Linux Audio

Check our new training course

Loading...
v6.8
   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		raw_spin_unlock(&acct->lock);
 568		if (work) {
 569			__io_worker_busy(wq, worker);
 570
 571			/*
 572			 * Make sure cancelation can find this, even before
 573			 * it becomes the active work. That avoids a window
 574			 * where the work has been removed from our general
 575			 * work list, but isn't yet discoverable as the
 576			 * current work item for this worker.
 577			 */
 578			raw_spin_lock(&worker->lock);
 579			worker->next_work = work;
 580			raw_spin_unlock(&worker->lock);
 581		} else {
 582			break;
 583		}
 
 
 
 
 
 
 
 
 584		io_assign_current_work(worker, work);
 585		__set_current_state(TASK_RUNNING);
 586
 587		/* handle a whole dependent link */
 588		do {
 589			struct io_wq_work *next_hashed, *linked;
 590			unsigned int hash = io_get_work_hash(work);
 591
 592			next_hashed = wq_next_work(work);
 593
 594			if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
 595				work->flags |= IO_WQ_WORK_CANCEL;
 
 596			wq->do_work(work);
 597			io_assign_current_work(worker, NULL);
 598
 599			linked = wq->free_work(work);
 600			work = next_hashed;
 601			if (!work && linked && !io_wq_is_hashed(linked)) {
 602				work = linked;
 603				linked = NULL;
 604			}
 605			io_assign_current_work(worker, work);
 606			if (linked)
 607				io_wq_enqueue(wq, linked);
 608
 609			if (hash != -1U && !next_hashed) {
 610				/* serialize hash clear with wake_up() */
 611				spin_lock_irq(&wq->hash->wait.lock);
 612				clear_bit(hash, &wq->hash->map);
 613				clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 614				spin_unlock_irq(&wq->hash->wait.lock);
 615				if (wq_has_sleeper(&wq->hash->wait))
 616					wake_up(&wq->hash->wait);
 617			}
 618		} while (work);
 619
 620		if (!__io_acct_run_queue(acct))
 621			break;
 622		raw_spin_lock(&acct->lock);
 623	} while (1);
 624}
 625
 626static int io_wq_worker(void *data)
 627{
 628	struct io_worker *worker = data;
 629	struct io_wq_acct *acct = io_wq_get_acct(worker);
 630	struct io_wq *wq = worker->wq;
 631	bool exit_mask = false, last_timeout = false;
 632	char buf[TASK_COMM_LEN];
 633
 634	worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 
 635
 636	snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
 637	set_task_comm(current, buf);
 638
 639	while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 640		long ret;
 641
 642		set_current_state(TASK_INTERRUPTIBLE);
 643
 644		/*
 645		 * If we have work to do, io_acct_run_queue() returns with
 646		 * the acct->lock held. If not, it will drop it.
 647		 */
 648		while (io_acct_run_queue(acct))
 649			io_worker_handle_work(acct, worker);
 650
 651		raw_spin_lock(&wq->lock);
 652		/*
