 582 if (dev->device_info->asic_family == CHIP_CARRIZO) { 583 pr_debug("kfd_ioctl_dbg_register not supported on CZ\n"); 584 return -EINVAL; 585 } 586 587 mutex_lock(&p->mutex); 588 mutex_lock(kfd_get_dbgmgr_mutex()); 589 590 /* 591 * make sure that we have pdd, if this the first queue created for 592 * this process 593 */ 594 pdd = kfd_bind_process_to_device(dev, p); 595 if (IS_ERR(pdd)) { 596 status = PTR_ERR(pdd); 597 goto out; 598 } 599 600 if (!dev->dbgmgr) { 601 /* In case of a legal call, we have no dbgmgr yet */ 602 create_ok = kfd_dbgmgr_create(&dbgmgr_ptr, dev); 603 if (create_ok) { 604 status = kfd_dbgmgr_register(dbgmgr_ptr, p); 605 if (status != 0) 606 kfd_dbgmgr_destroy(dbgmgr_ptr); 607 else 608 dev->dbgmgr = dbgmgr_ptr; 609 } 610 } else { 611 pr_debug("debugger already registered\n"); 612 status = -EINVAL; 613 } 614 615 out: 616 mutex_unlock(kfd_get_dbgmgr_mutex()); 617 mutex_unlock(&p->mutex); 618 619 return status; 620 } 621 622 static int kfd_ioctl_dbg_unregister(struct file *filep, 623 struct kfd_process *p, void *data) 624 { 625 struct kfd_ioctl_dbg_unregister_args *args = data; 626 struct kfd_dev *dev; 627 long status; 628 629 dev = kfd_device_by_id(args->gpu_id); 630 if (!dev || !dev->dbgmgr) 631 return -EINVAL; 632 633 if (dev->device_info->asic_family == CHIP_CARRIZO) { 634 pr_debug("kfd_ioctl_dbg_unregister not supported on CZ\n"); 635 return -EINVAL; 636 } 637 638 mutex_lock(kfd_get_dbgmgr_mutex()); 639 640 status = kfd_dbgmgr_unregister(dev->dbgmgr, p); 641 if (!status) { 642 kfd_dbgmgr_destroy(dev->dbgmgr); 643 dev->dbgmgr = NULL; 644 } 645 646 mutex_unlock(kfd_get_dbgmgr_mutex()); 647 648 return status; 649 } 650 651 /* 652 * Parse and generate variable size data structure for address watch. 653 * Total size of the buffer and # watch points is limited in order 654 * to prevent kernel abuse. (no bearing to the much smaller HW limitation 655 * which is enforced by dbgdev module) 656 * please also note that the watch address itself are not "copied from user", 657 * since it be set into the HW in user mode values. 658 * 659 */ 660 static int kfd_ioctl_dbg_address_watch(struct file *filep, 661 struct kfd_process *p, void *data) 662 { 663 struct kfd_ioctl_dbg_address_watch_args *args = data; 664 struct kfd_dev *dev; 665 struct dbg_address_watch_info aw_info; 666 unsigned char *args_buff; 667 long status; 668 void __user *cmd_from_user; 669 uint64_t watch_mask_value = 0; 670 unsigned int args_idx = 0; 671 672 memset((void *) &aw_info, 0, sizeof(struct dbg_address_watch_info)); 673 674 dev = kfd_device_by_id(args->gpu_id); 675 if (!dev) 676 return -EINVAL; 677 678 if (dev->device_info->asic_family == CHIP_CARRIZO) { 679 pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n"); 680 return -EINVAL; 681 } 682 683 cmd_from_user = (void __user *) args->content_ptr; 684 685 /* Validate arguments */ 686 687 if ((args->buf_size_in_bytes > MAX_ALLOWED_AW_BUFF_SIZE) || 688 (args->buf_size_in_bytes <= sizeof(*args) + sizeof(int) * 2) || 689 (cmd_from_user == NULL)) 690 return -EINVAL; 691 692 /* this is the actual buffer to work with */ 693 args_buff = memdup_user(cmd_from_user, 694 args->buf_size_in_bytes - sizeof(*args)); 695 if (IS_ERR(args_buff)) 696 return PTR_ERR(args_buff); 697 698 aw_info.process = p; 699 700 aw_info.num_watch_points = *((uint32_t *)(&args_buff[args_idx])); 701 args_idx += sizeof(aw_info.num_watch_points); 702 703 aw_info.watch_mode = (enum HSA_DBG_WATCH_MODE *) &args_buff[args_idx]; 704 args_idx += sizeof(enum HSA_DBG_WATCH_MODE) * aw_info.num_watch_points; 705 706 /* 707 * set watch address base pointer to point on the array base 708 * within args_buff 709 */ 710 aw_info.watch_address = (uint64_t *) &args_buff[args_idx]; 711 712 /* skip over the addresses buffer */ 713 args_idx += sizeof(aw_info.watch_address) * aw_info.num_watch_points; 714 715 if (args_idx >= args->buf_size_in_bytes - sizeof(*args)) { 716 status = -EINVAL; 717 goto out; 718 } 719 720 watch_mask_value = (uint64_t) args_buff[args_idx]; 721 722 if (watch_mask_value > 0) { 723 /* 724 * There is an array of masks. 725 * set watch mask base pointer to point on the array base 726 * within args_buff 727 */ 728 aw_info.watch_mask = (uint64_t *) &args_buff[args_idx]; 729 730 /* skip over the masks buffer */ 731 args_idx += sizeof(aw_info.watch_mask) * 732 aw_info.num_watch_points; 733 } else { 734 /* just the NULL mask, set to NULL and skip over it */ 735 aw_info.watch_mask = NULL; 736 args_idx += sizeof(aw_info.watch_mask); 737 } 738 739 if (args_idx >= args->buf_size_in_bytes - sizeof(args)) { 740 status = -EINVAL; 741 goto out; 742 } 743 744 /* Currently HSA Event is not supported for DBG */ 745 aw_info.watch_event = NULL; 746 747 mutex_lock(kfd_get_dbgmgr_mutex()); 748 749 status = kfd_dbgmgr_address_watch(dev->dbgmgr, &aw_info); 750 751 mutex_unlock(kfd_get_dbgmgr_mutex()); 752 753 out: 754 kfree(args_buff); 755 756 return status; 757 } 758 759 /* Parse and generate fixed size data structure for wave control */ 760 static int kfd_ioctl_dbg_wave_control(struct file *filep, 761 struct kfd_process *p, void *data) 762 { 763 struct kfd_ioctl_dbg_wave_control_args *args = data; 764 struct kfd_dev *dev; 765 struct dbg_wave_control_info wac_info; 766 unsigned char *args_buff; 767 uint32_t computed_buff_size; 768 long status; 769 void __user *cmd_from_user; 770 unsigned int args_idx = 0; 771 772 memset((void *) &wac_info, 0, sizeof(struct dbg_wave_control_info)); 773 774 /* we use compact form, independent of the packing attribute value */ 775 computed_buff_size = sizeof(*args) + 776 sizeof(wac_info.mode) + 777 sizeof(wac_info.operand) + 778 sizeof(wac_info.dbgWave_msg.DbgWaveMsg) + 779 sizeof(wac_info.dbgWave_msg.MemoryVA) + 780 sizeof(wac_info.trapId); 781 782 dev = kfd_device_by_id(args->gpu_id); 783 if (!dev) 784 return -EINVAL; 785 786 if (dev->device_info->asic_family == CHIP_CARRIZO) { 787 pr_debug("kfd_ioctl_dbg_wave_control not supported on CZ\n"); 788 return -EINVAL; 789 } 790 791 /* input size must match the computed "compact" size */ 792 if (args->buf_size_in_bytes != computed_buff_size) { 793 pr_debug("size mismatch, computed : actual %u : %u\n", 794 args->buf_size_in_bytes, computed_buff_size); 795 return -EINVAL; 796 } 797 798 cmd_from_user = (void __user *) args->content_ptr; 799 800 if (cmd_from_user == NULL) 801 return -EINVAL; 802 803 /* copy the entire buffer from user */ 804 805 args_buff = memdup_user(cmd_from_user, 806 args->buf_size_in_bytes - sizeof(*args)); 807 if (IS_ERR(args_buff)) 808 return PTR_ERR(args_buff); 809 810 /* move ptr to the start of the "pay-load" area */ 811 wac_info.process = p; 812 813 wac_info.operand = *((enum HSA_DBG_WAVEOP *)(&args_buff[args_idx])); 814 args_idx += sizeof(wac_info.operand); 815 816 wac_info.mode = *((enum HSA_DBG_WAVEMODE *)(&args_buff[args_idx])); 817 args_idx += sizeof(wac_info.mode); 818 819 wac_info.trapId = *((uint32_t *)(&args_buff[args_idx])); 820 args_idx += sizeof(wac_info.trapId); 821 822 wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value = 823 *((uint32_t *)(&args_buff[args_idx])); 824 wac_info.dbgWave_msg.MemoryVA = NULL; 825 826 mutex_lock(kfd_get_dbgmgr_mutex()); 827 828 pr_debug("Calling dbg manager process %p, operand %u, mode %u, trapId %u, message %u\n", 829 wac_info.process, wac_info.operand, 830 wac_info.mode, wac_info.trapId, 831 wac_info.dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value); 832 833 status = kfd_dbgmgr_wave_control(dev->dbgmgr, &wac_info); 834 835 pr_debug("Returned status of dbg manager is %ld\n", status); 836 837 mutex_unlock(kfd_get_dbgmgr_mutex()); 838 839 kfree(args_buff); 840 841 return status; 842 } 843 844 static int kfd_ioctl_get_clock_counters(struct file *filep, 845 struct kfd_process *p, void *data) 846 { 847 struct kfd_ioctl_get_clock_counters_args *args = data; 848 struct kfd_dev *dev; 849 850 dev = kfd_device_by_id(args->gpu_id); 851 if (dev) 852 /* Reading GPU clock counter from KGD */ 853 args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(dev->kgd); 854 else 855 /* Node without GPU resource */ 856 args->gpu_clock_counter = 0; 857 858 /* No access to rdtsc. Using raw monotonic time */ 859 args->cpu_clock_counter = ktime_get_raw_ns(); 860 args->system_clock_counter = ktime_get_boottime_ns(); 861 862 /* Since the counter is in nano-seconds we use 1GHz frequency */ 863 args->system_clock_freq = 1000000000; 864 865 return 0; 866 } 867 868 869 static int kfd_ioctl_get_process_apertures(struct file *filp, 870 struct kfd_process *p, void *data) 871 { 872 struct kfd_ioctl_get_process_apertures_args *args = data; 873 struct kfd_process_device_apertures *pAperture; 874 int i; 875 876 dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid); 877 878 args->num_of_nodes = 0; 879 880 mutex_lock(&p->mutex); 881 /* Run over all pdd of the process */ 882 for (i = 0; i < p->n_pdds; i++) { 883 struct kfd_process_device *pdd = p->pdds[i]; 884 885 pAperture = 886 &args->process_apertures[args->num_of_nodes]; 887 pAperture->gpu_id = pdd->dev->id; 888 pAperture->lds_base = pdd->lds_base; 889 pAperture->lds_limit = pdd->lds_limit; 890 pAperture->gpuvm_base = pdd->gpuvm_base; 891 pAperture->gpuvm_limit = pdd->gpuvm_limit; 892 pAperture->scratch_base = pdd->scratch_base; 893 pAperture->scratch_limit = pdd->scratch_limit; 894 895 dev_dbg(kfd_device, 896 "node id %u\n", args->num_of_nodes); 897 dev_dbg(kfd_device, 898 "gpu id %u\n", pdd->dev->id); 899 dev_dbg(kfd_device, 900 "lds_base %llX\n", pdd->lds_base); 901 dev_dbg(kfd_device, 902 "lds_limit %llX\n", pdd->lds_limit); 903 dev_dbg(kfd_device, 904 "gpuvm_base %llX\n", pdd->gpuvm_base); 905 dev_dbg(kfd_device, 906 "gpuvm_limit %llX\n", pdd->gpuvm_limit); 907 dev_dbg(kfd_device, 908 "scratch_base %llX\n", pdd->scratch_base); 909 dev_dbg(kfd_device, 910 "scratch_limit %llX\n", pdd->scratch_limit); 911 912 if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS) 913 break; 914 } 915 mutex_unlock(&p->mutex); 916 917 return 0; 918 } 919 920 static int kfd_ioctl_get_process_apertures_new(struct file *filp, 921 struct kfd_process *p, void *data) 922 { 923 struct kfd_ioctl_get_process_apertures_new_args *args = data; 924 struct kfd_process_device_apertures *pa; 925 int ret; 926 int i; 927 928 dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid); 929 930 if (args->num_of_nodes == 0) { 931 /* Return number of nodes, so that user space can alloacate 932 * sufficient memory 933 */ 934 mutex_lock(&p->mutex); 935 args->num_of_nodes = p->n_pdds; 936 goto out_unlock; 