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 = 0, /* up and active */ 29 IO_WORKER_F_RUNNING = 1, /* account as running */ 30 IO_WORKER_F_FREE = 2, /* worker on free list */ 31 IO_WORKER_F_BOUND = 3, /* 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 int create_index; 48 unsigned long flags; 49 struct hlist_nulls_node nulls_node; 50 struct list_head all_list; 51 struct task_struct *task; 52 struct io_wq *wq; 53 54 struct io_wq_work *cur_work; 55 struct io_wq_work *next_work; 56 raw_spinlock_t lock; 57 58 struct completion ref_done; 59 60 unsigned long create_state; 61 struct callback_head create_work; 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, test_bit(IO_WORKER_F_BOUND, &worker->flags)); 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 bool io_wq_worker_stopped(void) 178 { 179 struct io_worker *worker = current->worker_private; 180 181 if (WARN_ON_ONCE(!io_wq_current_is_worker())) 182 return true; 183 184 return test_bit(IO_WQ_BIT_EXIT, &worker->wq->state); 185 } 186 187 static void io_worker_cancel_cb(struct io_worker *worker) 188 { 189 struct io_wq_acct *acct = io_wq_get_acct(worker); 190 struct io_wq *wq = worker->wq; 191 192 atomic_dec(&acct->nr_running); 193 raw_spin_lock(&wq->lock); 194 acct->nr_workers--; 195 raw_spin_unlock(&wq->lock); 196 io_worker_ref_put(wq); 197 clear_bit_unlock(0, &worker->create_state); 198 io_worker_release(worker); 199 } 200 201 static bool io_task_worker_match(struct callback_head *cb, void *data) 202 { 203 struct io_worker *worker; 204 205 if (cb->func != create_worker_cb) 206 return false; 207 worker = container_of(cb, struct io_worker, create_work); 208 return worker == data; 209 } 210 211 static void io_worker_exit(struct io_worker *worker) 212 { 213 struct io_wq *wq = worker->wq; 214 215 while (1) { 216 struct callback_head *cb = task_work_cancel_match(wq->task, 217 io_task_worker_match, worker); 218 219 if (!cb) 220 break; 221 io_worker_cancel_cb(worker); 222 } 223 224 io_worker_release(worker); 225 wait_for_completion(&worker->ref_done); 226 227 raw_spin_lock(&wq->lock); 228 if (test_bit(IO_WORKER_F_FREE, &worker->flags)) 229 hlist_nulls_del_rcu(&worker->nulls_node); 230 list_del_rcu(&worker->all_list); 231 raw_spin_unlock(&wq->lock); 232 io_wq_dec_running(worker); 233 /* 234 * this worker is a goner, clear ->worker_private to avoid any 235 * inc/dec running calls that could happen as part of exit from 236 * touching 'worker'. 237 */ 238 current->worker_private = NULL; 239 240 kfree_rcu(worker, rcu); 241 io_worker_ref_put(wq); 242 do_exit(0); 243 } 244 245 static inline bool __io_acct_run_queue(struct io_wq_acct *acct) 246 { 247 return !test_bit(IO_ACCT_STALLED_BIT, &acct->flags) && 248 !wq_list_empty(&acct->work_list); 249 } 250 251 /* 252 * If there's work to do, returns true with acct->lock acquired. If not, 253 * returns false with no lock held. 254 */ 255 static inline bool io_acct_run_queue(struct io_wq_acct *acct) 256 __acquires(&acct->lock) 257 { 258 raw_spin_lock(&acct->lock); 259 if (__io_acct_run_queue(acct)) 260 return true; 261 262 raw_spin_unlock(&acct->lock); 263 return false; 264 } 265 266 /* 267 * Check head of free list for an available worker. If one isn't available, 268 * caller must create one. 269 */ 270 static bool io_wq_activate_free_worker(struct io_wq *wq, 271 struct io_wq_acct *acct) 272 __must_hold(RCU) 273 { 274 struct hlist_nulls_node *n; 275 struct io_worker *worker; 276 277 /* 278 * Iterate free_list and see if we can find an idle worker to 279 * activate. If a given worker is on the free_list but in the process 280 * of exiting, keep trying. 281 */ 282 hlist_nulls_for_each_entry_rcu(worker, n, &wq->free_list, nulls_node) { 283 if (!io_worker_get(worker)) 284 continue; 285 if (io_wq_get_acct(worker) != acct) { 286 io_worker_release(worker); 287 continue; 288 } 289 /* 290 * If the worker is already running, it's either already 291 * starting work or finishing work. In either case, if it does 292 * to go sleep, we'll kick off a new task for this work anyway. 293 */ 294 wake_up_process(worker->task); 295 io_worker_release(worker); 296 return true; 297 } 298 299 return false; 300 } 301 302 /* 303 * We need a worker. If we find a free one, we're good. If not, and we're 304 * below the max number of workers, create one. 305 */ 306 static bool io_wq_create_worker(struct io_wq *wq, struct io_wq_acct *acct) 307 { 308 /* 309 * Most likely an attempt to queue unbounded work on an io_wq that 310 * wasn't setup with any unbounded workers. 