1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Basic worker thread pool for io_uring 4 * 5 * Copyright (C) 2019 Jens Axboe 6 * 7 */ 8 #include <linux/kernel.h> 9 #include <linux/init.h> 10 #include <linux/errno.h> 11 #include <linux/sched/signal.h> 12 #include <linux/percpu.h> 13 #include <linux/slab.h> 14 #include <linux/rculist_nulls.h> 15 #include <linux/cpu.h> 16 #include <linux/task_work.h> 17 #include <linux/audit.h> 18 #include <linux/mmu_context.h> 19 #include <uapi/linux/io_uring.h> 20 21 #include "io-wq.h" 22 #include "slist.h" 23 #include "io_uring.h" 24 25 #define WORKER_IDLE_TIMEOUT (5 * HZ) 26 27 enum { 28 IO_WORKER_F_UP = 1, /* up and active */ 29 IO_WORKER_F_RUNNING = 2, /* account as running */ 30 IO_WORKER_F_FREE = 4, /* worker on free list */ 31 IO_WORKER_F_BOUND = 8, /* is doing bounded work */ 32 }; 33 34 enum { 35 IO_WQ_BIT_EXIT = 0, /* wq exiting */ 36 }; 37 38 enum { 39 IO_ACCT_STALLED_BIT = 0, /* stalled on hash */ 40 }; 41 42 /* 43 * One for each thread in a wq pool 44 */ 45 struct io_worker { 46 refcount_t ref; 47 unsigned flags; 48 struct hlist_nulls_node nulls_node; 49 struct list_head all_list; 50 struct task_struct *task; 51 struct io_wq *wq; 52 53 struct io_wq_work *cur_work; 54 struct io_wq_work *next_work; 55 raw_spinlock_t lock; 56 57 struct completion ref_done; 58 59 unsigned long create_state; 60 struct callback_head create_work; 61 int create_index; 62 63 union { 64 struct rcu_head rcu; 65 struct work_struct work; 66 }; 67 }; 68 69 #if BITS_PER_LONG == 64 70 #define IO_WQ_HASH_ORDER 6 71 #else 72 #define IO_WQ_HASH_ORDER 5 73 #endif 74 75 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER) 76 77 struct io_wq_acct { 78 unsigned nr_workers; 79 unsigned max_workers; 80 int index; 81 atomic_t nr_running; 82 raw_spinlock_t lock; 83 struct io_wq_work_list work_list; 84 unsigned long flags; 85 }; 86 87 enum { 88 IO_WQ_ACCT_BOUND, 89 IO_WQ_ACCT_UNBOUND, 90 IO_WQ_ACCT_NR, 91 }; 92 93 /* 94 * Per io_wq state 95 */ 96 struct io_wq { 97 unsigned long state; 98 99 free_work_fn *free_work; 100 io_wq_work_fn *do_work; 101 102 struct io_wq_hash *hash; 103 104 atomic_t worker_refs; 105 struct completion worker_done; 106 107 struct hlist_node cpuhp_node; 108 109 struct task_struct *task; 110 111 struct io_wq_acct acct[IO_WQ_ACCT_NR]; 112 113 /* lock protects access to elements below */ 114 raw_spinlock_t lock; 115 116 struct hlist_nulls_head free_list; 117 struct list_head all_list; 118 119 struct wait_queue_entry wait; 120 121 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS]; 122 123 cpumask_var_t cpu_mask; 124 }; 125 126 static enum cpuhp_state io_wq_online; 127 128 struct io_cb_cancel_data { 129 work_cancel_fn *fn; 130 void *data; 131 int nr_running; 132 int nr_pending; 133 bool cancel_all; 134 }; 135 136 static bool create_io_worker(struct io_wq *wq, int index); 137 static void io_wq_dec_running(struct io_worker *worker); 138 static bool io_acct_cancel_pending_work(struct io_wq *wq, 139 struct io_wq_acct *acct, 140 struct io_cb_cancel_data *match); 141 static void create_worker_cb(struct callback_head *cb); 142 static void io_wq_cancel_tw_create(struct io_wq *wq); 143 144 static bool io_worker_get(struct io_worker *worker) 145 { 146 return refcount_inc_not_zero(&worker->ref); 147 } 148 149 static void io_worker_release(struct io_worker *worker) 150 { 151 if (refcount_dec_and_test(&worker->ref)) 152 complete(&worker->ref_done); 153 } 154 155 static inline struct io_wq_acct *io_get_acct(struct io_wq *wq, bool bound) 156 { 157 return &wq->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND]; 158 } 159 160 static inline struct io_wq_acct *io_work_get_acct(struct io_wq *wq, 161 struct io_wq_work *work) 162 { 163 return io_get_acct(wq, !(work->flags & IO_WQ_WORK_UNBOUND)); 164 } 165 166 static inline struct io_wq_acct *io_wq_get_acct(struct io_worker *worker) 167 { 168 return io_get_acct(worker->wq, worker->flags & IO_WORKER_F_BOUND); 169 } 170 171 static void io_worker_ref_put(struct io_wq *wq) 172 { 173 if (atomic_dec_and_test(&wq->worker_refs)) 174 complete(&wq->worker_done); 175 } 176 177 static void io_worker_cancel_cb(struct io_worker *worker) 178 { 179 struct io_wq_acct *acct = io_wq_get_acct(worker); 180 struct io_wq *wq = worker->wq; 181 182 atomic_dec(&acct->nr_running); 183 raw_spin_lock(&wq->lock); 184 acct->nr_workers--; 185 raw_spin_unlock(&wq->lock); 186 io_worker_ref_put(wq); 187 clear_bit_unlock(0, &worker->create_state); 188 io_worker_release(worker); 189 } 190 191 static bool io_task_worker_match(struct