1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright 2014-2022 Advanced Micro Devices, Inc. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice shall be included in 13 * all copies or substantial portions of the Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 21 * OTHER DEALINGS IN THE SOFTWARE. 22 */ 23 24 #include <linux/capability.h> 25 #include <linux/device.h> 26 #include <linux/err.h> 27 #include <linux/fs.h> 28 #include <linux/file.h> 29 #include <linux/overflow.h> 30 #include <linux/sched.h> 31 #include <linux/slab.h> 32 #include <linux/uaccess.h> 33 #include <linux/compat.h> 34 #include <uapi/linux/kfd_ioctl.h> 35 #include <linux/time.h> 36 #include <linux/mm.h> 37 #include <linux/mman.h> 38 #include <linux/ptrace.h> 39 #include <linux/dma-buf.h> 40 #include <linux/processor.h> 41 #include "kfd_priv.h" 42 #include "kfd_device_queue_manager.h" 43 #include "kfd_svm.h" 44 #include "amdgpu_amdkfd.h" 45 #include "kfd_smi_events.h" 46 #include "amdgpu_dma_buf.h" 47 #include "kfd_debug.h" 48 #include "amdgpu_ptl.h" 49 50 static long kfd_ioctl(struct file *, unsigned int, unsigned long); 51 static int kfd_open(struct inode *, struct file *); 52 static int kfd_release(struct inode *, struct file *); 53 static int kfd_mmap(struct file *, struct vm_area_struct *); 54 55 static const char kfd_dev_name[] = "kfd"; 56 57 static const struct file_operations kfd_fops = { 58 .owner = THIS_MODULE, 59 .unlocked_ioctl = kfd_ioctl, 60 .compat_ioctl = compat_ptr_ioctl, 61 .open = kfd_open, 62 .release = kfd_release, 63 .mmap = kfd_mmap, 64 }; 65 66 static int kfd_char_dev_major = -1; 67 struct device *kfd_device; 68 static const struct class kfd_class = { 69 .name = kfd_dev_name, 70 }; 71 72 /* 73 * Cache the address space of the chardev on first open so that the reset 74 * path can drop all userspace mappings of doorbell and MMIO ranges via 75 * unmap_mapping_range(). 76 */ 77 static struct address_space *kfd_dev_mapping; 78 79 void kfd_dev_unmap_mapping_range(loff_t const holebegin, loff_t const holelen) 80 { 81 struct address_space *mapping = READ_ONCE(kfd_dev_mapping); 82 83 if (mapping) 84 unmap_mapping_range(mapping, holebegin, holelen, 1); 85 } 86 87 static inline struct kfd_process_device *kfd_lock_pdd_by_id(struct kfd_process *p, __u32 gpu_id) 88 { 89 struct kfd_process_device *pdd; 90 91 mutex_lock(&p->mutex); 92 pdd = kfd_process_device_data_by_id(p, gpu_id); 93 94 if (pdd) 95 return pdd; 96 97 mutex_unlock(&p->mutex); 98 return NULL; 99 } 100 101 static inline void kfd_unlock_pdd(struct kfd_process_device *pdd) 102 { 103 mutex_unlock(&pdd->process->mutex); 104 } 105 106 int kfd_chardev_init(void) 107 { 108 int err = 0; 109 110 kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops); 111 err = kfd_char_dev_major; 112 if (err < 0) 113 goto err_register_chrdev; 114 115 err = class_register(&kfd_class); 116 if (err) 117 goto err_class_create; 118 119 kfd_device = device_create(&kfd_class, NULL, 120 MKDEV(kfd_char_dev_major, 0), 121 NULL, kfd_dev_name); 122 err = PTR_ERR(kfd_device); 123 if (IS_ERR(kfd_device)) 124 goto err_device_create; 125 126 return 0; 127 128 err_device_create: 129 class_unregister(&kfd_class); 130 err_class_create: 131 unregister_chrdev(kfd_char_dev_major, kfd_dev_name); 132 err_register_chrdev: 133 return err; 134 } 135 136 void kfd_chardev_exit(void) 137 { 138 device_destroy(&kfd_class, MKDEV(kfd_char_dev_major, 0)); 139 class_unregister(&kfd_class); 140 unregister_chrdev(kfd_char_dev_major, kfd_dev_name); 141 kfd_device = NULL; 142 } 143 144 145 static int kfd_open(struct inode *inode, struct file *filep) 146 { 147 struct kfd_process *process; 148 bool is_32bit_user_mode; 149 150 if (iminor(inode) != 0) 151 return -ENODEV; 152 153 /* 154 * /dev/kfd is a single chardev so all opens share one inode. Cache 155 * its address_space on the first open for use by the reset path. 156 */ 157 if (!READ_ONCE(kfd_dev_mapping)) 158 cmpxchg(&kfd_dev_mapping, NULL, inode->i_mapping); 159 160 is_32bit_user_mode = in_compat_syscall(); 161 162 if (is_32bit_user_mode) { 163 dev_warn(kfd_device, 164 "Process %d (32-bit) failed to open /dev/kfd\n" 165 "32-bit processes are not supported by amdkfd\n", 166 current->pid); 167 return -EPERM; 168 } 169 170 process = kfd_create_process(current); 171 if (IS_ERR(process)) 172 return PTR_ERR(process); 173 174 /* filep now owns the reference returned by kfd_create_process */ 175 filep->private_data = process; 176 177 dev_dbg(kfd_device, "process pid %d opened kfd node, compat mode (32 bit) - %d\n", 178 process->lead_thread->pid, process->is_32bit_user_mode); 179 180 return 0; 181 } 182 183 static int kfd_release(struct inode *inode, struct file *filep) 184 { 185 struct kfd_process *process = filep->private_data; 186 187 if (!process) 188 return 0; 189 190 if (process->context_id != KFD_CONTEXT_ID_PRIMARY) 191 kfd_process_notifier_release_internal(process); 192 193 kfd_unref_process(process); 194 195 return 0; 196 } 197 198 static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p, 199 void *data) 200 { 201 struct kfd_ioctl_get_version_args *args = data; 202 203 args->major_version = KFD_IOCTL_MAJOR_VERSION; 204 args->minor_version = KFD_IOCTL_MINOR_VERSION; 205 206 return 0; 207 } 208 209 static int set_queue_properties_from_user(struct queue_properties *q_properties, 210 struct kfd_ioctl_create_queue_args *args) 211 { 212 /* 213 * Repurpose queue percentage to accommodate new features: 214 * bit 0-7: queue percentage 215 * bit 8-15: pm4_target_xcc 216 */ 217 if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) { 218 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n"); 219 return -EINVAL; 220 } 221 222 if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) { 223 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n"); 224 return -EINVAL; 225 } 226 227 if ((args->ring_base_address) && 228 (!access_ok((const void __user *) args->ring_base_address, 229 sizeof(uint64_t)))) { 230 pr_err("Can't access ring base address\n"); 231 return -EFAULT; 232 } 233 234 if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) { 235 pr_err("Ring size must be a power of 2 or 0\n"); 236 return -EINVAL; 237 } 238 239 if (args->ring_size < KFD_MIN_QUEUE_RING_SIZE) { 240 args->ring_size = KFD_MIN_QUEUE_RING_SIZE; 241 pr_debug("Size lower. clamped to KFD_MIN_QUEUE_RING_SIZE"); 242 } 243 244 if ((args->metadata_ring_size != 0) && !is_power_of_2(args->metadata_ring_size)) { 245 pr_err("Metadata ring size must be a power of 2 or 0\n"); 246 return -EINVAL; 247 } 248 249 if (!access_ok((const void __user *) args->read_pointer_address, 250 sizeof(uint32_t))) { 251 pr_err("Can't access read pointer\n"); 252 return -EFAULT; 253 } 254 255 if (!access_ok((const void __user *) args->write_pointer_address, 256 sizeof(uint32_t))) { 257 pr_err("Can't access write pointer\n"); 258 return -EFAULT; 259 } 260 261 if (args->eop_buffer_address && 262 !access_ok((const void __user *) args->eop_buffer_address, 263 sizeof(uint32_t))) { 264 pr_debug("Can't access eop buffer"); 265 return -EFAULT; 266 } 267 268 if (args->ctx_save_restore_address && 269 !access_ok((const void __user *) args->ctx_save_restore_address, 270 sizeof(uint32_t))) { 271 pr_debug("Can't access ctx save restore buffer"); 272 return -EFAULT; 273 } 274 275 q_properties->is_interop = false; 276 q_properties->is_gws = false; 277 q_properties->queue_percent = args->queue_percentage & 0xFF; 278 /* bit 8-15 are repurposed to be PM4 target XCC */ 279 q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF; 280 q_properties->priority = args->queue_priority; 281 q_properties->queue_address = args->ring_base_address; 282 q_properties->queue_size = args->ring_size; 283 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL) 284 q_properties->metadata_queue_size = args->metadata_ring_size; 285 286 q_properties->read_ptr = (void __user *)args->read_pointer_address; 287 q_properties->write_ptr = (void __user *)args->write_pointer_address; 288 q_properties->eop_ring_buffer_address = args->eop_buffer_address; 289 q_properties->eop_ring_buffer_size = args->eop_buffer_size; 290 q_properties->ctx_save_restore_area_address = 291 args->ctx_save_restore_address; 292 q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size; 293 q_properties->ctl_stack_size = args->ctl_stack_size; 294 q_properties->sdma_engine_id = args->sdma_engine_id; 295 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE || 296 args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL) 297 q_properties->type = KFD_QUEUE_TYPE_COMPUTE; 298 else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA) 299 q_properties->type = KFD_QUEUE_TYPE_SDMA; 300 else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_XGMI) 301 q_properties->type = KFD_QUEUE_TYPE_SDMA_XGMI; 302 else if (args->queue_type == KFD_IOC_QUEUE_TYPE_SDMA_BY_ENG_ID) 303 q_properties->type = KFD_QUEUE_TYPE_SDMA_BY_ENG_ID; 304 else 305 return -ENOTSUPP; 306 307 if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL) 308 q_properties->format = KFD_QUEUE_FORMAT_AQL; 309 else 310 q_properties->format = KFD_QUEUE_FORMAT_PM4; 311 312 pr_debug("Queue Percentage: %d, %d\n", 313 q_properties->queue_percent, args->queue_percentage); 314 315 pr_debug("Queue Priority: %d, %d\n", 316 q_properties->priority, args->queue_priority); 317 318 pr_debug("Queue Address: 0x%llX, 0x%llX\n", 319 q_properties->queue_address, args->ring_base_address); 320 321 pr_debug("Queue Size: 0x%llX, %u\n", 322 q_properties->queue_size, args->ring_size); 323 324 pr_debug("Queue r/w Pointers: %px, %px\n", 325 q_properties->read_ptr, 326 q_properties->write_ptr); 327 328 pr_debug("Queue Format: %d\n", q_properties->format); 329 330 pr_debug("Queue EOP: 0x%llX\n", q_properties->eop_ring_buffer_address); 331 332 pr_debug("Queue CTX save area: 0x%llX\n", 333 q_properties->ctx_save_restore_area_address); 334 335 return 0; 336 } 337 338 static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, 339 void *data) 340 { 341 struct kfd_ioctl_create_queue_args *args = data; 342 struct kfd_node *dev; 343 int err = 0; 344 unsigned int queue_id; 345 struct kfd_process_device *pdd; 346 struct queue_properties q_properties; 347 uint32_t doorbell_offset_in_process = 0; 348 349 memset(&q_properties, 0, sizeof(struct queue_properties)); 350 351 pr_debug("Creating queue ioctl\n"); 352 353 err = set_queue_properties_from_user(&q_properties, args); 354 if (err) 355 return err; 356 357 pr_debug("Looking for gpu id 0x%x\n", args->gpu_id); 358 359 mutex_lock(&p->mutex); 360 361 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 362 if (!pdd) { 363 pr_debug("Could not find gpu id 0x%x\n", args->gpu_id); 364 err = -EINVAL; 365 goto err_pdd; 366 } 367 dev = pdd->dev; 368 369 pdd = kfd_bind_process_to_device(dev, p); 370 if (IS_ERR(pdd)) { 371 err = -ESRCH; 372 goto err_bind_process; 373 } 374 375 if (q_properties.type == KFD_QUEUE_TYPE_SDMA_BY_ENG_ID) { 376 int max_sdma_eng_id = kfd_get_num_sdma_engines(dev) + 377 kfd_get_num_xgmi_sdma_engines(dev) - 1; 378 379 if (q_properties.sdma_engine_id > max_sdma_eng_id) { 380 err = -EINVAL; 381 pr_err("sdma_engine_id %i exceeds maximum id of %i\n", 382 q_properties.sdma_engine_id, max_sdma_eng_id); 383 goto err_sdma_engine_id; 384 } 385 } 386 387 if (!pdd->qpd.proc_doorbells) { 388 err = kfd_alloc_process_doorbells(dev->kfd, pdd); 389 if (err) { 390 pr_debug("failed to allocate process doorbells\n"); 391 goto err_bind_process; 392 } 393 } 394 395 err = kfd_queue_acquire_buffers(pdd, &q_properties); 396 if (err) { 397 pr_debug("failed to acquire user queue buffers\n"); 398 goto err_acquire_queue_buf; 399 } 400 401 pr_debug("Creating queue for process pid %d on gpu 0x%x\n", 402 p->lead_thread->pid, 403 dev->id); 404 405 err = pqm_create_queue(&p->pqm, dev, &q_properties, &queue_id, 406 NULL, NULL, NULL, &doorbell_offset_in_process); 407 if (err != 0) 408 goto err_create_queue; 409 410 args->queue_id = queue_id; 411 412 413 /* Return gpu_id as doorbell offset for mmap usage */ 414 args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL; 415 args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id); 416 if (KFD_IS_SOC15(dev)) 417 /* On SOC15 ASICs, include the doorbell offset within the 418 * process doorbell frame, which is 2 pages. 