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/mm_types.h> 25 #include <linux/slab.h> 26 #include <linux/types.h> 27 #include <linux/sched/signal.h> 28 #include <linux/sched/mm.h> 29 #include <linux/uaccess.h> 30 #include <linux/mman.h> 31 #include <linux/memory.h> 32 #include "kfd_priv.h" 33 #include "kfd_events.h" 34 #include "kfd_iommu.h" 35 #include <linux/device.h> 36 37 /* 38 * Wrapper around wait_queue_entry_t 39 */ 40 struct kfd_event_waiter { 41 wait_queue_entry_t wait; 42 struct kfd_event *event; /* Event to wait for */ 43 bool activated; /* Becomes true when event is signaled */ 44 }; 45 46 /* 47 * Each signal event needs a 64-bit signal slot where the signaler will write 48 * a 1 before sending an interrupt. (This is needed because some interrupts 49 * do not contain enough spare data bits to identify an event.) 50 * We get whole pages and map them to the process VA. 51 * Individual signal events use their event_id as slot index. 52 */ 53 struct kfd_signal_page { 54 uint64_t *kernel_address; 55 uint64_t __user *user_address; 56 bool need_to_free_pages; 57 }; 58 59 static uint64_t *page_slots(struct kfd_signal_page *page) 60 { 61 return page->kernel_address; 62 } 63 64 static struct kfd_signal_page *allocate_signal_page(struct kfd_process *p) 65 { 66 void *backing_store; 67 struct kfd_signal_page *page; 68 69 page = kzalloc(sizeof(*page), GFP_KERNEL); 70 if (!page) 71 return NULL; 72 73 backing_store = (void *) __get_free_pages(GFP_KERNEL, 74 get_order(KFD_SIGNAL_EVENT_LIMIT * 8)); 75 if (!backing_store) 76 goto fail_alloc_signal_store; 77 78 /* Initialize all events to unsignaled */ 79 memset(backing_store, (uint8_t) UNSIGNALED_EVENT_SLOT, 80 KFD_SIGNAL_EVENT_LIMIT * 8); 81 82 page->kernel_address = backing_store; 83 page->need_to_free_pages = true; 84 pr_debug("Allocated new event signal page at %p, for process %p\n", 85 page, p); 86 87 return page; 88 89 fail_alloc_signal_store: 90 kfree(page); 91 return NULL; 92 } 93 94 static int allocate_event_notification_slot(struct kfd_process *p, 95 struct kfd_event *ev, 96 const int *restore_id) 97 { 98 int id; 99 100 if (!p->signal_page) { 101 p->signal_page = allocate_signal_page(p); 102 if (!p->signal_page) 103 return -ENOMEM; 104 /* Oldest user mode expects 256 event slots */ 105 p->signal_mapped_size = 256*8; 106 } 107 108 if (restore_id) { 109 id = idr_alloc(&p->event_idr, ev, *restore_id, *restore_id + 1, 110 GFP_KERNEL); 111 } else { 112 /* 113 * Compatibility with old user mode: Only use signal slots 114 * user mode has mapped, may be less than 115 * KFD_SIGNAL_EVENT_LIMIT. This also allows future increase 116 * of the event limit without breaking user mode. 117 */ 118 id = idr_alloc(&p->event_idr, ev, 0, p->signal_mapped_size / 8, 119 GFP_KERNEL); 120 } 121 if (id < 0) 122 return id; 123 124 ev->event_id = id; 125 page_slots(p->signal_page)[id] = UNSIGNALED_EVENT_SLOT; 126 127 return 0; 128 } 129 130 /* 131 * Assumes that p->event_mutex or rcu_readlock is held and of course that p is 132 * not going away. 133 */ 134 static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id) 135 { 136 return idr_find(&p->event_idr, id); 137 } 138 139 /** 140 * lookup_signaled_event_by_partial_id - Lookup signaled event from partial ID 141 * @p: Pointer to struct kfd_process 142 * @id: ID to look up 143 * @bits: Number of valid bits in @id 144 * 145 * Finds the first signaled event with a matching partial ID. If no 146 * matching signaled event is found, returns NULL. In that case the 147 * caller should assume that the partial ID is invalid and do an 148 * exhaustive search of all siglaned events. 149 * 150 * If multiple events with the same partial ID signal at the same 151 * time, they will be found one interrupt at a time, not necessarily 152 * in the same order the interrupts occurred. As long as the number of 153 * interrupts is correct, all signaled events will be seen by the 154 * driver. 155 */ 156 static struct kfd_event *lookup_signaled_event_by_partial_id( 157 struct kfd_process *p, uint32_t id, uint32_t bits) 158 { 159 struct kfd_event *ev; 160 161 if (!p->signal_page || id >= KFD_SIGNAL_EVENT_LIMIT) 162 return NULL; 163 164 /* Fast path for the common case that @id is not a partial ID 165 * and we only need a single lookup. 166 */ 167 if (bits > 31 || (1U << bits) >= KFD_SIGNAL_EVENT_LIMIT) { 168 if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT) 169 return NULL; 170 171 return idr_find(&p->event_idr, id); 172 } 173 174 /* General case for partial IDs: Iterate over all matching IDs 175 * and find the first one that has signaled. 