callback_head *cb, void *data) 192 { 193 struct io_worker *worker; 194 195 if (cb->func != create_worker_cb) 196 return false; 197 worker = container_of(cb, struct io_worker, create_work); 198 return worker == data; 199 } 200 201 static void io_worker_exit(struct io_worker *worker) 202 { 203 struct io_wq *wq = worker->wq; 204 205 while (1) { 206 struct callback_head *cb = task_work_cancel_match(wq->task, 207 io_task_worker_match, worker); 208 209 if (!cb) 210 break; 211 io_worker_cancel_cb(worker); 212 } 213 214 io_worker_release(worker); 215 wait_for_completion(&worker->ref_done); 216 217 raw_spin_lock(&wq->lock); 218 if (worker->flags & IO_WORKER_F_FREE) 219 hlist_nulls_del_rcu(&worker->nulls_node); 220 list_del_rcu(&worker->all_list); 221 raw_spin_unlock(&wq->lock); 222 io_wq_dec_running(worker); 223 /* 224 * this worker is a goner, clear ->worker_private to avoid any 225 * inc/dec running calls that could happen as part of exit from 226 * touching 'worker'. 227 */ 228 current->worker_private = NULL; 229 230 kfree_rcu(worker, rcu); 231 io_worker_ref_put(wq); 232 do_exit(0); 233 } 234 235 static inline bool io_acct_run_queue(struct io_wq_acct *acct) 236 { 237 bool ret = false; 238 239 raw_spin_lock(&acct->lock); 240 if (!wq_list_empty(&acct->work_list) && 241 !test_bit(IO_ACCT_STALLED_BIT, &acct->flags)) 242 ret = true; 243 raw_spin_unlock(&acct->lock); 244 245 return ret; 246 } 247 248 /* 249 * Check head of free list for an available worker. If one isn't available, 250 * caller must create one. 251 */ 252 static bool io_wq_activate_free_worker(struct io_wq *wq, 253 struct io_wq_acct *acct) 254 __must_hold(RCU) 255 { 256 struct hlist_nulls_node *n; 257 struct io_worker *worker; 258 259 /* 260 * Iterate free_list and see if we can find an idle worker to 261 * activate. If a given worker is on the free_list but in the process 262 * of exiting, keep trying. 263 */ 264 hlist_nulls_for_each_entry_rcu(worker, n, &wq->free_list, nulls_node) { 265 if (!io_worker_get(worker)) 266 continue; 267 if (io_wq_get_acct(worker) != acct) { 268 io_worker_release(worker); 269 continue; 270 } 271 if (wake_up_process(worker->task)) { 272 io_worker_release(worker); 273 return true; 274 } 275 io_worker_release(worker); 276 } 277 278 return false; 279 } 280 281 /* 282 * We need a worker. If we find a free one, we're good. If not, and we're 283 * below the max number of workers, create one. 284 */ 285 static bool io_wq_create_worker(struct io_wq *wq, struct io_wq_acct *acct) 286 { 287 /* 288 * Most likely an attempt to queue unbounded work on an io_wq that 289 * wasn't setup with any unbounded workers. 290 */ 291 if (unlikely(!acct->max_workers)) 292 pr_warn_once("io-wq is not configured for unbound workers"); 293 294 raw_spin_lock(&wq->lock); 295 if (acct->nr_workers >= acct->max_workers) { 296 raw_spin_unlock(&wq->lock); 297 return true; 298 } 299 acct->nr_workers++; 300 raw_spin_unlock(&wq->lock); 301 atomic_inc(&acct->nr_running); 302 atomic_inc(&wq->worker_refs); 303 return create_io_worker(wq, acct->index); 304 } 305 306 static void io_wq_inc_running(struct io_worker *worker) 307 { 308 struct io_wq_acct *acct = io_wq_get_acct(worker); 309 310 atomic_inc(&acct->nr_running); 311 } 312 313 static void create_worker_cb(struct callback_head *cb) 314 { 315 struct io_worker *worker; 316 struct io_wq *wq; 317 318 struct io_wq_acct *acct; 319 bool do_create = false; 320 321 worker = container_of(cb, struct io_worker, create_work); 322 wq = worker->wq; 323 acct = &wq->acct[worker->create_index]; 324 raw_spin_lock(&wq->lock); 325 326 if (acct->nr_workers < acct->max_workers) { 327 acct->nr_workers++; 328 do_create = true; 329 } 330 raw_spin_unlock(&wq->lock); 331 if (do_create) { 332 create_io_worker(wq, worker->create_index); 333 } else { 334 atomic_dec(&acct->nr_running); 335 io_worker_ref_put(wq); 336 } 337 clear_bit_unlock(0, &worker->create_state); 338 io_worker_release(worker); 339 } 340 341 static bool io_queue_worker_create(struct io_worker *worker, 342 struct io_wq_acct *acct, 343 task_work_func_t func) 344 { 345 struct io_wq *wq = worker->wq; 346 347 /* raced with exit, just ignore create call */ 348 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) 349 goto fail; 350 if (!io_worker_get(worker)) 351 goto fail; 352 /* 353 * create_state manages ownership of create_work/index. We should 354 * only need one entry per worker, as the worker going to sleep 355 * will trigger the condition, and waking will clear it once it 356 * runs the task_work. 357 */ 358 if (test_bit(0, &worker->create_state) || 359 test_and_set_bit_lock(0, &worker->create_state)) 360 goto fail_release; 361 362 atomic_inc(&wq->worker_refs); 363 init_task_work(&worker->create_work, func); 364 worker->create_index = acct->index; 365 if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) { 366 /* 367 * EXIT may have been set after checking it above, check after 368 * adding the task_work and remove any creation item if it is 369 * now set. wq exit does that too, but we can have added this 370 * work item after we canceled in io_wq_exit_workers(). 