Linux Audio

Check our new training course

Loading...
Note: File does not exist in v5.4.
   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
  27enum {
  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
  34enum {
  35	IO_WQ_BIT_EXIT		= 0,	/* wq exiting */
  36};
  37
  38enum {
  39	IO_ACCT_STALLED_BIT	= 0,	/* stalled on hash */
  40};
  41
  42/*
  43 * One for each thread in a wq pool
  44 */
  45struct 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
  77struct 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
  87enum {
  88	IO_WQ_ACCT_BOUND,
  89	IO_WQ_ACCT_UNBOUND,
  90	IO_WQ_ACCT_NR,
  91};
  92
  93/*
  94 * Per io_wq state
  95  */
  96struct 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
 126static enum cpuhp_state io_wq_online;
 127
 128struct 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
 136static bool create_io_worker(struct io_wq *wq, int index);
 137static void io_wq_dec_running(struct io_worker *worker);
 138static bool io_acct_cancel_pending_work(struct io_wq *wq,
 139					struct io_wq_acct *acct,
 140					struct io_cb_cancel_data *match);
 141static void create_worker_cb(struct callback_head *cb);
 142static void io_wq_cancel_tw_create(struct io_wq *wq);
 143
 144static bool io_worker_get(struct io_worker *worker)
 145{
 146	return refcount_inc_not_zero(&worker->ref);
 147}
 148
 149static void io_worker_release(struct io_worker *worker)
 150{
 151	if (refcount_dec_and_test(&worker->ref))
 152		complete(&worker->ref_done);
 153}
 154
 155static 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
 160static 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
 166static 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
 171static 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
 177bool 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
 187static 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
 201static 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
 211static 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 (worker->flags & IO_WORKER_F_FREE)
 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
 245static 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 */
 255static 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 */
 270static 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 */
 306static 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
 327static 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
 334static 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
 362static 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);
 400fail_release:
 401	io_worker_release(worker);
 402fail:
 403	atomic_dec(&acct->nr_running);
 404	io_worker_ref_put(wq);
 405	return false;
 406}
 407
 408static 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 (!(worker->flags & IO_WORKER_F_UP))
 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 */
 431static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker)
 432{
 433	if (worker->flags & IO_WORKER_F_FREE) {
 434		worker->flags &= ~IO_WORKER_F_FREE;
 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 */
 444static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker)
 445	__must_hold(wq->lock)
 446{
 447	if (!(worker->flags & IO_WORKER_F_FREE)) {
 448		worker->flags |= IO_WORKER_F_FREE;
 449		hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
 450	}
 451}
 452
 453static inline unsigned int io_get_work_hash(struct io_wq_work *work)
 454{
 455	return work->flags >> IO_WQ_HASH_SHIFT;
 456}
 457
 458static 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
 475static 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
 532static 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 */
 549static 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
 629static 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	worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 638
 639	snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
 640	set_task_comm(current, buf);
 641
 642	while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 643		long ret;
 644
 645		set_current_state(TASK_INTERRUPTIBLE);
 646
 647		/*
 648		 * If we have work to do, io_acct_run_queue() returns with
 649		 * the acct->lock held. If not, it will drop it.
 650		 */
 651		while (io_acct_run_queue(acct))
 652			io_worker_handle_work(acct, worker);
 653
 654		raw_spin_lock(&wq->lock);
 655		/*
 656		 * Last sleep timed out. Exit if we're not the last worker,
 657		 * or if someone modified our affinity.
 658		 */
 659		if (last_timeout && (exit_mask || acct->nr_workers > 1)) {
 660			acct->nr_workers--;
 661			raw_spin_unlock(&wq->lock);
 662			__set_current_state(TASK_RUNNING);
 663			break;
 664		}
 665		last_timeout = false;
 666		__io_worker_idle(wq, worker);
 667		raw_spin_unlock(&wq->lock);
 668		if (io_run_task_work())
 669			continue;
 670		ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
 671		if (signal_pending(current)) {
 672			struct ksignal ksig;
 673
 674			if (!get_signal(&ksig))
 675				continue;
 676			break;
 677		}
 678		if (!ret) {
 679			last_timeout = true;
 680			exit_mask = !cpumask_test_cpu(raw_smp_processor_id(),
 681							wq->cpu_mask);
 682		}
 683	}
 684
 685	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && io_acct_run_queue(acct))
 686		io_worker_handle_work(acct, worker);
 687
 688	io_worker_exit(worker);
 689	return 0;
 690}
 691
 692/*
 693 * Called when a worker is scheduled in. Mark us as currently running.
