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 <linux/device.h> 35 36 /* 37 * Wrapper around wait_queue_entry_t 38 */ 39 struct kfd_event_waiter { 40 wait_queue_entry_t wait; 41 struct kfd_event *event; /* Event to wait for */ 42 bool activated; /* Becomes true when event is signaled */ 43 bool event_age_enabled; /* set to true when last_event_age is non-zero */ 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_node *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 ev->event_age = 1; 435 } else { 436 kfree(ev); 437 } 438 439 mutex_unlock(&p->event_mutex); 440 441 return ret; 442 } 443 444 int kfd_criu_restore_event(struct file *devkfd, 445 struct kfd_process *p, 446 uint8_t __user *user_priv_ptr, 447 uint64_t *priv_data_offset, 448 uint64_t max_priv_data_size) 449 { 450 struct kfd_criu_event_priv_data *ev_priv; 451 struct kfd_event *ev = NULL; 452 int ret = 0; 453 454 ev_priv = kmalloc(sizeof(*ev_priv), GFP_KERNEL); 455 if (!ev_priv) 456 return -ENOMEM; 457 458 ev = kzalloc(sizeof(*ev), GFP_KERNEL); 459 if (!ev) { 460 ret = -ENOMEM; 461 goto exit; 462 } 463 464 if (*priv_data_offset + sizeof(*ev_priv) > max_priv_data_size) { 465 ret = -EINVAL; 466 goto exit; 467 } 468 469 ret = copy_from_user(ev_priv, user_priv_ptr + *priv_data_offset, sizeof(*ev_priv)); 470 if (ret) { 471 ret = -EFAULT; 472 goto exit; 473 } 474 *priv_data_offset += sizeof(*ev_priv); 475 476 if (ev_priv->user_handle) { 477 ret = kfd_kmap_event_page(p, ev_priv->user_handle); 478 if (ret) 479 goto exit; 480 } 481 482 ev->type = ev_priv->type; 483 ev->auto_reset = ev_priv->auto_reset; 484 ev->signaled = ev_priv->signaled; 485 486 spin_lock_init(&ev->lock); 487 init_waitqueue_head(&ev->wq); 488 489 mutex_lock(&p->event_mutex); 490 switch (ev->type) { 491 case KFD_EVENT_TYPE_SIGNAL: 492 case KFD_EVENT_TYPE_DEBUG: 493 ret = create_signal_event(devkfd, p, ev, &ev_priv->event_id); 494 break; 495 case KFD_EVENT_TYPE_MEMORY: 496 memcpy(&ev->memory_exception_data, 497 &ev_priv->memory_exception_data, 498 sizeof(struct kfd_hsa_memory_exception_data)); 499 500 ret = create_other_event(p, ev, &ev_priv->event_id); 501 break; 502 case KFD_EVENT_TYPE_HW_EXCEPTION: 503 memcpy(&ev->hw_exception_data, 504 &ev_priv->hw_exception_data, 505 sizeof(struct kfd_hsa_hw_exception_data)); 506 507 ret = create_other_event(p, ev, &ev_priv->event_id); 508 break; 509 } 510 mutex_unlock(&p->event_mutex); 511 512 exit: 513 if (ret) 514 kfree(ev); 515 516 kfree(ev_priv); 517 518 return ret; 519 } 520 521 int kfd_criu_checkpoint_events(struct kfd_process *p, 522 uint8_t __user *user_priv_data, 523 uint64_t *priv_data_offset) 524 { 525 struct kfd_criu_event_priv_data *ev_privs; 526 int i = 0; 527 int ret = 0; 528 struct kfd_event *ev; 529 uint32_t ev_id; 530 531 uint32_t num_events = kfd_get_num_events(p); 532 533 if (!num_events) 534 return 0; 535 536 ev_privs = kvzalloc(num_events * sizeof(*ev_privs), GFP_KERNEL); 537 if (!ev_privs) 538 return -ENOMEM; 539 540 541 idr_for_each_entry(&p->event_idr, ev, ev_id) { 542 struct kfd_criu_event_priv_data *ev_priv; 543 544 /* 545 * Currently, all events have same size of private_data, but the current ioctl's 546 * and CRIU plugin supports private_data of variable sizes 547 */ 548 ev_priv = &ev_privs[i]; 549 550 ev_priv->object_type = KFD_CRIU_OBJECT_TYPE_EVENT; 551 552 /* We store the user_handle with the first event */ 553 if (i == 0 && p->signal_page) 554 ev_priv->user_handle = p->signal_handle; 555 556 ev_priv->event_id = ev->event_id; 557 ev_priv->auto_reset = ev->auto_reset; 558 ev_priv->type = ev->type; 559 ev_priv->signaled = ev->signaled; 560 561 if (ev_priv->type == KFD_EVENT_TYPE_MEMORY) 562 memcpy(&ev_priv->memory_exception_data, 