937 } 938 939 /* Fill in process-aperture information for all available 940 * nodes, but not more than args->num_of_nodes as that is 941 * the amount of memory allocated by user 942 */ 943 pa = kzalloc((sizeof(struct kfd_process_device_apertures) * 944 args->num_of_nodes), GFP_KERNEL); 945 if (!pa) 946 return -ENOMEM; 947 948 mutex_lock(&p->mutex); 949 950 if (!p->n_pdds) { 951 args->num_of_nodes = 0; 952 kfree(pa); 953 goto out_unlock; 954 } 955 956 /* Run over all pdd of the process */ 957 for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) { 958 struct kfd_process_device *pdd = p->pdds[i]; 959 960 pa[i].gpu_id = pdd->dev->id; 961 pa[i].lds_base = pdd->lds_base; 962 pa[i].lds_limit = pdd->lds_limit; 963 pa[i].gpuvm_base = pdd->gpuvm_base; 964 pa[i].gpuvm_limit = pdd->gpuvm_limit; 965 pa[i].scratch_base = pdd->scratch_base; 966 pa[i].scratch_limit = pdd->scratch_limit; 967 968 dev_dbg(kfd_device, 969 "gpu id %u\n", pdd->dev->id); 970 dev_dbg(kfd_device, 971 "lds_base %llX\n", pdd->lds_base); 972 dev_dbg(kfd_device, 973 "lds_limit %llX\n", pdd->lds_limit); 974 dev_dbg(kfd_device, 975 "gpuvm_base %llX\n", pdd->gpuvm_base); 976 dev_dbg(kfd_device, 977 "gpuvm_limit %llX\n", pdd->gpuvm_limit); 978 dev_dbg(kfd_device, 979 "scratch_base %llX\n", pdd->scratch_base); 980 dev_dbg(kfd_device, 981 "scratch_limit %llX\n", pdd->scratch_limit); 982 } 983 mutex_unlock(&p->mutex); 984 985 args->num_of_nodes = i; 986 ret = copy_to_user( 987 (void __user *)args->kfd_process_device_apertures_ptr, 988 pa, 989 (i * sizeof(struct kfd_process_device_apertures))); 990 kfree(pa); 991 return ret ? -EFAULT : 0; 992 993 out_unlock: 994 mutex_unlock(&p->mutex); 995 return 0; 996 } 997 998 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p, 999 void *data) 1000 { 1001 struct kfd_ioctl_create_event_args *args = data; 1002 int err; 1003 1004 /* For dGPUs the event page is allocated in user mode. The 1005 * handle is passed to KFD with the first call to this IOCTL 1006 * through the event_page_offset field. 1007 */ 1008 if (args->event_page_offset) { 1009 struct kfd_dev *kfd; 1010 struct kfd_process_device *pdd; 1011 void *mem, *kern_addr; 1012 uint64_t size; 1013 1014 if (p->signal_page) { 1015 pr_err("Event page is already set\n"); 1016 return -EINVAL; 1017 } 1018 1019 kfd = kfd_device_by_id(GET_GPU_ID(args->event_page_offset)); 1020 if (!kfd) { 1021 pr_err("Getting device by id failed in %s\n", __func__); 1022 return -EINVAL; 1023 } 1024 1025 mutex_lock(&p->mutex); 1026 pdd = kfd_bind_process_to_device(kfd, p); 1027 if (IS_ERR(pdd)) { 1028 err = PTR_ERR(pdd); 1029 goto out_unlock; 1030 } 1031 1032 mem = kfd_process_device_translate_handle(pdd, 1033 GET_IDR_HANDLE(args->event_page_offset)); 1034 if (!mem) { 1035 pr_err("Can't find BO, offset is 0x%llx\n", 1036 args->event_page_offset); 1037 err = -EINVAL; 1038 goto out_unlock; 1039 } 1040 mutex_unlock(&p->mutex); 1041 1042 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(kfd->kgd, 1043 mem, &kern_addr, &size); 1044 if (err) { 1045 pr_err("Failed to map event page to kernel\n"); 1046 return err; 1047 } 1048 1049 err = kfd_event_page_set(p, kern_addr, size); 1050 if (err) { 1051 pr_err("Failed to set event page\n"); 1052 return err; 1053 } 1054 } 1055 1056 err = kfd_event_create(filp, p, args->event_type, 1057 args->auto_reset != 0, args->node_id, 1058 &args->event_id, &args->event_trigger_data, 1059 &args->event_page_offset, 1060 &args->event_slot_index); 1061 1062 return err; 1063 1064 out_unlock: 1065 mutex_unlock(&p->mutex); 1066 return err; 1067 } 1068 1069 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p, 1070 void *data) 1071 { 1072 struct kfd_ioctl_destroy_event_args *args = data; 1073 1074 return kfd_event_destroy(p, args->event_id); 1075 } 1076 1077 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p, 1078 void *data) 1079 { 1080 struct kfd_ioctl_set_event_args *args = data; 1081 1082 return kfd_set_event(p, args->event_id); 1083 } 1084 1085 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p, 1086 void *data) 1087 { 1088 struct kfd_ioctl_reset_event_args *args = data; 1089 1090 return kfd_reset_event(p, args->event_id); 1091 } 1092 1093 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p, 1094 void *data) 1095 { 1096 struct kfd_ioctl_wait_events_args *args = data; 1097 int err; 1098 1099 err = kfd_wait_on_events(p, args->num_events, 1100 (void __user *)args->events_ptr, 1101 (args->wait_for_all != 0), 1102 args->timeout, &args->wait_result); 1103 1104 return err; 1105 } 1106 static int kfd_ioctl_set_scratch_backing_va(struct file *filep, 1107 struct kfd_process *p, void *data) 1108 { 1109 struct kfd_ioctl_set_scratch_backing_va_args *args = data; 1110 struct kfd_process_device *pdd; 1111 struct kfd_dev *dev; 1112 long err; 1113 1114 dev = kfd_device_by_id(args->gpu_id); 1115 if (!dev) 1116 return -EINVAL; 1117 1118 mutex_lock(&p->mutex); 1119 1120 pdd = kfd_bind_process_to_device(dev, p); 1121 if (IS_ERR(pdd)) { 1122 err = PTR_ERR(pdd); 1123 goto bind_process_to_device_fail; 1124 } 1125 1126 pdd->qpd.sh_hidden_private_base = args->va_addr; 1127 1128 mutex_unlock(&p->mutex); 1129 1130 if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS && 1131 pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va) 