371 */ 372 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) 373 io_wq_cancel_tw_create(wq); 374 io_worker_ref_put(wq); 375 return true; 376 } 377 io_worker_ref_put(wq); 378 clear_bit_unlock(0, &worker->create_state); 379 fail_release: 380 io_worker_release(worker); 381 fail: 382 atomic_dec(&acct->nr_running); 383 io_worker_ref_put(wq); 384 return false; 385 } 386 387 static void io_wq_dec_running(struct io_worker *worker) 388 { 389 struct io_wq_acct *acct = io_wq_get_acct(worker); 390 struct io_wq *wq = worker->wq; 391 392 if (!(worker->flags & IO_WORKER_F_UP)) 393 return; 394 395 if (!atomic_dec_and_test(&acct->nr_running)) 396 return; 397 if (!io_acct_run_queue(acct)) 398 return; 399 400 atomic_inc(&acct->nr_running); 401 atomic_inc(&wq->worker_refs); 402 io_queue_worker_create(worker, acct, create_worker_cb); 403 } 404 405 /* 406 * Worker will start processing some work. Move it to the busy list, if 407 * it's currently on the freelist 408 */ 409 static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker) 410 { 411 if (worker->flags & IO_WORKER_F_FREE) { 412 worker->flags &= ~IO_WORKER_F_FREE; 413 raw_spin_lock(&wq->lock); 414 hlist_nulls_del_init_rcu(&worker->nulls_node); 415 raw_spin_unlock(&wq->lock); 416 } 417 } 418 419 /* 420 * No work, worker going to sleep. Move to freelist. 421 */ 422 static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker) 423 __must_hold(wq->lock) 424 { 425 if (!(worker->flags & IO_WORKER_F_FREE)) { 426 worker->flags |= IO_WORKER_F_FREE; 427 hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list); 428 } 429 } 430 431 static inline unsigned int io_get_work_hash(struct io_wq_work *work) 432 { 433 return work->flags >> IO_WQ_HASH_SHIFT; 434 } 435 436 static bool io_wait_on_hash(struct io_wq *wq, unsigned int hash) 437 { 438 bool ret = false; 439 440 spin_lock_irq(&wq->hash->wait.lock); 441 if (list_empty(&wq->wait.entry)) { 442 __add_wait_queue(&wq->hash->wait, &wq->wait); 443 if (!test_bit(hash, &wq->hash->map)) { 444 __set_current_state(TASK_RUNNING); 445 list_del_init(&wq->wait.entry); 446 ret = true; 447 } 448 } 449 spin_unlock_irq(&wq->hash->wait.lock); 450 return ret; 451 } 452 453 static struct io_wq_work *io_get_next_work(struct io_wq_acct *acct, 454 struct io_worker *worker) 455 __must_hold(acct->lock) 456 { 457 struct io_wq_work_node *node, *prev; 458 struct io_wq_work *work, *tail; 459 unsigned int stall_hash = -1U; 460 struct io_wq *wq = worker->wq; 461 462 wq_list_for_each(node, prev, &acct->work_list) { 463 unsigned int hash; 464 465 work = container_of(node, struct io_wq_work, list); 466 467 /* not hashed, can run anytime */ 468 if (!io_wq_is_hashed(work)) { 469 wq_list_del(&acct->work_list, node, prev); 470 return work; 471 } 472 473 hash = io_get_work_hash(work); 474 /* all items with this hash lie in [work, tail] */ 475 tail = wq->hash_tail[hash]; 476 477 /* hashed, can run if not already running */ 478 if (!test_and_set_bit(hash, &wq->hash->map)) { 479 wq->hash_tail[hash] = NULL; 480 wq_list_cut(&acct->work_list, &tail->list, prev); 481 return work; 482 } 483 if (stall_hash == -1U) 484 stall_hash = hash; 485 /* fast forward to a next hash, for-each will fix up @prev */ 486 node = &tail->list; 487 } 488 489 if (stall_hash != -1U) { 490 bool unstalled; 491 492 /* 493 * Set this before dropping the lock to avoid racing with new 494 * work being added and clearing the stalled bit. 495 */ 496 set_bit(IO_ACCT_STALLED_BIT, &acct->flags); 497 raw_spin_unlock(&acct->lock); 498 unstalled = io_wait_on_hash(wq, stall_hash); 499 raw_spin_lock(&acct->lock); 500 if (unstalled) { 501 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); 502 if (wq_has_sleeper(&wq->hash->wait)) 503 wake_up(&wq->hash->wait); 504 } 505 } 506 507 return NULL; 508 } 509 510 static void io_assign_current_work(struct io_worker *worker, 511 struct io_wq_work *work) 512 { 513 if (work) { 514 io_run_task_work(); 515 cond_resched(); 516 } 517 518 raw_spin_lock(&worker->lock); 519 worker->cur_work = work; 520 