563 &ev->memory_exception_data, 564 sizeof(struct kfd_hsa_memory_exception_data)); 565 else if (ev_priv->type == KFD_EVENT_TYPE_HW_EXCEPTION) 566 memcpy(&ev_priv->hw_exception_data, 567 &ev->hw_exception_data, 568 sizeof(struct kfd_hsa_hw_exception_data)); 569 570 pr_debug("Checkpointed event[%d] id = 0x%08x auto_reset = %x type = %x signaled = %x\n", 571 i, 572 ev_priv->event_id, 573 ev_priv->auto_reset, 574 ev_priv->type, 575 ev_priv->signaled); 576 i++; 577 } 578 579 ret = copy_to_user(user_priv_data + *priv_data_offset, 580 ev_privs, num_events * sizeof(*ev_privs)); 581 if (ret) { 582 pr_err("Failed to copy events priv to user\n"); 583 ret = -EFAULT; 584 } 585 586 *priv_data_offset += num_events * sizeof(*ev_privs); 587 588 kvfree(ev_privs); 589 return ret; 590 } 591 592 int kfd_get_num_events(struct kfd_process *p) 593 { 594 struct kfd_event *ev; 595 uint32_t id; 596 u32 num_events = 0; 597 598 idr_for_each_entry(&p->event_idr, ev, id) 599 num_events++; 600 601 return num_events; 602 } 603 604 /* Assumes that p is current. */ 605 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id) 606 { 607 struct kfd_event *ev; 608 int ret = 0; 609 610 mutex_lock(&p->event_mutex); 611 612 ev = lookup_event_by_id(p, event_id); 613 614 if (ev) 615 destroy_event(p, ev); 616 else 617 ret = -EINVAL; 618 619 mutex_unlock(&p->event_mutex); 620 return ret; 621 } 622 623 static void set_event(struct kfd_event *ev) 624 { 625 struct kfd_event_waiter *waiter; 626 627 /* Auto reset if the list is non-empty and we're waking 628 * someone. waitqueue_active is safe here because we're 629 * protected by the ev->lock, which is also held when 630 * updating the wait queues in kfd_wait_on_events. 631 */ 632 ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq); 633 if (!(++ev->event_age)) { 634 /* Never wrap back to reserved/default event age 0/1 */ 635 ev->event_age = 2; 636 WARN_ONCE(1, "event_age wrap back!"); 637 } 638 639 list_for_each_entry(waiter, &ev->wq.head, wait.entry) 640 WRITE_ONCE(waiter->activated, true); 641 642 wake_up_all(&ev->wq); 643 } 644 645 /* Assumes that p is current. */ 646 int kfd_set_event(struct kfd_process *p, uint32_t event_id) 647 { 648 int ret = 0; 649 struct kfd_event *ev; 650 651 rcu_read_lock(); 652 653 ev = lookup_event_by_id(p, event_id); 654 if (!ev) { 655 ret = -EINVAL; 656 goto unlock_rcu; 657 } 658 spin_lock(&ev->lock); 659 660 if (event_can_be_cpu_signaled(ev)) 661 set_event(ev); 662 else 663 ret = -EINVAL; 664 665 spin_unlock(&ev->lock); 666 unlock_rcu: 667 rcu_read_unlock(); 668 return ret; 669 } 670 671 static void reset_event(struct kfd_event *ev) 672 { 673 ev->signaled = false; 674 } 675 676 /* Assumes that p is current. */ 677 int kfd_reset_event(struct kfd_process *p, uint32_t event_id) 678 { 679 int ret = 0; 680 struct kfd_event *ev; 681 682 rcu_read_lock(); 683 684 ev = lookup_event_by_id(p, event_id); 685 if (!ev) { 686 ret = -EINVAL; 687 goto unlock_rcu; 688 } 689 spin_lock(&ev->lock); 690 691 if (event_can_be_cpu_signaled(ev)) 692 reset_event(ev); 693 else 694 ret = -EINVAL; 695 696 spin_unlock(&ev->lock); 697 unlock_rcu: 698 rcu_read_unlock(); 699 return ret; 700 701 } 702 703 static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev) 704 { 705 WRITE_ONCE(page_slots(p->signal_page)[ev->event_id], UNSIGNALED_EVENT_SLOT); 706 } 707 708 static void set_event_from_interrupt(struct kfd_process *p, 709 struct kfd_event *ev) 710 { 711 if (ev && event_can_be_gpu_signaled(ev)) { 712 acknowledge_signal(p, ev); 713 spin_lock(&ev->lock); 714 set_event(ev); 715 spin_unlock(&ev->lock); 716 } 717 } 718 719 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, 720 uint32_t valid_id_bits) 721 { 722 struct kfd_event *ev = NULL; 723 724 /* 725 * Because we are called from arbitrary context (workqueue) as opposed 726 * to process context, kfd_process could attempt to exit while we are 727 * running so the lookup function increments the process ref count. 