 694 */
 695void io_wq_worker_running(struct task_struct *tsk)
 696{
 697	struct io_worker *worker = tsk->worker_private;
 698
 699	if (!worker)
 700		return;
 701	if (!(worker->flags & IO_WORKER_F_UP))
 702		return;
 703	if (worker->flags & IO_WORKER_F_RUNNING)
 704		return;
 705	worker->flags |= IO_WORKER_F_RUNNING;
 706	io_wq_inc_running(worker);
 707}
 708
 709/*
 710 * Called when worker is going to sleep. If there are no workers currently
 711 * running and we have work pending, wake up a free one or create a new one.
 712 */
 713void io_wq_worker_sleeping(struct task_struct *tsk)
 714{
 715	struct io_worker *worker = tsk->worker_private;
 716
 717	if (!worker)
 718		return;
 719	if (!(worker->flags & IO_WORKER_F_UP))
 720		return;
 721	if (!(worker->flags & IO_WORKER_F_RUNNING))
 722		return;
 723
 724	worker->flags &= ~IO_WORKER_F_RUNNING;
 725	io_wq_dec_running(worker);
 726}
 727
 728static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker,
 729			       struct task_struct *tsk)
 730{
 731	tsk->worker_private = worker;
 732	worker->task = tsk;
 733	set_cpus_allowed_ptr(tsk, wq->cpu_mask);
 734
 735	raw_spin_lock(&wq->lock);
 736	hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
 737	list_add_tail_rcu(&worker->all_list, &wq->all_list);
 738	worker->flags |= IO_WORKER_F_FREE;
 739	raw_spin_unlock(&wq->lock);
 740	wake_up_new_task(tsk);
 741}
 742
 743static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
 744{
 745	return true;
 746}
 747
 748static inline bool io_should_retry_thread(long err)
 749{
 750	/*
 751	 * Prevent perpetual task_work retry, if the task (or its group) is
 752	 * exiting.
 753	 */
 754	if (fatal_signal_pending(current))
 755		return false;
 756
 757	switch (err) {
 758	case -EAGAIN:
 759	case -ERESTARTSYS:
 760	case -ERESTARTNOINTR:
 761	case -ERESTARTNOHAND:
 762		return true;
 763	default:
 764		return false;
 765	}
 766}
 767
 768static void create_worker_cont(struct callback_head *cb)
 769{
 770	struct io_worker *worker;
 771	struct task_struct *tsk;
 772	struct io_wq *wq;
 773
 774	worker = container_of(cb, struct io_worker, create_work);
 775	clear_bit_unlock(0, &worker->create_state);
 776	wq = worker->wq;
 777	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
 778	if (!IS_ERR(tsk)) {
 779		io_init_new_worker(wq, worker, tsk);
 780		io_worker_release(worker);
 781		return;
 782	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
 783		struct io_wq_acct *acct = io_wq_get_acct(worker);
 784
 785		atomic_dec(&acct->nr_running);
 786		raw_spin_lock(&wq->lock);
 787		acct->nr_workers--;
 788		if (!acct->nr_workers) {
 789			struct io_cb_cancel_data match = {
 790				.fn		= io_wq_work_match_all,
 791				.cancel_all	= true,
 792			};
 793
 794			raw_spin_unlock(&wq->lock);
 795			while (io_acct_cancel_pending_work(wq, acct, &match))
 796				;
 797		} else {
 798			raw_spin_unlock(&wq->lock);
 799		}
 800		io_worker_ref_put(wq);
 801		kfree(worker);
 802		return;
 803	}
 804
 805	/* re-create attempts grab a new worker ref, drop the existing one */
 806	io_worker_release(worker);
 807	schedule_work(&worker->work);
 808}
 809
 810static void io_workqueue_create(struct work_struct *work)
 811{
 812	struct io_worker *worker = container_of(work, struct io_worker, work);
 813	struct io_wq_acct *acct = io_wq_get_acct(worker);
 814
 815	if (!io_queue_worker_create(worker, acct, create_worker_cont))
 816		kfree(worker);
 817}
 818
 819static bool create_io_worker(struct io_wq *wq, int index)
 820{
 821	struct io_wq_acct *acct = &wq->acct[index];
 822	struct io_worker *worker;
 823	struct task_struct *tsk;
 824
 825	__set_current_state(TASK_RUNNING);
 826
 827	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
 828	if (!worker) {
 829fail:
 830		atomic_dec(&acct->nr_running);
 831		raw_spin_lock(&wq->lock);
 832		acct->nr_workers--;
 833		raw_spin_unlock(&wq->lock);