 653		 * Last sleep timed out. Exit if we're not the last worker,
 654		 * or if someone modified our affinity.
 655		 */
 656		if (last_timeout && (exit_mask || acct->nr_workers > 1)) {
 657			acct->nr_workers--;
 658			raw_spin_unlock(&wq->lock);
 659			__set_current_state(TASK_RUNNING);
 660			break;
 661		}
 662		last_timeout = false;
 663		__io_worker_idle(wq, worker);
 664		raw_spin_unlock(&wq->lock);
 665		if (io_run_task_work())
 666			continue;
 667		ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
 668		if (signal_pending(current)) {
 669			struct ksignal ksig;
 670
 671			if (!get_signal(&ksig))
 672				continue;
 673			break;
 674		}
 675		if (!ret) {
 676			last_timeout = true;
 677			exit_mask = !cpumask_test_cpu(raw_smp_processor_id(),
 678							wq->cpu_mask);
 679		}
 680	}
 681
 682	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && io_acct_run_queue(acct))
 683		io_worker_handle_work(acct, worker);
 684
 685	io_worker_exit(worker);
 686	return 0;
 687}
 688
 689/*
 690 * Called when a worker is scheduled in. Mark us as currently running.
 691 */
 692void io_wq_worker_running(struct task_struct *tsk)
 693{
 694	struct io_worker *worker = tsk->worker_private;
 695
 696	if (!worker)
 697		return;
 698	if (!(worker->flags & IO_WORKER_F_UP))
 699		return;
 700	if (worker->flags & IO_WORKER_F_RUNNING)
 701		return;
 702	worker->flags |= IO_WORKER_F_RUNNING;
 703	io_wq_inc_running(worker);
 704}
 705
 706/*
 707 * Called when worker is going to sleep. If there are no workers currently
 708 * running and we have work pending, wake up a free one or create a new one.
 709 */
 710void io_wq_worker_sleeping(struct task_struct *tsk)
 711{
 712	struct io_worker *worker = tsk->worker_private;
 713
 714	if (!worker)
 715		return;
 716	if (!(worker->flags & IO_WORKER_F_UP))
 717		return;
 718	if (!(worker->flags & IO_WORKER_F_RUNNING))
 719		return;
 720
 721	worker->flags &= ~IO_WORKER_F_RUNNING;
 722	io_wq_dec_running(worker);
 723}
 724
 725static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker,
 726			       struct task_struct *tsk)
 727{
 728	tsk->worker_private = worker;
 729	worker->task = tsk;
 730	set_cpus_allowed_ptr(tsk, wq->cpu_mask);
 731
 732	raw_spin_lock(&wq->lock);
 733	hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
 734	list_add_tail_rcu(&worker->all_list, &wq->all_list);
 735	worker->flags |= IO_WORKER_F_FREE;
 736	raw_spin_unlock(&wq->lock);
 737	wake_up_new_task(tsk);
 738}
 739
 740static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
 741{
 742	return true;
 743}
 744
 745static inline bool io_should_retry_thread(long err)
 746{
 747	/*
 748	 * Prevent perpetual task_work retry, if the task (or its group) is
 749	 * exiting.
 750	 */
 751	if (fatal_signal_pending(current))
 752		return false;
 