728 */ 729 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 730 731 if (!p) 732 return; /* Presumably process exited. */ 733 734 rcu_read_lock(); 735 736 if (valid_id_bits) 737 ev = lookup_signaled_event_by_partial_id(p, partial_id, 738 valid_id_bits); 739 if (ev) { 740 set_event_from_interrupt(p, ev); 741 } else if (p->signal_page) { 742 /* 743 * Partial ID lookup failed. Assume that the event ID 744 * in the interrupt payload was invalid and do an 745 * exhaustive search of signaled events. 746 */ 747 uint64_t *slots = page_slots(p->signal_page); 748 uint32_t id; 749 750 if (valid_id_bits) 751 pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n", 752 partial_id, valid_id_bits); 753 754 if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) { 755 /* With relatively few events, it's faster to 756 * iterate over the event IDR 757 */ 758 idr_for_each_entry(&p->event_idr, ev, id) { 759 if (id >= KFD_SIGNAL_EVENT_LIMIT) 760 break; 761 762 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) 763 set_event_from_interrupt(p, ev); 764 } 765 } else { 766 /* With relatively many events, it's faster to 767 * iterate over the signal slots and lookup 768 * only signaled events from the IDR. 769 */ 770 for (id = 1; id < KFD_SIGNAL_EVENT_LIMIT; id++) 771 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) { 772 ev = lookup_event_by_id(p, id); 773 set_event_from_interrupt(p, ev); 774 } 775 } 776 } 777 778 rcu_read_unlock(); 779 kfd_unref_process(p); 780 } 781 782 static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events) 783 { 784 struct kfd_event_waiter *event_waiters; 785 uint32_t i; 786 787 event_waiters = kcalloc(num_events, sizeof(struct kfd_event_waiter), 788 GFP_KERNEL); 789 if (!event_waiters) 790 return NULL; 791 792 for (i = 0; i < num_events; i++) 793 init_wait(&event_waiters[i].wait); 794 795 return event_waiters; 796 } 797 798 static int init_event_waiter(struct kfd_process *p, 799 struct kfd_event_waiter *waiter, 800 struct kfd_event_data *event_data) 801 { 802 struct kfd_event *ev = lookup_event_by_id(p, event_data->event_id); 803 804 if (!ev) 805 return -EINVAL; 806 807 spin_lock(&ev->lock); 808 waiter->event = ev; 809 waiter->activated = ev->signaled; 810 ev->signaled = ev->signaled && !ev->auto_reset; 811 812 /* last_event_age = 0 reserved for backward compatible */ 813 if (waiter->event->type == KFD_EVENT_TYPE_SIGNAL && 814 event_data->signal_event_data.last_event_age) { 815 waiter->event_age_enabled = true; 816 if (ev->event_age != event_data->signal_event_data.last_event_age) 817 waiter->activated = true; 818 } 819 820 if (!waiter->activated) 821 add_wait_queue(&ev->wq, &waiter->wait); 822 spin_unlock(&ev->lock); 823 824 return 0; 825 } 826 827 /* test_event_condition - Test condition of events being waited for 828 * @all: Return completion only if all events have signaled 829 * @num_events: Number of events to wait for 830 * @event_waiters: Array of event waiters, one per event 831 * 832 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have 833 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all) 834 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of 835 * the events have been destroyed. 