 834		io_worker_ref_put(wq);
 835		return false;
 836	}
 837
 838	refcount_set(&worker->ref, 1);
 839	worker->wq = wq;
 840	raw_spin_lock_init(&worker->lock);
 841	init_completion(&worker->ref_done);
 842
 843	if (index == IO_WQ_ACCT_BOUND)
 844		worker->flags |= IO_WORKER_F_BOUND;
 845
 846	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
 847	if (!IS_ERR(tsk)) {
 848		io_init_new_worker(wq, worker, tsk);
 849	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
 850		kfree(worker);
 851		goto fail;
 852	} else {
 853		INIT_WORK(&worker->work, io_workqueue_create);
 854		schedule_work(&worker->work);
 855	}
 856
 857	return true;
 858}
 859
 860/*
 861 * Iterate the passed in list and call the specific function for each
 862 * worker that isn't exiting
 863 */
 864static bool io_wq_for_each_worker(struct io_wq *wq,
 865				  bool (*func)(struct io_worker *, void *),
 866				  void *data)
 867{
 868	struct io_worker *worker;
 869	bool ret = false;
 870
 871	list_for_each_entry_rcu(worker, &wq->all_list, all_list) {
 872		if (io_worker_get(worker)) {
 873			/* no task if node is/was offline */
 874			if (worker->task)
 875				ret = func(worker, data);
 876			io_worker_release(worker);
 877			if (ret)
 878				break;
 879		}
 880	}
 881
 882	return ret;
 883}
 884
 885static bool io_wq_worker_wake(struct io_worker *worker, void *data)
 886{
 887	__set_notify_signal(worker->task);
 888	wake_up_process(worker->task);
 889	return false;
 890}
 891
 892static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq)
 893{
 894	do {
 895		work->flags |= IO_WQ_WORK_CANCEL;
 896		wq->do_work(work);
 897		work = wq->free_work(work);
 898	} while (work);
 899}
 900
 901static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work)
 902{
 903	struct io_wq_acct *acct = io_work_get_acct(wq, work);
 904	unsigned int hash;
 905	struct io_wq_work *tail;
 906
 907	if (!io_wq_is_hashed(work)) {
 908append:
 909		wq_list_add_tail(&work->list, &acct->work_list);
 910		return;
 911	}
 912
 913	hash = io_get_work_hash(work);
 914	tail = wq->hash_tail[hash];
 915	wq->hash_tail[hash] = work;
 916	if (!tail)
 917		goto append;
 918
 919	wq_list_add_after(&work->list, &tail->list, &acct->work_list);
 920}
 921
 922static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
 923{
 924	return work == data;
 925}
 926
 927void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 928{
 929	struct io_wq_acct *acct = io_work_get_acct(wq, work);
 930	struct io_cb_cancel_data match;
 931	unsigned work_flags = work->flags;
 932	bool do_create;
 933
 934	/*
 935	 * If io-wq is exiting for this task, or if the request has explicitly
 936	 * been marked as one that should not get executed, cancel it here.
 937	 */
 938	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 939	    (work->flags & IO_WQ_WORK_CANCEL)) {
 940		io_run_cancel(work, wq);
 941		return;
 942	}
 943
 944	raw_spin_lock(&acct->lock);
 945	io_wq_insert_work(wq, work);
 946	clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 947	raw_spin_unlock(&acct->lock);
 948
 949	rcu_read_lock();
 950	do_create = !io_wq_activate_free_worker(wq, acct);
 951	rcu_read_unlock();
 952
 953	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 954	    !atomic_read(&acct->nr_running))) {
 955		bool did_create;
 956
 957		did_create = io_wq_create_worker(wq, acct);
 958		if (likely(did_create))
 959			return;
 960
 961		raw_spin_lock(&wq->lock);
 962		if (acct->nr_workers) {
 963			raw_spin_unlock(&wq->lock);
 964			return;
 965		}
 966		raw_spin_unlock(&wq->lock);
 967
 968		/* fatal condition, failed to create the first worker */
 969		match.fn		= io_wq_work_match_item,
 970		match.data		= work,
 971		match.cancel_all	= false,
 972
 973		io_acct_cancel_pending_work(wq, acct, &match);
 974	}
 975}
 976
 977/*
 978 * Work items that hash to the same value will not be done in parallel.