1132 dev->kfd2kgd->set_scratch_backing_va( 1133 dev->kgd, args->va_addr, pdd->qpd.vmid); 1134 1135 return 0; 1136 1137 bind_process_to_device_fail: 1138 mutex_unlock(&p->mutex); 1139 return err; 1140 } 1141 1142 static int kfd_ioctl_get_tile_config(struct file *filep, 1143 struct kfd_process *p, void *data) 1144 { 1145 struct kfd_ioctl_get_tile_config_args *args = data; 1146 struct kfd_dev *dev; 1147 struct tile_config config; 1148 int err = 0; 1149 1150 dev = kfd_device_by_id(args->gpu_id); 1151 if (!dev) 1152 return -EINVAL; 1153 1154 amdgpu_amdkfd_get_tile_config(dev->kgd, &config); 1155 1156 args->gb_addr_config = config.gb_addr_config; 1157 args->num_banks = config.num_banks; 1158 args->num_ranks = config.num_ranks; 1159 1160 if (args->num_tile_configs > config.num_tile_configs) 1161 args->num_tile_configs = config.num_tile_configs; 1162 err = copy_to_user((void __user *)args->tile_config_ptr, 1163 config.tile_config_ptr, 1164 args->num_tile_configs * sizeof(uint32_t)); 1165 if (err) { 1166 args->num_tile_configs = 0; 1167 return -EFAULT; 1168 } 1169 1170 if (args->num_macro_tile_configs > config.num_macro_tile_configs) 1171 args->num_macro_tile_configs = 1172 config.num_macro_tile_configs; 1173 err = copy_to_user((void __user *)args->macro_tile_config_ptr, 1174 config.macro_tile_config_ptr, 1175 args->num_macro_tile_configs * sizeof(uint32_t)); 1176 if (err) { 1177 args->num_macro_tile_configs = 0; 1178 return -EFAULT; 1179 } 1180 1181 return 0; 1182 } 1183 1184 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p, 1185 void *data) 1186 { 1187 struct kfd_ioctl_acquire_vm_args *args = data; 1188 struct kfd_process_device *pdd; 1189 struct kfd_dev *dev; 1190 struct file *drm_file; 1191 int ret; 1192 1193 dev = kfd_device_by_id(args->gpu_id); 1194 if (!dev) 1195 return -EINVAL; 1196 1197 drm_file = fget(args->drm_fd); 1198 if (!drm_file) 1199 return -EINVAL; 1200 1201 mutex_lock(&p->mutex); 1202 1203 pdd = kfd_get_process_device_data(dev, p); 1204 if (!pdd) { 1205 ret = -EINVAL; 1206 goto err_unlock; 1207 } 1208 1209 if (pdd->drm_file) { 1210 ret = pdd->drm_file == drm_file ? 0 : -EBUSY; 1211 goto err_unlock; 1212 } 1213 1214 ret = kfd_process_device_init_vm(pdd, drm_file); 1215 if (ret) 1216 goto err_unlock; 1217 /* On success, the PDD keeps the drm_file reference */ 1218 mutex_unlock(&p->mutex); 1219 1220 return 0; 1221 1222 err_unlock: 1223 mutex_unlock(&p->mutex); 1224 fput(drm_file); 1225 return ret; 1226 } 1227 1228 bool kfd_dev_is_large_bar(struct kfd_dev *dev) 1229 { 1230 struct kfd_local_mem_info mem_info; 1231 1232 if (debug_largebar) { 1233 pr_debug("Simulate large-bar allocation on non large-bar machine\n"); 1234 return true; 1235 } 1236 1237 if (dev->use_iommu_v2) 1238 return false; 1239 1240 amdgpu_amdkfd_get_local_mem_info(dev->kgd, &mem_info); 1241 if (mem_info.local_mem_size_private == 0 && 1242 mem_info.local_mem_size_public > 0) 1243 return true; 1244 return false; 1245 } 1246 1247 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, 1248 struct kfd_process *p, void *data) 1249 { 1250 struct kfd_ioctl_alloc_memory_of_gpu_args *args = data; 1251 struct kfd_process_device *pdd; 1252 void *mem; 1253 struct kfd_dev *dev; 1254 int idr_handle; 1255 long err; 1256 uint64_t offset = args->mmap_offset; 1257 uint32_t flags = args->flags; 1258 1259 if (args->size == 0) 1260 return -EINVAL; 1261 1262 dev = kfd_device_by_id(args->gpu_id); 1263 if (!dev) 1264 return -EINVAL; 1265 1266 if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) && 1267 (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) && 1268 !kfd_dev_is_large_bar(dev)) { 1269 pr_err("Alloc host visible vram on small bar is not allowed\n"); 1270 return -EINVAL; 1271 } 1272 1273 mutex_lock(&p->mutex); 1274 1275 pdd = kfd_bind_process_to_device(dev, p); 1276 if (IS_ERR(pdd)) { 1277 err = PTR_ERR(pdd); 1278 goto err_unlock; 1279 } 1280 1281 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) { 1282 if (args->size != kfd_doorbell_process_slice(dev)) { 1283 err = -EINVAL; 1284 goto err_unlock; 1285 } 1286 offset = kfd_get_process_doorbells(pdd); 1287 } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) { 1288 if (args->size != PAGE_SIZE) { 1289 err = -EINVAL; 1290 goto err_unlock; 1291 } 1292 offset = amdgpu_amdkfd_get_mmio_remap_phys_addr(dev->kgd); 1293 if (!offset) { 1294 err = -ENOMEM; 1295 goto err_unlock; 1296 } 1297 } 1298 1299 err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( 1300 dev->kgd, args->va_addr, args->size, 1301 pdd->drm_priv, (struct kgd_mem **) &mem, &offset, 1302 flags); 1303 1304 if (err) 1305 goto err_unlock; 1306 1307 idr_handle = kfd_process_device_create_obj_handle(pdd, mem); 1308 if (idr_handle < 0) { 1309 err = -EFAULT; 1310 goto err_free; 1311 } 1312 1313 /* Update the VRAM usage count */ 1314 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) 1315 WRITE_ONCE(pdd->vram_usage, pdd->vram_usage + args->size); 1316 1317 mutex_unlock(&p->mutex); 1318 1319 args->handle = MAKE_HANDLE(args->gpu_id, idr_handle); 1320 args->mmap_offset = offset; 1321 1322 /* MMIO is mapped through kfd device 1323 * Generate a kfd mmap offset 1324 */ 1325 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) 1326 args->mmap_offset = KFD_MMAP_TYPE_MMIO 1327 | KFD_MMAP_GPU_ID(args->gpu_id); 1328 1329 return 0; 1330 1331 err_free: 1332 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem, 1333 pdd->drm_priv, NULL); 1334 err_unlock: 1335 mutex_unlock(&p->mutex); 1336 return err; 1337 } 1338 1339 static int kfd_ioctl_free_memory_of_gpu(struct file *filep, 1340 struct kfd_process *p, void *data) 1341 { 1342 struct kfd_ioctl_free_memory_of_gpu_args *args = data; 1343 struct kfd_process_device *pdd; 1344 void *mem; 1345 struct kfd_dev *dev; 1346 int ret; 1347 uint64_t size = 0; 1348 1349 dev = kfd_device_by_id(GET_GPU_ID(args->handle)); 1350 if (!dev) 1351 return -EINVAL; 1352 1353 mutex_lock(&p->mutex); 1354 1355 pdd = kfd_get_process_device_data(dev, p); 1356 if (!pdd) { 1357 pr_err("Process device data doesn't exist\n"); 1358 ret = -EINVAL; 1359 goto err_unlock; 1360 } 1361 1362 mem = kfd_process_device_translate_handle( 1363 pdd, GET_IDR_HANDLE(args->handle)); 1364 if (!mem) { 1365 ret = -EINVAL; 1366 goto err_unlock; 1367 } 1368 1369 ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, 1370 (struct kgd_mem *)mem, pdd->drm_priv, &size); 1371 1372 /* If freeing the buffer failed, leave the handle in place for 1373 * clean-up during process tear-down. 1374 */ 1375 if (!ret) 1376 kfd_process_device_remove_obj_handle( 1377 pdd, GET_IDR_HANDLE(args->handle)); 1378 1379 WRITE_ONCE(pdd->vram_usage, pdd->vram_usage - size); 1380 1381 err_unlock: 1382 mutex_unlock(&p->mutex); 1383 return ret; 1384 } 1385 1386 static int kfd_ioctl_map_memory_to_gpu(struct file *filep, 1387 struct kfd_process *p, void *data) 1388 { 1389 struct kfd_ioctl_map_memory_to_gpu_args *args = data; 1390 struct kfd_process_device *pdd, *peer_pdd; 1391 void *mem; 1392 struct kfd_dev *dev, *peer; 1393 long err = 0; 1394 int i; 1395 uint32_t *devices_arr = NULL; 1396 bool table_freed = false; 1397 1398 dev = kfd_device_by_id(GET_GPU_ID(args->handle)); 1399 if (!dev) 1400 return -EINVAL; 1401 1402 if (!args->n_devices) { 1403 pr_debug("Device IDs array empty\n"); 1404 return -EINVAL; 1405 } 1406 if (args->n_success > args->n_devices) { 1407 pr_debug("n_success exceeds n_devices\n"); 1408 return -EINVAL; 1409 } 1410 1411 devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr), 1412 GFP_KERNEL); 1413 if (!devices_arr) 1414 return -ENOMEM; 1415 1416 err = copy_from_user(devices_arr, 1417 (void __user *)args->device_ids_array_ptr, 1418 args->n_devices * sizeof(*devices_arr)); 1419 if (err != 0) { 1420 err = -EFAULT; 1421 goto copy_from_user_failed; 1422 } 1423 1424 mutex_lock(&p->mutex); 1425 1426 pdd = kfd_bind_process_to_device(dev, p); 1427 if (IS_ERR(pdd)) { 1428 err = PTR_ERR(pdd); 1429 goto bind_process_to_device_failed; 1430 } 1431 1432 mem = kfd_process_device_translate_handle(pdd, 1433 GET_IDR_HANDLE(args->handle)); 1434 if (!mem) { 1435 err = -ENOMEM; 1436 goto get_mem_obj_from_handle_failed; 1437 } 1438 1439 for (i = args->n_success; i < args->n_devices; i++) { 1440 peer = kfd_device_by_id(devices_arr[i]); 1441 if (!peer) { 1442 pr_debug("Getting device by id failed for 0x%x\n", 1443 devices_arr[i]); 1444 err = -EINVAL; 1445 goto get_mem_obj_from_handle_failed; 1446 } 1447 1448 peer_pdd = kfd_bind_process_to_device(peer, p); 1449 if (IS_ERR(peer_pdd)) { 1450 err = PTR_ERR(peer_pdd); 1451 goto get_mem_obj_from_handle_failed; 1452 } 1453 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu( 1454 peer->kgd, (struct kgd_mem *)mem, 1455 peer_pdd->drm_priv, &table_freed); 1456 if (err) { 1457 pr_err("Failed to map to gpu %d/%d\n", 1458 i, args->n_devices); 1459 goto map_memory_to_gpu_failed; 1460 } 1461 args->n_success = i+1; 1462 } 1463 1464 mutex_unlock(&p->mutex); 1465 1466 err = amdgpu_amdkfd_gpuvm_sync_memory(dev->kgd, (struct kgd_mem *) mem, true); 1467 if (err) { 1468 pr_debug("Sync memory failed, wait interrupted by user signal\n"); 1469 goto sync_memory_failed; 1470 } 1471 1472 /* Flush TLBs after waiting for the page table updates to complete */ 1473 if (table_freed) { 1474 for (i = 0; i < args->n_devices; i++) { 1475 peer = kfd_device_by_id(devices_arr[i]); 1476 if (WARN_ON_ONCE(!peer)) 1477 continue; 1478 peer_pdd = kfd_get_process_device_data(peer, p); 1479 if (WARN_ON_ONCE(!peer_pdd)) 1480 continue; 1481 kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY); 1482 } 1483 } 1484 kfree(devices_arr); 1485 1486 return err; 1487 1488 bind_process_to_device_failed: 1489 get_mem_obj_from_handle_failed: 1490 map_memory_to_gpu_failed: 1491 mutex_unlock(&p->mutex); 1492 copy_from_user_failed: 1493 sync_memory_failed: 1494 kfree(devices_arr); 1495 1496 return err; 1497 } 1498 1499 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, 1500 struct kfd_process *p, void *data) 1501 { 1502 struct kfd_ioctl_unmap_memory_from_gpu_args *args = data; 1503 struct kfd_process_device *pdd, *peer_pdd; 1504 void *mem; 1505 struct kfd_dev *dev, *peer; 1506 long err = 0; 1507 uint32_t *devices_arr = NULL, i; 1508 1509 dev = kfd_device_by_id(GET_GPU_ID(args->handle)); 1510 if (!dev) 1511 return -EINVAL; 1512 1513 if (!args->n_devices) { 1514 pr_debug("Device IDs array empty\n"); 1515 return -EINVAL; 1516 } 1517 if (args->n_success > args->n_devices) { 1518 pr_debug("n_success exceeds n_devices\n"); 1519 return -EINVAL; 1520 } 1521 1522 devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr), 1523 GFP_KERNEL); 1524 if (!devices_arr) 1525 return -ENOMEM; 1526 1527 err = copy_from_user(devices_arr, 1528 (void __user *)args->device_ids_array_ptr, 1529 args->n_devices * sizeof(*devices_arr)); 1530 if (err != 0) { 1531 err = -EFAULT; 1532 goto copy_from_user_failed; 1533 } 1534 1535 mutex_lock(&p->mutex); 1536 1537 pdd = kfd_get_process_device_data(dev, p); 1538 if (!pdd) { 1539 err = -EINVAL; 1540 goto bind_process_to_device_failed; 1541 } 1542 1543 mem = kfd_process_device_translate_handle(pdd, 1544 GET_IDR_HANDLE(args->handle)); 1545 if (!mem) { 1546 err = -ENOMEM; 1547 goto get_mem_obj_from_handle_failed; 1548 } 1549 1550 for (i = args->n_success; i < args->n_devices; i++) { 1551 peer = kfd_device_by_id(devices_arr[i]); 1552 if (!peer) { 1553 err = -EINVAL; 1554 goto get_mem_obj_from_handle_failed; 1555 } 1556 1557 peer_pdd = kfd_get_process_device_data(peer, p); 1558 if (!peer_pdd) { 1559 err = -ENODEV; 1560 goto get_mem_obj_from_handle_failed; 1561 } 1562 err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( 1563 peer->kgd, (struct kgd_mem *)mem, peer_pdd->drm_priv); 1564 if (err) { 1565 pr_err("Failed to unmap from gpu %d/%d\n", 1566 i, args->n_devices); 1567 goto unmap_memory_from_gpu_failed; 1568 } 1569 args->n_success = i+1; 1570 } 1571 mutex_unlock(&p->mutex); 1572 1573 err = amdgpu_amdkfd_gpuvm_sync_memory(dev->kgd, (struct kgd_mem *) mem, true); 1574 if (err) { 1575 pr_debug("Sync memory failed, wait interrupted by user signal\n"); 1576 goto sync_memory_failed; 1577 } 1578 1579 /* Flush TLBs after waiting for the page table updates to complete */ 1580 for (i = 0; i < args->n_devices; i++) { 1581 peer = kfd_device_by_id(devices_arr[i]); 1582 if (WARN_ON_ONCE(!peer)) 1583 continue; 1584 peer_pdd = kfd_get_process_device_data(peer, p); 1585 if (WARN_ON_ONCE(!peer_pdd)) 1586 continue; 1587 kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT); 1588 } 1589 1590 kfree(devices_arr); 1591 1592 return 0; 1593 1594 bind_process_to_device_failed: 1595 get_mem_obj_from_handle_failed: 1596 unmap_memory_from_gpu_failed: 1597 mutex_unlock(&p->mutex); 1598 copy_from_user_failed: 1599 sync_memory_failed: 1600 kfree(devices_arr); 1601 return err; 1602 } 1603 1604 static int kfd_ioctl_alloc_queue_gws(struct file *filep, 1605 struct kfd_process *p, void *data) 1606 { 1607 int retval; 1608 struct kfd_ioctl_alloc_queue_gws_args *args = data; 1609 struct queue *q; 1610 struct kfd_dev *dev; 1611 1612 mutex_lock(&p->mutex); 1613 q = pqm_get_user_queue(&p->pqm, args->queue_id); 1614 1615 if (q) { 1616 dev = q->device; 1617 } else { 1618 retval = -EINVAL; 1619 goto out_unlock; 1620 } 1621 1622 if (!dev->gws) { 1623 retval = -ENODEV; 1624 goto out_unlock; 1625 } 1626 1627 if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { 1628 retval = -ENODEV; 1629 goto out_unlock; 1630 } 1631 1632 retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL); 1633 mutex_unlock(&p->mutex); 1634 1635 args->first_gws = 0; 1636 return retval; 1637 1638 out_unlock: 1639 mutex_unlock(&p->mutex); 1640 return retval; 1641 } 1642 1643 static int kfd_ioctl_get_dmabuf_info(struct file *filep, 1644 struct kfd_process *p, void *data) 1645 { 1646 struct kfd_ioctl_get_dmabuf_info_args *args = data; 1647 struct kfd_dev *dev = NULL; 1648 struct kgd_dev *dma_buf_kgd; 1649 void *metadata_buffer = NULL; 1650 uint32_t flags; 1651 unsigned int i; 1652 int r; 1653 1654 /* Find a KFD GPU device that supports the get_dmabuf_info query */ 1655 for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++) 1656 if (dev) 1657 break; 1658 if (!dev) 1659 return -EINVAL; 1660 1661 if (args->metadata_ptr) { 1662 metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL); 1663 if (!metadata_buffer) 1664 return -ENOMEM; 1665 } 1666 1667 /* Get dmabuf info from KGD */ 1668 r = amdgpu_amdkfd_get_dmabuf_info(dev->kgd, args->dmabuf_fd, 1669 &dma_buf_kgd, &args->size, 1670 metadata_buffer, args->metadata_size, 1671 &args->metadata_size, &flags); 1672 if (r) 1673 goto exit; 1674 1675 /* Reverse-lookup gpu_id from kgd pointer */ 1676 dev = kfd_device_by_kgd(dma_buf_kgd); 1677 if (!dev) { 1678 r = -EINVAL; 1679 goto exit; 1680 } 1681 args->gpu_id = dev->id; 1682 args->flags = flags; 1683 1684 /* Copy metadata buffer to user mode */ 1685 if (metadata_buffer) { 1686 r = copy_to_user((void __user *)args->metadata_ptr, 1687 metadata_buffer, args->metadata_size); 1688 if (r != 0) 1689 r = -EFAULT; 1690 } 1691 1692 exit: 1693 kfree(metadata_buffer); 1694 1695 return r; 1696 } 1697 1698 static int kfd_ioctl_import_dmabuf(struct file *filep, 1699 struct kfd_process *p, void *data) 1700 { 1701 struct kfd_ioctl_import_dmabuf_args *args = data; 1702 struct kfd_process_device *pdd; 1703 struct dma_buf *dmabuf; 1704 struct kfd_dev *dev; 1705 int idr_handle; 1706 uint64_t size; 1707 void *mem; 1708 int r; 1709 1710 dev = kfd_device_by_id(args->gpu_id); 1711 if (!dev) 1712 return -EINVAL; 1713 1714 dmabuf = dma_buf_get(args->dmabuf_fd); 1715 if (IS_ERR(dmabuf)) 1716 return PTR_ERR(dmabuf); 1717 1718 mutex_lock(&p->mutex); 1719 1720 pdd = kfd_bind_process_to_device(dev, p); 1721 if (IS_ERR(pdd)) { 1722 r = PTR_ERR(pdd); 1723 goto err_unlock; 1724 } 1725 1726 r = amdgpu_amdkfd_gpuvm_import_dmabuf(dev->kgd, dmabuf, 