836 */ 837 static uint32_t test_event_condition(bool all, uint32_t num_events, 838 struct kfd_event_waiter *event_waiters) 839 { 840 uint32_t i; 841 uint32_t activated_count = 0; 842 843 for (i = 0; i < num_events; i++) { 844 if (!READ_ONCE(event_waiters[i].event)) 845 return KFD_IOC_WAIT_RESULT_FAIL; 846 847 if (READ_ONCE(event_waiters[i].activated)) { 848 if (!all) 849 return KFD_IOC_WAIT_RESULT_COMPLETE; 850 851 activated_count++; 852 } 853 } 854 855 return activated_count == num_events ? 856 KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT; 857 } 858 859 /* 860 * Copy event specific data, if defined. 861 * Currently only memory exception events have additional data to copy to user 862 */ 863 static int copy_signaled_event_data(uint32_t num_events, 864 struct kfd_event_waiter *event_waiters, 865 struct kfd_event_data __user *data) 866 { 867 void *src; 868 void __user *dst; 869 struct kfd_event_waiter *waiter; 870 struct kfd_event *event; 871 uint32_t i, size = 0; 872 873 for (i = 0; i < num_events; i++) { 874 waiter = &event_waiters[i]; 875 event = waiter->event; 876 if (!event) 877 return -EINVAL; /* event was destroyed */ 878 if (waiter->activated) { 879 if (event->type == KFD_EVENT_TYPE_MEMORY) { 880 dst = &data[i].memory_exception_data; 881 src = &event->memory_exception_data; 882 size = sizeof(struct kfd_hsa_memory_exception_data); 883 } else if (event->type == KFD_EVENT_TYPE_SIGNAL && 884 waiter->event_age_enabled) { 885 dst = &data[i].signal_event_data.last_event_age; 886 src = &event->event_age; 887 size = sizeof(u64); 888 } 889 if (size && copy_to_user(dst, src, size)) 890 return -EFAULT; 891 } 892 } 893 894 return 0; 895 } 896 897 static long user_timeout_to_jiffies(uint32_t user_timeout_ms) 898 { 899 if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE) 900 return 0; 901 902 if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE) 903 return MAX_SCHEDULE_TIMEOUT; 904 905 /* 906 * msecs_to_jiffies interprets all values above 2^31-1 as infinite, 907 * but we consider them finite. 908 * This hack is wrong, but nobody is likely to notice. 909 */ 910 user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF); 911 912 return msecs_to_jiffies(user_timeout_ms) + 1; 913 } 914 915 static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters, 916 bool undo_auto_reset) 917 { 918 uint32_t i; 919 920 for (i = 0; i < num_events; i++) 921 if (waiters[i].event) { 922 spin_lock(&waiters[i].event->lock); 923 remove_wait_queue(&waiters[i].event->wq, 924 &waiters[i].wait); 925 if (undo_auto_reset && waiters[i].activated && 926 waiters[i].event && waiters[i].event->auto_reset) 927 set_event(waiters[i].event); 928 spin_unlock(&waiters[i].event->lock); 929 } 930 931 kfree(waiters); 932 } 933 934 int kfd_wait_on_events(struct kfd_process *p, 935 uint32_t num_events, void __user *data, 936 bool all, uint32_t *user_timeout_ms, 937 uint32_t *wait_result) 938 { 939 struct kfd_event_data __user *events = 940 (struct kfd_event_data __user *) data; 941 uint32_t i; 942 int ret = 0; 943 944 struct kfd_event_waiter *event_waiters = NULL; 945 long timeout = user_timeout_to_jiffies(*user_timeout_ms); 946 947 event_waiters = alloc_event_waiters(num_events); 948 if (!event_waiters) { 949 ret = -ENOMEM; 950 goto out; 951 } 952 953 /* Use p->event_mutex here to protect against concurrent creation and 954 * destruction of events while we initialize event_waiters. 