311 */ 312 if (unlikely(!acct->max_workers)) 313 pr_warn_once("io-wq is not configured for unbound workers"); 314 315 raw_spin_lock(&wq->lock); 316 if (acct->nr_workers >= acct->max_workers) { 317 raw_spin_unlock(&wq->lock); 318 return true; 319 } 320 acct->nr_workers++; 321 raw_spin_unlock(&wq->lock); 322 atomic_inc(&acct->nr_running); 323 atomic_inc(&wq->worker_refs); 324 return create_io_worker(wq, acct->index); 325 } 326 327 static void io_wq_inc_running(struct io_worker *worker) 328 { 329 struct io_wq_acct *acct = io_wq_get_acct(worker); 330 331 atomic_inc(&acct->nr_running); 332 } 333 334 static void create_worker_cb(struct callback_head *cb) 335 { 336 struct io_worker *worker; 337 struct io_wq *wq; 338 339 struct io_wq_acct *acct; 340 bool do_create = false; 341 342 worker = container_of(cb, struct io_worker, create_work); 343 wq = worker->wq; 344 acct = &wq->acct[worker->create_index]; 345 raw_spin_lock(&wq->lock); 346 347 if (acct->nr_workers < acct->max_workers) { 348 acct->nr_workers++; 349 do_create = true; 350 } 351 raw_spin_unlock(&wq->lock); 352 if (do_create) { 353 create_io_worker(wq, worker->create_index); 354 } else { 355 atomic_dec(&acct->nr_running); 356 io_worker_ref_put(wq); 357 } 358 clear_bit_unlock(0, &worker->create_state); 359 io_worker_release(worker); 360 } 361 362 static bool io_queue_worker_create(struct io_worker *worker, 363 struct io_wq_acct *acct, 364 task_work_func_t func) 365 { 366 struct io_wq *wq = worker->wq; 367 368 /* raced with exit, just ignore create call */ 369 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) 370 goto fail; 371 if (!io_worker_get(worker)) 372 goto fail; 373 /* 374 * create_state manages ownership of create_work/index. We should 375 * only need one entry per worker, as the worker going to sleep 376 * will trigger the condition, and waking will clear it once it 377 * runs the task_work. 378 */ 379 if (test_bit(0, &worker->create_state) || 380 test_and_set_bit_lock(0, &worker->create_state)) 381 goto fail_release; 382 383 atomic_inc(&wq->worker_refs); 384 init_task_work(&worker->create_work, func); 385 worker->create_index = acct->index; 386 if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) { 387 /* 388 * EXIT may have been set after checking it above, check after 389 * adding the task_work and remove any creation item if it is 390 * now set. wq exit does that too, but we can have added this 391 * work item after we canceled in io_wq_exit_workers(). 392 */ 393 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) 394 io_wq_cancel_tw_create(wq); 395 io_worker_ref_put(wq); 396 return true; 397 } 398 io_worker_ref_put(wq); 399 clear_bit_unlock(0, &worker->create_state); 400 fail_release: 401 io_worker_release(worker); 402 fail: 403 atomic_dec(&acct->nr_running); 404 io_worker_ref_put(wq); 405 return false; 406 } 407 408 static void io_wq_dec_running(struct io_worker *worker) 409 { 410 struct io_wq_acct *acct = io_wq_get_acct(worker); 411 struct io_wq *wq = worker->wq; 412 413 if (!test_bit(IO_WORKER_F_UP, &worker->flags)) 414 return; 415 416 if (!atomic_dec_and_test(&acct->nr_running)) 417 return; 418 if (!io_acct_run_queue(acct)) 419 return; 420 421 raw_spin_unlock(&acct->lock); 422 atomic_inc(&acct->nr_running); 423 atomic_inc(&wq->worker_refs); 424 io_queue_worker_create(worker, acct, create_worker_cb); 425 } 426 427 /* 428 * Worker will start processing some work. Move it to the busy list, if 429 * it's currently on the freelist 430 */ 431 static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker) 432 { 433 if (test_bit(IO_WORKER_F_FREE, &worker->flags)) { 434 clear_bit(IO_WORKER_F_FREE, &worker->flags); 435 raw_spin_lock(&wq->lock); 436 hlist_nulls_del_init_rcu(&worker->nulls_node); 437 raw_spin_unlock(&wq->lock); 438 } 439 } 440 441 /* 442 * No work, worker going to sleep. Move to freelist. 