419 */ 420 args->doorbell_offset |= doorbell_offset_in_process; 421 422 mutex_unlock(&p->mutex); 423 424 pr_debug("Queue id %d was created successfully\n", args->queue_id); 425 426 pr_debug("Ring buffer address == 0x%016llX\n", 427 args->ring_base_address); 428 429 pr_debug("Read ptr address == 0x%016llX\n", 430 args->read_pointer_address); 431 432 pr_debug("Write ptr address == 0x%016llX\n", 433 args->write_pointer_address); 434 435 kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0); 436 return 0; 437 438 err_create_queue: 439 kfd_queue_unref_bo_vas(pdd, &q_properties); 440 kfd_queue_release_buffers(pdd, &q_properties); 441 err_acquire_queue_buf: 442 err_sdma_engine_id: 443 err_bind_process: 444 err_pdd: 445 mutex_unlock(&p->mutex); 446 return err; 447 } 448 449 static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p, 450 void *data) 451 { 452 int retval; 453 struct kfd_ioctl_destroy_queue_args *args = data; 454 455 pr_debug("Destroying queue id %d for process pid %d\n", 456 args->queue_id, 457 p->lead_thread->pid); 458 459 mutex_lock(&p->mutex); 460 461 retval = pqm_destroy_queue(&p->pqm, args->queue_id); 462 463 mutex_unlock(&p->mutex); 464 return retval; 465 } 466 467 static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p, 468 void *data) 469 { 470 int retval; 471 struct kfd_ioctl_update_queue_args *args = data; 472 struct queue_properties properties; 473 474 /* 475 * Repurpose queue percentage to accommodate new features: 476 * bit 0-7: queue percentage 477 * bit 8-15: pm4_target_xcc 478 */ 479 if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) { 480 pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n"); 481 return -EINVAL; 482 } 483 484 if (args->queue_priority > KFD_MAX_QUEUE_PRIORITY) { 485 pr_err("Queue priority must be between 0 to KFD_MAX_QUEUE_PRIORITY\n"); 486 return -EINVAL; 487 } 488 489 if ((args->ring_base_address) && 490 (!access_ok((const void __user *) args->ring_base_address, 491 sizeof(uint64_t)))) { 492 pr_err("Can't access ring base address\n"); 493 return -EFAULT; 494 } 495 496 if (!is_power_of_2(args->ring_size) && (args->ring_size != 0)) { 497 pr_err("Ring size must be a power of 2 or 0\n"); 498 return -EINVAL; 499 } 500 501 if (args->ring_size < KFD_MIN_QUEUE_RING_SIZE) { 502 args->ring_size = KFD_MIN_QUEUE_RING_SIZE; 503 pr_debug("Size lower. clamped to KFD_MIN_QUEUE_RING_SIZE"); 504 } 505 506 properties.queue_address = args->ring_base_address; 507 properties.queue_size = args->ring_size; 508 properties.queue_percent = args->queue_percentage & 0xFF; 509 /* bit 8-15 are repurposed to be PM4 target XCC */ 510 properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF; 511 properties.priority = args->queue_priority; 512 513 pr_debug("Updating queue id %d for process pid %d\n", 514 args->queue_id, p->lead_thread->pid); 515 516 mutex_lock(&p->mutex); 517 518 retval = pqm_update_queue_properties(&p->pqm, args->queue_id, &properties); 519 520 mutex_unlock(&p->mutex); 521 522 return retval; 523 } 524 525 static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p, 526 void *data) 527 { 528 int retval; 529 const int max_num_cus = 1024; 530 struct kfd_ioctl_set_cu_mask_args *args = data; 531 struct mqd_update_info minfo = {0}; 532 uint32_t __user *cu_mask_ptr = (uint32_t __user *)args->cu_mask_ptr; 533 size_t cu_mask_size = sizeof(uint32_t) * (args->num_cu_mask / 32); 534 535 if ((args->num_cu_mask % 32) != 0) { 536 pr_debug("num_cu_mask 0x%x must be a multiple of 32", 537 args->num_cu_mask); 538 return -EINVAL; 539 } 540 541 minfo.cu_mask.count = args->num_cu_mask; 542 if (minfo.cu_mask.count == 0) { 543 pr_debug("CU mask cannot be 0"); 544 return -EINVAL; 545 } 546 547 /* To prevent an unreasonably large CU mask size, set an arbitrary 548 * limit of max_num_cus bits. We can then just drop any CU mask bits 549 * past max_num_cus bits and just use the first max_num_cus bits. 550 */ 551 if (minfo.cu_mask.count > max_num_cus) { 552 pr_debug("CU mask cannot be greater than 1024 bits"); 553 minfo.cu_mask.count = max_num_cus; 554 cu_mask_size = sizeof(uint32_t) * (max_num_cus/32); 555 } 556 557 minfo.cu_mask.ptr = memdup_user(cu_mask_ptr, cu_mask_size); 558 if (IS_ERR(minfo.cu_mask.ptr)) { 559 pr_debug("Could not copy CU mask from userspace"); 560 return PTR_ERR(minfo.cu_mask.ptr); 561 } 562 563 mutex_lock(&p->mutex); 564 565 retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo); 566 567 mutex_unlock(&p->mutex); 568 569 kfree(minfo.cu_mask.ptr); 570 return retval; 571 } 572 573 static int kfd_ioctl_get_queue_wave_state(struct file *filep, 574 struct kfd_process *p, void *data) 575 { 576 struct kfd_ioctl_get_queue_wave_state_args *args = data; 577 int r; 578 579 mutex_lock(&p->mutex); 580 581 r = pqm_get_wave_state(&p->pqm, args->queue_id, 582 (void __user *)args->ctl_stack_address, 583 &args->ctl_stack_used_size, 584 &args->save_area_used_size); 585 586 mutex_unlock(&p->mutex); 587 588 return r; 589 } 590 591 static int kfd_ioctl_set_memory_policy(struct file *filep, 592 struct kfd_process *p, void *data) 593 { 594 struct kfd_ioctl_set_memory_policy_args *args = data; 595 int err = 0; 596 struct kfd_process_device *pdd; 597 enum cache_policy default_policy, alternate_policy; 598 599 if (args->default_policy != KFD_IOC_CACHE_POLICY_COHERENT 600 && args->default_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) { 601 return -EINVAL; 602 } 603 604 if (args->alternate_policy != KFD_IOC_CACHE_POLICY_COHERENT 605 && args->alternate_policy != KFD_IOC_CACHE_POLICY_NONCOHERENT) { 606 return -EINVAL; 607 } 608 609 mutex_lock(&p->mutex); 610 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 611 if (!pdd) { 612 pr_debug("Could not find gpu id 0x%x\n", args->gpu_id); 613 err = -EINVAL; 614 goto err_pdd; 615 } 616 617 pdd = kfd_bind_process_to_device(pdd->dev, p); 618 if (IS_ERR(pdd)) { 619 err = -ESRCH; 620 goto out; 621 } 622 623 default_policy = (args->default_policy == KFD_IOC_CACHE_POLICY_COHERENT) 624 ? cache_policy_coherent : cache_policy_noncoherent; 625 626 alternate_policy = 627 (args->alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT) 628 ? cache_policy_coherent : cache_policy_noncoherent; 629 630 if (!pdd->dev->dqm->ops.set_cache_memory_policy(pdd->dev->dqm, 631 &pdd->qpd, 632 default_policy, 633 alternate_policy, 634 (void __user *)args->alternate_aperture_base, 635 args->alternate_aperture_size, 636 args->misc_process_flag)) 637 err = -EINVAL; 638 639 out: 640 err_pdd: 641 mutex_unlock(&p->mutex); 642 643 return err; 644 } 645 646 static int kfd_ioctl_set_trap_handler(struct file *filep, 647 struct kfd_process *p, void *data) 648 { 649 struct kfd_ioctl_set_trap_handler_args *args = data; 650 int err = 0; 651 struct kfd_process_device *pdd; 652 653 mutex_lock(&p->mutex); 654 655 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 656 if (!pdd) { 657 err = -EINVAL; 658 goto err_pdd; 659 } 660 661 pdd = kfd_bind_process_to_device(pdd->dev, p); 662 if (IS_ERR(pdd)) { 663 err = -ESRCH; 664 goto out; 665 } 666 667 kfd_process_set_trap_handler(&pdd->qpd, args->tba_addr, args->tma_addr); 668 669 out: 670 err_pdd: 671 mutex_unlock(&p->mutex); 672 673 return err; 674 } 675 676 static int kfd_ioctl_dbg_register(struct file *filep, 677 struct kfd_process *p, void *data) 678 { 679 return -EPERM; 680 } 681 682 static int kfd_ioctl_dbg_unregister(struct file *filep, 683 struct kfd_process *p, void *data) 684 { 685 return -EPERM; 686 } 687 688 static int kfd_ioctl_dbg_address_watch(struct file *filep, 689 struct kfd_process *p, void *data) 690 { 691 return -EPERM; 692 } 693 694 /* Parse and generate fixed size data structure for wave control */ 695 static int kfd_ioctl_dbg_wave_control(struct file *filep, 696 struct kfd_process *p, void *data) 697 { 698 return -EPERM; 699 } 700 701 static int kfd_ioctl_get_clock_counters(struct file *filep, 702 struct kfd_process *p, void *data) 703 { 704 struct kfd_ioctl_get_clock_counters_args *args = data; 705 struct kfd_process_device *pdd; 706 707 mutex_lock(&p->mutex); 708 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 709 mutex_unlock(&p->mutex); 710 if (pdd) 711 /* Reading GPU clock counter from KGD */ 712 args->gpu_clock_counter = amdgpu_amdkfd_get_gpu_clock_counter(pdd->dev->adev); 713 else 714 /* Node without GPU resource */ 715 args->gpu_clock_counter = 0; 716 717 /* No access to rdtsc. Using raw monotonic time */ 718 args->cpu_clock_counter = ktime_get_raw_ns(); 719 args->system_clock_counter = ktime_get_boottime_ns(); 720 721 /* Since the counter is in nano-seconds we use 1GHz frequency */ 722 args->system_clock_freq = 1000000000; 723 724 return 0; 725 } 726 727 728 static int kfd_ioctl_get_process_apertures(struct file *filp, 729 struct kfd_process *p, void *data) 730 { 731 struct kfd_ioctl_get_process_apertures_args *args = data; 732 struct kfd_process_device_apertures *pAperture; 733 int i; 734 735 dev_dbg(kfd_device, "get apertures for process pid %d", p->lead_thread->pid); 736 737 args->num_of_nodes = 0; 738 739 mutex_lock(&p->mutex); 740 /* Run over all pdd of the process */ 741 for (i = 0; i < p->n_pdds; i++) { 742 struct kfd_process_device *pdd = p->pdds[i]; 743 744 pAperture = 745 &args->process_apertures[args->num_of_nodes]; 746 pAperture->gpu_id = pdd->dev->id; 747 pAperture->lds_base = pdd->lds_base; 748 pAperture->lds_limit = pdd->lds_limit; 749 pAperture->gpuvm_base = pdd->gpuvm_base; 750 pAperture->gpuvm_limit = pdd->gpuvm_limit; 751 pAperture->scratch_base = pdd->scratch_base; 752 pAperture->scratch_limit = pdd->scratch_limit; 753 754 dev_dbg(kfd_device, 755 "node id %u\n", args->num_of_nodes); 756 dev_dbg(kfd_device, 757 "gpu id %u\n", pdd->dev->id); 758 dev_dbg(kfd_device, 759 "lds_base %llX\n", pdd->lds_base); 760 dev_dbg(kfd_device, 761 "lds_limit %llX\n", pdd->lds_limit); 762 dev_dbg(kfd_device, 763 "gpuvm_base %llX\n", pdd->gpuvm_base); 764 dev_dbg(kfd_device, 765 "gpuvm_limit %llX\n", pdd->gpuvm_limit); 766 dev_dbg(kfd_device, 767 "scratch_base %llX\n", pdd->scratch_base); 768 dev_dbg(kfd_device, 769 "scratch_limit %llX\n", pdd->scratch_limit); 770 771 if (++args->num_of_nodes >= NUM_OF_SUPPORTED_GPUS) 772 break; 773 } 774 mutex_unlock(&p->mutex); 775 776 return 0; 777 } 778 779 static int kfd_ioctl_get_process_apertures_new(struct file *filp, 780 struct kfd_process *p, void *data) 781 { 782 struct kfd_ioctl_get_process_apertures_new_args *args = data; 783 struct kfd_process_device_apertures *pa; 784 int ret; 785 int i; 786 787 dev_dbg(kfd_device, "get apertures for process pid %d", 788 p->lead_thread->pid); 789 790 if (args->num_of_nodes == 0) { 791 /* Return number of nodes, so that user space can alloacate 792 * sufficient memory 793 */ 794 mutex_lock(&p->mutex); 795 args->num_of_nodes = p->n_pdds; 796 goto out_unlock; 797 } 798 799 if (args->num_of_nodes > kfd_topology_get_num_devices()) 800 return -EINVAL; 801 802 /* Fill in process-aperture information for all available 803 * nodes, but not more than args->num_of_nodes as that is 804 * the amount of memory allocated by user 805 */ 806 pa = kzalloc_objs(struct kfd_process_device_apertures, 807 args->num_of_nodes); 808 if (!pa) 809 return -ENOMEM; 810 811 mutex_lock(&p->mutex); 812 813 if (!p->n_pdds) { 814 args->num_of_nodes = 0; 815 kfree(pa); 816 goto out_unlock; 817 } 818 819 /* Run over all pdd of the process */ 820 for (i = 0; i < min(p->n_pdds, args->num_of_nodes); i++) { 821 struct kfd_process_device *pdd = p->pdds[i]; 822 823 pa[i].gpu_id = pdd->dev->id; 824 pa[i].lds_base = pdd->lds_base; 825 pa[i].lds_limit = pdd->lds_limit; 826 pa[i].gpuvm_base = pdd->gpuvm_base; 827 pa[i].gpuvm_limit = pdd->gpuvm_limit; 828 pa[i].scratch_base = pdd->scratch_base; 829 pa[i].scratch_limit = pdd->scratch_limit; 830 831 dev_dbg(kfd_device, 832 "gpu id %u\n", pdd->dev->id); 833 dev_dbg(kfd_device, 834 "lds_base %llX\n", pdd->lds_base); 835 dev_dbg(kfd_device, 836 "lds_limit %llX\n", pdd->lds_limit); 837 dev_dbg(kfd_device, 838 "gpuvm_base %llX\n", pdd->gpuvm_base); 839 dev_dbg(kfd_device, 840 "gpuvm_limit %llX\n", pdd->gpuvm_limit); 841 dev_dbg(kfd_device, 842 "scratch_base %llX\n", pdd->scratch_base); 843 dev_dbg(kfd_device, 844 "scratch_limit %llX\n", pdd->scratch_limit); 845 } 846 mutex_unlock(&p->mutex); 847 848 args->num_of_nodes = i; 849 ret = copy_to_user( 850 (void __user *)args->kfd_process_device_apertures_ptr, 851 pa, 852 (i * sizeof(struct kfd_process_device_apertures))); 853 kfree(pa); 854 return ret ? -EFAULT : 0; 855 856 out_unlock: 857 mutex_unlock(&p->mutex); 858 return 0; 859 } 860 861 static int kfd_ioctl_create_event(struct file *filp, struct kfd_process *p, 862 void *data) 863 { 864 struct kfd_ioctl_create_event_args *args = data; 865 int err; 866 867 /* For dGPUs the event page is allocated in user mode. The 868 * handle is passed to KFD with the first call to this IOCTL 869 * through the event_page_offset field. 870 */ 871 if (args->event_page_offset) { 872 mutex_lock(&p->mutex); 873 err = kfd_kmap_event_page(p, args->event_page_offset); 874 mutex_unlock(&p->mutex); 875 if (err) 876 return err; 877 } 878 879 err = kfd_event_create(filp, p, args->event_type, 880 args->auto_reset != 0, args->node_id, 881 &args->event_id, &args->event_trigger_data, 882 &args->event_page_offset, 883 &args->event_slot_index); 884 885 pr_debug("Created event (id:0x%08x) (%s)\n", args->event_id, __func__); 886 return err; 887 } 888 889 static int kfd_ioctl_destroy_event(struct file *filp, struct kfd_process *p, 890 void *data) 891 { 892 struct kfd_ioctl_destroy_event_args *args = data; 893 894 return kfd_event_destroy(p, args->event_id); 895 } 896 897 static int kfd_ioctl_set_event(struct file *filp, struct kfd_process *p, 898 void *data) 899 { 900 struct kfd_ioctl_set_event_args *args = data; 901 902 return kfd_set_event(p, args->event_id); 903 } 904 905 static int kfd_ioctl_reset_event(struct file *filp, struct kfd_process *p, 906 void *data) 907 { 908 struct kfd_ioctl_reset_event_args *args = data; 909 910 return kfd_reset_event(p, args->event_id); 911 } 912 913 static int kfd_ioctl_wait_events(struct file *filp, struct kfd_process *p, 914 void *data) 915 { 916 struct kfd_ioctl_wait_events_args *args = data; 917 918 return kfd_wait_on_events(p, args->num_events, 919 (void __user *)args->events_ptr, 920 (args->wait_for_all != 0), 921 &args->timeout, &args->wait_result); 922 } 923 static int kfd_ioctl_set_scratch_backing_va(struct file *filep, 924 struct kfd_process *p, void *data) 925 { 926 struct kfd_ioctl_set_scratch_backing_va_args *args = data; 927 struct kfd_process_device *pdd; 928 struct kfd_node *dev; 929 long err; 930 931 mutex_lock(&p->mutex); 932 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 933 if (!pdd) { 934 err = -EINVAL; 935 goto err_pdd; 936 } 937 dev = pdd->dev; 938 939 pdd = kfd_bind_process_to_device(dev, p); 940 if (IS_ERR(pdd)) { 941 err = PTR_ERR(pdd); 942 goto bind_process_to_device_fail; 943 } 944 945 