176 */ 177 for (ev = NULL; id < KFD_SIGNAL_EVENT_LIMIT && !ev; id += 1U << bits) { 178 if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT) 179 continue; 180 181 ev = idr_find(&p->event_idr, id); 182 } 183 184 return ev; 185 } 186 187 static int create_signal_event(struct file *devkfd, struct kfd_process *p, 188 struct kfd_event *ev, const int *restore_id) 189 { 190 int ret; 191 192 if (p->signal_mapped_size && 193 p->signal_event_count == p->signal_mapped_size / 8) { 194 if (!p->signal_event_limit_reached) { 195 pr_debug("Signal event wasn't created because limit was reached\n"); 196 p->signal_event_limit_reached = true; 197 } 198 return -ENOSPC; 199 } 200 201 ret = allocate_event_notification_slot(p, ev, restore_id); 202 if (ret) { 203 pr_warn("Signal event wasn't created because out of kernel memory\n"); 204 return ret; 205 } 206 207 p->signal_event_count++; 208 209 ev->user_signal_address = &p->signal_page->user_address[ev->event_id]; 210 pr_debug("Signal event number %zu created with id %d, address %p\n", 211 p->signal_event_count, ev->event_id, 212 ev->user_signal_address); 213 214 return 0; 215 } 216 217 static int create_other_event(struct kfd_process *p, struct kfd_event *ev, const int *restore_id) 218 { 219 int id; 220 221 if (restore_id) 222 id = idr_alloc(&p->event_idr, ev, *restore_id, *restore_id + 1, 223 GFP_KERNEL); 224 else 225 /* Cast KFD_LAST_NONSIGNAL_EVENT to uint32_t. This allows an 226 * intentional integer overflow to -1 without a compiler 227 * warning. idr_alloc treats a negative value as "maximum 228 * signed integer". 229 */ 230 id = idr_alloc(&p->event_idr, ev, KFD_FIRST_NONSIGNAL_EVENT_ID, 231 (uint32_t)KFD_LAST_NONSIGNAL_EVENT_ID + 1, 232 GFP_KERNEL); 233 234 if (id < 0) 235 return id; 236 ev->event_id = id; 237 238 return 0; 239 } 240 241 int kfd_event_init_process(struct kfd_process *p) 242 { 243 int id; 244 245 mutex_init(&p->event_mutex); 246 idr_init(&p->event_idr); 247 p->signal_page = NULL; 248 p->signal_event_count = 1; 249 /* Allocate event ID 0. It is used for a fast path to ignore bogus events 250 * that are sent by the CP without a context ID 251 */ 252 id = idr_alloc(&p->event_idr, NULL, 0, 1, GFP_KERNEL); 253 if (id < 0) { 254 idr_destroy(&p->event_idr); 255 mutex_destroy(&p->event_mutex); 256 return id; 257 } 258 return 0; 259 } 260 261 static void destroy_event(struct kfd_process *p, struct kfd_event *ev) 262 { 263 struct kfd_event_waiter *waiter; 264 265 /* Wake up pending waiters. They will return failure */ 266 spin_lock(&ev->lock); 267 list_for_each_entry(waiter, &ev->wq.head, wait.entry) 268 WRITE_ONCE(waiter->event, NULL); 269 wake_up_all(&ev->wq); 270 spin_unlock(&ev->lock); 271 272 if (ev->type == KFD_EVENT_TYPE_SIGNAL || 273 ev->type == KFD_EVENT_TYPE_DEBUG) 274 p->signal_event_count--; 275 276 idr_remove(&p->event_idr, ev->event_id); 277 kfree_rcu(ev, rcu); 278 } 279 280 static void destroy_events(struct kfd_process *p) 281 { 282 struct kfd_event *ev; 283 uint32_t id; 284 285 idr_for_each_entry(&p->event_idr, ev, id) 286 if (ev) 287 destroy_event(p, ev); 288 idr_destroy(&p->event_idr); 289 mutex_destroy(&p->event_mutex); 290 } 291 292 /* 293 * We assume that the process is being destroyed and there is no need to 294 * unmap the pages or keep bookkeeping data in order. 295 */ 296 static void shutdown_signal_page(struct kfd_process *p) 297 { 298 struct kfd_signal_page *page = p->signal_page; 299 300 if (page) { 301 if (page->need_to_free_pages) 302 free_pages((unsigned long)page->kernel_address, 303 get_order(KFD_SIGNAL_EVENT_LIMIT * 8)); 304 kfree(page); 305 } 306 } 307 308 void kfd_event_free_process(struct kfd_process *p) 309 { 310 destroy_events(p); 311 shutdown_signal_page(p); 312 } 313 314 static bool event_can_be_gpu_signaled(const struct kfd_event *ev) 315 { 316 return ev->type == KFD_EVENT_TYPE_SIGNAL || 317 ev->type == KFD_EVENT_TYPE_DEBUG; 318 } 319 320 static bool event_can_be_cpu_signaled(const struct kfd_event *ev) 321 { 322 return ev->type == KFD_EVENT_TYPE_SIGNAL; 323 } 324 325 static int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 326 uint64_t size, uint64_t user_handle) 327 { 328 struct kfd_signal_page *page; 329 330 if (p->signal_page) 331 return -EBUSY; 332 333 page = kzalloc(sizeof(*page), GFP_KERNEL); 334 if (!page) 335 return -ENOMEM; 336 337 /* Initialize all events to unsignaled */ 338 memset(kernel_address, (uint8_t) UNSIGNALED_EVENT_SLOT, 339 KFD_SIGNAL_EVENT_LIMIT * 8); 340 341 page->kernel_address = kernel_address; 342 343 p->signal_page = page; 344 p->signal_mapped_size = size; 345 p->signal_handle = user_handle; 346 return 0; 347 } 348 349 int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset) 350 { 351 struct kfd_dev *kfd; 352 struct kfd_process_device *pdd; 353 void *mem, *kern_addr; 354 uint64_t size; 355 int err = 0; 356 357 if (p->signal_page) { 358 pr_err("Event page is already set\n"); 359 return -EINVAL; 360 } 361 362 pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(event_page_offset)); 363 if (!pdd) { 364 pr_err("Getting device by id failed in %s\n", __func__); 365 return -EINVAL; 