955 */ 956 mutex_lock(&p->event_mutex); 957 958 for (i = 0; i < num_events; i++) { 959 struct kfd_event_data event_data; 960 961 if (copy_from_user(&event_data, &events[i], 962 sizeof(struct kfd_event_data))) { 963 ret = -EFAULT; 964 goto out_unlock; 965 } 966 967 ret = init_event_waiter(p, &event_waiters[i], &event_data); 968 if (ret) 969 goto out_unlock; 970 } 971 972 /* Check condition once. */ 973 *wait_result = test_event_condition(all, num_events, event_waiters); 974 if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) { 975 ret = copy_signaled_event_data(num_events, 976 event_waiters, events); 977 goto out_unlock; 978 } else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) { 979 /* This should not happen. Events shouldn't be 980 * destroyed while we're holding the event_mutex 981 */ 982 goto out_unlock; 983 } 984 985 mutex_unlock(&p->event_mutex); 986 987 while (true) { 988 if (fatal_signal_pending(current)) { 989 ret = -EINTR; 990 break; 991 } 992 993 if (signal_pending(current)) { 994 ret = -ERESTARTSYS; 995 if (*user_timeout_ms != KFD_EVENT_TIMEOUT_IMMEDIATE && 996 *user_timeout_ms != KFD_EVENT_TIMEOUT_INFINITE) 997 *user_timeout_ms = jiffies_to_msecs( 998 max(0l, timeout-1)); 999 break; 1000 } 1001 1002 /* Set task state to interruptible sleep before 1003 * checking wake-up conditions. A concurrent wake-up 1004 * will put the task back into runnable state. In that 1005 * case schedule_timeout will not put the task to 1006 * sleep and we'll get a chance to re-check the 1007 * updated conditions almost immediately. Otherwise, 1008 * this race condition would lead to a soft hang or a 1009 * very long sleep. 1010 */ 1011 set_current_state(TASK_INTERRUPTIBLE); 1012 1013 *wait_result = test_event_condition(all, num_events, 1014 event_waiters); 1015 if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT) 1016 break; 1017 1018 if (timeout <= 0) 1019 break; 1020 1021 timeout = schedule_timeout(timeout); 1022 } 1023 __set_current_state(TASK_RUNNING); 1024 1025 mutex_lock(&p->event_mutex); 1026 /* copy_signaled_event_data may sleep. So this has to happen 1027 * after the task state is set back to RUNNING. 1028 * 1029 * The event may also have been destroyed after signaling. So 1030 * copy_signaled_event_data also must confirm that the event 1031 * still exists. Therefore this must be under the p->event_mutex 1032 * which is also held when events are destroyed. 1033 */ 1034 if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) 1035 ret = copy_signaled_event_data(num_events, 1036 event_waiters, events); 1037 1038 out_unlock: 1039 free_waiters(num_events, event_waiters, ret == -ERESTARTSYS); 1040 mutex_unlock(&p->event_mutex); 1041 out: 1042 if (ret) 1043 *wait_result = KFD_IOC_WAIT_RESULT_FAIL; 1044 else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL) 1045 ret = -EIO; 1046 1047 return ret; 1048 } 1049 1050 int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma) 1051 { 1052 unsigned long pfn; 1053 struct kfd_signal_page *page; 1054 int ret; 1055 1056 /* check required size doesn't exceed the allocated size */ 1057 if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) < 1058 get_order(vma->vm_end - vma->vm_start)) { 1059 pr_err("Event page mmap requested illegal size\n"); 1060 return -EINVAL; 1061 } 1062 1063 page = p->signal_page; 1064 if (!page) { 1065 /* Probably KFD bug, but mmap is user-accessible. */ 1066 pr_debug("Signal page could not be found\n"); 1067 return -EINVAL; 1068 } 1069 1070 pfn = __pa(page->kernel_address); 1071 pfn >>= PAGE_SHIFT; 1072 1073 vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE 1074 | VM_DONTDUMP | VM_PFNMAP); 1075 1076 pr_debug("Mapping signal page\n"); 1077 pr_debug(" start user address == 0x%08lx\n", vma->vm_start); 1078 pr_debug(" end user address == 0x%08lx\n", vma->vm_end); 1079 pr_debug(" pfn == 0x%016lX\n", pfn); 1080 pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags); 1081 pr_debug(" size == 0x%08lX\n", 1082 vma->vm_end - vma->vm_start); 1083 1084 page->user_address = (uint64_t __user *)vma->vm_start; 1085 1086 /* mapping the page to user process */ 1087 ret = remap_pfn_range(vma, vma->vm_start, pfn, 