pdd->qpd.sh_hidden_private_base = args->va_addr; 946 947 mutex_unlock(&p->mutex); 948 949 if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS && 950 pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va) 951 dev->kfd2kgd->set_scratch_backing_va( 952 dev->adev, args->va_addr, pdd->qpd.vmid); 953 954 return 0; 955 956 bind_process_to_device_fail: 957 err_pdd: 958 mutex_unlock(&p->mutex); 959 return err; 960 } 961 962 static int kfd_ioctl_get_tile_config(struct file *filep, 963 struct kfd_process *p, void *data) 964 { 965 struct kfd_ioctl_get_tile_config_args *args = data; 966 struct kfd_process_device *pdd; 967 struct tile_config config; 968 int err = 0; 969 970 mutex_lock(&p->mutex); 971 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 972 mutex_unlock(&p->mutex); 973 if (!pdd) 974 return -EINVAL; 975 976 amdgpu_amdkfd_get_tile_config(pdd->dev->adev, &config); 977 978 args->gb_addr_config = config.gb_addr_config; 979 args->num_banks = config.num_banks; 980 args->num_ranks = config.num_ranks; 981 982 if (args->num_tile_configs > config.num_tile_configs) 983 args->num_tile_configs = config.num_tile_configs; 984 err = copy_to_user((void __user *)args->tile_config_ptr, 985 config.tile_config_ptr, 986 args->num_tile_configs * sizeof(uint32_t)); 987 if (err) { 988 args->num_tile_configs = 0; 989 return -EFAULT; 990 } 991 992 if (args->num_macro_tile_configs > config.num_macro_tile_configs) 993 args->num_macro_tile_configs = 994 config.num_macro_tile_configs; 995 err = copy_to_user((void __user *)args->macro_tile_config_ptr, 996 config.macro_tile_config_ptr, 997 args->num_macro_tile_configs * sizeof(uint32_t)); 998 if (err) { 999 args->num_macro_tile_configs = 0; 1000 return -EFAULT; 1001 } 1002 1003 return 0; 1004 } 1005 1006 static int kfd_ioctl_acquire_vm(struct file *filep, struct kfd_process *p, 1007 void *data) 1008 { 1009 struct kfd_ioctl_acquire_vm_args *args = data; 1010 struct kfd_process_device *pdd; 1011 struct file *drm_file; 1012 int ret; 1013 1014 drm_file = fget(args->drm_fd); 1015 if (!drm_file) 1016 return -EINVAL; 1017 1018 mutex_lock(&p->mutex); 1019 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 1020 if (!pdd) { 1021 ret = -EINVAL; 1022 goto err_pdd; 1023 } 1024 1025 if (pdd->drm_file) { 1026 ret = pdd->drm_file == drm_file ? 0 : -EBUSY; 1027 goto err_drm_file; 1028 } 1029 1030 ret = kfd_process_device_init_vm(pdd, drm_file); 1031 if (ret) 1032 goto err_unlock; 1033 1034 /* On success, the PDD keeps the drm_file reference */ 1035 mutex_unlock(&p->mutex); 1036 1037 return 0; 1038 1039 err_unlock: 1040 err_pdd: 1041 err_drm_file: 1042 mutex_unlock(&p->mutex); 1043 fput(drm_file); 1044 return ret; 1045 } 1046 1047 bool kfd_dev_is_large_bar(struct kfd_node *dev) 1048 { 1049 if (dev->kfd->adev->debug_largebar) { 1050 pr_debug("Simulate large-bar allocation on non large-bar machine\n"); 1051 return true; 1052 } 1053 1054 if (dev->local_mem_info.local_mem_size_private == 0 && 1055 dev->local_mem_info.local_mem_size_public > 0) 1056 return true; 1057 1058 if (dev->local_mem_info.local_mem_size_public == 0 && 1059 dev->kfd->adev->gmc.is_app_apu) { 1060 pr_debug("APP APU, Consider like a large bar system\n"); 1061 return true; 1062 } 1063 1064 return false; 1065 } 1066 1067 static int kfd_ioctl_get_available_memory(struct file *filep, 1068 struct kfd_process *p, void *data) 1069 { 1070 struct kfd_ioctl_get_available_memory_args *args = data; 1071 struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id); 1072 1073 if (!pdd) 1074 return -EINVAL; 1075 args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev, 1076 pdd->dev->node_id); 1077 kfd_unlock_pdd(pdd); 1078 return 0; 1079 } 1080 1081 static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, 1082 struct kfd_process *p, void *data) 1083 { 1084 struct kfd_ioctl_alloc_memory_of_gpu_args *args = data; 1085 struct kfd_process_device *pdd; 1086 void *mem; 1087 struct kfd_node *dev; 1088 int idr_handle; 1089 long err; 1090 uint64_t offset = args->mmap_offset; 1091 uint32_t flags = args->flags; 1092 1093 if (args->size == 0) 1094 return -EINVAL; 1095 1096 if (p->context_id != KFD_CONTEXT_ID_PRIMARY && (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)) { 1097 pr_debug("USERPTR is not supported on non-primary kfd_process\n"); 1098 1099 return -EOPNOTSUPP; 1100 } 1101 1102 #if IS_ENABLED(CONFIG_HSA_AMD_SVM) 1103 /* Flush pending deferred work to avoid racing with deferred actions 1104 * from previous memory map changes (e.g. munmap). 1105 */ 1106 svm_range_list_lock_and_flush_work(&p->svms, current->mm); 1107 mutex_lock(&p->svms.lock); 1108 mmap_write_unlock(current->mm); 1109 1110 /* Skip a special case that allocates VRAM without VA, 1111 * VA will be invalid of 0. 1112 */ 1113 if (!(!args->va_addr && (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)) && 1114 interval_tree_iter_first(&p->svms.objects, 1115 args->va_addr >> PAGE_SHIFT, 1116 (args->va_addr + args->size - 1) >> PAGE_SHIFT)) { 1117 pr_err("Address: 0x%llx already allocated by SVM\n", 1118 args->va_addr); 1119 mutex_unlock(&p->svms.lock); 1120 return -EADDRINUSE; 1121 } 1122 1123 /* When register user buffer check if it has been registered by svm by 1124 * buffer cpu virtual address. 1125 */ 1126 if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) && 1127 interval_tree_iter_first(&p->svms.objects, 1128 args->mmap_offset >> PAGE_SHIFT, 1129 (args->mmap_offset + args->size - 1) >> PAGE_SHIFT)) { 1130 pr_err("User Buffer Address: 0x%llx already allocated by SVM\n", 1131 args->mmap_offset); 1132 mutex_unlock(&p->svms.lock); 1133 return -EADDRINUSE; 1134 } 1135 1136 mutex_unlock(&p->svms.lock); 1137 #endif 1138 mutex_lock(&p->mutex); 1139 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 1140 if (!pdd) { 1141 err = -EINVAL; 1142 goto err_pdd; 1143 } 1144 1145 dev = pdd->dev; 1146 1147 if ((flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) && 1148 (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) && 1149 !kfd_dev_is_large_bar(dev)) { 1150 pr_err("Alloc host visible vram on small bar is not allowed\n"); 1151 err = -EINVAL; 1152 goto err_large_bar; 1153 } 1154 1155 pdd = kfd_bind_process_to_device(dev, p); 1156 if (IS_ERR(pdd)) { 1157 err = PTR_ERR(pdd); 1158 goto err_unlock; 1159 } 1160 1161 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) { 1162 if (args->size != kfd_doorbell_process_slice(dev->kfd)) { 1163 err = -EINVAL; 1164 goto err_unlock; 1165 } 1166 offset = kfd_get_process_doorbells(pdd); 1167 if (!offset) { 1168 err = -ENOMEM; 1169 goto err_unlock; 1170 } 1171 } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) { 1172 if (args->size != PAGE_SIZE) { 1173 err = -EINVAL; 1174 goto err_unlock; 1175 } 1176 offset = dev->adev->rmmio_remap.bus_addr; 1177 if (!offset || (PAGE_SIZE > 4096)) { 1178 err = -ENOMEM; 1179 goto err_unlock; 1180 } 1181 } 1182 1183 err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( 1184 dev->adev, args->va_addr, args->size, 1185 pdd->drm_priv, (struct kgd_mem **) &mem, &offset, 1186 flags, false); 1187 1188 if (err) 1189 goto err_unlock; 1190 1191 idr_handle = kfd_process_device_create_obj_handle(pdd, mem); 1192 if (idr_handle < 0) { 1193 err = -EFAULT; 1194 goto err_free; 1195 } 1196 1197 /* Update the VRAM usage count */ 1198 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { 1199 uint64_t size = args->size; 1200 1201 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM) 1202 size >>= 1; 1203 atomic64_add(PAGE_ALIGN(size), &pdd->vram_usage); 1204 } 1205 1206 mutex_unlock(&p->mutex); 1207 1208 args->handle = MAKE_HANDLE(args->gpu_id, idr_handle); 1209 args->mmap_offset = offset; 1210 1211 /* MMIO is mapped through kfd device 1212 * Generate a kfd mmap offset 1213 */ 1214 if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) 1215 args->mmap_offset = KFD_MMAP_TYPE_MMIO 1216 | KFD_MMAP_GPU_ID(args->gpu_id); 1217 1218 return 0; 1219 1220 err_free: 1221 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, (struct kgd_mem *)mem, 1222 pdd->drm_priv, NULL); 1223 err_unlock: 1224 err_pdd: 1225 err_large_bar: 1226 mutex_unlock(&p->mutex); 1227 return err; 1228 } 1229 1230 static int kfd_ioctl_free_memory_of_gpu(struct file *filep, 1231 struct kfd_process *p, void *data) 1232 { 1233 struct kfd_ioctl_free_memory_of_gpu_args *args = data; 1234 struct kfd_process_device *pdd; 1235 void *mem; 1236 int ret; 1237 uint64_t size = 0; 1238 1239 mutex_lock(&p->mutex); 1240 /* 1241 * Safeguard to prevent user space from freeing signal BO. 1242 * It will be freed at process termination. 1243 */ 1244 if (p->signal_handle && (p->signal_handle == args->handle)) { 1245 pr_err("Free signal BO is not allowed\n"); 1246 ret = -EPERM; 1247 goto err_unlock; 1248 } 1249 1250 pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle)); 1251 if (!pdd) { 1252 pr_err("Process device data doesn't exist\n"); 1253 ret = -EINVAL; 1254 goto err_pdd; 1255 } 1256 1257 mem = kfd_process_device_translate_handle( 1258 pdd, GET_IDR_HANDLE(args->handle)); 1259 if (!mem) { 1260 ret = -EINVAL; 1261 goto err_unlock; 1262 } 1263 1264 ret = amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, 1265 (struct kgd_mem *)mem, pdd->drm_priv, &size); 1266 1267 /* If freeing the buffer failed, leave the handle in place for 1268 * clean-up during process tear-down. 1269 */ 1270 if (!ret) 1271 kfd_process_device_remove_obj_handle( 1272 pdd, GET_IDR_HANDLE(args->handle)); 1273 1274 atomic64_sub(size, &pdd->vram_usage); 1275 1276 err_unlock: 1277 err_pdd: 1278 mutex_unlock(&p->mutex); 1279 return ret; 1280 } 1281 1282 static int kfd_ioctl_map_memory_to_gpu(struct file *filep, 1283 struct kfd_process *p, void *data) 1284 { 1285 struct kfd_ioctl_map_memory_to_gpu_args *args = data; 1286 struct kfd_process_device *pdd, *peer_pdd; 1287 void *mem; 1288 struct kfd_node *dev; 1289 long err = 0; 1290 int i; 1291 uint32_t *devices_arr = NULL; 1292 1293 if (!args->n_devices) { 1294 pr_debug("Device IDs array empty\n"); 1295 return -EINVAL; 1296 } 1297 if (args->n_success > args->n_devices) { 1298 pr_debug("n_success exceeds n_devices\n"); 1299 return -EINVAL; 1300 } 1301 1302 devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr), 1303 GFP_KERNEL); 1304 if (!devices_arr) 1305 return -ENOMEM; 1306 1307 err = copy_from_user(devices_arr, 1308 (void __user *)args->device_ids_array_ptr, 1309 args->n_devices * sizeof(*devices_arr)); 1310 if (err != 0) { 1311 err = -EFAULT; 1312 goto copy_from_user_failed; 1313 } 1314 1315 mutex_lock(&p->mutex); 1316 pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle)); 1317 if (!pdd) { 1318 err = -EINVAL; 1319 goto get_process_device_data_failed; 1320 } 1321 dev = pdd->dev; 1322 1323 pdd = kfd_bind_process_to_device(dev, p); 1324 if (IS_ERR(pdd)) { 1325 err = PTR_ERR(pdd); 1326 goto bind_process_to_device_failed; 1327 } 1328 1329 mem = kfd_process_device_translate_handle(pdd, 1330 GET_IDR_HANDLE(args->handle)); 1331 if (!mem) { 1332 err = -ENOMEM; 1333 goto get_mem_obj_from_handle_failed; 1334 } 1335 1336 for (i = args->n_success; i < args->n_devices; i++) { 1337 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]); 1338 if (!peer_pdd) { 1339 pr_debug("Getting device by id failed for 0x%x\n", 1340 devices_arr[i]); 1341 err = -EINVAL; 1342 goto get_mem_obj_from_handle_failed; 1343 } 1344 1345 peer_pdd = kfd_bind_process_to_device(peer_pdd->dev, p); 1346 if (IS_ERR(peer_pdd)) { 1347 err = PTR_ERR(peer_pdd); 1348 goto get_mem_obj_from_handle_failed; 1349 } 1350 1351 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu( 1352 peer_pdd->dev->adev, (struct kgd_mem *)mem, 1353 peer_pdd->drm_priv); 1354 if (err) { 1355 struct pci_dev *pdev = peer_pdd->dev->adev->pdev; 1356 1357 dev_err(dev->adev->dev, 1358 "Failed to map peer:%04x:%02x:%02x.