366 } 367 kfd = pdd->dev; 368 369 pdd = kfd_bind_process_to_device(kfd, p); 370 if (IS_ERR(pdd)) 371 return PTR_ERR(pdd); 372 373 mem = kfd_process_device_translate_handle(pdd, 374 GET_IDR_HANDLE(event_page_offset)); 375 if (!mem) { 376 pr_err("Can't find BO, offset is 0x%llx\n", event_page_offset); 377 return -EINVAL; 378 } 379 380 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(mem, &kern_addr, &size); 381 if (err) { 382 pr_err("Failed to map event page to kernel\n"); 383 return err; 384 } 385 386 err = kfd_event_page_set(p, kern_addr, size, event_page_offset); 387 if (err) { 388 pr_err("Failed to set event page\n"); 389 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem); 390 return err; 391 } 392 return err; 393 } 394 395 int kfd_event_create(struct file *devkfd, struct kfd_process *p, 396 uint32_t event_type, bool auto_reset, uint32_t node_id, 397 uint32_t *event_id, uint32_t *event_trigger_data, 398 uint64_t *event_page_offset, uint32_t *event_slot_index) 399 { 400 int ret = 0; 401 struct kfd_event *ev = kzalloc(sizeof(*ev), GFP_KERNEL); 402 403 if (!ev) 404 return -ENOMEM; 405 406 ev->type = event_type; 407 ev->auto_reset = auto_reset; 408 ev->signaled = false; 409 410 spin_lock_init(&ev->lock); 411 init_waitqueue_head(&ev->wq); 412 413 *event_page_offset = 0; 414 415 mutex_lock(&p->event_mutex); 416 417 switch (event_type) { 418 case KFD_EVENT_TYPE_SIGNAL: 419 case KFD_EVENT_TYPE_DEBUG: 420 ret = create_signal_event(devkfd, p, ev, NULL); 421 if (!ret) { 422 *event_page_offset = KFD_MMAP_TYPE_EVENTS; 423 *event_slot_index = ev->event_id; 424 } 425 break; 426 default: 427 ret = create_other_event(p, ev, NULL); 428 break; 429 } 430 431 if (!ret) { 432 *event_id = ev->event_id; 433 *event_trigger_data = ev->event_id; 434 } else { 435 kfree(ev); 436 } 437 438 mutex_unlock(&p->event_mutex); 439 440 return ret; 441 } 442 443 int kfd_criu_restore_event(struct file *devkfd, 444 struct kfd_process *p, 445 uint8_t __user *user_priv_ptr, 446 uint64_t *priv_data_offset, 447 uint64_t max_priv_data_size) 448 { 449 struct kfd_criu_event_priv_data *ev_priv; 450 struct kfd_event *ev = NULL; 451 int ret = 0; 452 453 ev_priv = kmalloc(sizeof(*ev_priv), GFP_KERNEL); 454 if (!ev_priv) 455 return -ENOMEM; 456 457 ev = kzalloc(sizeof(*ev), GFP_KERNEL); 458 if (!ev) { 459 ret = -ENOMEM; 460 goto exit; 461 } 462 463 if (*priv_data_offset + sizeof(*ev_priv) > max_priv_data_size) { 464 ret = -EINVAL; 465 goto exit; 466 } 467 468 ret = copy_from_user(ev_priv, user_priv_ptr + *priv_data_offset, sizeof(*ev_priv)); 469 if (ret) { 470 ret = -EFAULT; 471 goto exit; 472 } 473 *priv_data_offset += sizeof(*ev_priv); 474 475 if (ev_priv->user_handle) { 476 ret = kfd_kmap_event_page(p, ev_priv->user_handle); 477 if (ret) 478 goto exit; 479 } 480 481 ev->type = ev_priv->type; 482 ev->auto_reset = ev_priv->auto_reset; 483 ev->signaled = ev_priv->signaled; 484 485 spin_lock_init(&ev->lock); 486 init_waitqueue_head(&ev->wq); 487 488 mutex_lock(&p->event_mutex); 489 switch (ev->type) { 490 case KFD_EVENT_TYPE_SIGNAL: 491 case KFD_EVENT_TYPE_DEBUG: 492 ret = create_signal_event(devkfd, p, ev, &ev_priv->event_id); 493 break; 494 case KFD_EVENT_TYPE_MEMORY: 495 memcpy(&ev->memory_exception_data, 496 &ev_priv->memory_exception_data, 497 sizeof(struct kfd_hsa_memory_exception_data)); 498 499 ret = create_other_event(p, ev, &ev_priv->event_id); 500 break; 501 case KFD_EVENT_TYPE_HW_EXCEPTION: 502 memcpy(&ev->hw_exception_data, 503 &ev_priv->hw_exception_data, 504 sizeof(struct kfd_hsa_hw_exception_data)); 505 506 ret = create_other_event(p, ev, &ev_priv->event_id); 507 break; 508 } 509 mutex_unlock(&p->event_mutex); 510 511 exit: 512 if (ret) 513 kfree(ev); 514 515 kfree(ev_priv); 516 517 return ret; 518 } 519 520 int kfd_criu_checkpoint_events(struct kfd_process *p, 521 uint8_t __user *user_priv_data, 522 uint64_t *priv_data_offset) 523 { 524 struct kfd_criu_event_priv_data *ev_privs; 525 int i = 0; 526 int ret = 0; 527 struct kfd_event *ev; 528 uint32_t ev_id; 529 530 uint32_t num_events = kfd_get_num_events(p); 531 532 if (!num_events) 533 return 0; 534 535 ev_privs = kvzalloc(num_events * sizeof(*ev_privs), GFP_KERNEL); 536 if (!ev_privs) 537 return -ENOMEM; 538 539 540 idr_for_each_entry(&p->event_idr, ev, ev_id) { 541 struct kfd_criu_event_priv_data *ev_priv; 542 543 /* 544 * Currently, all events have same size of private_data, but the current ioctl's 545 * and CRIU plugin supports private_data of variable sizes 546 */ 547 ev_priv = &ev_privs[i]; 548 549 ev_priv->object_type = KFD_CRIU_OBJECT_TYPE_EVENT; 550 551 /* We store the user_handle with the first event */ 552 if (i == 0 && p->signal_page) 553 ev_priv->user_handle = p->signal_handle; 554 555 ev_priv->event_id = ev->event_id; 556 ev_priv->auto_reset = ev->auto_reset; 557 ev_priv->type = ev->type; 558 ev_priv->signaled = ev->signaled; 559 560 if (ev_priv->type == KFD_EVENT_TYPE_MEMORY) 561 memcpy(&ev_priv->memory_exception_data, 562 &ev->memory_exception_data, 563 sizeof(struct kfd_hsa_memory_exception_data)); 564 else if (ev_priv->type == KFD_EVENT_TYPE_HW_EXCEPTION) 565 memcpy(&ev_priv->hw_exception_data, 566 &ev->hw_exception_data, 567 sizeof(struct kfd_hsa_hw_exception_data)); 568 569 pr_debug("Checkpointed event[%d] id = 0x%08x auto_reset = %x type = %x signaled = %x\n", 570 i, 571 ev_priv->event_id, 572 ev_priv->auto_reset, 573 ev_priv->type, 574 ev_priv->signaled); 575 i++; 576 } 577 578 ret = copy_to_user(user_priv_data + *priv_data_offset, 579 ev_privs, num_events * sizeof(*ev_privs)); 580 if (ret) { 581 pr_err("Failed to copy events priv to user\n"); 582 ret = -EFAULT; 583 } 584 585 *priv_data_offset += num_events * sizeof(*ev_privs); 586 587 kvfree(ev_privs); 588 return ret; 589 } 590 591 int kfd_get_num_events(struct kfd_process *p) 592 { 593 struct kfd_event *ev; 594 uint32_t id; 595 u32 num_events = 0; 596 597 idr_for_each_entry(&p->event_idr, ev, id) 598 num_events++; 599 600 return num_events; 601 } 602 603 /* Assumes that p is current. */ 604 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id) 605 { 606 struct kfd_event *ev; 607 int ret = 0; 608 609 mutex_lock(&p->event_mutex); 610 611 ev = lookup_event_by_id(p, event_id); 612 613 if (ev) 614 destroy_event(p, ev); 615 else 616 ret = -EINVAL; 617 618 mutex_unlock(&p->event_mutex); 619 return ret; 620 } 621 622 static void set_event(struct kfd_event *ev) 623 { 624 struct kfd_event_waiter *waiter; 625 626 /* Auto reset if the list is non-empty and we're waking 627 * someone. waitqueue_active is safe here because we're 628 * protected by the ev->lock, which is also held when 629 * updating the wait queues in kfd_wait_on_events. 630 */ 631 ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq); 632 633 list_for_each_entry(waiter, &ev->wq.head, wait.entry) 634 WRITE_ONCE(waiter->activated, true); 635 636 wake_up_all(&ev->wq); 637 } 638 639 /* Assumes that p is current. */ 640 int kfd_set_event(struct kfd_process *p, uint32_t event_id) 641 { 642 int ret = 0; 643 struct kfd_event *ev; 644 645 rcu_read_lock(); 646 647 ev = lookup_event_by_id(p, event_id); 648 if (!ev) { 649 ret = -EINVAL; 650 goto unlock_rcu; 651 } 652 spin_lock(&ev->lock); 653 654 if (event_can_be_cpu_signaled(ev)) 655 set_event(ev); 656 else 657 ret = -EINVAL; 658 659 spin_unlock(&ev->lock); 660 unlock_rcu: 661 rcu_read_unlock(); 662 return ret; 663 } 664 665 static void reset_event(struct kfd_event *ev) 666 { 667 ev->signaled = false; 668 } 669 670 /* Assumes that p is current. */ 671 int kfd_reset_event(struct kfd_process *p, uint32_t event_id) 672 { 673 int ret = 0; 674 struct kfd_event *ev; 675 676 rcu_read_lock(); 677 678 ev = lookup_event_by_id(p, event_id); 679 if (!ev) { 680 ret = -EINVAL; 681 goto unlock_rcu; 682 } 683 spin_lock(&ev->lock); 684 685 if (event_can_be_cpu_signaled(ev)) 686 reset_event(ev); 687 else 688 ret = -EINVAL; 689 690 spin_unlock(&ev->lock); 691 unlock_rcu: 692 rcu_read_unlock(); 693 return ret; 694 695 } 696 697 static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev) 698 { 699 WRITE_ONCE(page_slots(p->signal_page)[ev->event_id], UNSIGNALED_EVENT_SLOT); 700 } 701 702 static void set_event_from_interrupt(struct kfd_process *p, 703 struct kfd_event *ev) 704 { 705 if (ev && event_can_be_gpu_signaled(ev)) { 706 acknowledge_signal(p, ev); 707 spin_lock(&ev->lock); 708 set_event(ev); 709 spin_unlock(&ev->lock); 710 } 711 } 712 713 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, 714 uint32_t valid_id_bits) 715 { 716 struct kfd_event *ev = NULL; 717 718 /* 719 * Because we are called from arbitrary context (workqueue) as opposed 720 * to process context, kfd_process could attempt to exit while we are 721 * running so the lookup function increments the process ref count. 722 */ 723 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 724 725 if (!p) 726 return; /* Presumably process exited. */ 727 728 rcu_read_lock(); 729 730 if (valid_id_bits) 731 ev = lookup_signaled_event_by_partial_id(p, partial_id, 732 valid_id_bits); 733 if (ev) { 734 set_event_from_interrupt(p, ev); 735 } else if (p->signal_page) { 736 /* 737 * Partial ID lookup failed. Assume that the event ID 738 * in the interrupt payload was invalid and do an 739 * exhaustive search of signaled events. 