1088 vma->vm_end - vma->vm_start, vma->vm_page_prot); 1089 if (!ret) 1090 p->signal_mapped_size = vma->vm_end - vma->vm_start; 1091 1092 return ret; 1093 } 1094 1095 /* 1096 * Assumes that p is not going away. 1097 */ 1098 static void lookup_events_by_type_and_signal(struct kfd_process *p, 1099 int type, void *event_data) 1100 { 1101 struct kfd_hsa_memory_exception_data *ev_data; 1102 struct kfd_event *ev; 1103 uint32_t id; 1104 bool send_signal = true; 1105 1106 ev_data = (struct kfd_hsa_memory_exception_data *) event_data; 1107 1108 rcu_read_lock(); 1109 1110 id = KFD_FIRST_NONSIGNAL_EVENT_ID; 1111 idr_for_each_entry_continue(&p->event_idr, ev, id) 1112 if (ev->type == type) { 1113 send_signal = false; 1114 dev_dbg(kfd_device, 1115 "Event found: id %X type %d", 1116 ev->event_id, ev->type); 1117 spin_lock(&ev->lock); 1118 set_event(ev); 1119 if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data) 1120 ev->memory_exception_data = *ev_data; 1121 spin_unlock(&ev->lock); 1122 } 1123 1124 if (type == KFD_EVENT_TYPE_MEMORY) { 1125 dev_warn(kfd_device, 1126 "Sending SIGSEGV to process %d (pasid 0x%x)", 1127 p->lead_thread->pid, p->pasid); 1128 send_sig(SIGSEGV, p->lead_thread, 0); 1129 } 1130 1131 /* Send SIGTERM no event of type "type" has been found*/ 1132 if (send_signal) { 1133 if (send_sigterm) { 1134 dev_warn(kfd_device, 1135 "Sending SIGTERM to process %d (pasid 0x%x)", 1136 p->lead_thread->pid, p->pasid); 1137 send_sig(SIGTERM, p->lead_thread, 0); 1138 } else { 1139 dev_err(kfd_device, 1140 "Process %d (pasid 0x%x) got unhandled exception", 1141 p->lead_thread->pid, p->pasid); 1142 } 1143 } 1144 1145 rcu_read_unlock(); 1146 } 1147 1148 void kfd_signal_hw_exception_event(u32 pasid) 1149 { 1150 /* 1151 * Because we are called from arbitrary context (workqueue) as opposed 1152 * to process context, kfd_process could attempt to exit while we are 1153 * running so the lookup function increments the process ref count. 1154 */ 1155 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 1156 1157 if (!p) 1158 return; /* Presumably process exited. */ 1159 1160 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL); 1161 kfd_unref_process(p); 1162 } 1163 1164 void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid, 1165 struct kfd_vm_fault_info *info, 1166 struct kfd_hsa_memory_exception_data *data) 1167 { 1168 struct kfd_event *ev; 1169 uint32_t id; 1170 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 1171 struct kfd_hsa_memory_exception_data memory_exception_data; 1172 int user_gpu_id; 1173 1174 if (!p) 1175 return; /* Presumably process exited. */ 1176 1177 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id); 1178 if (unlikely(user_gpu_id == -EINVAL)) { 1179 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id); 1180 return; 1181 } 1182 1183 /* SoC15 chips and onwards will pass in data from now on. */ 1184 if (!data) { 1185 memset(&memory_exception_data, 0, sizeof(memory_exception_data)); 1186 memory_exception_data.gpu_id = user_gpu_id; 1187 memory_exception_data.failure.imprecise = true; 1188 1189 /* Set failure reason */ 1190 if (info) { 1191 memory_exception_data.va = (info->page_addr) << 1192 PAGE_SHIFT; 1193 memory_exception_data.failure.NotPresent = 1194 info->prot_valid ? 1 : 0; 1195 memory_exception_data.failure.NoExecute = 1196 info->prot_exec ? 1 : 0; 1197 memory_exception_data.failure.ReadOnly = 1198 info->prot_write ? 