%d mem_domain:%d\n", 1359 pci_domain_nr(pdev->bus), 1360 pdev->bus->number, 1361 PCI_SLOT(pdev->devfn), 1362 PCI_FUNC(pdev->devfn), 1363 ((struct kgd_mem *)mem)->domain); 1364 goto map_memory_to_gpu_failed; 1365 } 1366 args->n_success = i+1; 1367 } 1368 1369 err = amdgpu_amdkfd_gpuvm_sync_memory(dev->adev, (struct kgd_mem *) mem, true); 1370 if (err) { 1371 pr_debug("Sync memory failed, wait interrupted by user signal\n"); 1372 goto sync_memory_failed; 1373 } 1374 1375 mutex_unlock(&p->mutex); 1376 1377 /* Flush TLBs after waiting for the page table updates to complete */ 1378 for (i = 0; i < args->n_devices; i++) { 1379 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]); 1380 if (WARN_ON_ONCE(!peer_pdd)) 1381 continue; 1382 kfd_flush_tlb(peer_pdd); 1383 } 1384 kfree(devices_arr); 1385 1386 return err; 1387 1388 get_process_device_data_failed: 1389 bind_process_to_device_failed: 1390 get_mem_obj_from_handle_failed: 1391 map_memory_to_gpu_failed: 1392 sync_memory_failed: 1393 mutex_unlock(&p->mutex); 1394 copy_from_user_failed: 1395 kfree(devices_arr); 1396 1397 return err; 1398 } 1399 1400 static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, 1401 struct kfd_process *p, void *data) 1402 { 1403 struct kfd_ioctl_unmap_memory_from_gpu_args *args = data; 1404 struct kfd_process_device *pdd, *peer_pdd; 1405 void *mem; 1406 long err = 0; 1407 uint32_t *devices_arr = NULL, i; 1408 bool flush_tlb; 1409 1410 if (!args->n_devices) { 1411 pr_debug("Device IDs array empty\n"); 1412 return -EINVAL; 1413 } 1414 if (args->n_success > args->n_devices) { 1415 pr_debug("n_success exceeds n_devices\n"); 1416 return -EINVAL; 1417 } 1418 1419 devices_arr = kmalloc_array(args->n_devices, sizeof(*devices_arr), 1420 GFP_KERNEL); 1421 if (!devices_arr) 1422 return -ENOMEM; 1423 1424 err = copy_from_user(devices_arr, 1425 (void __user *)args->device_ids_array_ptr, 1426 args->n_devices * sizeof(*devices_arr)); 1427 if (err != 0) { 1428 err = -EFAULT; 1429 goto copy_from_user_failed; 1430 } 1431 1432 mutex_lock(&p->mutex); 1433 pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(args->handle)); 1434 if (!pdd) { 1435 err = -EINVAL; 1436 goto bind_process_to_device_failed; 1437 } 1438 1439 mem = kfd_process_device_translate_handle(pdd, 1440 GET_IDR_HANDLE(args->handle)); 1441 if (!mem) { 1442 err = -ENOMEM; 1443 goto get_mem_obj_from_handle_failed; 1444 } 1445 1446 for (i = args->n_success; i < args->n_devices; i++) { 1447 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]); 1448 if (!peer_pdd) { 1449 err = -EINVAL; 1450 goto get_mem_obj_from_handle_failed; 1451 } 1452 err = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( 1453 peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv); 1454 if (err) { 1455 pr_debug("Failed to unmap from gpu %d/%d\n", i, args->n_devices); 1456 goto unmap_memory_from_gpu_failed; 1457 } 1458 args->n_success = i+1; 1459 } 1460 1461 flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd); 1462 if (flush_tlb) { 1463 err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev, 1464 (struct kgd_mem *) mem, true); 1465 if (err) { 1466 pr_debug("Sync memory failed, wait interrupted by user signal\n"); 1467 goto sync_memory_failed; 1468 } 1469 } 1470 1471 /* Flush TLBs after waiting for the page table updates to complete */ 1472 for (i = 0; i < args->n_devices; i++) { 1473 peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]); 1474 if (WARN_ON_ONCE(!peer_pdd)) 1475 continue; 1476 if (flush_tlb) 1477 kfd_flush_tlb(peer_pdd); 1478 1479 /* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */ 1480 err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd->drm_priv); 1481 if (err) 1482 goto sync_memory_failed; 1483 } 1484 1485 mutex_unlock(&p->mutex); 1486 1487 kfree(devices_arr); 1488 1489 return 0; 1490 1491 bind_process_to_device_failed: 1492 get_mem_obj_from_handle_failed: 1493 unmap_memory_from_gpu_failed: 1494 sync_memory_failed: 1495 mutex_unlock(&p->mutex); 1496 copy_from_user_failed: 1497 kfree(devices_arr); 1498 return err; 1499 } 1500 1501 static int kfd_ioctl_alloc_queue_gws(struct file *filep, 1502 struct kfd_process *p, void *data) 1503 { 1504 int retval; 1505 struct kfd_ioctl_alloc_queue_gws_args *args = data; 1506 struct queue *q; 1507 struct kfd_node *dev; 1508 1509 mutex_lock(&p->mutex); 1510 q = pqm_get_user_queue(&p->pqm, args->queue_id); 1511 1512 if (q) { 1513 dev = q->device; 1514 } else { 1515 retval = -EINVAL; 1516 goto out_unlock; 1517 } 1518 1519 if (!dev->gws) { 1520 retval = -ENODEV; 1521 goto out_unlock; 1522 } 1523 1524 if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { 1525 retval = -ENODEV; 1526 goto out_unlock; 1527 } 1528 1529 if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) || 1530 kfd_dbg_has_cwsr_workaround(dev))) { 1531 retval = -EBUSY; 1532 goto out_unlock; 1533 } 1534 1535 retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL); 1536 mutex_unlock(&p->mutex); 1537 1538 args->first_gws = 0; 1539 return retval; 1540 1541 out_unlock: 1542 mutex_unlock(&p->mutex); 1543 return retval; 1544 } 1545 1546 static int kfd_ioctl_get_dmabuf_info(struct file *filep, 1547 struct kfd_process *p, void *data) 1548 { 1549 struct kfd_ioctl_get_dmabuf_info_args *args = data; 1550 struct kfd_node *dev = NULL; 1551 struct amdgpu_device *dmabuf_adev; 1552 void *metadata_buffer = NULL; 1553 uint32_t flags; 1554 int8_t xcp_id; 1555 unsigned int i; 1556 int r; 1557 1558 /* Find a KFD GPU device that supports the get_dmabuf_info query */ 1559 for (i = 0; kfd_topology_enum_kfd_devices(i, &dev) == 0; i++) 1560 if (dev && !kfd_devcgroup_check_permission(dev)) 1561 break; 1562 if (!dev) 1563 return -EINVAL; 1564 1565 if (args->metadata_ptr) { 1566 metadata_buffer = kzalloc(args->metadata_size, GFP_KERNEL); 1567 if (!metadata_buffer) 1568 return -ENOMEM; 1569 } 1570 1571 /* Get dmabuf info from KGD */ 1572 r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd, 1573 &dmabuf_adev, &args->size, 1574 metadata_buffer, args->metadata_size, 1575 &args->metadata_size, &flags, &xcp_id); 1576 if (r) 1577 goto exit; 1578 1579 if (xcp_id >= 0) 1580 args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id; 1581 else 1582 args->gpu_id = dev->id; 1583 args->flags = flags; 1584 1585 /* Copy metadata buffer to user mode */ 1586 if (metadata_buffer) { 1587 r = copy_to_user((void __user *)args->metadata_ptr, 1588 metadata_buffer, args->metadata_size); 1589 if (r != 0) 1590 r = -EFAULT; 1591 } 1592 1593 exit: 1594 kfree(metadata_buffer); 1595 1596 return r; 1597 } 1598 1599 static int kfd_ioctl_import_dmabuf(struct file *filep, 1600 struct kfd_process *p, void *data) 1601 { 1602 struct kfd_ioctl_import_dmabuf_args *args = data; 1603 struct kfd_process_device *pdd; 1604 int idr_handle; 1605 uint64_t size; 1606 void *mem; 1607 int r; 1608 1609 mutex_lock(&p->mutex); 1610 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 1611 if (!pdd) { 1612 r = -EINVAL; 1613 goto err_unlock; 1614 } 1615 1616 pdd = kfd_bind_process_to_device(pdd->dev, p); 1617 if (IS_ERR(pdd)) { 1618 r = PTR_ERR(pdd); 1619 goto err_unlock; 1620 } 1621 1622 r = amdgpu_amdkfd_gpuvm_import_dmabuf_fd(pdd->dev->adev, args->dmabuf_fd, 1623 args->va_addr, pdd->drm_priv, 1624 (struct kgd_mem **)&mem, &size, 1625 NULL); 1626 if (r) 1627 goto err_unlock; 1628 1629 idr_handle = kfd_process_device_create_obj_handle(pdd, mem); 1630 if (idr_handle < 0) { 1631 r = -EFAULT; 1632 goto err_free; 1633 } 1634 1635 mutex_unlock(&p->mutex); 1636 1637 args->handle = MAKE_HANDLE(args->gpu_id, idr_handle); 1638 1639 return 0; 1640 1641 err_free: 1642 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, (struct kgd_mem *)mem, 1643 pdd->drm_priv, NULL); 1644 err_unlock: 1645 mutex_unlock(&p->mutex); 1646 return r; 1647 } 1648 1649 static int kfd_ioctl_export_dmabuf(struct file *filep, 1650 struct kfd_process *p, void *data) 1651 { 1652 struct kfd_ioctl_export_dmabuf_args *args = data; 1653 struct kfd_process_device *pdd; 1654 struct dma_buf *dmabuf; 1655 struct kfd_node *dev; 1656 void *mem; 1657 int ret = 0; 1658 1659 dev = kfd_device_by_id(GET_GPU_ID(args->handle)); 1660 if (!dev) 1661 return -EINVAL; 1662 1663 mutex_lock(&p->mutex); 1664 1665 pdd = kfd_get_process_device_data(dev, p); 1666 if (!pdd) { 1667 ret = -EINVAL; 1668 goto err_unlock; 1669 } 1670 1671 mem = kfd_process_device_translate_handle(pdd, 1672 GET_IDR_HANDLE(args->handle)); 1673 if (!mem) { 1674 ret = -EINVAL; 1675 goto err_unlock; 1676 } 1677 1678 ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf); 1679 mutex_unlock(&p->mutex); 1680 if (ret) 1681 goto err_out; 1682 1683 ret = dma_buf_fd(dmabuf, args->flags); 1684 if (ret < 0) { 1685 dma_buf_put(dmabuf); 1686 goto err_out; 1687 } 1688 /* dma_buf_fd assigns the reference count to the fd, no need to 1689 * put the reference here. 1690 */ 1691 args->dmabuf_fd = ret; 1692 1693 return 0; 1694 1695 err_unlock: 1696 mutex_unlock(&p->mutex); 1697 err_out: 1698 return ret; 1699 } 1700 1701 /* Handle requests for watching SMI events */ 1702 static int kfd_ioctl_smi_events(struct file *filep, 1703 struct kfd_process *p, void *data) 1704 { 1705 struct kfd_ioctl_smi_events_args *args = data; 1706 struct kfd_process_device *pdd; 1707 1708 mutex_lock(&p->mutex); 1709 1710 pdd = kfd_process_device_data_by_id(p, args->gpuid); 1711 mutex_unlock(&p->mutex); 1712 if (!pdd) 1713 return -EINVAL; 1714 1715 return kfd_smi_event_open(pdd->dev, &args->anon_fd); 1716 } 1717 1718 static int kfd_ioctl_svm_validate(void *kdata, unsigned int usize) 1719 { 1720 struct kfd_ioctl_svm_args *args = kdata; 1721 size_t expected = struct_size(args, attrs, args->nattr); 1722 1723 if (expected == SIZE_MAX || usize < expected) 1724 return -EINVAL; 1725 return 0; 1726 } 1727 1728 #if IS_ENABLED(CONFIG_HSA_AMD_SVM) 1729 1730 static int kfd_ioctl_set_xnack_mode(struct file *filep, 1731 struct kfd_process *p, void *data) 1732 { 1733 struct kfd_ioctl_set_xnack_mode_args *args = data; 1734 int r = 0; 1735 1736 mutex_lock(&p->mutex); 1737 if (args->xnack_enabled >= 0) { 1738 if (!list_empty(&p->pqm.queues)) { 1739 pr_debug("Process has user queues running\n"); 1740 r = -EBUSY; 1741 goto out_unlock; 1742 } 1743 1744 if (p->xnack_enabled == args->xnack_enabled) 1745 goto out_unlock; 1746 1747 if (args->xnack_enabled && !kfd_process_xnack_mode(p, true)) { 1748 r = -EPERM; 1749 goto out_unlock; 1750 } 1751 1752 r = svm_range_switch_xnack_reserve_mem(p, args->xnack_enabled); 1753 } else { 1754 args->xnack_enabled = p->xnack_enabled; 1755 } 1756 1757 out_unlock: 1758 mutex_unlock(&p->mutex); 1759 1760 return r; 1761 } 1762 1763 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data) 1764 { 1765 struct kfd_ioctl_svm_args *args = data; 1766 int r = 0; 1767 1768 if (p->context_id != KFD_CONTEXT_ID_PRIMARY) { 1769 pr_debug("SVM ioctl not supported on non-primary kfd process\n"); 1770 1771 return -EOPNOTSUPP; 1772 } 1773 1774 pr_debug("start 0x%llx size 0x%llx op 0x%x nattr 0x%x\n", 1775 args->start_addr, args->size, args->op, args->nattr); 1776 1777 if ((args->start_addr & ~PAGE_MASK) || (args->size & ~PAGE_MASK)) 1778 return -EINVAL; 1779 if (!args->start_addr || !args->size) 1780 return -EINVAL; 1781 1782 r = svm_ioctl(p, args->op, args->start_addr, args->size, args->nattr, 1783 args->attrs); 1784 1785 return r; 1786 } 1787 #else 1788 static int kfd_ioctl_set_xnack_mode(struct file *filep, 1789 struct kfd_process *p, void *data) 1790 { 1791 return -EPERM; 1792 } 1793 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data) 1794 { 1795 return -EPERM; 1796 } 1797 #endif 1798 1799 static int kfd_ptl_control(struct kfd_process_device *pdd, bool enable) 1800 { 1801 struct amdgpu_device *adev = pdd->dev->adev; 1802 struct amdgpu_ptl *ptl = &adev->psp.ptl; 1803 enum amdgpu_ptl_fmt pref_format1 = ptl->fmt1; 1804 enum amdgpu_ptl_fmt pref_format2 = ptl->fmt2; 1805 uint32_t ptl_state = enable ? 1 : 0; 1806 int ret; 1807 1808 if (!ptl->hw_supported) 1809 return -EOPNOTSUPP; 1810 1811 if (!pdd->dev->kfd2kgd || !pdd->dev->kfd2kgd->ptl_ctrl) 1812 return -EOPNOTSUPP; 1813 1814 ret = pdd->dev->kfd2kgd->ptl_ctrl(adev, PSP_PTL_PERF_MON_SET, 1815 &ptl_state, 1816 &pref_format1, 1817 &pref_format2); 1818 1819 return ret; 1820 } 1821 1822 int kfd_ptl_disable_request(struct kfd_process_device *pdd, 1823 struct kfd_process *p) 1824 { 1825 struct amdgpu_device *adev = pdd->dev->adev; 1826 struct amdgpu_ptl *ptl = &adev->psp.ptl; 1827 int ret = 0; 1828 1829 mutex_lock(&ptl->mutex); 1830 1831 if (pdd->ptl_disable_req) 1832 goto out; 1833 1834 if (atomic_inc_return(&ptl->disable_ref) == 1) { 1835 ret = kfd_ptl_control(pdd, false); 1836 if (ret) { 1837 atomic_dec(&ptl->disable_ref); 1838 dev_warn(pdd->dev->adev->dev, 1839 "failed to disable PTL\n"); 1840 goto out; 1841 } 1842 } 1843 set_bit(AMDGPU_PTL_DISABLE_PROFILER, ptl->disable_bitmap); 1844 pdd->ptl_disable_req = true; 1845 1846 out: 1847 mutex_unlock(&ptl->mutex); 1848 return ret; 1849 } 1850 1851 int kfd_ptl_disable_release(struct kfd_process_device *pdd, 1852 struct kfd_process *p) 1853 { 1854 struct amdgpu_device *adev = pdd->dev->adev; 1855 struct amdgpu_ptl *ptl = &adev->psp.ptl; 1856 int ret = 0; 1857 1858 mutex_lock(&ptl->mutex); 1859 1860 if (!pdd->ptl_disable_req) 1861 goto out; 1862 1863 if (atomic_dec_return(&ptl->disable_ref) == 0) { 1864 clear_bit(AMDGPU_PTL_DISABLE_PROFILER, ptl->disable_bitmap); 1865 ret = kfd_ptl_control(pdd, true); 1866 if (ret) { 1867 atomic_inc(&ptl->disable_ref); 1868 set_bit(AMDGPU_PTL_DISABLE_PROFILER, ptl->disable_bitmap); 1869 dev_warn(adev->dev, "Failed to enable PTL on release: %d\n", ret); 1870 goto out; 1871 } 1872 } 1873 pdd->ptl_disable_req = false; 1874 1875 out: 1876 mutex_unlock(&ptl->mutex); 1877 return ret; 1878 } 1879 1880 static int kfd_profiler_ptl_control(struct kfd_process *p, 1881 struct kfd_ioctl_ptl_control *args) 1882 { 1883 struct kfd_process_device *pdd; 1884 int ret; 1885 1886 mutex_lock(&p->mutex); 1887 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 1888 mutex_unlock(&p->mutex); 1889 1890 if (!pdd || !pdd->dev || !pdd->dev->kfd) 1891 return -EINVAL; 1892 1893 if (args->enable == 0) 1894 ret = kfd_ptl_disable_request(pdd, p); 1895 else 1896 ret = kfd_ptl_disable_release(pdd, p); 1897 1898 return ret; 1899 } 1900 1901 static int criu_checkpoint_process(struct kfd_process *p, 1902 uint8_t __user *user_priv_data, 1903 uint64_t *priv_offset) 1904 { 1905 struct kfd_criu_process_priv_data process_priv; 1906 int ret; 1907 1908 memset(&process_priv, 0, sizeof(process_priv)); 1909 1910 process_priv.version = KFD_CRIU_PRIV_VERSION; 1911 /* For CR, we don't consider negative xnack mode which is used for 1912 * querying without changing it, here 0 simply means disabled and 1 1913 * means enabled so retry for finding a valid PTE. 1914 */ 1915 process_priv.xnack_mode = p->xnack_enabled ? 