740 */ 741 uint64_t *slots = page_slots(p->signal_page); 742 uint32_t id; 743 744 if (valid_id_bits) 745 pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n", 746 partial_id, valid_id_bits); 747 748 if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) { 749 /* With relatively few events, it's faster to 750 * iterate over the event IDR 751 */ 752 idr_for_each_entry(&p->event_idr, ev, id) { 753 if (id >= KFD_SIGNAL_EVENT_LIMIT) 754 break; 755 756 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) 757 set_event_from_interrupt(p, ev); 758 } 759 } else { 760 /* With relatively many events, it's faster to 761 * iterate over the signal slots and lookup 762 * only signaled events from the IDR. 763 */ 764 for (id = 1; id < KFD_SIGNAL_EVENT_LIMIT; id++) 765 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) { 766 ev = lookup_event_by_id(p, id); 767 set_event_from_interrupt(p, ev); 768 } 769 } 770 } 771 772 rcu_read_unlock(); 773 kfd_unref_process(p); 774 } 775 776 static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events) 777 { 778 struct kfd_event_waiter *event_waiters; 779 uint32_t i; 780 781 event_waiters = kcalloc(num_events, sizeof(struct kfd_event_waiter), 782 GFP_KERNEL); 783 if (!event_waiters) 784 return NULL; 785 786 for (i = 0; i < num_events; i++) 787 init_wait(&event_waiters[i].wait); 788 789 return event_waiters; 790 } 791 792 static int init_event_waiter(struct kfd_process *p, 793 struct kfd_event_waiter *waiter, 794 uint32_t event_id) 795 { 796 struct kfd_event *ev = lookup_event_by_id(p, event_id); 797 798 if (!ev) 799 return -EINVAL; 800 801 spin_lock(&ev->lock); 802 waiter->event = ev; 803 waiter->activated = ev->signaled; 804 ev->signaled = ev->signaled && !ev->auto_reset; 805 if (!waiter->activated) 806 add_wait_queue(&ev->wq, &waiter->wait); 807 spin_unlock(&ev->lock); 808 809 return 0; 810 } 811 812 /* test_event_condition - Test condition of events being waited for 813 * @all: Return completion only if all events have signaled 814 * @num_events: Number of events to wait for 815 * @event_waiters: Array of event waiters, one per event 816 * 817 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have 818 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all) 819 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of 820 * the events have been destroyed. 821 */ 822 static uint32_t test_event_condition(bool all, uint32_t num_events, 823 struct kfd_event_waiter *event_waiters) 824 { 825 uint32_t i; 826 uint32_t activated_count = 0; 827 828 for (i = 0; i < num_events; i++) { 829 if (!READ_ONCE(event_waiters[i].event)) 830 return KFD_IOC_WAIT_RESULT_FAIL; 831 832 if (READ_ONCE(event_waiters[i].activated)) { 833 if (!all) 834 return KFD_IOC_WAIT_RESULT_COMPLETE; 835 836 activated_count++; 837 } 838 } 839 840 return activated_count == num_events ? 841 KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT; 842 } 843 844 /* 845 * Copy event specific data, if defined. 846 * Currently only memory exception events have additional data to copy to user 847 */ 848 static int copy_signaled_event_data(uint32_t num_events, 849 struct kfd_event_waiter *event_waiters, 850 struct kfd_event_data __user *data) 851 { 852 struct kfd_hsa_memory_exception_data *src; 853 struct kfd_hsa_memory_exception_data __user *dst; 854 struct kfd_event_waiter *waiter; 855 struct kfd_event *event; 856 uint32_t i; 857 858 for (i = 0; i < num_events; i++) { 859 waiter = &event_waiters[i]; 860 event = waiter->event; 861 if (!event) 862 return -EINVAL; /* event was destroyed */ 863 if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) { 864 dst = &data[i].memory_exception_data; 865 src = &event->memory_exception_data; 866 if (copy_to_user(dst, src, 867 sizeof(struct kfd_hsa_memory_exception_data))) 868 return -EFAULT; 869 } 870 } 871 872 return 0; 873 } 874 875 static long user_timeout_to_jiffies(uint32_t user_timeout_ms) 876 { 877 if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE) 878 return 0; 879 880 if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE) 881 return MAX_SCHEDULE_TIMEOUT; 882 883 /* 884 * msecs_to_jiffies interprets all values above 2^31-1 as infinite, 885 * but we consider them finite. 886 * This hack is wrong, but nobody is likely to notice. 887 */ 888 user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF); 889 890 return msecs_to_jiffies(user_timeout_ms) + 1; 891 } 892 893 static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters, 894 bool undo_auto_reset) 895 { 896 uint32_t i; 897 898 for (i = 0; i < num_events; i++) 899 if (waiters[i].event) { 900 spin_lock(&waiters[i].event->lock); 901 remove_wait_queue(&waiters[i].event->wq, 902 &waiters[i].wait); 903 if (undo_auto_reset && waiters[i].activated && 904 waiters[i].event && waiters[i].event->auto_reset) 905 set_event(waiters[i].event); 906 spin_unlock(&waiters[i].event->lock); 907 } 908 909 kfree(waiters); 910 } 911 912 int kfd_wait_on_events(struct kfd_process *p, 913 uint32_t num_events, void __user *data, 914 bool all, uint32_t *user_timeout_ms, 915 uint32_t *wait_result) 916 { 917 struct kfd_event_data __user *events = 918 (struct kfd_event_data __user *) data; 919 uint32_t i; 920 int ret = 0; 921 922 struct kfd_event_waiter *event_waiters = NULL; 923 long timeout = user_timeout_to_jiffies(*user_timeout_ms); 924 925 event_waiters = alloc_event_waiters(num_events); 926 if (!event_waiters) { 927 ret = -ENOMEM; 928 goto out; 929 } 930 931 /* Use p->event_mutex here to protect against concurrent creation and 932 * destruction of events while we initialize event_waiters. 