1 : 0; 1199 memory_exception_data.failure.imprecise = 0; 1200 } 1201 } 1202 1203 rcu_read_lock(); 1204 1205 id = KFD_FIRST_NONSIGNAL_EVENT_ID; 1206 idr_for_each_entry_continue(&p->event_idr, ev, id) 1207 if (ev->type == KFD_EVENT_TYPE_MEMORY) { 1208 spin_lock(&ev->lock); 1209 ev->memory_exception_data = data ? *data : 1210 memory_exception_data; 1211 set_event(ev); 1212 spin_unlock(&ev->lock); 1213 } 1214 1215 rcu_read_unlock(); 1216 kfd_unref_process(p); 1217 } 1218 1219 void kfd_signal_reset_event(struct kfd_node *dev) 1220 { 1221 struct kfd_hsa_hw_exception_data hw_exception_data; 1222 struct kfd_hsa_memory_exception_data memory_exception_data; 1223 struct kfd_process *p; 1224 struct kfd_event *ev; 1225 unsigned int temp; 1226 uint32_t id, idx; 1227 int reset_cause = atomic_read(&dev->sram_ecc_flag) ? 1228 KFD_HW_EXCEPTION_ECC : 1229 KFD_HW_EXCEPTION_GPU_HANG; 1230 1231 /* Whole gpu reset caused by GPU hang and memory is lost */ 1232 memset(&hw_exception_data, 0, sizeof(hw_exception_data)); 1233 hw_exception_data.memory_lost = 1; 1234 hw_exception_data.reset_cause = reset_cause; 1235 1236 memset(&memory_exception_data, 0, sizeof(memory_exception_data)); 1237 memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC; 1238 memory_exception_data.failure.imprecise = true; 1239 1240 idx = srcu_read_lock(&kfd_processes_srcu); 1241 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { 1242 int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id); 1243 1244 if (unlikely(user_gpu_id == -EINVAL)) { 1245 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id); 1246 continue; 1247 } 1248 1249 rcu_read_lock(); 1250 1251 id = KFD_FIRST_NONSIGNAL_EVENT_ID; 1252 idr_for_each_entry_continue(&p->event_idr, ev, id) { 1253 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) { 1254 spin_lock(&ev->lock); 1255 ev->hw_exception_data = hw_exception_data; 1256 ev->hw_exception_data.gpu_id = user_gpu_id; 1257 set_event(ev); 1258 spin_unlock(&ev->lock); 1259 } 1260 if (ev->type == KFD_EVENT_TYPE_MEMORY && 1261 reset_cause == KFD_HW_EXCEPTION_ECC) { 1262 spin_lock(&ev->lock); 1263 ev->memory_exception_data = memory_exception_data; 1264 ev->memory_exception_data.gpu_id = user_gpu_id; 1265 set_event(ev); 1266 spin_unlock(&ev->lock); 1267 } 1268 } 1269 1270 rcu_read_unlock(); 1271 } 1272 srcu_read_unlock(&kfd_processes_srcu, idx); 1273 } 1274 1275 void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid) 1276 { 1277 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 1278 struct kfd_hsa_memory_exception_data memory_exception_data; 1279 struct kfd_hsa_hw_exception_data hw_exception_data; 1280 struct kfd_event *ev; 1281 uint32_t id = KFD_FIRST_NONSIGNAL_EVENT_ID; 1282 int user_gpu_id; 1283 1284 if (!p) 1285 return; /* Presumably process exited. */ 1286 1287 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id); 1288 if (unlikely(user_gpu_id == -EINVAL)) { 1289 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id); 1290 return; 1291 } 1292 1293 memset(&hw_exception_data, 0, sizeof(hw_exception_data)); 1294 hw_exception_data.gpu_id = user_gpu_id; 1295 hw_exception_data.memory_lost = 1; 1296 hw_exception_data.reset_cause = KFD_HW_EXCEPTION_ECC; 1297 1298 memset(&memory_exception_data, 0, sizeof(memory_exception_data)); 1299 memory_exception_data.ErrorType = KFD_MEM_ERR_POISON_CONSUMED; 1300 memory_exception_data.gpu_id = user_gpu_id; 1301 memory_exception_data.failure.imprecise = true; 1302 1303 rcu_read_lock(); 1304 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 set_event(ev); 1310 spin_unlock(&ev->lock); 1311 } 1312 1313 if (ev->type == KFD_EVENT_TYPE_MEMORY) { 1314 spin_lock(&ev->lock); 1315 ev->memory_exception_data = memory_exception_data; 1316 set_event(ev); 1317 spin_unlock(&ev->lock); 1318 } 1319 } 1320 1321 rcu_read_unlock(); 1322 1323 /* user application will handle SIGBUS signal */ 1324 send_sig(SIGBUS, p->lead_thread, 0); 1325 1326 kfd_unref_process(p); 1327 } 1328