1 : 0; 1916 1917 ret = copy_to_user(user_priv_data + *priv_offset, 1918 &process_priv, sizeof(process_priv)); 1919 1920 if (ret) { 1921 pr_err("Failed to copy process information to user\n"); 1922 ret = -EFAULT; 1923 } 1924 1925 *priv_offset += sizeof(process_priv); 1926 return ret; 1927 } 1928 1929 static int criu_checkpoint_devices(struct kfd_process *p, 1930 uint32_t num_devices, 1931 uint8_t __user *user_addr, 1932 uint8_t __user *user_priv_data, 1933 uint64_t *priv_offset) 1934 { 1935 struct kfd_criu_device_priv_data *device_priv = NULL; 1936 struct kfd_criu_device_bucket *device_buckets = NULL; 1937 int ret = 0, i; 1938 1939 device_buckets = kvzalloc(num_devices * sizeof(*device_buckets), GFP_KERNEL); 1940 if (!device_buckets) { 1941 ret = -ENOMEM; 1942 goto exit; 1943 } 1944 1945 device_priv = kvzalloc(num_devices * sizeof(*device_priv), GFP_KERNEL); 1946 if (!device_priv) { 1947 ret = -ENOMEM; 1948 goto exit; 1949 } 1950 1951 for (i = 0; i < num_devices; i++) { 1952 struct kfd_process_device *pdd = p->pdds[i]; 1953 1954 device_buckets[i].user_gpu_id = pdd->user_gpu_id; 1955 device_buckets[i].actual_gpu_id = pdd->dev->id; 1956 1957 /* 1958 * priv_data does not contain useful information for now and is reserved for 1959 * future use, so we do not set its contents. 1960 */ 1961 } 1962 1963 ret = copy_to_user(user_addr, device_buckets, num_devices * sizeof(*device_buckets)); 1964 if (ret) { 1965 pr_err("Failed to copy device information to user\n"); 1966 ret = -EFAULT; 1967 goto exit; 1968 } 1969 1970 ret = copy_to_user(user_priv_data + *priv_offset, 1971 device_priv, 1972 num_devices * sizeof(*device_priv)); 1973 if (ret) { 1974 pr_err("Failed to copy device information to user\n"); 1975 ret = -EFAULT; 1976 } 1977 *priv_offset += num_devices * sizeof(*device_priv); 1978 1979 exit: 1980 kvfree(device_buckets); 1981 kvfree(device_priv); 1982 return ret; 1983 } 1984 1985 static uint32_t get_process_num_bos(struct kfd_process *p) 1986 { 1987 uint32_t num_of_bos = 0; 1988 int i; 1989 1990 /* Run over all PDDs of the process */ 1991 for (i = 0; i < p->n_pdds; i++) { 1992 struct kfd_process_device *pdd = p->pdds[i]; 1993 void *mem; 1994 int id; 1995 1996 idr_for_each_entry(&pdd->alloc_idr, mem, id) { 1997 struct kgd_mem *kgd_mem = (struct kgd_mem *)mem; 1998 1999 if (!kgd_mem->va || kgd_mem->va > pdd->gpuvm_base) 2000 num_of_bos++; 2001 } 2002 } 2003 return num_of_bos; 2004 } 2005 2006 static int criu_get_prime_handle(struct kgd_mem *mem, 2007 int flags, u32 *shared_fd, 2008 struct file **file) 2009 { 2010 struct dma_buf *dmabuf; 2011 int ret; 2012 2013 ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf); 2014 if (ret) { 2015 pr_err("dmabuf export failed for the BO\n"); 2016 return ret; 2017 } 2018 2019 ret = get_unused_fd_flags(flags); 2020 if (ret < 0) { 2021 pr_err("dmabuf create fd failed, ret:%d\n", ret); 2022 goto out_free_dmabuf; 2023 } 2024 2025 *shared_fd = ret; 2026 *file = dmabuf->file; 2027 return 0; 2028 2029 out_free_dmabuf: 2030 dma_buf_put(dmabuf); 2031 return ret; 2032 } 2033 2034 static void commit_files(struct file **files, 2035 struct kfd_criu_bo_bucket *bo_buckets, 2036 unsigned int count, 2037 int err) 2038 { 2039 while (count--) { 2040 struct file *file = files[count]; 2041 2042 if (!file) 2043 continue; 2044 if (err) { 2045 fput(file); 2046 put_unused_fd(bo_buckets[count].dmabuf_fd); 2047 } else { 2048 fd_install(bo_buckets[count].dmabuf_fd, file); 2049 } 2050 } 2051 } 2052 2053 static int criu_checkpoint_bos(struct kfd_process *p, 2054 uint32_t num_bos, 2055 uint8_t __user *user_bos, 2056 uint8_t __user *user_priv_data, 2057 uint64_t *priv_offset) 2058 { 2059 struct kfd_criu_bo_bucket *bo_buckets; 2060 struct kfd_criu_bo_priv_data *bo_privs; 2061 struct file **files = NULL; 2062 int ret = 0, pdd_index, bo_index = 0, id; 2063 void *mem; 2064 2065 bo_buckets = kvzalloc(num_bos * sizeof(*bo_buckets), GFP_KERNEL); 2066 if (!bo_buckets) 2067 return -ENOMEM; 2068 2069 bo_privs = kvzalloc(num_bos * sizeof(*bo_privs), GFP_KERNEL); 2070 if (!bo_privs) { 2071 ret = -ENOMEM; 2072 goto exit; 2073 } 2074 2075 files = kvzalloc(num_bos * sizeof(struct file *), GFP_KERNEL); 2076 if (!files) { 2077 ret = -ENOMEM; 2078 goto exit; 2079 } 2080 2081 for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) { 2082 struct kfd_process_device *pdd = p->pdds[pdd_index]; 2083 struct amdgpu_bo *dumper_bo; 2084 struct kgd_mem *kgd_mem; 2085 2086 idr_for_each_entry(&pdd->alloc_idr, mem, id) { 2087 struct kfd_criu_bo_bucket *bo_bucket; 2088 struct kfd_criu_bo_priv_data *bo_priv; 2089 int i, dev_idx = 0; 2090 2091 kgd_mem = (struct kgd_mem *)mem; 2092 dumper_bo = kgd_mem->bo; 2093 2094 /* Skip checkpointing BOs that are used for Trap handler 2095 * code and state. Currently, these BOs have a VA that 2096 * is less GPUVM Base 2097 */ 2098 if (kgd_mem->va && kgd_mem->va <= pdd->gpuvm_base) 2099 continue; 2100 2101 bo_bucket = &bo_buckets[bo_index]; 2102 bo_priv = &bo_privs[bo_index]; 2103 2104 bo_bucket->gpu_id = pdd->user_gpu_id; 2105 bo_bucket->addr = (uint64_t)kgd_mem->va; 2106 bo_bucket->size = amdgpu_bo_size(dumper_bo); 2107 bo_bucket->alloc_flags = (uint32_t)kgd_mem->alloc_flags; 2108 bo_priv->idr_handle = id; 2109 2110 if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { 2111 ret = amdgpu_ttm_tt_get_userptr(&dumper_bo->tbo, 2112 &bo_priv->user_addr); 2113 if (ret) { 2114 pr_err("Failed to obtain user address for user-pointer bo\n"); 2115 goto exit; 2116 } 2117 } 2118 if (bo_bucket->alloc_flags 2119 & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) { 2120 ret = criu_get_prime_handle(kgd_mem, 2121 bo_bucket->alloc_flags & 2122 KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DRM_RDWR : 0, 2123 &bo_bucket->dmabuf_fd, &files[bo_index]); 2124 if (ret) 2125 goto exit; 2126 } else { 2127 bo_bucket->dmabuf_fd = KFD_INVALID_FD; 2128 } 2129 2130 if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) 2131 bo_bucket->offset = KFD_MMAP_TYPE_DOORBELL | 2132 KFD_MMAP_GPU_ID(pdd->dev->id); 2133 else if (bo_bucket->alloc_flags & 2134 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) 2135 bo_bucket->offset = KFD_MMAP_TYPE_MMIO | 2136 KFD_MMAP_GPU_ID(pdd->dev->id); 2137 else 2138 bo_bucket->offset = amdgpu_bo_mmap_offset(dumper_bo); 2139 2140 for (i = 0; i < p->n_pdds; i++) { 2141 if (amdgpu_amdkfd_bo_mapped_to_dev(p->pdds[i]->drm_priv, kgd_mem)) 2142 bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id; 2143 } 2144 2145 pr_debug("bo_size = 0x%llx, bo_addr = 0x%llx bo_offset = 0x%llx\n" 2146 "gpu_id = 0x%x alloc_flags = 0x%x idr_handle = 0x%x", 2147 bo_bucket->size, 2148 bo_bucket->addr, 2149 bo_bucket->offset, 2150 bo_bucket->gpu_id, 2151 bo_bucket->alloc_flags, 2152 bo_priv->idr_handle); 2153 bo_index++; 2154 } 2155 } 2156 2157 ret = copy_to_user(user_bos, bo_buckets, num_bos * sizeof(*bo_buckets)); 2158 if (ret) { 2159 pr_err("Failed to copy BO information to user\n"); 2160 ret = -EFAULT; 2161 goto exit; 2162 } 2163 2164 ret = copy_to_user(user_priv_data + *priv_offset, bo_privs, num_bos * sizeof(*bo_privs)); 2165 if (ret) { 2166 pr_err("Failed to copy BO priv information to user\n"); 2167 ret = -EFAULT; 2168 goto exit; 2169 } 2170 2171 *priv_offset += num_bos * sizeof(*bo_privs); 2172 2173 exit: 2174 commit_files(files, bo_buckets, bo_index, ret); 2175 kvfree(files); 2176 kvfree(bo_buckets); 2177 kvfree(bo_privs); 2178 return ret; 2179 } 2180 2181 static int criu_get_process_object_info(struct kfd_process *p, 2182 uint32_t *num_devices, 2183 uint32_t *num_bos, 2184 uint32_t *num_objects, 2185 uint64_t *objs_priv_size) 2186 { 2187 uint64_t queues_priv_data_size, svm_priv_data_size, priv_size; 2188 uint32_t num_queues, num_events, num_svm_ranges; 2189 int ret; 2190 2191 *num_devices = p->n_pdds; 2192 *num_bos = get_process_num_bos(p); 2193 2194 ret = kfd_process_get_queue_info(p, &num_queues, &queues_priv_data_size); 2195 if (ret) 2196 return ret; 2197 2198 num_events = kfd_get_num_events(p); 2199 2200 svm_range_get_info(p, &num_svm_ranges, &svm_priv_data_size); 2201 2202 *num_objects = num_queues + num_events + num_svm_ranges; 2203 2204 if (objs_priv_size) { 2205 priv_size = sizeof(struct kfd_criu_process_priv_data); 2206 priv_size += *num_devices * sizeof(struct kfd_criu_device_priv_data); 2207 priv_size += *num_bos * sizeof(struct kfd_criu_bo_priv_data); 2208 priv_size += queues_priv_data_size; 2209 priv_size += num_events * sizeof(struct kfd_criu_event_priv_data); 2210 priv_size += svm_priv_data_size; 2211 *objs_priv_size = priv_size; 2212 } 2213 return 0; 2214 } 2215 2216 static int criu_checkpoint(struct file *filep, 2217 struct kfd_process *p, 2218 struct kfd_ioctl_criu_args *args) 2219 { 2220 int ret; 2221 uint32_t num_devices, num_bos, num_objects; 2222 uint64_t priv_size, priv_offset = 0, bo_priv_offset; 2223 2224 if (!args->devices || !args->bos || !args->priv_data) 2225 return -EINVAL; 2226 2227 mutex_lock(&p->mutex); 2228 2229 if (!p->n_pdds) { 2230 pr_err("No pdd for given process\n"); 2231 ret = -ENODEV; 2232 goto exit_unlock; 2233 } 2234 2235 /* Confirm all process queues are evicted */ 2236 if (!p->queues_paused) { 2237 pr_err("Cannot dump process when queues are not in evicted state\n"); 2238 /* CRIU plugin did not call op PROCESS_INFO before checkpointing */ 2239 ret = -EINVAL; 2240 goto exit_unlock; 2241 } 2242 2243 ret = criu_get_process_object_info(p, &num_devices, &num_bos, &num_objects, &priv_size); 2244 if (ret) 2245 goto exit_unlock; 2246 2247 if (num_devices != args->num_devices || 2248 num_bos != args->num_bos || 2249 num_objects != args->num_objects || 2250 priv_size != args->priv_data_size) { 2251 2252 ret = -EINVAL; 2253 goto exit_unlock; 2254 } 2255 2256 /* each function will store private data inside priv_data and adjust priv_offset */ 2257 ret = criu_checkpoint_process(p, (uint8_t __user *)args->priv_data, &priv_offset); 2258 if (ret) 2259 goto exit_unlock; 2260 2261 ret = criu_checkpoint_devices(p, num_devices, (uint8_t __user *)args->devices, 2262 (uint8_t __user *)args->priv_data, &priv_offset); 2263 if (ret) 2264 goto exit_unlock; 2265 2266 /* Leave room for BOs in the private data. They need to be restored 2267 * before events, but we checkpoint them last to simplify the error 2268 * handling. 2269 */ 2270 bo_priv_offset = priv_offset; 2271 priv_offset += num_bos * sizeof(struct kfd_criu_bo_priv_data); 2272 2273 if (num_objects) { 2274 ret = kfd_criu_checkpoint_queues(p, (uint8_t __user *)args->priv_data, 2275 &priv_offset); 2276 if (ret) 2277 goto exit_unlock; 2278 2279 ret = kfd_criu_checkpoint_events(p, (uint8_t __user *)args->priv_data, 2280 &priv_offset); 2281 if (ret) 2282 goto exit_unlock; 2283 2284 ret = kfd_criu_checkpoint_svm(p, (uint8_t __user *)args->priv_data, &priv_offset); 2285 if (ret) 2286 goto exit_unlock; 2287 } 2288 2289 /* This must be the last thing in this function that can fail. 2290 * Otherwise we leak dmabuf file descriptors. 2291 */ 2292 ret = criu_checkpoint_bos(p, num_bos, (uint8_t __user *)args->bos, 2293 (uint8_t __user *)args->priv_data, &bo_priv_offset); 2294 2295 exit_unlock: 2296 mutex_unlock(&p->mutex); 2297 if (ret) 2298 pr_err("Failed to dump CRIU ret:%d\n", ret); 2299 else 2300 pr_debug("CRIU dump ret:%d\n", ret); 2301 2302 return ret; 2303 } 2304 2305 static int criu_restore_process(struct kfd_process *p, 2306 struct kfd_ioctl_criu_args *args, 2307 uint64_t *priv_offset, 2308 uint64_t max_priv_data_size) 2309 { 2310 int ret = 0; 2311 struct kfd_criu_process_priv_data process_priv; 2312 2313 if (*priv_offset + sizeof(process_priv) > max_priv_data_size) 2314 return -EINVAL; 2315 2316 ret = copy_from_user(&process_priv, 2317 (void __user *)(args->priv_data + *priv_offset), 2318 sizeof(process_priv)); 2319 if (ret) { 2320 pr_err("Failed to copy process private information from user\n"); 2321 ret = -EFAULT; 2322 goto exit; 2323 } 2324 *priv_offset += sizeof(process_priv); 2325 2326 if (process_priv.version != KFD_CRIU_PRIV_VERSION) { 2327 pr_err("Invalid CRIU API version (checkpointed:%d current:%d)\n", 2328 process_priv.version, KFD_CRIU_PRIV_VERSION); 2329 return -EINVAL; 2330 } 2331 2332 pr_debug("Setting XNACK mode\n"); 2333 if (process_priv.xnack_mode && !kfd_process_xnack_mode(p, true)) { 2334 pr_err("xnack mode cannot be set\n"); 2335 ret = -EPERM; 2336 goto exit; 2337 } else { 2338 pr_debug("set xnack mode: %d\n", process_priv.xnack_mode); 2339 p->xnack_enabled = process_priv.xnack_mode; 2340 } 2341 2342 exit: 2343 return ret; 2344 } 2345 2346 static int criu_restore_devices(struct kfd_process *p, 2347 struct kfd_ioctl_criu_args *args, 2348 uint64_t *priv_offset, 2349 uint64_t max_priv_data_size) 2350 { 2351 struct kfd_criu_device_bucket *device_buckets; 2352 struct kfd_criu_device_priv_data *device_privs; 2353 int ret = 0; 2354 uint32_t i; 2355 2356 if (args->num_devices != p->n_pdds) 2357 return -EINVAL; 2358 2359 if (*priv_offset + (args->num_devices * sizeof(*device_privs)) > max_priv_data_size) 2360 return -EINVAL; 2361 2362 device_buckets = kmalloc_objs(*device_buckets, args->num_devices); 2363 if (!device_buckets) 2364 return -ENOMEM; 2365 2366 ret = copy_from_user(device_buckets, (void __user *)args->devices, 2367 args->num_devices * sizeof(*device_buckets)); 2368 if (ret) { 2369 pr_err("Failed to copy devices buckets from user\n"); 2370 ret = -EFAULT; 2371 goto exit; 2372 } 2373 2374 for (i = 0; i < args->num_devices; i++) { 2375 struct kfd_node *dev; 2376 struct kfd_process_device *pdd; 2377 struct file *drm_file; 2378 2379 /* device private data is not currently used */ 2380 2381 if (!device_buckets[i].user_gpu_id) { 2382 pr_err("Invalid user gpu_id\n"); 2383 ret = -EINVAL; 2384 goto exit; 2385 } 2386 2387 dev = kfd_device_by_id(device_buckets[i].actual_gpu_id); 2388 if (!dev) { 2389 pr_err("Failed to find device with gpu_id = %x\n", 2390 device_buckets[i].actual_gpu_id); 2391 ret = -EINVAL; 2392 goto exit; 2393 } 2394 2395 pdd = kfd_get_process_device_data(dev, p); 2396 if (!pdd) { 2397 pr_err("Failed to get pdd for gpu_id = %x\n", 2398 device_buckets[i].actual_gpu_id); 2399 ret = -EINVAL; 2400 goto exit; 2401 } 2402 pdd->user_gpu_id = device_buckets[i].user_gpu_id; 2403 2404 drm_file = fget(device_buckets[i].drm_fd); 2405 if (!drm_file) { 