1727 args->va_addr, pdd->drm_priv, 1728 (struct kgd_mem **)&mem, &size, 1729 NULL); 1730 if (r) 1731 goto err_unlock; 1732 1733 idr_handle = kfd_process_device_create_obj_handle(pdd, mem); 1734 if (idr_handle < 0) { 1735 r = -EFAULT; 1736 goto err_free; 1737 } 1738 1739 mutex_unlock(&p->mutex); 1740 dma_buf_put(dmabuf); 1741 1742 args->handle = MAKE_HANDLE(args->gpu_id, idr_handle); 1743 1744 return 0; 1745 1746 err_free: 1747 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, (struct kgd_mem *)mem, 1748 pdd->drm_priv, NULL); 1749 err_unlock: 1750 mutex_unlock(&p->mutex); 1751 dma_buf_put(dmabuf); 1752 return r; 1753 } 1754 1755 /* Handle requests for watching SMI events */ 1756 static int kfd_ioctl_smi_events(struct file *filep, 1757 struct kfd_process *p, void *data) 1758 { 1759 struct kfd_ioctl_smi_events_args *args = data; 1760 struct kfd_dev *dev; 1761 1762 dev = kfd_device_by_id(args->gpuid); 1763 if (!dev) 1764 return -EINVAL; 1765 1766 return kfd_smi_event_open(dev, &args->anon_fd); 1767 } 1768 1769 static int kfd_ioctl_set_xnack_mode(struct file *filep, 1770 struct kfd_process *p, void *data) 1771 { 1772 struct kfd_ioctl_set_xnack_mode_args *args = data; 1773 int r = 0; 1774 1775 mutex_lock(&p->mutex); 1776 if (args->xnack_enabled >= 0) { 1777 if (!list_empty(&p->pqm.queues)) { 1778 pr_debug("Process has user queues running\n"); 1779 mutex_unlock(&p->mutex); 1780 return -EBUSY; 1781 } 1782 if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) 1783 r = -EPERM; 1784 else 1785 p->xnack_enabled = args->xnack_enabled; 1786 } else { 1787 args->xnack_enabled = p->xnack_enabled; 1788 } 1789 mutex_unlock(&p->mutex); 1790 1791 return r; 1792 } 1793 1794 #if IS_ENABLED(CONFIG_HSA_AMD_SVM) 1795 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data) 1796 { 1797 struct kfd_ioctl_svm_args *args = data; 1798 int r = 0; 1799 1800 pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n", 1801 args->start_addr, args->size, args->op, args->nattr); 1802 1803 if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK)) 1804 return -EINVAL; 1805 if (!args->start_addr || !args->size) 1806 return -EINVAL; 1807 1808 mutex_lock(&p->mutex); 1809 1810 r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr, 1811 args->attrs); 1812 1813 mutex_unlock(&p->mutex); 1814 1815 return r; 1816 } 1817 #else 1818 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data) 1819 { 1820 return -EPERM; 1821 } 1822 #endif 1823 1824 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \ 1825 [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \ 1826 .cmd_drv = 0, .name = #ioctl} 1827 1828 /** Ioctl table */ 1829 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = { 1830 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION, 1831 kfd_ioctl_get_version, 0), 1832 1833 AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE, 1834 kfd_ioctl_create_queue, 0), 1835 1836 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE, 1837 kfd_ioctl_destroy_queue, 0), 1838 1839 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY, 1840 kfd_ioctl_set_memory_policy, 0), 1841 1842 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS, 1843 kfd_ioctl_get_clock_counters, 0), 1844 1845 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES, 1846 kfd_ioctl_get_process_apertures, 0), 1847 1848 AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE, 1849 kfd_ioctl_update_queue, 0), 1850 1851 AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT, 1852 kfd_ioctl_create_event, 0), 1853 1854 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT, 1855 kfd_ioctl_destroy_event, 0), 1856 1857 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT, 1858 kfd_ioctl_set_event, 0), 1859 1860 AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT, 1861 kfd_ioctl_reset_event, 0), 1862 1863 AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS, 1864 kfd_ioctl_wait_events, 0), 1865 1866 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER, 1867 kfd_ioctl_dbg_register, 0), 1868 1869 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER, 1870 kfd_ioctl_dbg_unregister, 0), 1871 1872 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH, 1873 kfd_ioctl_dbg_address_watch, 0), 1874 1875 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL, 1876 kfd_ioctl_dbg_wave_control, 0), 1877 1878 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA, 1879 kfd_ioctl_set_scratch_backing_va, 0), 1880 1881 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG, 1882 kfd_ioctl_get_tile_config, 0), 1883 1884 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER, 1885 kfd_ioctl_set_trap_handler, 0), 1886 1887 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW, 1888 kfd_ioctl_get_process_apertures_new, 0), 1889 1890 AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM, 1891 kfd_ioctl_acquire_vm, 0), 1892 1893 AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU, 1894 kfd_ioctl_alloc_memory_of_gpu, 0), 1895 1896 AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU, 1897 kfd_ioctl_free_memory_of_gpu, 0), 1898 1899 AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU, 1900 kfd_ioctl_map_memory_to_gpu, 0), 1901 1902 AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU, 1903 kfd_ioctl_unmap_memory_from_gpu, 