443 */ 444 static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker) 445 __must_hold(wq->lock) 446 { 447 if (!test_bit(IO_WORKER_F_FREE, &worker->flags)) { 448 set_bit(IO_WORKER_F_FREE, &worker->flags); 449 hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list); 450 } 451 } 452 453 static inline unsigned int io_get_work_hash(struct io_wq_work *work) 454 { 455 return work->flags >> IO_WQ_HASH_SHIFT; 456 } 457 458 static bool io_wait_on_hash(struct io_wq *wq, unsigned int hash) 459 { 460 bool ret = false; 461 462 spin_lock_irq(&wq->hash->wait.lock); 463 if (list_empty(&wq->wait.entry)) { 464 __add_wait_queue(&wq->hash->wait, &wq->wait); 465 if (!test_bit(hash, &wq->hash->map)) { 466 __set_current_state(TASK_RUNNING); 467 list_del_init(&wq->wait.entry); 468 ret = true; 469 } 470 } 471 spin_unlock_irq(&wq->hash->wait.lock); 472 return ret; 473 } 474 475 static struct io_wq_work *io_get_next_work(struct io_wq_acct *acct, 476 struct io_worker *worker) 477 __must_hold(acct->lock) 478 { 479 struct io_wq_work_node *node, *prev; 480 struct io_wq_work *work, *tail; 481 unsigned int stall_hash = -1U; 482 struct io_wq *wq = worker->wq; 483 484 wq_list_for_each(node, prev, &acct->work_list) { 485 unsigned int hash; 486 487 work = container_of(node, struct io_wq_work, list); 488 489 /* not hashed, can run anytime */ 490 if (!io_wq_is_hashed(work)) { 491 wq_list_del(&acct->work_list, node, prev); 492 return work; 493 } 494 495 hash = io_get_work_hash(work); 496 /* all items with this hash lie in [work, tail] */ 497 tail = wq->hash_tail[hash]; 498 499 /* hashed, can run if not already running */ 500 if (!test_and_set_bit(hash, &wq->hash->map)) { 501 wq->hash_tail[hash] = NULL; 502 wq_list_cut(&acct->work_list, &tail->list, prev); 503 return work; 504 } 505 if (stall_hash == -1U) 506 stall_hash = hash; 507 /* fast forward to a next hash, for-each will fix up @prev */ 508 node = &tail->list; 509 } 510 511 if (stall_hash != -1U) { 512 bool unstalled; 513 514 /* 515 * Set this before dropping the lock to avoid racing with new 516 * work being added and clearing the stalled bit. 517 */ 518 set_bit(IO_ACCT_STALLED_BIT, &acct->flags); 519 raw_spin_unlock(&acct->lock); 520 unstalled = io_wait_on_hash(wq, stall_hash); 521 raw_spin_lock(&acct->lock); 522 if (unstalled) { 523 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); 524 if (wq_has_sleeper(&wq->hash->wait)) 525 wake_up(&wq->hash->wait); 526 } 527 } 528 529 return NULL; 530 } 531 532 static void io_assign_current_work(struct io_worker *worker, 533 struct io_wq_work *work) 534 { 535 if (work) { 536 io_run_task_work(); 537 cond_resched(); 538 } 539 540 raw_spin_lock(&worker->lock); 541 worker->cur_work = work; 542 worker->next_work = NULL; 543 raw_spin_unlock(&worker->lock); 544 } 545 546 /* 547 * Called with acct->lock held, drops it before returning 548 */ 549 static void io_worker_handle_work(struct io_wq_acct *acct, 550 struct io_worker *worker) 551 __releases(&acct->lock) 552 { 553 struct io_wq *wq = worker->wq; 554 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state); 555 556 do { 557 struct io_wq_work *work; 558 559 /* 560 * If we got some work, mark us as busy. If we didn't, but 561 * the list isn't empty, it means we stalled on hashed work. 562 * Mark us stalled so we don't keep looking for work when we 563 * can't make progress, any work completion or insertion will 564 * clear the stalled flag. 565 */ 566 work = io_get_next_work(acct, worker); 567 if (work) { 568 /* 569 * Make sure cancelation can find this, even before 570 * it becomes the active work. That avoids a window 571 * where the work has been removed from our general 572 * work list, but isn't yet discoverable as the 573 * current work item for this worker. 574 */ 575 raw_spin_lock(&worker->lock); 576 worker->next_work = work; 577 raw_spin_unlock(&worker->lock); 578 } 579 580 raw_spin_unlock(&acct->lock); 581 582 if (!work) 583 break; 584 585 __io_worker_busy(wq, worker); 586 587 io_assign_current_work(worker, work); 588 __set_current_state(TASK_RUNNING); 589 590 /* handle a whole dependent link */ 591 do { 592 struct io_wq_work *next_hashed, *linked; 593 unsigned int hash = io_get_work_hash(work); 594 595 next_hashed = wq_next_work(work); 596 597 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND)) 598 work->flags |= IO_WQ_WORK_CANCEL; 599 wq->do_work(work); 600 io_assign_current_work(worker, NULL); 601 602 linked = wq->free_work(work); 603 work = next_hashed; 604 if (!work && linked && !io_wq_is_hashed(linked)) { 605 work = linked; 606 linked = NULL; 607 } 608 io_assign_current_work(worker, work); 609 if (linked) 610 io_wq_enqueue(wq, linked); 611 612 if (hash != -1U && !next_hashed) { 613 /* serialize hash clear with wake_up() */ 614 spin_lock_irq(&wq->hash->wait.lock); 615 clear_bit(hash, &wq->hash->map); 616 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); 617 spin_unlock_irq(&wq->hash->wait.lock); 618 if (wq_has_sleeper(&wq->hash->wait)) 619 wake_up(&wq->hash->wait); 620 } 621 } while (work); 622 623 if (!