2406 pr_err("Invalid render node file descriptor sent from plugin (%d)\n", 2407 device_buckets[i].drm_fd); 2408 ret = -EINVAL; 2409 goto exit; 2410 } 2411 2412 if (pdd->drm_file) { 2413 ret = -EINVAL; 2414 goto exit; 2415 } 2416 2417 /* create the vm using render nodes for kfd pdd */ 2418 if (kfd_process_device_init_vm(pdd, drm_file)) { 2419 pr_err("could not init vm for given pdd\n"); 2420 /* On success, the PDD keeps the drm_file reference */ 2421 fput(drm_file); 2422 ret = -EINVAL; 2423 goto exit; 2424 } 2425 /* 2426 * pdd now already has the vm bound to render node so below api won't create a new 2427 * exclusive kfd mapping but use existing one with renderDXXX but is still needed 2428 * for iommu v2 binding and runtime pm. 2429 */ 2430 pdd = kfd_bind_process_to_device(dev, p); 2431 if (IS_ERR(pdd)) { 2432 ret = PTR_ERR(pdd); 2433 goto exit; 2434 } 2435 2436 if (!pdd->qpd.proc_doorbells) { 2437 ret = kfd_alloc_process_doorbells(dev->kfd, pdd); 2438 if (ret) 2439 goto exit; 2440 } 2441 } 2442 2443 /* 2444 * We are not copying device private data from user as we are not using the data for now, 2445 * but we still adjust for its private data. 2446 */ 2447 *priv_offset += args->num_devices * sizeof(*device_privs); 2448 2449 exit: 2450 kfree(device_buckets); 2451 return ret; 2452 } 2453 2454 static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd, 2455 struct kfd_criu_bo_bucket *bo_bucket, 2456 struct kfd_criu_bo_priv_data *bo_priv, 2457 struct kgd_mem **kgd_mem) 2458 { 2459 int idr_handle; 2460 int ret; 2461 const bool criu_resume = true; 2462 u64 offset; 2463 2464 if (bo_priv->idr_handle > INT_MAX) 2465 return -EINVAL; 2466 2467 if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) { 2468 if (bo_bucket->size != 2469 kfd_doorbell_process_slice(pdd->dev->kfd)) 2470 return -EINVAL; 2471 2472 offset = kfd_get_process_doorbells(pdd); 2473 if (!offset) 2474 return -ENOMEM; 2475 } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) { 2476 /* MMIO BOs need remapped bus address */ 2477 if (bo_bucket->size != PAGE_SIZE) { 2478 pr_err("Invalid page size\n"); 2479 return -EINVAL; 2480 } 2481 offset = pdd->dev->adev->rmmio_remap.bus_addr; 2482 if (!offset || (PAGE_SIZE > 4096)) { 2483 pr_err("amdgpu_amdkfd_get_mmio_remap_phys_addr failed\n"); 2484 return -ENOMEM; 2485 } 2486 } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { 2487 offset = bo_priv->user_addr; 2488 } 2489 /* Create the BO */ 2490 ret = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(pdd->dev->adev, bo_bucket->addr, 2491 bo_bucket->size, pdd->drm_priv, kgd_mem, 2492 &offset, bo_bucket->alloc_flags, criu_resume); 2493 if (ret) { 2494 pr_err("Could not create the BO\n"); 2495 return ret; 2496 } 2497 pr_debug("New BO created: size:0x%llx addr:0x%llx offset:0x%llx\n", 2498 bo_bucket->size, bo_bucket->addr, offset); 2499 2500 /* Restore previous IDR handle */ 2501 pr_debug("Restoring old IDR handle for the BO"); 2502 idr_handle = idr_alloc(&pdd->alloc_idr, *kgd_mem, bo_priv->idr_handle, 2503 bo_priv->idr_handle + 1, GFP_KERNEL); 2504 2505 if (idr_handle < 0) { 2506 pr_err("Could not allocate idr\n"); 2507 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, *kgd_mem, pdd->drm_priv, 2508 NULL); 2509 return -ENOMEM; 2510 } 2511 2512 if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) 2513 bo_bucket->restored_offset = KFD_MMAP_TYPE_DOORBELL | KFD_MMAP_GPU_ID(pdd->dev->id); 2514 if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) { 2515 bo_bucket->restored_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(pdd->dev->id); 2516 } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) { 2517 bo_bucket->restored_offset = offset; 2518 } else if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { 2519 bo_bucket->restored_offset = offset; 2520 /* Update the VRAM usage count */ 2521 atomic64_add(bo_bucket->size, &pdd->vram_usage); 2522 } 2523 return 0; 2524 } 2525 2526 static int criu_restore_bo(struct kfd_process *p, 2527 struct kfd_criu_bo_bucket *bo_bucket, 2528 struct kfd_criu_bo_priv_data *bo_priv, 2529 struct file **file) 2530 { 2531 struct kfd_process_device *pdd; 2532 struct kgd_mem *kgd_mem; 2533 int ret; 2534 int j; 2535 2536 pr_debug("Restoring BO size:0x%llx addr:0x%llx gpu_id:0x%x flags:0x%x idr_handle:0x%x\n", 2537 bo_bucket->size, bo_bucket->addr, bo_bucket->gpu_id, bo_bucket->alloc_flags, 2538 bo_priv->idr_handle); 2539 2540 pdd = kfd_process_device_data_by_id(p, bo_bucket->gpu_id); 2541 if (!pdd) { 2542 pr_err("Failed to get pdd\n"); 2543 return -ENODEV; 2544 } 2545 2546 ret = criu_restore_memory_of_gpu(pdd, bo_bucket, bo_priv, &kgd_mem); 2547 if (ret) 2548 return ret; 2549 2550 /* now map these BOs to GPU/s */ 2551 for (j = 0; j < p->n_pdds; j++) { 2552 struct kfd_node *peer; 2553 struct kfd_process_device *peer_pdd; 2554 2555 if (!bo_priv->mapped_gpuids[j]) 2556 break; 2557 2558 peer_pdd = kfd_process_device_data_by_id(p, bo_priv->mapped_gpuids[j]); 2559 if (!peer_pdd) 2560 return -EINVAL; 2561 2562 peer = peer_pdd->dev; 2563 2564 peer_pdd = kfd_bind_process_to_device(peer, p); 2565 if (IS_ERR(peer_pdd)) 2566 return PTR_ERR(peer_pdd); 2567 2568 ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem, 2569 peer_pdd->drm_priv); 2570 if (ret) { 2571 pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds); 2572 return ret; 2573 } 2574 } 2575 2576 pr_debug("map memory was successful for the BO\n"); 2577 /* create the dmabuf object and export the bo */ 2578 if (bo_bucket->alloc_flags 2579 & (KFD_IOC_ALLOC_MEM_FLAGS_VRAM | KFD_IOC_ALLOC_MEM_FLAGS_GTT)) { 2580 ret = criu_get_prime_handle(kgd_mem, DRM_RDWR, 2581 &bo_bucket->dmabuf_fd, file); 2582 if (ret) 2583 return ret; 2584 } else { 2585 bo_bucket->dmabuf_fd = KFD_INVALID_FD; 2586 } 2587 2588 return 0; 2589 } 2590 2591 static int criu_restore_bos(struct kfd_process *p, 2592 struct kfd_ioctl_criu_args *args, 2593 uint64_t *priv_offset, 2594 uint64_t max_priv_data_size) 2595 { 2596 struct kfd_criu_bo_bucket *bo_buckets = NULL; 2597 struct kfd_criu_bo_priv_data *bo_privs = NULL; 2598 struct file **files = NULL; 2599 int ret = 0; 2600 uint32_t i = 0; 2601 2602 if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > max_priv_data_size) 2603 return -EINVAL; 2604 2605 /* Prevent MMU notifications until stage-4 IOCTL (CRIU_RESUME) is received */ 2606 amdgpu_amdkfd_block_mmu_notifications(p->kgd_process_info); 2607 2608 bo_buckets = kvmalloc_objs(*bo_buckets, args->num_bos); 2609 if (!bo_buckets) 2610 return -ENOMEM; 2611 2612 files = kvzalloc(args->num_bos * sizeof(struct file *), GFP_KERNEL); 2613 if (!files) { 2614 ret = -ENOMEM; 2615 goto exit; 2616 } 2617 2618 ret = copy_from_user(bo_buckets, (void __user *)args->bos, 2619 args->num_bos * sizeof(*bo_buckets)); 2620 if (ret) { 2621 pr_err("Failed to copy BOs information from user\n"); 2622 ret = -EFAULT; 2623 goto exit; 2624 } 2625 2626 bo_privs = kvmalloc_objs(*bo_privs, args->num_bos); 2627 if (!bo_privs) { 2628 ret = -ENOMEM; 2629 goto exit; 2630 } 2631 2632 ret = copy_from_user(bo_privs, (void __user *)args->priv_data + *priv_offset, 2633 args->num_bos * sizeof(*bo_privs)); 2634 if (ret) { 2635 pr_err("Failed to copy BOs information from user\n"); 2636 ret = -EFAULT; 2637 goto exit; 2638 } 2639 *priv_offset += args->num_bos * sizeof(*bo_privs); 2640 2641 /* Create and map new BOs */ 2642 for (; i < args->num_bos; i++) { 2643 ret = criu_restore_bo(p, &bo_buckets[i], &bo_privs[i], &files[i]); 2644 if (ret) { 2645 pr_debug("Failed to restore BO[%d] ret%d\n", i, ret); 2646 goto exit; 2647 } 2648 } /* done */ 2649 2650 /* Copy only the buckets back so user can read bo_buckets[N].restored_offset */ 2651 ret = copy_to_user((void __user *)args->bos, 2652 bo_buckets, 2653 (args->num_bos * sizeof(*bo_buckets))); 2654 if (ret) 2655 ret = -EFAULT; 2656 2657 exit: 2658 commit_files(files, bo_buckets, i, ret); 2659 kvfree(files); 2660 kvfree(bo_buckets); 2661 kvfree(bo_privs); 2662 return ret; 2663 } 2664 2665 static int criu_restore_objects(struct file *filep, 2666 struct kfd_process *p, 2667 struct kfd_ioctl_criu_args *args, 2668 uint64_t *priv_offset, 2669 uint64_t max_priv_data_size) 2670 { 2671 int ret = 0; 2672 uint32_t i; 2673 2674 BUILD_BUG_ON(offsetof(struct kfd_criu_queue_priv_data, object_type)); 2675 BUILD_BUG_ON(offsetof(struct kfd_criu_event_priv_data, object_type)); 2676 BUILD_BUG_ON(offsetof(struct kfd_criu_svm_range_priv_data, object_type)); 2677 2678 for (i = 0; i < args->num_objects; i++) { 2679 uint32_t object_type; 2680 2681 if (*priv_offset + sizeof(object_type) > max_priv_data_size) { 2682 pr_err("Invalid private data size\n"); 2683 return -EINVAL; 2684 } 2685 2686 ret = get_user(object_type, (uint32_t __user *)(args->priv_data + *priv_offset)); 2687 if (ret) { 2688 pr_err("Failed to copy private information from user\n"); 2689 goto exit; 2690 } 2691 2692 switch (object_type) { 2693 case KFD_CRIU_OBJECT_TYPE_QUEUE: 2694 ret = kfd_criu_restore_queue(p, (uint8_t __user *)args->priv_data, 2695 priv_offset, max_priv_data_size); 2696 if (ret) 2697 goto exit; 2698 break; 2699 case KFD_CRIU_OBJECT_TYPE_EVENT: 2700 ret = kfd_criu_restore_event(filep, p, (uint8_t __user *)args->priv_data, 2701 priv_offset, max_priv_data_size); 2702 if (ret) 2703 goto exit; 2704 break; 2705 case KFD_CRIU_OBJECT_TYPE_SVM_RANGE: 2706 ret = kfd_criu_restore_svm(p, (uint8_t __user *)args->priv_data, 2707 priv_offset, max_priv_data_size); 2708 if (ret) 2709 goto exit; 2710 break; 2711 default: 2712 pr_err("Invalid object type:%u at index:%d\n", object_type, i); 2713 ret = -EINVAL; 2714 goto exit; 2715 } 2716 } 2717 exit: 2718 return ret; 2719 } 2720 2721 static int criu_restore(struct file *filep, 2722 struct kfd_process *p, 2723 struct kfd_ioctl_criu_args *args) 2724 { 2725 uint64_t priv_offset = 0; 2726 int ret = 0; 2727 2728 pr_debug("CRIU restore (num_devices:%u num_bos:%u num_objects:%u priv_data_size:%llu)\n", 2729 args->num_devices, args->num_bos, args->num_objects, args->priv_data_size); 2730 2731 if ((args->num_bos > 0 && !args->bos) || !args->devices || !args->priv_data || 2732 !args->priv_data_size || !args->num_devices) 2733 return -EINVAL; 2734 2735 mutex_lock(&p->mutex); 2736 2737 /* 2738 * Set the process to evicted state to avoid running any new queues before all the memory 2739 * mappings are ready. 2740 */ 2741 ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_RESTORE); 2742 if (ret) 2743 goto exit_unlock; 2744 2745 /* Each function will adjust priv_offset based on how many bytes they consumed */ 2746 ret = criu_restore_process(p, args, &priv_offset, args->priv_data_size); 2747 if (ret) 2748 goto exit_unlock; 2749 2750 ret = criu_restore_devices(p, args, &priv_offset, args->priv_data_size); 2751 if (ret) 2752 goto exit_unlock; 2753 2754 ret = criu_restore_bos(p, args, &priv_offset, args->priv_data_size); 2755 if (ret) 2756 goto exit_unlock; 2757 2758 ret = criu_restore_objects(filep, p, args, &priv_offset, args->priv_data_size); 2759 if (ret) 2760 goto exit_unlock; 2761 2762 if (priv_offset != args->priv_data_size) { 2763 pr_err("Invalid private data size\n"); 2764 ret = -EINVAL; 2765 } 2766 2767 exit_unlock: 2768 mutex_unlock(&p->mutex); 2769 if (ret) 2770 pr_err("Failed to restore CRIU ret:%d\n", ret); 2771 else 2772 pr_debug("CRIU restore successful\n"); 2773 2774 return ret; 2775 } 2776 2777 static int criu_unpause(struct file *filep, 2778 struct kfd_process *p, 2779 struct kfd_ioctl_criu_args *args) 2780 { 2781 int ret; 2782 2783 mutex_lock(&p->mutex); 2784 2785 if (!p->queues_paused) { 2786 mutex_unlock(&p->mutex); 2787 return -EINVAL; 2788 } 2789 2790 ret = kfd_process_restore_queues(p); 2791 if (ret) 2792 pr_err("Failed to unpause queues ret:%d\n", ret); 2793 else 2794 p->queues_paused = false; 2795 2796 mutex_unlock(&p->mutex); 2797 2798 return ret; 2799 } 2800 2801 static int criu_resume(struct file *filep, 2802 struct kfd_process *p, 2803 struct kfd_ioctl_criu_args *args) 2804 { 2805 struct kfd_process *target = NULL; 2806 struct pid *pid = NULL; 2807 int ret = 0; 2808 2809 pr_debug("Inside %s, target pid for criu restore: %d\n", __func__, 2810 args->pid); 2811 2812 pid = find_get_pid(args->pid); 2813 if (!pid) { 2814 pr_err("Cannot find pid info for %i\n", args->pid); 2815 return -ESRCH; 2816 } 2817 2818 pr_debug("calling kfd_lookup_process_by_pid\n"); 2819 target = kfd_lookup_process_by_pid(pid); 2820 2821 put_pid(pid); 2822 2823 if (!target) { 2824 pr_debug("Cannot find process info for %i\n", args->pid); 2825 return -ESRCH; 2826 } 2827 2828 mutex_lock(&target->mutex); 2829 ret = kfd_criu_resume_svm(target); 2830 if (ret) { 2831 pr_err("kfd_criu_resume_svm failed for %i\n", args->pid); 2832 goto exit; 2833 } 2834 2835 ret = amdgpu_amdkfd_criu_resume(target->kgd_process_info); 2836 if (ret) 2837 pr_err("amdgpu_amdkfd_criu_resume failed for %i\n", args->pid); 2838 2839 exit: 2840 mutex_unlock(&target->mutex); 2841 2842 kfd_unref_process(target); 2843 return ret; 2844 } 2845 2846 static int criu_process_info(struct file *filep, 2847 struct kfd_process *p, 2848 struct kfd_ioctl_criu_args *args) 2849 { 2850 int ret = 0; 2851 2852 mutex_lock(&p->mutex); 2853 2854 if (!p->n_pdds) { 2855 pr_err("No pdd for given process\n"); 2856 ret = -ENODEV; 2857 goto err_unlock; 2858 } 2859 2860 ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_CRIU_CHECKPOINT); 2861 if (ret) 2862 goto err_unlock; 2863 2864 p->queues_paused = true; 2865 2866 args->pid = task_pid_nr_ns(p->lead_thread, 2867 task_active_pid_ns(p->lead_thread)); 2868 2869 ret = criu_get_process_object_info(p, &args->num_devices, &args->num_bos, 2870 &args->num_objects, &args->priv_data_size); 2871 if (ret) 2872 goto err_unlock; 2873 2874 dev_dbg(kfd_device, "Num of devices:%u bos:%u objects:%u priv_data_size:%lld\n", 2875 args->num_devices, args->num_bos, args->num_objects, 2876 args->priv_data_size); 2877 2878 err_unlock: 2879 if (ret) { 2880 kfd_process_restore_queues(p); 2881 p->queues_paused = false; 2882 } 2883 mutex_unlock(&p->mutex); 2884 return ret; 2885 } 2886 2887 static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data) 2888 { 2889 struct kfd_ioctl_criu_args *args = data; 2890 int ret; 2891 2892 dev_dbg(kfd_device, "CRIU operation: %d\n", args->op); 2893 switch (args->op) { 2894 case KFD_CRIU_OP_PROCESS_INFO: 2895 ret = criu_process_info(filep, p, args); 2896 break; 2897 case KFD_CRIU_OP_CHECKPOINT: 2898 ret = criu_checkpoint(filep, p, args); 2899 break; 2900 case KFD_CRIU_OP_UNPAUSE: 2901 ret = criu_unpause(filep, p, args); 2902 break; 2903 case KFD_CRIU_OP_RESTORE: 2904 ret = criu_restore(filep, p, args); 2905 break; 2906 case KFD_CRIU_OP_RESUME: 2907 ret = criu_resume(filep, p, args); 2908 break; 2909 default: 2910 dev_dbg(kfd_device, "Unsupported CRIU operation:%d\n", args->op); 2911 ret = -EINVAL; 2912 break; 2913 } 2914 2915 if (ret) 2916 dev_dbg(kfd_device, "CRIU operation:%d err:%d\n", args->op, ret); 2917 2918 return ret; 2919 } 2920 2921 static int runtime_enable(struct kfd_process *p, uint64_t r_debug, 2922 bool enable_ttmp_setup) 2923 { 2924 int i = 0, ret = 0; 2925 2926 if (p->is_runtime_retry) 2927 goto retry; 2928 2929 if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED) 2930 return -EBUSY; 2931 2932 for (i = 0; i < p->n_pdds; i++) { 2933 struct kfd_process_device *pdd = p->pdds[i]; 2934 2935 if (pdd->qpd.queue_count) 2936 return -EEXIST; 2937 2938 /* 2939 * Setup TTMPs by default. 