worker->next_work = NULL; 521 raw_spin_unlock(&worker->lock); 522 } 523 524 static void io_worker_handle_work(struct io_worker *worker) 525 { 526 struct io_wq_acct *acct = io_wq_get_acct(worker); 527 struct io_wq *wq = worker->wq; 528 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state); 529 530 do { 531 struct io_wq_work *work; 532 533 /* 534 * If we got some work, mark us as busy. If we didn't, but 535 * the list isn't empty, it means we stalled on hashed work. 536 * Mark us stalled so we don't keep looking for work when we 537 * can't make progress, any work completion or insertion will 538 * clear the stalled flag. 539 */ 540 raw_spin_lock(&acct->lock); 541 work = io_get_next_work(acct, worker); 542 raw_spin_unlock(&acct->lock); 543 if (work) { 544 __io_worker_busy(wq, worker); 545 546 /* 547 * Make sure cancelation can find this, even before 548 * it becomes the active work. That avoids a window 549 * where the work has been removed from our general 550 * work list, but isn't yet discoverable as the 551 * current work item for this worker. 552 */ 553 raw_spin_lock(&worker->lock); 554 worker->next_work = work; 555 raw_spin_unlock(&worker->lock); 556 } else { 557 break; 558 } 559 io_assign_current_work(worker, work); 560 __set_current_state(TASK_RUNNING); 561 562 /* handle a whole dependent link */ 563 do { 564 struct io_wq_work *next_hashed, *linked; 565 unsigned int hash = io_get_work_hash(work); 566 567 next_hashed = wq_next_work(work); 568 569 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND)) 570 work->flags |= IO_WQ_WORK_CANCEL; 571 wq->do_work(work); 572 io_assign_current_work(worker, NULL); 573 574 linked = wq->free_work(work); 575 work = next_hashed; 576 if (!work && linked && !io_wq_is_hashed(linked)) { 577 work = linked; 578 linked = NULL; 579 } 580 io_assign_current_work(worker, work); 581 if (linked) 582 io_wq_enqueue(wq, linked); 583 584 if (hash != -1U && !next_hashed) { 585 /* serialize hash clear with wake_up() */ 586 spin_lock_irq(&wq->hash->wait.lock); 587 clear_bit(hash, &wq->hash->map); 588 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); 589 spin_unlock_irq(&wq->hash->wait.lock); 590 if (wq_has_sleeper(&wq->hash->wait)) 591 wake_up(&wq->hash->wait); 592 } 593 } while (work); 594 } while (1); 595 } 596 597 static int io_wq_worker(void *data) 598 { 599 struct io_worker *worker = data; 600 struct io_wq_acct *acct = io_wq_get_acct(worker); 601 struct io_wq *wq = worker->wq; 602 bool exit_mask = false, last_timeout = false; 603 char buf[TASK_COMM_LEN]; 604 605 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); 606 607 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid); 608 set_task_comm(current, buf); 609 610 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { 611 long ret; 612 613 set_current_state(TASK_INTERRUPTIBLE); 614 while (io_acct_run_queue(acct)) 615 io_worker_handle_work(worker); 616 617 raw_spin_lock(&wq->lock); 618 /* 619 * Last sleep timed out. Exit if we're not the last worker, 620 * or if someone modified our affinity. 621 */ 622 if (last_timeout && (exit_mask || acct->nr_workers > 1)) { 623 acct->nr_workers--; 624 raw_spin_unlock(&wq->lock); 625 __set_current_state(TASK_RUNNING); 626 break; 627 } 628 last_timeout = false; 629 __io_worker_idle(wq, worker); 630 raw_spin_unlock(&wq->lock); 631 if (io_run_task_work()) 632 continue; 633 ret = schedule_timeout(WORKER_IDLE_TIMEOUT); 634 if (signal_pending(current)) { 635 struct ksignal ksig; 636 637 if (!get_signal(&ksig)) 638 continue; 639 break; 640 } 641 if (!ret) { 642 last_timeout = true; 643 exit_mask = !cpumask_test_cpu(raw_smp_processor_id(), 644 wq->cpu_mask); 645 } 646 } 647 648 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) 649 io_worker_handle_work(worker); 650 651 io_worker_exit(worker); 652 return 0; 653 } 654 655 /* 656 * Called when a worker is scheduled in. Mark us as currently running. 657 */ 658 void io_wq_worker_running(struct task_struct *tsk) 659 { 660 struct io_worker *worker = tsk->worker_private; 661 662 if (!worker) 663 return; 664 if (!(worker->flags & IO_WORKER_F_UP)) 665 return; 666 if (worker->flags & IO_WORKER_F_RUNNING) 667 return; 668 worker->flags |= IO_WORKER_F_RUNNING; 669 io_wq_inc_running(worker); 670 } 671 672 /* 673 * Called when worker is going to sleep. If there are no workers currently 674 * running and we have work pending, wake up a free one or create a new one. 675 */ 676 void io_wq_worker_sleeping(struct task_struct *tsk) 677 { 678 struct io_worker *worker = tsk->worker_private; 679 680 if (!worker) 681 return; 682 if (!