__io_acct_run_queue(acct)) 624 break; 625 raw_spin_lock(&acct->lock); 626 } while (1); 627 } 628 629 static int io_wq_worker(void *data) 630 { 631 struct io_worker *worker = data; 632 struct io_wq_acct *acct = io_wq_get_acct(worker); 633 struct io_wq *wq = worker->wq; 634 bool exit_mask = false, last_timeout = false; 635 char buf[TASK_COMM_LEN]; 636 637 set_mask_bits(&worker->flags, 0, 638 BIT(IO_WORKER_F_UP) | BIT(IO_WORKER_F_RUNNING)); 639 640 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid); 641 set_task_comm(current, buf); 642 643 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { 644 long ret; 645 646 set_current_state(TASK_INTERRUPTIBLE); 647 648 /* 649 * If we have work to do, io_acct_run_queue() returns with 650 * the acct->lock held. If not, it will drop it. 651 */ 652 while (io_acct_run_queue(acct)) 653 io_worker_handle_work(acct, worker); 654 655 raw_spin_lock(&wq->lock); 656 /* 657 * Last sleep timed out. Exit if we're not the last worker, 658 * or if someone modified our affinity. 659 */ 660 if (last_timeout && (exit_mask || acct->nr_workers > 1)) { 661 acct->nr_workers--; 662 raw_spin_unlock(&wq->lock); 663 __set_current_state(TASK_RUNNING); 664 break; 665 } 666 last_timeout = false; 667 __io_worker_idle(wq, worker); 668 raw_spin_unlock(&wq->lock); 669 if (io_run_task_work()) 670 continue; 671 ret = schedule_timeout(WORKER_IDLE_TIMEOUT); 672 if (signal_pending(current)) { 673 struct ksignal ksig; 674 675 if (!get_signal(&ksig)) 676 continue; 677 break; 678 } 679 if (!ret) { 680 last_timeout = true; 681 exit_mask = !cpumask_test_cpu(raw_smp_processor_id(), 682 wq->cpu_mask); 683 } 684 } 685 686 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && io_acct_run_queue(acct)) 687 io_worker_handle_work(acct, worker); 688 689 io_worker_exit(worker); 690 return 0; 691 } 692 693 /* 694 * Called when a worker is scheduled in. Mark us as currently running. 695 */ 696 void io_wq_worker_running(struct task_struct *tsk) 697 { 698 struct io_worker *worker = tsk->worker_private; 699 700 if (!worker) 701 return; 702 if (!test_bit(IO_WORKER_F_UP, &worker->flags)) 703 return; 704 if (test_bit(IO_WORKER_F_RUNNING, &worker->flags)) 705 return; 706 set_bit(IO_WORKER_F_RUNNING, &worker->flags); 707 io_wq_inc_running(worker); 708 } 709 710 /* 711 * Called when worker is going to sleep. If there are no workers currently 712 * running and we have work pending, wake up a free one or create a new one. 713 */ 714 void io_wq_worker_sleeping(struct task_struct *tsk) 715 { 716 struct io_worker *worker = tsk->worker_private; 717 718 if (!worker) 719 return; 720 if (!test_bit(IO_WORKER_F_UP, &worker->flags)) 721 return; 722 if (!test_bit(IO_WORKER_F_RUNNING, &worker->flags)) 723 return; 724 725 clear_bit(IO_WORKER_F_RUNNING, &worker->flags); 726 io_wq_dec_running(worker); 727 } 728 729 static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker, 730 struct task_struct *tsk) 731 { 732 tsk->worker_private = worker; 733 worker->task = tsk; 734 set_cpus_allowed_ptr(tsk, wq->cpu_mask); 735 736 raw_spin_lock(&wq->lock); 737 hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list); 738 list_add_tail_rcu(&worker->all_list, &wq->all_list); 739 set_bit(IO_WORKER_F_FREE, &worker->flags); 740 raw_spin_unlock(&wq->lock); 741 wake_up_new_task(tsk); 742 } 743 744 static bool io_wq_work_match_all(struct io_wq_work *work, void *data) 745 { 746 return true; 747 } 748 749 static inline bool io_should_retry_thread(long err) 750 { 751 /* 752 * Prevent perpetual task_work retry, if the task (or its group) is 753 * exiting. 