 
 753
 754	switch (err) {
 755	case -EAGAIN:
 756	case -ERESTARTSYS:
 757	case -ERESTARTNOINTR:
 758	case -ERESTARTNOHAND:
 759		return true;
 760	default:
 761		return false;
 762	}
 763}
 764
 765static void create_worker_cont(struct callback_head *cb)
 766{
 767	struct io_worker *worker;
 768	struct task_struct *tsk;
 769	struct io_wq *wq;
 770
 771	worker = container_of(cb, struct io_worker, create_work);
 772	clear_bit_unlock(0, &worker->create_state);
 773	wq = worker->wq;
 774	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
 775	if (!IS_ERR(tsk)) {
 776		io_init_new_worker(wq, worker, tsk);
 777		io_worker_release(worker);
 778		return;
 779	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
 780		struct io_wq_acct *acct = io_wq_get_acct(worker);
 781
 782		atomic_dec(&acct->nr_running);
 783		raw_spin_lock(&wq->lock);
 784		acct->nr_workers--;
 785		if (!acct->nr_workers) {
 786			struct io_cb_cancel_data match = {
 787				.fn		= io_wq_work_match_all,
 788				.cancel_all	= true,
 789			};
 790
 791			raw_spin_unlock(&wq->lock);
 792			while (io_acct_cancel_pending_work(wq, acct, &match))
 793				;
 794		} else {
 795			raw_spin_unlock(&wq->lock);
 796		}
 797		io_worker_ref_put(wq);
 798		kfree(worker);
 799		return;
 800	}
 801
 802	/* re-create attempts grab a new worker ref, drop the existing one */
 803	io_worker_release(worker);
 804	schedule_work(&worker->work);
 805}
 806
 807static void io_workqueue_create(struct work_struct *work)
 808{
 809	struct io_worker *worker = container_of(work, struct io_worker, work);
 810	struct io_wq_acct *acct = io_wq_get_acct(worker);
 811
 812	if (!io_queue_worker_create(worker, acct, create_worker_cont))
 813		kfree(worker);
 814}
 815
 816static bool create_io_worker(struct io_wq *wq, int index)
 817{
 818	struct io_wq_acct *acct = &wq->acct[index];
 819	struct io_worker *worker;
 820	struct task_struct *tsk;
 821
 822	__set_current_state(TASK_RUNNING);
 823
 824	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
 825	if (!worker) {
 826fail:
 827		atomic_dec(&acct->nr_running);
 828		raw_spin_lock(&wq->lock);
 829		acct->nr_workers--;
 830		raw_spin_unlock(&wq->lock);
 831		io_worker_ref_put(wq);
 832		return false;
 833	}
 834
 835	refcount_set(&worker->ref, 1);
 836	worker->wq = wq;
 837	raw_spin_lock_init(&worker->lock);
 838	init_completion(&worker->ref_done);
 839
 840	if (index == IO_WQ_ACCT_BOUND)
 841		worker->flags |= IO_WORKER_F_BOUND;
 842
 843	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
 844	if (!IS_ERR(tsk)) {
 845		io_init_new_worker(wq, worker, tsk);
 846	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
 847		kfree(worker);
 848		goto fail;
 849	} else {
 850		INIT_WORK(&worker->work, io_workqueue_create);
 851		schedule_work(&worker->work);
 852	}
 853
 854	return true;
 855}
 856
 857/*
 858 * Iterate the passed in list and call the specific function for each
 859 * worker that isn't exiting
 860 */
 861static bool io_wq_for_each_worker(struct io_wq *wq,
 862				  bool (*func)(struct io_worker *, void *),
 863				  void *data)
 864{
 865	struct io_worker *worker;
 866	bool ret = false;
 867
 868	list_for_each_entry_rcu(worker, &wq->all_list, all_list) {
 869		if (io_worker_get(worker)) {
 870			/* no task if node is/was offline */
 871			if (worker->task)
 872				ret = func(worker, data);
 873			io_worker_release(worker);
 874			if (ret)
 875				break;
 876		}
 877	}
 878
 879	return ret;
 880}
 881
 882static bool io_wq_worker_wake(struct io_worker *worker, void *data)
 883{
 884	__set_notify_signal(worker->task);
 885	wake_up_process(worker->task);
 886	return false;
 887}
 888
 889static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq)
 890{
 891	do {
 892		work->flags |= IO_WQ_WORK_CANCEL;
 893		wq->do_work(work);
 894		work = wq->free_work(work);
 895	} while (work);
 896}
 897
 898static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work)
 899{
 900	struct io_wq_acct *acct = io_work_get_acct(wq, work);
 901	unsigned int hash;
 902	struct io_wq_work *tail;
 903
 904	if (!io_wq_is_hashed(work)) {
 905append:
 906		wq_list_add_tail(&work->list, &acct->work_list);
 907		return;
 908	}
 909
 910	hash = io_get_work_hash(work);
 911	tail = wq->hash_tail[hash];
 912	wq->hash_tail[hash] = work;
 913	if (!tail)
 914		goto append;
 915
 916	wq_list_add_after(&work->list, &tail->list, &acct->work_list);
 917}
 918
 919static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
 920{
 921	return work == data;
 922}
 923
 924void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 925{
 926	struct io_wq_acct *acct = io_work_get_acct(wq, work);
 927	struct io_cb_cancel_data match;
 928	unsigned work_flags = work->flags;
 