(worker->flags & IO_WORKER_F_UP)) 683 return; 684 if (!(worker->flags & IO_WORKER_F_RUNNING)) 685 return; 686 687 worker->flags &= ~IO_WORKER_F_RUNNING; 688 io_wq_dec_running(worker); 689 } 690 691 static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker, 692 struct task_struct *tsk) 693 { 694 tsk->worker_private = worker; 695 worker->task = tsk; 696 set_cpus_allowed_ptr(tsk, wq->cpu_mask); 697 698 raw_spin_lock(&wq->lock); 699 hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list); 700 list_add_tail_rcu(&worker->all_list, &wq->all_list); 701 worker->flags |= IO_WORKER_F_FREE; 702 raw_spin_unlock(&wq->lock); 703 wake_up_new_task(tsk); 704 } 705 706 static bool io_wq_work_match_all(struct io_wq_work *work, void *data) 707 { 708 return true; 709 } 710 711 static inline bool io_should_retry_thread(long err) 712 { 713 /* 714 * Prevent perpetual task_work retry, if the task (or its group) is 715 * exiting. 716 */ 717 if (fatal_signal_pending(current)) 718 return false; 719 720 switch (err) { 721 case -EAGAIN: 722 case -ERESTARTSYS: 723 case -ERESTARTNOINTR: 724 case -ERESTARTNOHAND: 725 return true; 726 default: 727 return false; 728 } 729 } 730 731 static void create_worker_cont(struct callback_head *cb) 732 { 733 struct io_worker *worker; 734 struct task_struct *tsk; 735 struct io_wq *wq; 736 737 worker = container_of(cb, struct io_worker, create_work); 738 clear_bit_unlock(0, &worker->create_state); 739 wq = worker->wq; 740 tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE); 741 if (!IS_ERR(tsk)) { 742 io_init_new_worker(wq, worker, tsk); 743 io_worker_release(worker); 744 return; 745 } else if (!io_should_retry_thread(PTR_ERR(tsk))) { 746 struct io_wq_acct *acct = io_wq_get_acct(worker); 747 748 atomic_dec(&acct->nr_running); 749 raw_spin_lock(&wq->lock); 750 acct->nr_workers--; 751 if (!acct->nr_workers) { 752 struct io_cb_cancel_data match = { 753 .fn = io_wq_work_match_all, 754 .cancel_all = true, 755 }; 756 757 raw_spin_unlock(&wq->lock); 758 while (io_acct_cancel_pending_work(wq, acct, &match)) 759 ; 760 } else { 761 raw_spin_unlock(&wq->lock); 762 } 763 io_worker_ref_put(wq); 764 kfree(worker); 765 return; 766 } 767 768 /* re-create attempts grab a new worker ref, drop the existing one */ 769 io_worker_release(worker); 770 schedule_work(&worker->work); 771 } 772 773 static void io_workqueue_create(struct work_struct *work) 774 { 775 struct io_worker *worker = container_of(work, struct io_worker, work); 776 struct io_wq_acct *acct = io_wq_get_acct(worker); 777 778 if (!io_queue_worker_create(worker, acct, create_worker_cont)) 779 kfree(worker); 780 } 781 782 static bool create_io_worker(struct io_wq *wq, int index) 783 { 784 struct io_wq_acct *acct = &wq->acct[index]; 785 struct io_worker *worker; 786 struct task_struct *tsk; 787 788 __set_current_state(TASK_RUNNING); 789 790 worker = kzalloc(sizeof(*worker), GFP_KERNEL); 791 if (!worker) { 792 fail: 793 atomic_dec(&acct->nr_running); 794 raw_spin_lock(&wq->lock); 795 acct->nr_workers--; 796 raw_spin_unlock(&wq->lock); 797 io_worker_ref_put(wq); 798 return false; 799 } 800 801 refcount_set(&worker->ref, 1); 802 worker->wq = wq; 803 raw_spin_lock_init(&worker->lock); 804 init_completion(&worker->ref_done); 805 806 if (index == IO_WQ_ACCT_BOUND) 807 worker->flags |= IO_WORKER_F_BOUND; 808 809 tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE); 810 if (!IS_ERR(tsk)) { 811 io_init_new_worker(wq, worker, tsk); 812 } else if (!io_should_retry_thread(PTR_ERR(tsk))) { 813 kfree(worker); 814 goto fail; 815 } else { 816 INIT_WORK(&worker->work, io_workqueue_create); 817 schedule_work(&worker->work); 818 } 819 820 return true; 821 } 822 823 /* 824 * Iterate the passed in list and call the specific function for each 825 * worker that isn't exiting 826 */ 827 static bool io_wq_for_each_worker(struct io_wq *wq, 828 bool (*func)(struct io_worker *, void *), 829 void *data) 830 { 831 struct io_worker *worker; 832 bool ret = false; 833 834 list_for_each_entry_rcu(worker, &wq->all_list, all_list) { 835 if (io_worker_get(worker)) { 836 /* no task if node is/was offline */ 837 if (worker->task) 838 ret = func(worker, data); 839 io_worker_release(worker); 840 if (ret) 841 break; 842 } 843 } 844 845 return ret; 846 } 847 848 static bool io_wq_worker_wake(struct io_worker *worker, void *data) 849 { 850 __set_notify_signal(worker->task); 851 wake_up_process(worker->task); 852 return false; 853 } 854 855 static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq) 856 { 857 do { 858 work->flags |= IO_WQ_WORK_CANCEL; 859 wq->do_work(work); 860 work = wq->free_work(work); 861 } while (work); 862 } 863 864 static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work) 865 { 866 struct io_wq_acct *acct = io_work_get_acct(wq, work); 867 unsigned int hash; 868 struct io_wq_work *tail; 869 870 if (!io_wq_is_hashed(work)) { 871 append: 872 wq_list_add_tail(&work->list, &acct->work_list); 873 return; 874 } 875 876 hash = io_get_work_hash(work); 877 tail = wq->hash_tail[hash]; 878 wq->hash_tail[hash] = work; 879 if (!tail) 880 goto append; 881 882 wq_list_add_after(&work->list, &tail->list, &acct->work_list); 883 } 884 885 static bool io_wq_work_match_item(struct io_wq_work *work, void *data) 886 { 887 return work == data; 888 } 889 890 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) 891 { 892 struct io_wq_acct *acct = io_work_get_acct(wq, work); 893 struct io_cb_cancel_data match; 894 unsigned work_flags = work->flags; 895 bool do_create; 896 897 /* 898 * If io-wq is exiting for this task, or if the request has explicitly 899 * been marked as one that should not get executed, cancel it here. 900 */ 901 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || 902 (work->flags & IO_WQ_WORK_CANCEL)) { 903 io_run_cancel(work, wq); 904 return; 905 } 906 907 raw_spin_lock(&acct->lock); 908 io_wq_insert_work(wq, work); 909 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); 910 raw_spin_unlock(&acct->lock); 911 912 raw_spin_lock(&wq->lock); 913 rcu_read_lock(); 914 do_create = !io_wq_activate_free_worker(wq, acct); 915 rcu_read_unlock(); 916 917 raw_spin_unlock(&wq->lock); 918 919 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) || 920 !atomic_read(&acct->nr_running))) { 921 bool did_create; 922 923 did_create = io_wq_create_worker(wq, acct); 924 if (likely(did_create)) 925 return; 926 927 raw_spin_lock(&wq->lock); 928 if (acct->nr_workers) { 929 raw_spin_unlock(&wq->lock); 930 return; 931 } 932 raw_spin_unlock(&wq->lock); 933 934 /* fatal condition, failed to create the first worker */ 935 match.fn = io_wq_work_match_item, 936 match.data = work, 937 match.cancel_all = false, 938 939 io_acct_cancel_pending_work(wq, acct, &match); 940 } 941 } 942 943 /* 944 * Work items that hash to the same value will not be done in parallel. 945 * Used to limit concurrent writes, generally hashed by inode. 946 */ 947 void io_wq_hash_work(struct io_wq_work *work, void *val) 948 { 949 unsigned int bit; 950 951 bit = hash_ptr(val, IO_WQ_HASH_ORDER); 952 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); 953 } 954 955 static bool __io_wq_worker_cancel(struct io_worker *worker, 956 struct io_cb_cancel_data *match, 957 struct io_wq_work *work) 958 { 959 if (work && match->fn(work, match->data)) { 960 work->flags |= IO_WQ_WORK_CANCEL; 961 __set_notify_signal(worker->task); 962 return true; 963 } 964 965 return false; 966 } 967 968 static bool io_wq_worker_cancel(struct io_worker *worker, void *data) 969 { 970 struct io_cb_cancel_data *match = data; 971 972 /* 973 * Hold the lock to avoid ->cur_work going out of scope, caller 974 * may dereference the passed in work. 975 */ 976 raw_spin_lock(&worker->lock); 977 if (__io_wq_worker_cancel(worker, match, worker->cur_work) || 978 __io_wq_worker_cancel(worker, match, worker->next_work)) 979 match->nr_running++; 980 raw_spin_unlock(&worker->lock); 981 982 return match->nr_running && !match->cancel_all; 983 } 984 985 static inline void io_wq_remove_pending(struct io_wq *wq, 986 struct io_wq_work *work, 987 struct io_wq_work_node *prev) 988 { 989 struct io_wq_acct *acct = io_work_get_acct(wq, work); 990 unsigned int hash = io_get_work_hash(work); 991 struct io_wq_work *prev_work = NULL; 992 993 if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) { 994 if (prev) 995 prev_work = container_of(prev, struct io_wq_work, list); 996 if (prev_work && io_get_work_hash(prev_work) == hash) 997 wq->hash_tail[hash] = prev_work; 998 else 999 wq->hash_tail[hash] = NULL; 1000 } 1001 wq_list_del(&acct->work_list, &work->list, prev); 1002 } 1003 1004 static bool io_acct_cancel_pending_work(struct io_wq *wq, 1005 struct io_wq_acct *acct, 1006 struct io_cb_cancel_data *match) 1007 { 1008 struct io_wq_work_node *node, *prev; 1009 struct io_wq_work *work; 1010 1011 raw_spin_lock(&acct->lock); 1012 wq_list_for_each(node, prev, &acct->work_list) { 1013 work = container_of(node, struct io_wq_work, list); 1014 if (!match->fn(work, match->data)) 1015 continue; 1016 io_wq_remove_pending(wq, work, prev); 1017 raw_spin_unlock(&acct->lock); 1018 io_run_cancel(work, wq); 