 979 * Used to limit concurrent writes, generally hashed by inode.
 980 */
 981void io_wq_hash_work(struct io_wq_work *work, void *val)
 982{
 983	unsigned int bit;
 984
 985	bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 986	work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 987}
 988
 989static bool __io_wq_worker_cancel(struct io_worker *worker,
 990				  struct io_cb_cancel_data *match,
 991				  struct io_wq_work *work)
 992{
 993	if (work && match->fn(work, match->data)) {
 994		work->flags |= IO_WQ_WORK_CANCEL;
 995		__set_notify_signal(worker->task);
 996		return true;
 997	}
 998
 999	return false;
1000}
1001
1002static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
1003{
1004	struct io_cb_cancel_data *match = data;
1005
1006	/*
1007	 * Hold the lock to avoid ->cur_work going out of scope, caller
1008	 * may dereference the passed in work.
1009	 */
1010	raw_spin_lock(&worker->lock);
1011	if (__io_wq_worker_cancel(worker, match, worker->cur_work) ||
1012	    __io_wq_worker_cancel(worker, match, worker->next_work))
1013		match->nr_running++;
1014	raw_spin_unlock(&worker->lock);
1015
1016	return match->nr_running && !match->cancel_all;
1017}
1018
1019static inline void io_wq_remove_pending(struct io_wq *wq,
1020					 struct io_wq_work *work,
1021					 struct io_wq_work_node *prev)
1022{
1023	struct io_wq_acct *acct = io_work_get_acct(wq, work);
1024	unsigned int hash = io_get_work_hash(work);
1025	struct io_wq_work *prev_work = NULL;
1026
1027	if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) {
1028		if (prev)
1029			prev_work = container_of(prev, struct io_wq_work, list);
1030		if (prev_work && io_get_work_hash(prev_work) == hash)
1031			wq->hash_tail[hash] = prev_work;
1032		else
1033			wq->hash_tail[hash] = NULL;
1034	}
1035	wq_list_del(&acct->work_list, &work->list, prev);
1036}
1037
1038static bool io_acct_cancel_pending_work(struct io_wq *wq,
1039					struct io_wq_acct *acct,
1040					struct io_cb_cancel_data *match)
1041{
1042	struct io_wq_work_node *node, *prev;
1043	struct io_wq_work *work;
1044
1045	raw_spin_lock(&acct->lock);
1046	wq_list_for_each(node, prev, &acct->work_list) {
1047		work = container_of(node, struct io_wq_work, list);
1048		if (!match->fn(work, match->data))
1049			continue;
1050		io_wq_remove_pending(wq, work, prev);
1051		raw_spin_unlock(&acct->lock);
1052		io_run_cancel(work, wq);
1053		match->nr_pending++;
1054		/* not safe to continue after unlock */
1055		return true;
1056	}
1057	raw_spin_unlock(&acct->lock);
1058
1059	return false;
1060}
1061
1062static void io_wq_cancel_pending_work(struct io_wq *wq,
1063				      struct io_cb_cancel_data *match)
1064{
1065	int i;
1066retry:
1067	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1068		struct io_wq_acct *acct = io_get_acct(wq, i == 0);
1069
1070		if (io_acct_cancel_pending_work(wq, acct, match)) {
1071			if (match->cancel_all)
1072				goto retry;
1073			break;
1074		}
1075	}
1076}
1077
1078static void io_wq_cancel_running_work(struct io_wq *wq,
1079				       struct io_cb_cancel_data *match)
1080{
1081	rcu_read_lock();
1082	io_wq_for_each_worker(wq, io_wq_worker_cancel, match);
1083	rcu_read_unlock();
1084}
1085
1086enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1087				  void *data, bool cancel_all)
1088{
1089	struct io_cb_cancel_data match = {
1090		.fn		= cancel,
1091		.data		= data,
1092		.cancel_all	= cancel_all,
1093	};
1094
1095	/*
1096	 * First check pending list, if we're lucky we can just remove it
1097	 * from there. CANCEL_OK means that the work is returned as-new,
1098	 * no completion will be posted for it.