754 */ 755 if (fatal_signal_pending(current)) 756 return false; 757 758 switch (err) { 759 case -EAGAIN: 760 case -ERESTARTSYS: 761 case -ERESTARTNOINTR: 762 case -ERESTARTNOHAND: 763 return true; 764 default: 765 return false; 766 } 767 } 768 769 static void create_worker_cont(struct callback_head *cb) 770 { 771 struct io_worker *worker; 772 struct task_struct *tsk; 773 struct io_wq *wq; 774 775 worker = container_of(cb, struct io_worker, create_work); 776 clear_bit_unlock(0, &worker->create_state); 777 wq = worker->wq; 778 tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE); 779 if (!IS_ERR(tsk)) { 780 io_init_new_worker(wq, worker, tsk); 781 io_worker_release(worker); 782 return; 783 } else if (!io_should_retry_thread(PTR_ERR(tsk))) { 784 struct io_wq_acct *acct = io_wq_get_acct(worker); 785 786 atomic_dec(&acct->nr_running); 787 raw_spin_lock(&wq->lock); 788 acct->nr_workers--; 789 if (!acct->nr_workers) { 790 struct io_cb_cancel_data match = { 791 .fn = io_wq_work_match_all, 792 .cancel_all = true, 793 }; 794 795 raw_spin_unlock(&wq->lock); 796 while (io_acct_cancel_pending_work(wq, acct, &match)) 797 ; 798 } else { 799 raw_spin_unlock(&wq->lock); 800 } 801 io_worker_ref_put(wq); 802 kfree(worker); 803 return; 804 } 805 806 /* re-create attempts grab a new worker ref, drop the existing one */ 807 io_worker_release(worker); 808 schedule_work(&worker->work); 809 } 810 811 static void io_workqueue_create(struct work_struct *work) 812 { 813 struct io_worker *worker = container_of(work, struct io_worker, work); 814 struct io_wq_acct *acct = io_wq_get_acct(worker); 815 816 if (!io_queue_worker_create(worker, acct, create_worker_cont)) 817 kfree(worker); 818 } 819 820 static bool create_io_worker(struct io_wq *wq, int index) 821 { 822 struct io_wq_acct *acct = &wq->acct[index]; 823 struct io_worker *worker; 824 struct task_struct *tsk; 825 826 __set_current_state(TASK_RUNNING); 827 828 worker = kzalloc(sizeof(*worker), GFP_KERNEL); 829 if (!worker) { 830 fail: 831 atomic_dec(&acct->nr_running); 832 raw_spin_lock(&wq->lock); 833 acct->nr_workers--; 834 raw_spin_unlock(&wq->lock); 835 io_worker_ref_put(wq); 836 return false; 837 } 838 839 refcount_set(&worker->ref, 1); 840 worker->wq = wq; 841 raw_spin_lock_init(&worker->lock); 842 init_completion(&worker->ref_done); 843 844 if (index == IO_WQ_ACCT_BOUND) 845 set_bit(IO_WORKER_F_BOUND, &worker->flags); 846 847 tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE); 848 if (!IS_ERR(tsk)) { 849 io_init_new_worker(wq, worker, tsk); 850 } else if (!io_should_retry_thread(PTR_ERR(tsk))) { 851 kfree(worker); 852 goto fail; 853 } else { 854 INIT_WORK(&worker->work, io_workqueue_create); 855 schedule_work(&worker->work); 856 } 857 858 return true; 859 } 860 861 /* 862 * Iterate the passed in list and call the specific function for each 863 * worker that isn't exiting 864 */ 865 static bool io_wq_for_each_worker(struct io_wq *wq, 866 bool (*func)(struct io_worker *, void *), 867 void *data) 868 { 869 struct io_worker *worker; 870 bool ret = false; 871 872 list_for_each_entry_rcu(worker, &wq->all_list, all_list) { 873 if (io_worker_get(worker)) { 874 /* no task if node is/was offline */ 875 if (worker->task) 876 ret = func(worker, data); 877 io_worker_release(worker); 878 if (ret) 879 break; 880 } 881 } 882 883 return ret; 884 } 885 886 static bool io_wq_worker_wake(struct io_worker *worker, void *data) 887 { 888 __set_notify_signal(worker->task); 889 wake_up_process(worker->task); 890 return false; 891 } 892 893 static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq) 894 { 895 do { 896 work->flags |= IO_WQ_WORK_CANCEL; 897 wq->do_work(work); 898 work = wq->free_work(work); 899 } while (work); 900 } 901 902 static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work) 903 { 904 struct io_wq_acct *acct = io_work_get_acct(wq, work); 905 unsigned int hash; 906 struct io_wq_work *tail; 907 908 if (!io_wq_is_hashed(work)) { 909 append: 910 wq_list_add_tail(&work->list, &acct->work_list); 911 return; 912 } 913 914 hash = io_get_work_hash(work); 915 tail = wq->hash_tail[hash]; 916 wq->hash_tail[hash] = work; 917 if (!tail) 918 goto append; 919 920 wq_list_add_after(&work->list, &tail->list, &acct->work_list); 921 } 922 923 static bool io_wq_work_match_item(struct io_wq_work *work, void *data) 924 { 925 return work == data; 926 } 927 928 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) 929 { 930 struct io_wq_acct *acct = io_work_get_acct(wq, work); 931 unsigned long work_flags = work->flags; 932 struct io_cb_cancel_data match = { 933 .fn = io_wq_work_match_item, 934 .data = work, 935 .cancel_all = false, 936 }; 937 bool do_create; 938 939 /* 940 * If io-wq is exiting for this task, or if the request has explicitly 941 * been marked as one that should not get executed, cancel it here. 942 */ 943 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || 944 (work->flags & IO_WQ_WORK_CANCEL)) { 945 io_run_cancel(work, wq); 946 return; 947 } 948 949 raw_spin_lock(&acct->lock); 950 io_wq_insert_work(wq, work); 951 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); 952 raw_spin_unlock(&acct->lock); 953 954 rcu_read_lock(); 955 do_create = !io_wq_activate_free_worker(wq, acct); 956 rcu_read_unlock(); 957 958 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) || 959 !atomic_read(&acct->nr_running))) { 960 bool did_create; 961 962 did_create = io_wq_create_worker(wq, acct); 963 if (likely(did_create)) 964 return; 965 966 raw_spin_lock(&wq->lock); 967 if (acct->nr_workers) { 968 raw_spin_unlock(&wq->lock); 969 return; 970 } 971 raw_spin_unlock(&wq->lock); 972 973 /* fatal condition, failed to create the first worker */ 974 io_acct_cancel_pending_work(wq, acct, &match); 975 } 976 } 977 978 /* 979 * Work items that hash to the same value will not be done in parallel. 980 * Used to limit concurrent writes, generally hashed by inode. 981 */ 982 void io_wq_hash_work(struct io_wq_work *work, void *val) 983 { 984 unsigned int bit; 985 986 bit = hash_ptr(val, IO_WQ_HASH_ORDER); 987 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); 988 } 989 990 static bool __io_wq_worker_cancel(struct io_worker *worker, 991 struct io_cb_cancel_data *match, 992 struct io_wq_work *work) 993 { 994 if (work && match->fn(work, match->data)) { 995 work->flags |= IO_WQ_WORK_CANCEL; 996 __set_notify_signal(worker->task); 997 return true; 998 } 999 1000 return false; 1001 } 1002 1003 static bool io_wq_worker_cancel(struct io_worker *worker, void *data) 1004 { 1005 struct io_cb_cancel_data *match = data; 1006 1007 /* 1008 * Hold the lock to avoid ->cur_work going out of scope, caller 1009 * may dereference the passed in work. 1010 */ 1011 raw_spin_lock(&worker->lock); 1012 if (__io_wq_worker_cancel(worker, match, worker->cur_work) || 1013 __io_wq_worker_cancel(worker, match, worker->next_work)) 1014 match->nr_running++; 1015 raw_spin_unlock(&worker->lock); 1016 1017 return match->nr_running && !match->cancel_all; 1018 } 1019 1020 static inline void io_wq_remove_pending(struct io_wq *wq, 1021 struct io_wq_work *work, 1022 struct io_wq_work_node *prev) 1023 { 1024 struct io_wq_acct *acct = io_work_get_acct(wq, work); 1025 unsigned int hash = io_get_work_hash(work); 1026 struct io_wq_work *prev_work = NULL; 1027 1028 if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) { 1029 if (prev) 1030 prev_work = container_of(prev, struct io_wq_work, list); 1031 if (prev_work && io_get_work_hash(prev_work) == hash) 1032 wq->hash_tail[hash] = prev_work; 1033 else 1034 wq->hash_tail[hash] = NULL; 1035 } 1036 wq_list_del(&acct->work_list, &work->list, prev); 1037 } 1038 1039 static bool io_acct_cancel_pending_work(struct io_wq *wq, 1040 struct io_wq_acct *acct, 1041 struct io_cb_cancel_data *match) 1042 { 1043 struct io_wq_work_node *node, *prev; 1044 struct io_wq_work *work; 1045 1046 raw_spin_lock(&acct->lock); 1047 wq_list_for_each(node, prev, &acct->work_list) { 1048 work = container_of(node, struct io_wq_work, list); 1049 if (!match->fn(work, match->data)) 1050 continue; 1051 io_wq_remove_pending(wq, work, prev); 1052 raw_spin_unlock(&acct->lock); 1053 io_run_cancel(work, wq); 1054 match->nr_pending++; 1055 /* not safe to continue after unlock */ 1056 return true; 1057 } 1058 raw_spin_unlock(&acct->lock); 1059 1060 return false; 1061 } 1062 1063 static void io_wq_cancel_pending_work(struct io_wq *wq, 1064 struct io_cb_cancel_data *match) 1065 { 1066 int i; 1067 retry: 1068 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1069 struct io_wq_acct *acct = io_get_acct(wq, i == 0); 1070 1071 if (io_acct_cancel_pending_work(wq, acct, match)) { 1072 if (match->cancel_all) 1073 goto retry; 1074 break; 1075 } 1076 } 1077 } 1078 