1019 match->nr_pending++; 1020 /* not safe to continue after unlock */ 1021 return true; 1022 } 1023 raw_spin_unlock(&acct->lock); 1024 1025 return false; 1026 } 1027 1028 static void io_wq_cancel_pending_work(struct io_wq *wq, 1029 struct io_cb_cancel_data *match) 1030 { 1031 int i; 1032 retry: 1033 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1034 struct io_wq_acct *acct = io_get_acct(wq, i == 0); 1035 1036 if (io_acct_cancel_pending_work(wq, acct, match)) { 1037 if (match->cancel_all) 1038 goto retry; 1039 break; 1040 } 1041 } 1042 } 1043 1044 static void io_wq_cancel_running_work(struct io_wq *wq, 1045 struct io_cb_cancel_data *match) 1046 { 1047 rcu_read_lock(); 1048 io_wq_for_each_worker(wq, io_wq_worker_cancel, match); 1049 rcu_read_unlock(); 1050 } 1051 1052 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, 1053 void *data, bool cancel_all) 1054 { 1055 struct io_cb_cancel_data match = { 1056 .fn = cancel, 1057 .data = data, 1058 .cancel_all = cancel_all, 1059 }; 1060 1061 /* 1062 * First check pending list, if we're lucky we can just remove it 1063 * from there. CANCEL_OK means that the work is returned as-new, 1064 * no completion will be posted for it. 1065 * 1066 * Then check if a free (going busy) or busy worker has the work 1067 * currently running. If we find it there, we'll return CANCEL_RUNNING 1068 * as an indication that we attempt to signal cancellation. The 1069 * completion will run normally in this case. 1070 * 1071 * Do both of these while holding the wq->lock, to ensure that 1072 * we'll find a work item regardless of state. 1073 */ 1074 io_wq_cancel_pending_work(wq, &match); 1075 if (match.nr_pending && !match.cancel_all) 1076 return IO_WQ_CANCEL_OK; 1077 1078 raw_spin_lock(&wq->lock); 1079 io_wq_cancel_running_work(wq, &match); 1080 raw_spin_unlock(&wq->lock); 1081 if (match.nr_running && !match.cancel_all) 1082 return IO_WQ_CANCEL_RUNNING; 1083 1084 if (match.nr_running) 1085 return IO_WQ_CANCEL_RUNNING; 1086 if (match.nr_pending) 1087 return IO_WQ_CANCEL_OK; 1088 return IO_WQ_CANCEL_NOTFOUND; 1089 } 1090 1091 static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode, 1092 int sync, void *key) 1093 { 1094 struct io_wq *wq = container_of(wait, struct io_wq, wait); 1095 int i; 1096 1097 list_del_init(&wait->entry); 1098 1099 rcu_read_lock(); 1100 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1101 struct io_wq_acct *acct = &wq->acct[i]; 1102 1103 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags)) 1104 io_wq_activate_free_worker(wq, acct); 1105 } 1106 rcu_read_unlock(); 1107 return 1; 1108 } 1109 1110 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) 1111 { 1112 int ret, i; 1113 struct io_wq *wq; 1114 1115 if (WARN_ON_ONCE(!data->free_work || !data->do_work)) 1116 return ERR_PTR(-EINVAL); 1117 if (WARN_ON_ONCE(!bounded)) 1118 return ERR_PTR(-EINVAL); 1119 1120 wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL); 1121 if (!wq) 1122 return ERR_PTR(-ENOMEM); 1123 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node); 1124 if (ret) 1125 goto err_wq; 1126 1127 refcount_inc(&data->hash->refs); 1128 wq->hash = data->hash; 1129 wq->free_work = data->free_work; 1130 wq->do_work = data->do_work; 1131 1132 ret = -ENOMEM; 1133 1134 if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL)) 1135 goto err; 1136 cpumask_copy(wq->cpu_mask, cpu_possible_mask); 1137 wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; 1138 wq->acct[IO_WQ_ACCT_UNBOUND].max_workers = 1139 task_rlimit(current, RLIMIT_NPROC); 1140 INIT_LIST_HEAD(&wq->wait.entry); 1141 wq->wait.func = io_wq_hash_wake; 1142 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1143 struct io_wq_acct *acct = &wq->acct[i]; 1144 1145 acct->index = i; 1146 atomic_set(&acct->nr_running, 0); 1147 INIT_WQ_LIST(&acct->work_list); 1148 raw_spin_lock_init(&acct->lock); 1149 } 1150 1151 raw_spin_lock_init(&wq->lock); 1152 INIT_HLIST_NULLS_HEAD(&wq->free_list, 0); 1153 INIT_LIST_HEAD(&wq->all_list); 1154 1155 wq->task = get_task_struct(data->task); 1156 atomic_set(&wq->worker_refs, 1); 1157 init_completion(&wq->worker_done); 1158 return wq; 1159 err: 1160 io_wq_put_hash(data->hash); 1161 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); 1162 1163 free_cpumask_var(wq->cpu_mask); 1164 err_wq: 1165 kfree(wq); 1166 return ERR_PTR(ret); 1167 } 1168 1169 static bool io_task_work_match(struct callback_head *cb, void *data) 1170 { 1171 struct io_worker *worker; 1172 1173 if (cb->func != create_worker_cb && cb->func != create_worker_cont) 1174 return false; 1175 worker = container_of(cb, struct io_worker, create_work); 1176 return worker->wq == data; 1177 } 1178 1179 void io_wq_exit_start(struct io_wq *wq) 1180 { 1181 set_bit(IO_WQ_BIT_EXIT, &wq->state); 1182 } 1183 1184 static void io_wq_cancel_tw_create(struct io_wq *wq) 1185 { 1186 struct callback_head *cb; 1187 1188 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) { 1189 struct io_worker *worker; 1190 1191 worker = container_of(cb, struct io_worker, create_work); 1192 io_worker_cancel_cb(worker); 1193 /* 1194 * Only the worker continuation helper has worker allocated and 1195 * hence needs freeing. 