2940 * Note that this call must remain here for MES ADD QUEUE to 2941 * skip_process_ctx_clear unconditionally as the first call to 2942 * SET_SHADER_DEBUGGER clears any stale process context data 2943 * saved in MES. 2944 */ 2945 if (pdd->dev->kfd->shared_resources.enable_mes) { 2946 ret = kfd_dbg_set_mes_debug_mode( 2947 pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev)); 2948 if (ret) 2949 return ret; 2950 } 2951 } 2952 2953 p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED; 2954 p->runtime_info.r_debug = r_debug; 2955 p->runtime_info.ttmp_setup = enable_ttmp_setup; 2956 2957 if (p->runtime_info.ttmp_setup) { 2958 for (i = 0; i < p->n_pdds; i++) { 2959 struct kfd_process_device *pdd = p->pdds[i]; 2960 2961 if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) { 2962 amdgpu_gfx_off_ctrl(pdd->dev->adev, false); 2963 pdd->dev->kfd2kgd->enable_debug_trap( 2964 pdd->dev->adev, 2965 true, 2966 pdd->dev->vm_info.last_vmid_kfd); 2967 } else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) { 2968 pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap( 2969 pdd->dev->adev, 2970 false, 2971 0); 2972 } 2973 } 2974 } 2975 2976 retry: 2977 if (p->debug_trap_enabled) { 2978 if (!p->is_runtime_retry) { 2979 kfd_dbg_trap_activate(p); 2980 kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME), 2981 p, NULL, 0, false, NULL, 0); 2982 } 2983 2984 mutex_unlock(&p->mutex); 2985 ret = down_interruptible(&p->runtime_enable_sema); 2986 mutex_lock(&p->mutex); 2987 2988 p->is_runtime_retry = !!ret; 2989 } 2990 2991 return ret; 2992 } 2993 2994 static int runtime_disable(struct kfd_process *p) 2995 { 2996 int i = 0, ret = 0; 2997 bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED; 2998 2999 p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED; 3000 p->runtime_info.r_debug = 0; 3001 3002 if (p->debug_trap_enabled) { 3003 if (was_enabled) 3004 kfd_dbg_trap_deactivate(p, false, 0); 3005 3006 if (!p->is_runtime_retry) 3007 kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME), 3008 p, NULL, 0, false, NULL, 0); 3009 3010 mutex_unlock(&p->mutex); 3011 ret = down_interruptible(&p->runtime_enable_sema); 3012 mutex_lock(&p->mutex); 3013 3014 p->is_runtime_retry = !!ret; 3015 if (ret) 3016 return ret; 3017 } 3018 3019 if (was_enabled && p->runtime_info.ttmp_setup) { 3020 for (i = 0; i < p->n_pdds; i++) { 3021 struct kfd_process_device *pdd = p->pdds[i]; 3022 3023 if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) 3024 amdgpu_gfx_off_ctrl(pdd->dev->adev, true); 3025 } 3026 } 3027 3028 p->runtime_info.ttmp_setup = false; 3029 3030 /* disable ttmp setup */ 3031 for (i = 0; i < p->n_pdds; i++) { 3032 struct kfd_process_device *pdd = p->pdds[i]; 3033 int last_err = 0; 3034 3035 if (kfd_dbg_is_per_vmid_supported(pdd->dev)) { 3036 pdd->spi_dbg_override = 3037 pdd->dev->kfd2kgd->disable_debug_trap( 3038 pdd->dev->adev, 3039 false, 3040 pdd->dev->vm_info.last_vmid_kfd); 3041 3042 if (!pdd->dev->kfd->shared_resources.enable_mes) 3043 last_err = debug_refresh_runlist(pdd->dev->dqm); 3044 else 3045 last_err = kfd_dbg_set_mes_debug_mode(pdd, 3046 !kfd_dbg_has_cwsr_workaround(pdd->dev)); 3047 3048 if (last_err) 3049 ret = last_err; 3050 } 3051 } 3052 3053 return ret; 3054 } 3055 3056 static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data) 3057 { 3058 struct kfd_ioctl_runtime_enable_args *args = data; 3059 int r; 3060 3061 mutex_lock(&p->mutex); 3062 3063 if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK) 3064 r = runtime_enable(p, args->r_debug, 3065 !!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK)); 3066 else 3067 r = runtime_disable(p); 3068 3069 mutex_unlock(&p->mutex); 3070 3071 return r; 3072 } 3073 3074 static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data) 3075 { 3076 struct kfd_ioctl_dbg_trap_args *args = data; 3077 struct task_struct *thread = NULL; 3078 struct mm_struct *mm = NULL; 3079 struct pid *pid = NULL; 3080 struct kfd_process *target = NULL; 3081 struct kfd_process_device *pdd = NULL; 3082 int r = 0; 3083 3084 if (p->context_id != KFD_CONTEXT_ID_PRIMARY) { 3085 pr_debug("Set debug trap ioctl can not be invoked on non-primary kfd process\n"); 3086 3087 return -EOPNOTSUPP; 3088 } 3089 3090 if (sched_policy == KFD_SCHED_POLICY_NO_HWS) { 3091 pr_err("Debugging does not support sched_policy %i", sched_policy); 3092 return -EINVAL; 3093 } 3094 3095 pid = find_get_pid(args->pid); 3096 if (!pid) { 3097 pr_debug("Cannot find pid info for %i\n", args->pid); 3098 r = -ESRCH; 3099 goto out; 3100 } 3101 3102 thread = get_pid_task(pid, PIDTYPE_PID); 3103 if (!thread) { 3104 r = -ESRCH; 3105 goto out; 3106 } 3107 3108 mm = get_task_mm(thread); 3109 if (!mm) { 3110 r = -ESRCH; 3111 goto out; 3112 } 3113 3114 if (args->op == KFD_IOC_DBG_TRAP_ENABLE) { 3115 bool create_process; 3116 3117 rcu_read_lock(); 3118 create_process = thread && thread != current && ptrace_parent(thread) == current; 3119 rcu_read_unlock(); 3120 3121 target = create_process ? kfd_create_process(thread) : 3122 kfd_lookup_process_by_pid(pid); 3123 } else { 3124 target = kfd_lookup_process_by_pid(pid); 3125 } 3126 3127 if (IS_ERR_OR_NULL(target)) { 3128 pr_debug("Cannot find process PID %i to debug\n", args->pid); 3129 r = target ? PTR_ERR(target) : -ESRCH; 3130 target = NULL; 3131 goto out; 3132 } 3133 3134 if (target->context_id != KFD_CONTEXT_ID_PRIMARY) { 3135 pr_debug("Set debug trap ioctl not supported on non-primary kfd process\n"); 3136 r = -EOPNOTSUPP; 3137 goto out; 3138 } 3139 3140 /* Check if target is still PTRACED. */ 3141 rcu_read_lock(); 3142 if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE 3143 && ptrace_parent(target->lead_thread) != current) { 3144 pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid); 3145 r = -EPERM; 3146 } 3147 rcu_read_unlock(); 3148 3149 if (r) 3150 goto out; 3151 3152 mutex_lock(&target->mutex); 3153 3154 if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) { 3155 pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op); 3156 r = -EINVAL; 3157 goto unlock_out; 3158 } 3159 3160 if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED && 3161 (args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE || 3162 args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE || 3163 args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES || 3164 args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES || 3165 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH || 3166 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH || 3167 args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) { 3168 r = -EPERM; 3169 goto unlock_out; 3170 } 3171 3172 if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH || 3173 args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) { 3174 int user_gpu_id = kfd_process_get_user_gpu_id(target, 3175 args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ? 3176 args->set_node_address_watch.gpu_id : 3177 args->clear_node_address_watch.gpu_id); 3178 3179 pdd = kfd_process_device_data_by_id(target, user_gpu_id); 3180 if (user_gpu_id == -EINVAL || !pdd) { 3181 r = -ENODEV; 3182 goto unlock_out; 3183 } 3184 } 3185 3186 switch (args->op) { 3187 case KFD_IOC_DBG_TRAP_ENABLE: 3188 if (target != p) 3189 target->debugger_process = p; 3190 3191 r = kfd_dbg_trap_enable(target, 3192 args->enable.dbg_fd, 3193 (void __user *)args->enable.rinfo_ptr, 3194 &args->enable.rinfo_size); 3195 if (!r) 3196 target->exception_enable_mask = args->enable.exception_mask; 3197 3198 break; 3199 case KFD_IOC_DBG_TRAP_DISABLE: 3200 r = kfd_dbg_trap_disable(target); 3201 break; 3202 case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT: 3203 r = kfd_dbg_send_exception_to_runtime(target, 3204 args->send_runtime_event.gpu_id, 3205 args->send_runtime_event.queue_id, 3206 args->send_runtime_event.exception_mask); 3207 break; 3208 case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED: 3209 kfd_dbg_set_enabled_debug_exception_mask(target, 3210 args->set_exceptions_enabled.exception_mask); 3211 break; 3212 case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE: 3213 r = kfd_dbg_trap_set_wave_launch_override(target, 3214 args->launch_override.override_mode, 3215 args->launch_override.enable_mask, 3216 args->launch_override.support_request_mask, 3217 &args->launch_override.enable_mask, 3218 &args->launch_override.support_request_mask); 3219 break; 3220 case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE: 3221 r = kfd_dbg_trap_set_wave_launch_mode(target, 3222 args->launch_mode.launch_mode); 3223 break; 3224 case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES: 3225 r = suspend_queues(target, 3226 args->suspend_queues.num_queues, 3227 args->suspend_queues.grace_period, 3228 args->suspend_queues.exception_mask, 3229 (uint32_t *)args->suspend_queues.queue_array_ptr); 3230 3231 break; 3232 case KFD_IOC_DBG_TRAP_RESUME_QUEUES: 3233 r = resume_queues(target, args->resume_queues.num_queues, 3234 (uint32_t *)args->resume_queues.queue_array_ptr); 3235 break; 3236 case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH: 3237 r = kfd_dbg_trap_set_dev_address_watch(pdd, 3238 args->set_node_address_watch.address, 3239 args->set_node_address_watch.mask, 3240 &args->set_node_address_watch.id, 3241 args->set_node_address_watch.mode); 3242 break; 3243 case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH: 3244 r = kfd_dbg_trap_clear_dev_address_watch(pdd, 3245 args->clear_node_address_watch.id); 3246 break; 3247 case KFD_IOC_DBG_TRAP_SET_FLAGS: 3248 r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags); 3249 break; 3250 case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT: 3251 r = kfd_dbg_ev_query_debug_event(target, 3252 &args->query_debug_event.queue_id, 3253 &args->query_debug_event.gpu_id, 3254 args->query_debug_event.exception_mask, 3255 &args->query_debug_event.exception_mask); 3256 break; 3257 case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO: 3258 r = kfd_dbg_trap_query_exception_info(target, 3259 args->query_exception_info.source_id, 3260 args->query_exception_info.exception_code, 3261 args->query_exception_info.clear_exception, 3262 (void __user *)args->query_exception_info.info_ptr, 3263 &args->query_exception_info.info_size); 3264 break; 3265 case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT: 3266 r = pqm_get_queue_snapshot(&target->pqm, 3267 args->queue_snapshot.exception_mask, 3268 (void __user *)args->queue_snapshot.snapshot_buf_ptr, 3269 &args->queue_snapshot.num_queues, 3270 &args->queue_snapshot.entry_size); 3271 break; 3272 case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT: 3273 r = kfd_dbg_trap_device_snapshot(target, 3274 args->device_snapshot.exception_mask, 3275 (void __user *)args->device_snapshot.snapshot_buf_ptr, 3276 &args->device_snapshot.num_devices, 3277 &args->device_snapshot.entry_size); 3278 break; 3279 default: 3280 pr_err("Invalid option: %i\n", args->op); 3281 r = -EINVAL; 3282 } 3283 3284 unlock_out: 3285 mutex_unlock(&target->mutex); 3286 3287 out: 3288 if (thread) 3289 put_task_struct(thread); 3290 3291 if (mm) 3292 mmput(mm); 3293 3294 if (pid) 3295 put_pid(pid); 3296 3297 if (target) 3298 kfd_unref_process(target); 3299 3300 return r; 3301 } 3302 3303 /* userspace programs need to invoke this ioctl explicitly on a FD to 3304 * create a secondary kfd_process which replacing its primary kfd_process 3305 */ 3306 static int kfd_ioctl_create_process(struct file *filep, struct kfd_process *p, void *data) 3307 { 3308 struct kfd_process *process; 3309 int ret; 3310 3311 if (!filep->private_data || !p) 3312 return -EINVAL; 3313 3314 /* Each FD owns only one kfd_process */ 3315 if (p->context_id != KFD_CONTEXT_ID_PRIMARY) 3316 return -EINVAL; 3317 3318 mutex_lock(&kfd_processes_mutex); 3319 if (p != filep->private_data) { 3320 