933 */ 934 mutex_lock(&p->event_mutex); 935 936 for (i = 0; i < num_events; i++) { 937 struct kfd_event_data event_data; 938 939 if (copy_from_user(&event_data, &events[i], 940 sizeof(struct kfd_event_data))) { 941 ret = -EFAULT; 942 goto out_unlock; 943 } 944 945 ret = init_event_waiter(p, &event_waiters[i], 946 event_data.event_id); 947 if (ret) 948 goto out_unlock; 949 } 950 951 /* Check condition once. */ 952 *wait_result = test_event_condition(all, num_events, event_waiters); 953 if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) { 954 ret = copy_signaled_event_data(num_events, 955 event_waiters, events); 956 goto out_unlock; 957 } else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) { 958 /* This should not happen. Events shouldn't be 959 * destroyed while we're holding the event_mutex 960 */ 961 goto out_unlock; 962 } 963 964 mutex_unlock(&p->event_mutex); 965 966 while (true) { 967 if (fatal_signal_pending(current)) { 968 ret = -EINTR; 969 break; 970 } 971 972 if (signal_pending(current)) { 973 ret = -ERESTARTSYS; 974 if (*user_timeout_ms != KFD_EVENT_TIMEOUT_IMMEDIATE && 975 *user_timeout_ms != KFD_EVENT_TIMEOUT_INFINITE) 976 *user_timeout_ms = jiffies_to_msecs( 977 max(0l, timeout-1)); 978 break; 979 } 980 981 /* Set task state to interruptible sleep before 982 * checking wake-up conditions. A concurrent wake-up 983 * will put the task back into runnable state. In that 984 * case schedule_timeout will not put the task to 985 * sleep and we'll get a chance to re-check the 986 * updated conditions almost immediately. Otherwise, 987 * this race condition would lead to a soft hang or a 988 * very long sleep. 989 */ 990 set_current_state(TASK_INTERRUPTIBLE); 991 992 *wait_result = test_event_condition(all, num_events, 993 event_waiters); 994 if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT) 995 break; 996 997 if (timeout <= 0) 998 break; 999 1000 timeout = schedule_timeout(timeout); 1001 } 1002 __set_current_state(TASK_RUNNING); 1003 1004 mutex_lock(&p->event_mutex); 1005 /* copy_signaled_event_data may sleep. So this has to happen 1006 * after the task state is set back to RUNNING. 1007 * 1008 * The event may also have been destroyed after signaling. So 1009 * copy_signaled_event_data also must confirm that the event 1010 * still exists. Therefore this must be under the p->event_mutex 1011 * which is also held when events are destroyed. 1012 */ 1013 if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) 1014 ret = copy_signaled_event_data(num_events, 1015 event_waiters, events); 1016 1017 out_unlock: 1018 free_waiters(num_events, event_waiters, ret == -ERESTARTSYS); 1019 mutex_unlock(&p->event_mutex); 1020 out: 1021 if (ret) 1022 *wait_result = KFD_IOC_WAIT_RESULT_FAIL; 1023 else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL) 1024 ret = -EIO; 1025 1026 return ret; 1027 } 1028 1029 int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma) 1030 { 1031 unsigned long pfn; 1032 struct kfd_signal_page *page; 1033 int ret; 1034 1035 /* check required size doesn't exceed the allocated size */ 1036 if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) < 1037 get_order(vma->vm_end - vma->vm_start)) { 1038 pr_err("Event page mmap requested illegal size\n"); 1039 return -EINVAL; 1040 } 1041 1042 page = p->signal_page; 1043 if (!page) { 1044 /* Probably KFD bug, but mmap is user-accessible. */ 1045 pr_debug("Signal page could not be found\n"); 1046 return -EINVAL; 1047 } 1048 1049 pfn = __pa(page->kernel_address); 1050 pfn >>= PAGE_SHIFT; 1051 1052 vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE 1053 | VM_DONTDUMP | VM_PFNMAP); 1054 1055 pr_debug("Mapping signal page\n"); 1056 pr_debug(" start user address == 0x%08lx\n", vma->vm_start); 1057 pr_debug(" end user address == 0x%08lx\n", vma->vm_end); 1058 pr_debug(" pfn == 0x%016lX\n", pfn); 1059 pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags); 1060 pr_debug(" size == 0x%08lX\n", 1061 vma->vm_end - vma->vm_start); 1062 1063 page->user_address = (uint64_t __user *)vma->vm_start; 1064 1065 /* mapping the page to user process */ 1066 ret = remap_pfn_range(vma, vma->vm_start, pfn, 1067 vma->vm_end - vma->vm_start, vma->vm_page_prot); 1068 if (!ret) 1069 p->signal_mapped_size = vma->vm_end - vma->vm_start; 1070 1071 return ret; 1072 } 1073 1074 /* 1075 * Assumes that p is not going away. 1076 */ 1077 static void lookup_events_by_type_and_signal(struct kfd_process *p, 1078 int type, void *event_data) 1079 { 1080 struct kfd_hsa_memory_exception_data *ev_data; 1081 struct kfd_event *ev; 1082 