1196 */ 1197 if (cb->func == create_worker_cont) 1198 kfree(worker); 1199 } 1200 } 1201 1202 static void io_wq_exit_workers(struct io_wq *wq) 1203 { 1204 if (!wq->task) 1205 return; 1206 1207 io_wq_cancel_tw_create(wq); 1208 1209 rcu_read_lock(); 1210 io_wq_for_each_worker(wq, io_wq_worker_wake, NULL); 1211 rcu_read_unlock(); 1212 io_worker_ref_put(wq); 1213 wait_for_completion(&wq->worker_done); 1214 1215 spin_lock_irq(&wq->hash->wait.lock); 1216 list_del_init(&wq->wait.entry); 1217 spin_unlock_irq(&wq->hash->wait.lock); 1218 1219 put_task_struct(wq->task); 1220 wq->task = NULL; 1221 } 1222 1223 static void io_wq_destroy(struct io_wq *wq) 1224 { 1225 struct io_cb_cancel_data match = { 1226 .fn = io_wq_work_match_all, 1227 .cancel_all = true, 1228 }; 1229 1230 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); 1231 io_wq_cancel_pending_work(wq, &match); 1232 free_cpumask_var(wq->cpu_mask); 1233 io_wq_put_hash(wq->hash); 1234 kfree(wq); 1235 } 1236 1237 void io_wq_put_and_exit(struct io_wq *wq) 1238 { 1239 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state)); 1240 1241 io_wq_exit_workers(wq); 1242 io_wq_destroy(wq); 1243 } 1244 1245 struct online_data { 1246 unsigned int cpu; 1247 bool online; 1248 }; 1249 1250 static bool io_wq_worker_affinity(struct io_worker *worker, void *data) 1251 { 1252 struct online_data *od = data; 1253 1254 if (od->online) 1255 cpumask_set_cpu(od->cpu, worker->wq->cpu_mask); 1256 else 1257 cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask); 1258 return false; 1259 } 1260 1261 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online) 1262 { 1263 struct online_data od = { 1264 .cpu = cpu, 1265 .online = online 1266 }; 1267 1268 rcu_read_lock(); 1269 io_wq_for_each_worker(wq, io_wq_worker_affinity, &od); 1270 rcu_read_unlock(); 1271 return 0; 1272 } 1273 1274 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node) 1275 { 1276 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); 1277 1278 return __io_wq_cpu_online(wq, cpu, true); 1279 } 1280 1281 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node) 1282 { 1283 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); 1284 1285 return __io_wq_cpu_online(wq, cpu, false); 1286 } 1287 1288 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask) 1289 { 1290 rcu_read_lock(); 1291 if (mask) 1292 cpumask_copy(wq->cpu_mask, mask); 1293 else 1294 cpumask_copy(wq->cpu_mask, cpu_possible_mask); 1295 rcu_read_unlock(); 1296 1297 return 0; 1298 } 1299 1300 /* 1301 * Set max number of unbounded workers, returns old value. If new_count is 0, 1302 * then just return the old value. 1303 */ 1304 int io_wq_max_workers(struct io_wq *wq, int *new_count) 1305 { 1306 struct io_wq_acct *acct; 1307 int prev[IO_WQ_ACCT_NR]; 1308 int i; 1309 1310 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND); 1311 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND); 1312 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2); 1313 1314 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1315 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC)) 1316 new_count[i] = task_rlimit(current, RLIMIT_NPROC); 1317 } 1318 1319 for (i = 0; i < IO_WQ_ACCT_NR; i++) 1320 prev[i] = 0; 1321 1322 rcu_read_lock(); 1323 1324 raw_spin_lock(&wq->lock); 1325 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1326 acct = &wq->acct[i]; 1327 prev[i] = max_t(int, acct->max_workers, prev[i]); 1328 if (new_count[i]) 1329 acct->max_workers = new_count[i]; 1330 } 1331 raw_spin_unlock(&wq->lock); 1332 rcu_read_unlock(); 1333 1334 for (i = 0; i < IO_WQ_ACCT_NR; i++) 1335 new_count[i] = prev[i]; 1336 1337 return 0; 1338 } 1339 1340 static __init int io_wq_init(void) 1341 { 1342 int ret; 1343 1344 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online", 1345 io_wq_cpu_online, io_wq_cpu_offline); 1346 if (ret < 0) 1347 return ret; 1348 io_wq_online = ret; 1349 return 0; 1350 } 1351 subsys_initcall(io_wq_init); 1352