1099	 *
1100	 * Then check if a free (going busy) or busy worker has the work
1101	 * currently running. If we find it there, we'll return CANCEL_RUNNING
1102	 * as an indication that we attempt to signal cancellation. The
1103	 * completion will run normally in this case.
1104	 *
1105	 * Do both of these while holding the wq->lock, to ensure that
1106	 * we'll find a work item regardless of state.
1107	 */
1108	io_wq_cancel_pending_work(wq, &match);
1109	if (match.nr_pending && !match.cancel_all)
1110		return IO_WQ_CANCEL_OK;
1111
1112	raw_spin_lock(&wq->lock);
1113	io_wq_cancel_running_work(wq, &match);
1114	raw_spin_unlock(&wq->lock);
1115	if (match.nr_running && !match.cancel_all)
1116		return IO_WQ_CANCEL_RUNNING;
1117
1118	if (match.nr_running)
1119		return IO_WQ_CANCEL_RUNNING;
1120	if (match.nr_pending)
1121		return IO_WQ_CANCEL_OK;
1122	return IO_WQ_CANCEL_NOTFOUND;
1123}
1124
1125static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1126			    int sync, void *key)
1127{
1128	struct io_wq *wq = container_of(wait, struct io_wq, wait);
1129	int i;
1130
1131	list_del_init(&wait->entry);
1132
1133	rcu_read_lock();
1134	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1135		struct io_wq_acct *acct = &wq->acct[i];
1136
1137		if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1138			io_wq_activate_free_worker(wq, acct);
1139	}
1140	rcu_read_unlock();
1141	return 1;
1142}
1143
1144struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1145{
1146	int ret, i;
1147	struct io_wq *wq;
1148
1149	if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1150		return ERR_PTR(-EINVAL);
1151	if (WARN_ON_ONCE(!bounded))
1152		return ERR_PTR(-EINVAL);
1153
1154	wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL);
1155	if (!wq)
1156		return ERR_PTR(-ENOMEM);
1157
1158	refcount_inc(&data->hash->refs);
1159	wq->hash = data->hash;
1160	wq->free_work = data->free_work;
1161	wq->do_work = data->do_work;
1162
1163	ret = -ENOMEM;
1164
1165	if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL))
1166		goto err;
1167	cpumask_copy(wq->cpu_mask, cpu_possible_mask);
1168	wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1169	wq->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1170				task_rlimit(current, RLIMIT_NPROC);
1171	INIT_LIST_HEAD(&wq->wait.entry);
1172	wq->wait.func = io_wq_hash_wake;
1173	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1174		struct io_wq_acct *acct = &wq->acct[i];
1175
1176		acct->index = i;
1177		atomic_set(&acct->nr_running, 0);
1178		INIT_WQ_LIST(&acct->work_list);
1179		raw_spin_lock_init(&acct->lock);
1180	}
1181
1182	raw_spin_lock_init(&wq->lock);
1183	INIT_HLIST_NULLS_HEAD(&wq->free_list, 0);
1184	INIT_LIST_HEAD(&wq->all_list);
1185
1186	wq->task = get_task_struct(data->task);
1187	atomic_set(&wq->worker_refs, 1);
1188	init_completion(&wq->worker_done);
1189	ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1190	if (ret)
1191		goto err;
1192
1193	return wq;
1194err:
1195	io_wq_put_hash(data->hash);
1196	free_cpumask_var(wq->cpu_mask);
1197	kfree(wq);
1198	return ERR_PTR(ret);
1199}
1200
1201static bool io_task_work_match(struct callback_head *cb, void *data)
1202{
1203	struct io_worker *worker;
1204
1205	if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1206		return false;
1207	worker = container_of(cb, struct io_worker, create_work);
1208	return worker->wq == data;
1209}
1210
1211void io_wq_exit_start(struct io_wq *wq)
1212{
1213	set_bit(IO_WQ_BIT_EXIT, &wq->state);
1214}
1215
1216static void io_wq_cancel_tw_create(struct io_wq *wq)
1217{
1218	struct callback_head *cb;
1219
1220	while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1221		struct io_worker *worker;
1222
1223		worker = container_of(cb, struct io_worker, create_work);
1224		io_worker_cancel_cb(worker);
1225		/*
1226		 * Only the worker continuation helper has worker allocated and
1227		 * hence needs freeing.