uint32_t id; 1083 bool send_signal = true; 1084 1085 ev_data = (struct kfd_hsa_memory_exception_data *) event_data; 1086 1087 rcu_read_lock(); 1088 1089 id = KFD_FIRST_NONSIGNAL_EVENT_ID; 1090 idr_for_each_entry_continue(&p->event_idr, ev, id) 1091 if (ev->type == type) { 1092 send_signal = false; 1093 dev_dbg(kfd_device, 1094 "Event found: id %X type %d", 1095 ev->event_id, ev->type); 1096 spin_lock(&ev->lock); 1097 set_event(ev); 1098 if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data) 1099 ev->memory_exception_data = *ev_data; 1100 spin_unlock(&ev->lock); 1101 } 1102 1103 if (type == KFD_EVENT_TYPE_MEMORY) { 1104 dev_warn(kfd_device, 1105 "Sending SIGSEGV to process %d (pasid 0x%x)", 1106 p->lead_thread->pid, p->pasid); 1107 send_sig(SIGSEGV, p->lead_thread, 0); 1108 } 1109 1110 /* Send SIGTERM no event of type "type" has been found*/ 1111 if (send_signal) { 1112 if (send_sigterm) { 1113 dev_warn(kfd_device, 1114 "Sending SIGTERM to process %d (pasid 0x%x)", 1115 p->lead_thread->pid, p->pasid); 1116 send_sig(SIGTERM, p->lead_thread, 0); 1117 } else { 1118 dev_err(kfd_device, 1119 "Process %d (pasid 0x%x) got unhandled exception", 1120 p->lead_thread->pid, p->pasid); 1121 } 1122 } 1123 1124 rcu_read_unlock(); 1125 } 1126 1127 #ifdef KFD_SUPPORT_IOMMU_V2 1128 void kfd_signal_iommu_event(struct kfd_dev *dev, u32 pasid, 1129 unsigned long address, bool is_write_requested, 1130 bool is_execute_requested) 1131 { 1132 struct kfd_hsa_memory_exception_data memory_exception_data; 1133 struct vm_area_struct *vma; 1134 int user_gpu_id; 1135 1136 /* 1137 * Because we are called from arbitrary context (workqueue) as opposed 1138 * to process context, kfd_process could attempt to exit while we are 1139 * running so the lookup function increments the process ref count. 1140 */ 1141 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 1142 struct mm_struct *mm; 1143 1144 if (!p) 1145 return; /* Presumably process exited. */ 1146 1147 /* Take a safe reference to the mm_struct, which may otherwise 1148 * disappear even while the kfd_process is still referenced. 1149 */ 1150 mm = get_task_mm(p->lead_thread); 1151 if (!mm) { 1152 kfd_unref_process(p); 1153 return; /* Process is exiting */ 1154 } 1155 1156 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id); 1157 if (unlikely(user_gpu_id == -EINVAL)) { 1158 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id); 1159 return; 1160 } 1161 memset(&memory_exception_data, 0, sizeof(memory_exception_data)); 1162 1163 mmap_read_lock(mm); 1164 vma = find_vma(mm, address); 1165 1166 memory_exception_data.gpu_id = user_gpu_id; 1167 memory_exception_data.va = address; 1168 /* Set failure reason */ 1169 memory_exception_data.failure.NotPresent = 1; 1170 memory_exception_data.failure.NoExecute = 0; 1171 memory_exception_data.failure.ReadOnly = 0; 1172 if (vma && address >= vma->vm_start) { 1173 memory_exception_data.failure.NotPresent = 0; 1174 1175 if (is_write_requested && !(vma->vm_flags & VM_WRITE)) 1176 memory_exception_data.failure.ReadOnly = 1; 1177 else 1178 memory_exception_data.failure.ReadOnly = 0; 1179 1180 if (is_execute_requested && !(vma->vm_flags & VM_EXEC)) 1181 memory_exception_data.failure.NoExecute = 1; 1182 else 1183 memory_exception_data.failure.NoExecute = 0; 1184 } 1185 1186 mmap_read_unlock(mm); 1187 mmput(mm); 1188 1189 pr_debug("notpresent %d, noexecute %d, readonly %d\n", 1190 memory_exception_data.failure.NotPresent, 1191 memory_exception_data.failure.NoExecute, 1192 memory_exception_data.failure.ReadOnly); 1193 1194 /* Workaround on Raven to not kill the process when memory is freed 1195 * before IOMMU is able to finish processing all the excessive PPRs 1196 */ 1197 1198 if (KFD_GC_VERSION(dev) != IP_VERSION(9, 1, 0) && 1199 KFD_GC_VERSION(dev) != IP_VERSION(9, 2, 2) && 1200 KFD_GC_VERSION(dev) != IP_VERSION(9, 3, 0)) 1201 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY, 1202 &memory_exception_data); 1203 1204 kfd_unref_process(p); 1205 } 1206 #endif /* KFD_SUPPORT_IOMMU_V2 */ 1207 1208 void kfd_signal_hw_exception_event(u32 pasid) 1209 { 1210 /* 1211 * Because we are called from arbitrary context (workqueue) as opposed 1212 * to process context, kfd_process could attempt to exit while we are 1213 * running so the lookup function increments the process ref count. 