1079 static void io_wq_cancel_running_work(struct io_wq *wq, 1080 struct io_cb_cancel_data *match) 1081 { 1082 rcu_read_lock(); 1083 io_wq_for_each_worker(wq, io_wq_worker_cancel, match); 1084 rcu_read_unlock(); 1085 } 1086 1087 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel, 1088 void *data, bool cancel_all) 1089 { 1090 struct io_cb_cancel_data match = { 1091 .fn = cancel, 1092 .data = data, 1093 .cancel_all = cancel_all, 1094 }; 1095 1096 /* 1097 * First check pending list, if we're lucky we can just remove it 1098 * from there. CANCEL_OK means that the work is returned as-new, 1099 * no completion will be posted for it. 1100 * 1101 * Then check if a free (going busy) or busy worker has the work 1102 * currently running. If we find it there, we'll return CANCEL_RUNNING 1103 * as an indication that we attempt to signal cancellation. The 1104 * completion will run normally in this case. 1105 * 1106 * Do both of these while holding the wq->lock, to ensure that 1107 * we'll find a work item regardless of state. 1108 */ 1109 io_wq_cancel_pending_work(wq, &match); 1110 if (match.nr_pending && !match.cancel_all) 1111 return IO_WQ_CANCEL_OK; 1112 1113 raw_spin_lock(&wq->lock); 1114 io_wq_cancel_running_work(wq, &match); 1115 raw_spin_unlock(&wq->lock); 1116 if (match.nr_running && !match.cancel_all) 1117 return IO_WQ_CANCEL_RUNNING; 1118 1119 if (match.nr_running) 1120 return IO_WQ_CANCEL_RUNNING; 1121 if (match.nr_pending) 1122 return IO_WQ_CANCEL_OK; 1123 return IO_WQ_CANCEL_NOTFOUND; 1124 } 1125 1126 static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode, 1127 int sync, void *key) 1128 { 1129 struct io_wq *wq = container_of(wait, struct io_wq, wait); 1130 int i; 1131 1132 list_del_init(&wait->entry); 1133 1134 rcu_read_lock(); 1135 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1136 struct io_wq_acct *acct = &wq->acct[i]; 1137 1138 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags)) 1139 io_wq_activate_free_worker(wq, acct); 1140 } 1141 rcu_read_unlock(); 1142 return 1; 1143 } 1144 1145 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data) 1146 { 1147 int ret, i; 1148 struct io_wq *wq; 1149 1150 if (WARN_ON_ONCE(!data->free_work || !data->do_work)) 1151 return ERR_PTR(-EINVAL); 1152 if (WARN_ON_ONCE(!bounded)) 1153 return ERR_PTR(-EINVAL); 1154 1155 wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL); 1156 if (!wq) 1157 return ERR_PTR(-ENOMEM); 1158 1159 refcount_inc(&data->hash->refs); 1160 wq->hash = data->hash; 1161 wq->free_work = data->free_work; 1162 wq->do_work = data->do_work; 1163 1164 ret = -ENOMEM; 1165 1166 if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL)) 1167 goto err; 1168 cpumask_copy(wq->cpu_mask, cpu_possible_mask); 1169 wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded; 1170 wq->acct[IO_WQ_ACCT_UNBOUND].max_workers = 1171 task_rlimit(current, RLIMIT_NPROC); 1172 INIT_LIST_HEAD(&wq->wait.entry); 1173 wq->wait.func = io_wq_hash_wake; 1174 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1175 struct io_wq_acct *acct = &wq->acct[i]; 1176 1177 acct->index = i; 1178 atomic_set(&acct->nr_running, 0); 1179 INIT_WQ_LIST(&acct->work_list); 1180 raw_spin_lock_init(&acct->lock); 1181 } 1182 1183 raw_spin_lock_init(&wq->lock); 1184 INIT_HLIST_NULLS_HEAD(&wq->free_list, 0); 1185 INIT_LIST_HEAD(&wq->all_list); 1186 1187 wq->task = get_task_struct(data->task); 1188 atomic_set(&wq->worker_refs, 1); 1189 init_completion(&wq->worker_done); 1190 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node); 1191 if (ret) 1192 goto err; 1193 1194 return wq; 1195 err: 1196 io_wq_put_hash(data->hash); 1197 free_cpumask_var(wq->cpu_mask); 1198 kfree(wq); 1199 return ERR_PTR(ret); 1200 } 1201 1202 static bool io_task_work_match(struct callback_head *cb, void *data) 1203 { 1204 struct io_worker *worker; 1205 1206 if (cb->func != create_worker_cb && cb->func != create_worker_cont) 1207 return false; 1208 worker = container_of(cb, struct io_worker, create_work); 1209 return worker->wq == data; 1210 } 1211 1212 void io_wq_exit_start(struct io_wq *wq) 1213 { 1214 set_bit(IO_WQ_BIT_EXIT, &wq->state); 1215 } 1216 1217 static void io_wq_cancel_tw_create(struct io_wq *wq) 1218 { 1219 struct callback_head *cb; 1220 1221 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) { 1222 struct io_worker *worker; 1223 1224 worker = container_of(cb, struct io_worker, create_work); 1225 io_worker_cancel_cb(worker); 1226 /* 1227 * Only the worker continuation helper has worker allocated and 1228 * hence needs freeing. 