1228		 */
1229		if (cb->func == create_worker_cont)
1230			kfree(worker);
1231	}
1232}
1233
1234static void io_wq_exit_workers(struct io_wq *wq)
1235{
1236	if (!wq->task)
1237		return;
1238
1239	io_wq_cancel_tw_create(wq);
1240
1241	rcu_read_lock();
1242	io_wq_for_each_worker(wq, io_wq_worker_wake, NULL);
1243	rcu_read_unlock();
1244	io_worker_ref_put(wq);
1245	wait_for_completion(&wq->worker_done);
1246
1247	spin_lock_irq(&wq->hash->wait.lock);
1248	list_del_init(&wq->wait.entry);
1249	spin_unlock_irq(&wq->hash->wait.lock);
1250
1251	put_task_struct(wq->task);
1252	wq->task = NULL;
1253}
1254
1255static void io_wq_destroy(struct io_wq *wq)
1256{
1257	struct io_cb_cancel_data match = {
1258		.fn		= io_wq_work_match_all,
1259		.cancel_all	= true,
1260	};
1261
1262	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1263	io_wq_cancel_pending_work(wq, &match);
1264	free_cpumask_var(wq->cpu_mask);
1265	io_wq_put_hash(wq->hash);
1266	kfree(wq);
1267}
1268
1269void io_wq_put_and_exit(struct io_wq *wq)
1270{
1271	WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1272
1273	io_wq_exit_workers(wq);
1274	io_wq_destroy(wq);
1275}
1276
1277struct online_data {
1278	unsigned int cpu;
1279	bool online;
1280};
1281
1282static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1283{
1284	struct online_data *od = data;
1285
1286	if (od->online)
1287		cpumask_set_cpu(od->cpu, worker->wq->cpu_mask);
1288	else
1289		cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask);
1290	return false;
1291}
1292
1293static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1294{
1295	struct online_data od = {
1296		.cpu = cpu,
1297		.online = online
1298	};
1299
1300	rcu_read_lock();
1301	io_wq_for_each_worker(wq, io_wq_worker_affinity, &od);
1302	rcu_read_unlock();
1303	return 0;
1304}
1305
1306static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1307{
1308	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1309
1310	return __io_wq_cpu_online(wq, cpu, true);
1311}
1312
1313static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1314{
1315	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1316
1317	return __io_wq_cpu_online(wq, cpu, false);
1318}
1319
1320int io_wq_cpu_affinity(struct io_uring_task *tctx, cpumask_var_t mask)
1321{
1322	if (!tctx || !tctx->io_wq)
1323		return -EINVAL;
1324
1325	rcu_read_lock();
1326	if (mask)
1327		cpumask_copy(tctx->io_wq->cpu_mask, mask);
1328	else
1329		cpumask_copy(tctx->io_wq->cpu_mask, cpu_possible_mask);
1330	rcu_read_unlock();
1331
1332	return 0;
1333}
1334
1335/*
1336 * Set max number of unbounded workers, returns old value. If new_count is 0,
1337 * then just return the old value.
1338 */
1339int io_wq_max_workers(struct io_wq *wq, int *new_count)
1340{
1341	struct io_wq_acct *acct;
1342	int prev[IO_WQ_ACCT_NR];
1343	int i;
1344
1345	BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND   != (int) IO_WQ_BOUND);
1346	BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1347	BUILD_BUG_ON((int) IO_WQ_ACCT_NR      != 2);
1348
1349	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1350		if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1351			new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1352	}
1353
1354	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1355		prev[i] = 0;
1356
1357	rcu_read_lock();
1358
1359	raw_spin_lock(&wq->lock);
1360	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1361		acct = &wq->acct[i];
1362		prev[i] = max_t(int, acct->max_workers, prev[i]);
1363		if (new_count[i])
1364			acct->max_workers = new_count[i];
1365	}
1366	raw_spin_unlock(&wq->lock);
1367	rcu_read_unlock();
1368
1369	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1370		new_count[i] = prev[i];
1371
1372	return 0;
1373}
1374
1375static __init int io_wq_init(void)
1376{
1377	int ret;
1378
1379	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1380					io_wq_cpu_online, io_wq_cpu_offline);
1381	if (ret < 0)
1382		return ret;
1383	io_wq_online = ret;
1384	return 0;
1385}
1386subsys_initcall(io_wq_init);