mutex_unlock(&kfd_processes_mutex); 3321 return -EINVAL; 3322 } 3323 3324 process = create_process(current, false); 3325 if (IS_ERR(process)) { 3326 mutex_unlock(&kfd_processes_mutex); 3327 return PTR_ERR(process); 3328 } 3329 3330 filep->private_data = process; 3331 mutex_unlock(&kfd_processes_mutex); 3332 3333 ret = kfd_create_process_sysfs(process); 3334 if (ret) 3335 pr_warn("Failed to create sysfs entry for the kfd_process"); 3336 3337 /* Each open() increases kref of the primary kfd_process, 3338 * so we need to reduce it here when we create a new secondary process replacing it 3339 */ 3340 kfd_unref_process(p); 3341 3342 return 0; 3343 } 3344 3345 static inline uint32_t profile_lock_device(struct kfd_process *p, 3346 uint32_t gpu_id, uint32_t op) 3347 { 3348 struct kfd_process_device *pdd; 3349 struct kfd_dev *kfd; 3350 int status = -EINVAL; 3351 struct amdgpu_ptl *ptl; 3352 3353 if (!p) 3354 return -EINVAL; 3355 3356 mutex_lock(&p->mutex); 3357 pdd = kfd_process_device_data_by_id(p, gpu_id); 3358 mutex_unlock(&p->mutex); 3359 3360 if (!pdd || !pdd->dev || !pdd->dev->kfd) 3361 return -EINVAL; 3362 3363 kfd = pdd->dev->kfd; 3364 ptl = &pdd->dev->adev->psp.ptl; 3365 3366 mutex_lock(&kfd->profiler_lock); 3367 if (op == 1) { 3368 if (!kfd->profiler_process) { 3369 kfd->profiler_process = p; 3370 status = 0; 3371 mutex_unlock(&kfd->profiler_lock); 3372 if (ptl->hw_supported) { 3373 status = kfd_ptl_disable_request(pdd, p); 3374 if (status != 0) 3375 dev_err(kfd_device, 3376 "Failed to lock device %d for profiling, error %d\n", 3377 gpu_id, status); 3378 } 3379 return status; 3380 } else if (kfd->profiler_process == p) { 3381 status = -EALREADY; 3382 } else { 3383 status = -EBUSY; 3384 } 3385 } else if (op == 0 && kfd->profiler_process == p) { 3386 kfd->profiler_process = NULL; 3387 status = 0; 3388 mutex_unlock(&kfd->profiler_lock); 3389 3390 if (ptl->hw_supported) { 3391 status = kfd_ptl_disable_release(pdd, p); 3392 if (status) 3393 dev_err(kfd_device, 3394 "Failed to unlock device %d for profiling, error %d\n", 3395 gpu_id, status); 3396 } 3397 return status; 3398 } 3399 mutex_unlock(&kfd->profiler_lock); 3400 3401 return status; 3402 } 3403 3404 static inline int kfd_profiler_pmc(struct kfd_process *p, 3405 struct kfd_ioctl_pmc_settings *args) 3406 { 3407 struct kfd_process_device *pdd; 3408 struct device_queue_manager *dqm; 3409 int status; 3410 3411 /* Check if we have the correct permissions. */ 3412 if (!perfmon_capable()) 3413 return -EPERM; 3414 3415 /* Lock/Unlock the device based on the parameter given in OP */ 3416 status = profile_lock_device(p, args->gpu_id, args->lock); 3417 if (status != 0) 3418 return status; 3419 3420 /* Enable/disable perfcount if requested */ 3421 mutex_lock(&p->mutex); 3422 pdd = kfd_process_device_data_by_id(p, args->gpu_id); 3423 dqm = pdd->dev->dqm; 3424 mutex_unlock(&p->mutex); 3425 3426 dqm->ops.set_perfcount(dqm, args->perfcount_enable); 3427 return status; 3428 } 3429 3430 static int kfd_ioctl_profiler(struct file *filep, struct kfd_process *p, void *data) 3431 { 3432 struct kfd_ioctl_profiler_args *args = data; 3433 3434 switch (args->op) { 3435 case KFD_IOC_PROFILER_VERSION: 3436 args->version = KFD_IOC_PROFILER_VERSION_NUM; 3437 return 0; 3438 case KFD_IOC_PROFILER_PMC: 3439 return kfd_profiler_pmc(p, &args->pmc); 3440 case KFD_IOC_PROFILER_PTL_CONTROL: 3441 return kfd_profiler_ptl_control(p, &args->ptl); 3442 } 3443 return -EINVAL; 3444 } 3445 3446 #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \ 3447 [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \ 3448 .validate = NULL, .cmd_drv = 0, .name = #ioctl} 3449 3450 #define AMDKFD_IOCTL_DEF_V(ioctl, _func, _validate, _flags) \ 3451 [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \ 3452 .validate = _validate, .cmd_drv = 0, .name = #ioctl} 3453 3454 /** Ioctl table */ 3455 static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = { 3456 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_VERSION, 3457 kfd_ioctl_get_version, 0), 3458 3459 AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_QUEUE, 3460 kfd_ioctl_create_queue, 0), 3461 3462 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_QUEUE, 3463 kfd_ioctl_destroy_queue, 0), 3464 3465 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_MEMORY_POLICY, 3466 kfd_ioctl_set_memory_policy, 0), 3467 3468 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_CLOCK_COUNTERS, 3469 kfd_ioctl_get_clock_counters, 0), 3470 3471 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES, 3472 kfd_ioctl_get_process_apertures, 0), 3473 3474 AMDKFD_IOCTL_DEF(AMDKFD_IOC_UPDATE_QUEUE, 3475 kfd_ioctl_update_queue, 0), 3476 3477 AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_EVENT, 3478 kfd_ioctl_create_event, 0), 3479 3480 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DESTROY_EVENT, 3481 kfd_ioctl_destroy_event, 0), 3482 3483 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_EVENT, 3484 kfd_ioctl_set_event, 0), 3485 3486 AMDKFD_IOCTL_DEF(AMDKFD_IOC_RESET_EVENT, 3487 kfd_ioctl_reset_event, 0), 3488 3489 AMDKFD_IOCTL_DEF(AMDKFD_IOC_WAIT_EVENTS, 3490 kfd_ioctl_wait_events, 0), 3491 3492 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_REGISTER_DEPRECATED, 3493 kfd_ioctl_dbg_register, 0), 3494 3495 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_UNREGISTER_DEPRECATED, 3496 kfd_ioctl_dbg_unregister, 0), 3497 3498 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_ADDRESS_WATCH_DEPRECATED, 3499 kfd_ioctl_dbg_address_watch, 0), 3500 3501 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_WAVE_CONTROL_DEPRECATED, 3502 kfd_ioctl_dbg_wave_control, 0), 3503 3504 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_SCRATCH_BACKING_VA, 3505 kfd_ioctl_set_scratch_backing_va, 0), 3506 3507 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG, 3508 kfd_ioctl_get_tile_config, 0), 3509 3510 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER, 3511 kfd_ioctl_set_trap_handler, 0), 3512 3513 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_PROCESS_APERTURES_NEW, 3514 kfd_ioctl_get_process_apertures_new, 0), 3515 3516 AMDKFD_IOCTL_DEF(AMDKFD_IOC_ACQUIRE_VM, 3517 kfd_ioctl_acquire_vm, 0), 3518 3519 AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_MEMORY_OF_GPU, 3520 kfd_ioctl_alloc_memory_of_gpu, 0), 3521 3522 AMDKFD_IOCTL_DEF(AMDKFD_IOC_FREE_MEMORY_OF_GPU, 3523 kfd_ioctl_free_memory_of_gpu, 0), 3524 3525 AMDKFD_IOCTL_DEF(AMDKFD_IOC_MAP_MEMORY_TO_GPU, 3526 kfd_ioctl_map_memory_to_gpu, 0), 3527 3528 AMDKFD_IOCTL_DEF(AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU, 3529 kfd_ioctl_unmap_memory_from_gpu, 0), 3530 3531 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_CU_MASK, 3532 kfd_ioctl_set_cu_mask, 0), 3533 3534 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_QUEUE_WAVE_STATE, 3535 kfd_ioctl_get_queue_wave_state, 0), 3536 3537 AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_DMABUF_INFO, 3538 kfd_ioctl_get_dmabuf_info, 0), 3539 3540 AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF, 3541 kfd_ioctl_import_dmabuf, 0), 3542 3543 AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS, 3544 kfd_ioctl_alloc_queue_gws, 0), 3545 3546 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SMI_EVENTS, 3547 kfd_ioctl_smi_events, 0), 3548 3549 AMDKFD_IOCTL_DEF_V(AMDKFD_IOC_SVM, kfd_ioctl_svm, 3550 kfd_ioctl_svm_validate, 0), 3551 3552 AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE, 3553 kfd_ioctl_set_xnack_mode, 0), 3554 3555 AMDKFD_IOCTL_DEF(AMDKFD_IOC_CRIU_OP, 3556 kfd_ioctl_criu, KFD_IOC_FLAG_CHECKPOINT_RESTORE), 3557 3558 AMDKFD_IOCTL_DEF(AMDKFD_IOC_AVAILABLE_MEMORY, 3559 kfd_ioctl_get_available_memory, 0), 3560 3561 AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF, 3562 kfd_ioctl_export_dmabuf, 0), 3563 3564 AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE, 3565 kfd_ioctl_runtime_enable, 0), 3566 3567 AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP, 3568 kfd_ioctl_set_debug_trap, 0), 3569 3570 AMDKFD_IOCTL_DEF(AMDKFD_IOC_CREATE_PROCESS, 3571 kfd_ioctl_create_process, 0), 3572 3573 AMDKFD_IOCTL_DEF(AMDKFD_IOC_PROFILER, 3574 kfd_ioctl_profiler, 0), 3575 }; 3576 3577 #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls) 3578 3579 static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg) 3580 { 3581 struct kfd_process *process; 3582 amdkfd_ioctl_t *func; 3583 const struct amdkfd_ioctl_desc *ioctl = NULL; 3584 unsigned int nr = _IOC_NR(cmd); 3585 char stack_kdata[128]; 3586 char *kdata = NULL; 3587 unsigned int usize, asize; 3588 int retcode = -EINVAL; 3589 bool ptrace_attached = false; 3590 3591 if (nr >= AMDKFD_CORE_IOCTL_COUNT) { 3592 retcode = -ENOTTY; 3593 goto err_i1; 3594 } 3595 3596 if ((nr >= AMDKFD_COMMAND_START) && (nr < AMDKFD_COMMAND_END)) { 3597 u32 amdkfd_size; 3598 3599 ioctl = &amdkfd_ioctls[nr]; 3600 3601 amdkfd_size = _IOC_SIZE(ioctl->cmd); 3602 usize = asize = _IOC_SIZE(cmd); 3603 if (amdkfd_size > asize) 3604 asize = amdkfd_size; 3605 3606 cmd = ioctl->cmd; 3607 } else { 3608 retcode = -ENOTTY; 3609 goto err_i1; 3610 } 3611 3612 dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg); 3613 3614 /* Get the process struct from the filep. Only the process 3615 * that opened /dev/kfd can use the file descriptor. Child 3616 * processes need to create their own KFD device context. 3617 */ 3618 process = filep->private_data; 3619 3620 rcu_read_lock(); 3621 if ((ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE) && 3622 ptrace_parent(process->lead_thread) == current) 3623 ptrace_attached = true; 3624 rcu_read_unlock(); 3625 3626 if (process->lead_thread != current->group_leader 3627 && !ptrace_attached) { 3628 dev_dbg(kfd_device, "Using KFD FD in wrong process\n"); 3629 retcode = -EBADF; 3630 goto err_i1; 3631 } 3632 3633 /* Do not trust userspace, use our own definition */ 3634 func = ioctl->func; 3635 3636 if (unlikely(!func)) { 3637 dev_dbg(kfd_device, "no function\n"); 3638 retcode = -EINVAL; 3639 goto err_i1; 3640 } 3641 3642 /* 3643 * Versions of docker shipped in Ubuntu 18.xx and 20.xx do not support 3644 * CAP_CHECKPOINT_RESTORE, so we also allow access if CAP_SYS_ADMIN as CAP_SYS_ADMIN is a 3645 * more priviledged access. 3646 */ 3647 if (unlikely(ioctl->flags & KFD_IOC_FLAG_CHECKPOINT_RESTORE)) { 3648 if (!capable(CAP_CHECKPOINT_RESTORE) && 3649 !capable(CAP_SYS_ADMIN)) { 3650 retcode = -EACCES; 3651 goto err_i1; 3652 } 3653 } 3654 3655 if (cmd & (IOC_IN | IOC_OUT)) { 3656 if (asize <= sizeof(stack_kdata)) { 3657 kdata = stack_kdata; 3658 } else { 3659 kdata = kmalloc(asize, GFP_KERNEL); 3660 if (!kdata) { 3661 retcode = -ENOMEM; 3662 goto err_i1; 3663 } 3664 } 3665 if (asize > usize) 3666 memset(kdata + usize, 0, asize - usize); 3667 } 3668 3669 if (cmd & IOC_IN) { 3670 if (copy_from_user(kdata, (void __user *)arg, usize) != 0) { 3671 retcode = -EFAULT; 3672 goto err_i1; 3673 } 3674 } else if (cmd & IOC_OUT) { 3675 memset(kdata, 0, usize); 3676 } 3677 3678 if (ioctl->validate) { 3679 retcode = ioctl->validate(kdata, usize); 3680 if (retcode) 3681 goto err_i1; 3682 } 3683 3684 retcode = func(filep, process, kdata); 3685 3686 if (cmd & IOC_OUT) 3687 if (copy_to_user((void __user *)arg, kdata, usize) != 0) 3688 retcode = -EFAULT; 3689 3690 err_i1: 3691 if (!ioctl) 3692 dev_dbg(kfd_device, "invalid ioctl: pid=%d, cmd=0x%02x, nr=0x%02x\n", 3693 task_pid_nr(current), cmd, nr); 3694 3695 if (kdata != stack_kdata) 3696 kfree(kdata); 3697 3698 if (retcode) 3699 dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n", 3700 nr, arg, retcode); 3701 3702 return retcode; 3703 } 3704 3705 static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process, 3706 struct vm_area_struct *vma) 3707 { 3708 phys_addr_t address; 3709 3710 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 3711 return -EINVAL; 3712 3713 if (PAGE_SIZE > 4096) 3714 return -EINVAL; 3715 3716 address = dev->adev->rmmio_remap.bus_addr; 3717 3718 vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE | 3719 VM_DONTDUMP | VM_PFNMAP); 3720 3721 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 3722 3723 pr_debug("process pid %d mapping mmio page\n" 3724 " target user address == 0x%08llX\n" 3725 " physical address == 0x%08llX\n" 3726 " vm_flags == 0x%04lX\n" 3727 " size == 0x%04lX\n", 3728 process->lead_thread->pid, (unsigned long long) vma->vm_start, 3729 address, vma->vm_flags, PAGE_SIZE); 3730 3731 return io_remap_pfn_range(vma, 3732 vma->vm_start, 3733 address >> PAGE_SHIFT, 3734 PAGE_SIZE, 3735 vma->vm_page_prot); 3736 } 3737 3738 3739 static int kfd_mmap(struct file *filep, struct vm_area_struct *vma) 3740 { 3741 struct kfd_process *process; 3742 struct kfd_node *dev = NULL; 3743 unsigned long mmap_offset; 3744 unsigned int gpu_id; 3745 3746 process = filep->private_data; 3747 if (!process) 3748 return -ESRCH; 3749 3750 if (process->lead_thread != current->group_leader) 3751 return -EBADF; 3752 3753 mmap_offset = vma->vm_pgoff << PAGE_SHIFT; 3754 gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset); 3755 if (gpu_id) 3756 dev = kfd_device_by_id(gpu_id); 3757 3758 switch (mmap_offset & KFD_MMAP_TYPE_MASK) { 3759 case KFD_MMAP_TYPE_DOORBELL: 3760 if (!dev) 3761 return -ENODEV; 3762 return kfd_doorbell_mmap(dev, process, vma); 3763 3764 case KFD_MMAP_TYPE_EVENTS: 3765 return kfd_event_mmap(process, vma); 3766 3767 case KFD_MMAP_TYPE_RESERVED_MEM: 3768 pr_warn("KFD_MMAP_TYPE_RESERVED_MEM is no longer supported\n"); 3769 return -EINVAL; 3770 case KFD_MMAP_TYPE_MMIO: 3771 if (!dev) 3772 return -ENODEV; 3773 return kfd_mmio_mmap(dev, process, vma); 3774 } 3775 3776 return -EFAULT; 3777 } 3778