1229 */ 1230 if (cb->func == create_worker_cont) 1231 kfree(worker); 1232 } 1233 } 1234 1235 static void io_wq_exit_workers(struct io_wq *wq) 1236 { 1237 if (!wq->task) 1238 return; 1239 1240 io_wq_cancel_tw_create(wq); 1241 1242 rcu_read_lock(); 1243 io_wq_for_each_worker(wq, io_wq_worker_wake, NULL); 1244 rcu_read_unlock(); 1245 io_worker_ref_put(wq); 1246 wait_for_completion(&wq->worker_done); 1247 1248 spin_lock_irq(&wq->hash->wait.lock); 1249 list_del_init(&wq->wait.entry); 1250 spin_unlock_irq(&wq->hash->wait.lock); 1251 1252 put_task_struct(wq->task); 1253 wq->task = NULL; 1254 } 1255 1256 static void io_wq_destroy(struct io_wq *wq) 1257 { 1258 struct io_cb_cancel_data match = { 1259 .fn = io_wq_work_match_all, 1260 .cancel_all = true, 1261 }; 1262 1263 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node); 1264 io_wq_cancel_pending_work(wq, &match); 1265 free_cpumask_var(wq->cpu_mask); 1266 io_wq_put_hash(wq->hash); 1267 kfree(wq); 1268 } 1269 1270 void io_wq_put_and_exit(struct io_wq *wq) 1271 { 1272 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state)); 1273 1274 io_wq_exit_workers(wq); 1275 io_wq_destroy(wq); 1276 } 1277 1278 struct online_data { 1279 unsigned int cpu; 1280 bool online; 1281 }; 1282 1283 static bool io_wq_worker_affinity(struct io_worker *worker, void *data) 1284 { 1285 struct online_data *od = data; 1286 1287 if (od->online) 1288 cpumask_set_cpu(od->cpu, worker->wq->cpu_mask); 1289 else 1290 cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask); 1291 return false; 1292 } 1293 1294 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online) 1295 { 1296 struct online_data od = { 1297 .cpu = cpu, 1298 .online = online 1299 }; 1300 1301 rcu_read_lock(); 1302 io_wq_for_each_worker(wq, io_wq_worker_affinity, &od); 1303 rcu_read_unlock(); 1304 return 0; 1305 } 1306 1307 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node) 1308 { 1309 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); 1310 1311 return __io_wq_cpu_online(wq, cpu, true); 1312 } 1313 1314 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node) 1315 { 1316 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node); 1317 1318 return __io_wq_cpu_online(wq, cpu, false); 1319 } 1320 1321 int io_wq_cpu_affinity(struct io_uring_task *tctx, cpumask_var_t mask) 1322 { 1323 if (!tctx || !tctx->io_wq) 1324 return -EINVAL; 1325 1326 rcu_read_lock(); 1327 if (mask) 1328 cpumask_copy(tctx->io_wq->cpu_mask, mask); 1329 else 1330 cpumask_copy(tctx->io_wq->cpu_mask, cpu_possible_mask); 1331 rcu_read_unlock(); 1332 1333 return 0; 1334 } 1335 1336 /* 1337 * Set max number of unbounded workers, returns old value. If new_count is 0, 1338 * then just return the old value. 1339 */ 1340 int io_wq_max_workers(struct io_wq *wq, int *new_count) 1341 { 1342 struct io_wq_acct *acct; 1343 int prev[IO_WQ_ACCT_NR]; 1344 int i; 1345 1346 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND); 1347 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND); 1348 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2); 1349 1350 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1351 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC)) 1352 new_count[i] = task_rlimit(current, RLIMIT_NPROC); 1353 } 1354 1355 for (i = 0; i < IO_WQ_ACCT_NR; i++) 1356 prev[i] = 0; 1357 1358 rcu_read_lock(); 1359 1360 raw_spin_lock(&wq->lock); 1361 for (i = 0; i < IO_WQ_ACCT_NR; i++) { 1362 acct = &wq->acct[i]; 1363 prev[i] = max_t(int, acct->max_workers, prev[i]); 1364 if (new_count[i]) 1365 acct->max_workers = new_count[i]; 1366 } 1367 raw_spin_unlock(&wq->lock); 1368 rcu_read_unlock(); 1369 1370 for (i = 0; i < IO_WQ_ACCT_NR; i++) 1371 new_count[i] = prev[i]; 1372 1373 return 0; 1374 } 1375 1376 static __init int io_wq_init(void) 1377 { 1378 int ret; 1379 1380 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online", 1381 io_wq_cpu_online, io_wq_cpu_offline); 1382 if (ret < 0) 1383 return ret; 1384 io_wq_online = ret; 1385 return 0; 1386 } 1387 subsys_initcall(io_wq_init); 1388