1214 */ 1215 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 1216 1217 if (!p) 1218 return; /* Presumably process exited. */ 1219 1220 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL); 1221 kfd_unref_process(p); 1222 } 1223 1224 void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, 1225 struct kfd_vm_fault_info *info) 1226 { 1227 struct kfd_event *ev; 1228 uint32_t id; 1229 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 1230 struct kfd_hsa_memory_exception_data memory_exception_data; 1231 int user_gpu_id; 1232 1233 if (!p) 1234 return; /* Presumably process exited. */ 1235 1236 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id); 1237 if (unlikely(user_gpu_id == -EINVAL)) { 1238 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id); 1239 return; 1240 } 1241 1242 memset(&memory_exception_data, 0, sizeof(memory_exception_data)); 1243 memory_exception_data.gpu_id = user_gpu_id; 1244 memory_exception_data.failure.imprecise = true; 1245 /* Set failure reason */ 1246 if (info) { 1247 memory_exception_data.va = (info->page_addr) << PAGE_SHIFT; 1248 memory_exception_data.failure.NotPresent = 1249 info->prot_valid ? 1 : 0; 1250 memory_exception_data.failure.NoExecute = 1251 info->prot_exec ? 1 : 0; 1252 memory_exception_data.failure.ReadOnly = 1253 info->prot_write ? 1 : 0; 1254 memory_exception_data.failure.imprecise = 0; 1255 } 1256 1257 rcu_read_lock(); 1258 1259 id = KFD_FIRST_NONSIGNAL_EVENT_ID; 1260 idr_for_each_entry_continue(&p->event_idr, ev, id) 1261 if (ev->type == KFD_EVENT_TYPE_MEMORY) { 1262 spin_lock(&ev->lock); 1263 ev->memory_exception_data = memory_exception_data; 1264 set_event(ev); 1265 spin_unlock(&ev->lock); 1266 } 1267 1268 rcu_read_unlock(); 1269 kfd_unref_process(p); 1270 } 1271 1272 void kfd_signal_reset_event(struct kfd_dev *dev) 1273 { 1274 struct kfd_hsa_hw_exception_data hw_exception_data; 1275 struct kfd_hsa_memory_exception_data memory_exception_data; 1276 struct kfd_process *p; 1277 struct kfd_event *ev; 1278 unsigned int temp; 1279 uint32_t id, idx; 1280 int reset_cause = atomic_read(&dev->sram_ecc_flag) ? 1281 KFD_HW_EXCEPTION_ECC : 1282 KFD_HW_EXCEPTION_GPU_HANG; 1283 1284 /* Whole gpu reset caused by GPU hang and memory is lost */ 1285 memset(&hw_exception_data, 0, sizeof(hw_exception_data)); 1286 hw_exception_data.memory_lost = 1; 1287 hw_exception_data.reset_cause = reset_cause; 1288 1289 memset(&memory_exception_data, 0, sizeof(memory_exception_data)); 1290 memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC; 1291 memory_exception_data.failure.imprecise = true; 1292 1293 idx = srcu_read_lock(&kfd_processes_srcu); 1294 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { 1295 int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id); 1296 1297 if (unlikely(user_gpu_id == -EINVAL)) { 1298 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id); 1299 continue; 1300 } 1301 1302 rcu_read_lock(); 1303 1304 id = KFD_FIRST_NONSIGNAL_EVENT_ID; 1305 idr_for_each_entry_continue(&p->event_idr, ev, id) { 1306 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) { 1307 spin_lock(&ev->lock); 1308 ev->hw_exception_data = hw_exception_data; 1309 ev->hw_exception_data.gpu_id = user_gpu_id; 1310 set_event(ev); 1311 spin_unlock(&ev->lock); 1312 } 1313 if (ev->type == KFD_EVENT_TYPE_MEMORY && 1314 reset_cause == KFD_HW_EXCEPTION_ECC) { 1315 spin_lock(&ev->lock); 1316 ev->memory_exception_data = memory_exception_data; 1317 ev->memory_exception_data.gpu_id = user_gpu_id; 1318 set_event(ev); 1319 spin_unlock(&ev->lock); 1320 } 1321 } 1322 1323 rcu_read_unlock(); 1324 } 1325 srcu_read_unlock(&kfd_processes_srcu, idx); 1326 } 1327 1328 void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid) 1329 { 1330 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 1331 struct kfd_hsa_memory_exception_data memory_exception_data; 1332 struct kfd_hsa_hw_exception_data hw_exception_data; 1333 struct kfd_event *ev; 1334 uint32_t id = KFD_FIRST_NONSIGNAL_EVENT_ID; 1335 int user_gpu_id; 1336 1337 if (!p) 1338 return; /* Presumably process exited. */ 1339 1340 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id); 1341 if (unlikely(user_gpu_id == -EINVAL)) { 1342 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id); 1343 return; 1344 } 1345 1346 memset(&hw_exception_data, 0, sizeof(hw_exception_data)); 1347 hw_exception_data.gpu_id = user_gpu_id; 1348 hw_exception_data.memory_lost = 1; 1349 hw_exception_data.reset_cause = KFD_HW_EXCEPTION_ECC; 1350 1351 memset(&memory_exception_data, 0, sizeof(memory_exception_data)); 1352 memory_exception_data.ErrorType = KFD_MEM_ERR_POISON_CONSUMED; 1353 memory_exception_data.gpu_id = user_gpu_id; 1354 memory_exception_data.failure.imprecise = true; 1355 1356 rcu_read_lock(); 1357 1358 idr_for_each_entry_continue(&p->event_idr, ev, id) { 1359 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) { 1360 spin_lock(&ev->lock); 1361 ev->hw_exception_data = hw_exception_data; 1362 set_event(ev); 1363 spin_unlock(&ev->lock); 1364 } 1365 1366 if (ev->type == KFD_EVENT_TYPE_MEMORY) { 1367 spin_lock(&ev->lock); 1368 ev->memory_exception_data = memory_exception_data; 1369 set_event(ev); 1370 spin_unlock(&ev->lock); 1371 } 1372 } 1373 1374 rcu_read_unlock(); 1375 1376 /* user application will handle SIGBUS signal */ 1377 send_sig(SIGBUS, p->lead_thread, 0); 1378 1379 kfd_unref_process(p); 1380 } 1381