0), 1904 1905 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK, 1906 kfd_ioctl_set_cu_mask, 0), 1907 1908 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE, 1909 kfd_ioctl_get_queue_wave_state, 0), 1910 1911 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO, 1912 kfd_ioctl_get_dmabuf_info, 0), 1913 1914 AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF, 1915 kfd_ioctl_import_dmabuf, 0), 1916 1917 AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS, 1918 kfd_ioctl_alloc_queue_gws, 0), 1919 1920 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS, 1921 kfd_ioctl_smi_events, 0), 1922 1923 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0), 1924 1925 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE, 1926 kfd_ioctl_set_xnack_mode, 0), 1927 }; 1928 1929 #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls) 1930 1931 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg) 1932 { 1933 struct kfd_process *process; 1934 amdkfd_ioctl_t *func; 1935 const struct amdkfd_ioctl_desc *ioctl = NULL; 1936 unsigned int nr = _IOC_NR(cmd); 1937 char stack_kdata[128]; 1938 char *kdata = NULL; 1939 unsigned int usize, asize; 1940 int retcode = -EINVAL; 1941 1942 if (nr >= AMDKFD_CORE_IOCTL_COUNT) 1943 goto err_i1; 1944 1945 if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) { 1946 u32 amdkfd_size; 1947 1948 ioctl = &amdkfd_ioctls[nr]; 1949 1950 amdkfd_size = _IOC_SIZE(ioctl->cmd); 1951 usize = asize = _IOC_SIZE(cmd); 1952 if (amdkfd_size > asize) 1953 asize = amdkfd_size; 1954 1955 cmd = ioctl->cmd; 1956 } else 1957 goto err_i1; 1958 1959 dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg); 1960 1961 /* Get the process struct from the filep. Only the process 1962 * that opened /dev/kfd can use the file descriptor. Child 1963 * processes need to create their own KFD device context. 1964 */ 1965 process = filep->private_data; 1966 if (process->lead_thread != current->group_leader) { 1967 dev_dbg(kfd_device, "Using KFD FD in wrong process\n"); 1968 retcode = -EBADF; 1969 goto err_i1; 1970 } 1971 1972 /* Do not trust userspace, use our own definition */ 1973 func = ioctl->func; 1974 1975 if (unlikely(!func)) { 1976 dev_dbg(kfd_device, "no function\n"); 1977 retcode = -EINVAL; 1978 goto err_i1; 1979 } 1980 1981 if (cmd & (IOC_IN | IOC_OUT)) { 1982 if (asize <= sizeof(stack_kdata)) { 1983 kdata = stack_kdata; 1984 } else { 1985 kdata = kmalloc(asize, GFP_KERNEL); 1986 if (!kdata) { 1987 retcode = -ENOMEM; 1988 goto err_i1; 1989 } 1990 } 1991 if (asize > usize) 1992 memset(kdata + usize, 0, asize - usize); 1993 } 1994 1995 if (cmd & IOC_IN) { 1996 if (copy_from_user(kdata, (void __user *)arg, usize) != 0) { 1997 retcode = -EFAULT; 1998 goto err_i1; 1999 } 2000 } else if (cmd & IOC_OUT) { 2001 memset(kdata, 0, usize); 2002 } 2003 2004 retcode = func(filep, process, kdata); 2005 2006 if (cmd & IOC_OUT) 2007 if (copy_to_user((void __user *)arg, kdata, usize) != 0) 2008 retcode = -EFAULT; 2009 2010 err_i1: 2011 if (!ioctl) 2012 dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n", 2013 task_pid_nr(current), cmd, nr); 2014 2015 if (kdata != stack_kdata) 2016 kfree(kdata); 2017 2018 if (retcode) 2019 dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n", 2020 nr, arg, retcode); 2021 2022 return retcode; 2023 } 2024 2025 static int kfd_mmio_mmap(struct kfd_dev *dev, struct kfd_process *process, 2026 struct vm_area_struct *vma) 2027 { 2028 phys_addr_t address; 2029 int ret; 2030 2031 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 2032 return -EINVAL; 2033 2034 address = amdgpu_amdkfd_get_mmio_remap_phys_addr(dev->kgd); 2035 2036 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE | 2037 VM_DONTDUMP | VM_PFNMAP; 2038 2039 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 2040 2041 pr_debug("pasid 0x%x mapping mmio page\n" 2042 " target user address == 0x%08llX\n" 2043 " physical address == 0x%08llX\n" 2044 " vm_flags == 0x%04lX\n" 2045 " size == 0x%04lX\n", 2046 process->pasid, (unsigned long long) vma->vm_start, 2047 address, vma->vm_flags, PAGE_SIZE); 2048 2049 ret = io_remap_pfn_range(vma, 2050 vma->vm_start, 2051 address >> PAGE_SHIFT, 2052 PAGE_SIZE, 2053 vma->vm_page_prot); 2054 return ret; 2055 } 2056 2057 2058 static int kfd_mmap(struct file *filp, struct vm_area_struct *vma) 2059 { 2060 struct kfd_process *process; 2061 struct kfd_dev *dev = NULL; 2062 unsigned long mmap_offset; 2063 unsigned int gpu_id; 2064 2065 process = kfd_get_process(current); 2066 if (IS_ERR(process)) 2067 return PTR_ERR(process); 2068 2069 mmap_offset = vma->vm_pgoff << PAGE_SHIFT; 2070 gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset); 2071 if (gpu_id) 2072 dev = kfd_device_by_id(gpu_id); 2073 2074 switch (mmap_offset & KFD_MMAP_TYPE_MASK) { 2075 case KFD_MMAP_TYPE_DOORBELL: 2076 if (!dev) 2077 return -ENODEV; 2078 return kfd_doorbell_mmap(dev, process, vma); 2079 2080 case KFD_MMAP_TYPE_EVENTS: 2081 return kfd_event_mmap(process, vma); 2082 2083 case KFD_MMAP_TYPE_RESERVED_MEM: 2084 if (!dev) 2085 return -ENODEV; 2086 return kfd_reserved_mem_mmap(dev, process, vma); 2087 case KFD_MMAP_TYPE_MMIO: 2088 if (!dev) 2089 return -ENODEV; 2090 return kfd_mmio_mmap(dev, process, vma); 2091 } 2092 2093 return -EFAULT; 2094 } 2095