 
 
 
 929	bool do_create;
 930
 931	/*
 932	 * If io-wq is exiting for this task, or if the request has explicitly
 933	 * been marked as one that should not get executed, cancel it here.
 934	 */
 935	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 936	    (work->flags & IO_WQ_WORK_CANCEL)) {
 937		io_run_cancel(work, wq);
 938		return;
 939	}
 940
 941	raw_spin_lock(&acct->lock);
 942	io_wq_insert_work(wq, work);
 943	clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 944	raw_spin_unlock(&acct->lock);
 945
 946	rcu_read_lock();
 947	do_create = !io_wq_activate_free_worker(wq, acct);
 948	rcu_read_unlock();
 949
 950	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 951	    !atomic_read(&acct->nr_running))) {
 952		bool did_create;
 953
 954		did_create = io_wq_create_worker(wq, acct);
 955		if (likely(did_create))
 956			return;
 957
 958		raw_spin_lock(&wq->lock);
 959		if (acct->nr_workers) {
 960			raw_spin_unlock(&wq->lock);
 961			return;
 962		}
 963		raw_spin_unlock(&wq->lock);
 964
 965		/* fatal condition, failed to create the first worker */
 966		match.fn		= io_wq_work_match_item,
 967		match.data		= work,
 968		match.cancel_all	= false,
 969
 970		io_acct_cancel_pending_work(wq, acct, &match);
 971	}
 972}
 973
 974/*
 975 * Work items that hash to the same value will not be done in parallel.
 976 * Used to limit concurrent writes, generally hashed by inode.
 977 */
 978void io_wq_hash_work(struct io_wq_work *work, void *val)
 979{
 980	unsigned int bit;
 981
 982	bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 983	work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 984}
 985
 986static bool __io_wq_worker_cancel(struct io_worker *worker,
 987				  struct io_cb_cancel_data *match,
 988				  struct io_wq_work *work)
 989{
 990	if (work && match->fn(work, match->data)) {
 991		work->flags |= IO_WQ_WORK_CANCEL;
 992		__set_notify_signal(worker->task);
 993		return true;
 994	}
 995
 996	return false;
 997}
 998
 999static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
1000{
1001	struct io_cb_cancel_data *match = data;
1002
1003	/*
1004	 * Hold the lock to avoid ->cur_work going out of scope, caller
1005	 * may dereference the passed in work.
1006	 */
1007	raw_spin_lock(&worker->lock);
1008	if (__io_wq_worker_cancel(worker, match, worker->cur_work) ||
1009	    __io_wq_worker_cancel(worker, match, worker->next_work))
1010		match->nr_running++;
1011	raw_spin_unlock(&worker->lock);
1012
1013	return match->nr_running && !match->cancel_all;
1014}
1015
1016static inline void io_wq_remove_pending(struct io_wq *wq,
1017					 struct io_wq_work *work,
1018					 struct io_wq_work_node *prev)
1019{
1020	struct io_wq_acct *acct = io_work_get_acct(wq, work);
1021	unsigned int hash = io_get_work_hash(work);
1022	struct io_wq_work *prev_work = NULL;
1023
1024	if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) {
1025		if (prev)
1026			prev_work = container_of(prev, struct io_wq_work, list);
1027		if (prev_work && io_get_work_hash(prev_work) == hash)
1028			wq->hash_tail[hash] = prev_work;
1029		else
1030			wq->hash_tail[hash] = NULL;
1031	}
1032	wq_list_del(&acct->work_list, &work->list, prev);
1033}
1034
1035static bool io_acct_cancel_pending_work(struct io_wq *wq,
1036					struct io_wq_acct *acct,
1037					struct io_cb_cancel_data *match)
1038{
1039	struct io_wq_work_node *node, *prev;
1040	struct io_wq_work *work;
1041
1042	raw_spin_lock(&acct->lock);
1043	wq_list_for_each(node, prev, &acct->work_list) {
1044		work = container_of(node, struct io_wq_work, list);
1045		if (!match->fn(work, match->data))
1046			continue;
1047		io_wq_remove_pending(wq, work, prev);
1048		raw_spin_unlock(&acct->lock);
1049		io_run_cancel(work, wq);
1050		match->nr_pending++;
1051		/* not safe to continue after unlock */
1052		return true;
1053	}
1054	raw_spin_unlock(&acct->lock);
1055
1056	return false;
1057}
1058
1059static void io_wq_cancel_pending_work(struct io_wq *wq,
1060				      struct io_cb_cancel_data *match)
1061{
1062	int i;
1063retry:
1064	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1065		struct io_wq_acct *acct = io_get_acct(wq, i == 0);
1066
1067		if (io_acct_cancel_pending_work(wq, acct, match)) {
1068			if (match->cancel_all)
1069				goto retry;
1070			break;
1071		}
1072	}
1073}
1074
1075static void io_wq_cancel_running_work(struct io_wq *wq,
1076				       struct io_cb_cancel_data *match)
1077{
1078	rcu_read_lock();
1079	io_wq_for_each_worker(wq, io_wq_worker_cancel, match);
1080	rcu_read_unlock();
1081}
1082
1083enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1084				  void *data, bool cancel_all)
1085{
1086	struct io_cb_cancel_data match = {
1087		.fn		= cancel,
1088		.data		= data,
1089		.cancel_all	= cancel_all,
1090	};
1091
1092	/*
1093	 * First check pending list, if we're lucky we can just remove it
1094	 * from there. CANCEL_OK means that the work is returned as-new,
1095	 * no completion will be posted for it.
1096	 *
1097	 * Then check if a free (going busy) or busy worker has the work
1098	 * currently running. If we find it there, we'll return CANCEL_RUNNING
1099	 * as an indication that we attempt to signal cancellation. The
1100	 * completion will run normally in this case.
1101	 *
1102	 * Do both of these while holding the wq->lock, to ensure that
1103	 * we'll find a work item regardless of state.
1104	 */
1105	io_wq_cancel_pending_work(wq, &match);
1106	if (match.nr_pending && !match.cancel_all)
1107		return IO_WQ_CANCEL_OK;
1108
1109	raw_spin_lock(&wq->lock);
1110	io_wq_cancel_running_work(wq, &match);
1111	raw_spin_unlock(&wq->lock);
1112	if (match.nr_running && !match.cancel_all)
1113		return IO_WQ_CANCEL_RUNNING;
1114
1115	if (match.nr_running)
1116		return IO_WQ_CANCEL_RUNNING;
1117	if (match.nr_pending)
1118		return IO_WQ_CANCEL_OK;
1119	return IO_WQ_CANCEL_NOTFOUND;
1120}
1121
1122static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1123			    int sync, void *key)
1124{
1125	struct io_wq *wq = container_of(wait, struct io_wq, wait);
1126	int i;
1127
1128	list_del_init(&wait->entry);
1129
1130	rcu_read_lock();
1131	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1132		struct io_wq_acct *acct = &wq->acct[i];
1133
1134		if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1135			io_wq_activate_free_worker(wq, acct);
1136	}
1137	rcu_read_unlock();
1138	return 1;
1139}
1140
1141struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1142{
1143	int ret, i;
1144	struct io_wq *wq;
1145
1146	if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1147		return ERR_PTR(-EINVAL);
1148	if (WARN_ON_ONCE(!bounded))
1149		return ERR_PTR(-EINVAL);
1150
1151	wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL);
1152	if (!wq)
1153		return ERR_PTR(-ENOMEM);
1154
1155	refcount_inc(&data->hash->refs);
1156	wq->hash = data->hash;
1157	wq->free_work = data->free_work;
1158	wq->do_work = data->do_work;
1159
1160	ret = -ENOMEM;
1161
1162	if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL))
1163		goto err;
1164	cpumask_copy(wq->cpu_mask, cpu_possible_mask);
1165	wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1166	wq->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1167				task_rlimit(current, RLIMIT_NPROC);
1168	INIT_LIST_HEAD(&wq->wait.entry);
1169	wq->wait.func = io_wq_hash_wake;
1170	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1171		struct io_wq_acct *acct = &wq->acct[i];
1172
1173		acct->index = i;
1174		atomic_set(&acct->nr_running, 0);
1175		INIT_WQ_LIST(&acct->work_list);
1176		raw_spin_lock_init(&acct->lock);
1177	}
1178
1179	raw_spin_lock_init(&wq->lock);
1180	INIT_HLIST_NULLS_HEAD(&wq->free_list, 0);
1181	INIT_LIST_HEAD(&wq->all_list);
1182
1183	wq->task = get_task_struct(data->task);
1184	atomic_set(&wq->worker_refs, 1);
1185	init_completion(&wq->worker_done);
1186	ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1187	if (ret)
1188		goto err;
1189
1190	return wq;
1191err:
1192	io_wq_put_hash(data->hash);
1193	free_cpumask_var(wq->cpu_mask);
1194	kfree(wq);
1195	return ERR_PTR(ret);
1196}
1197
1198static bool io_task_work_match(struct callback_head *cb, void *data)
1199{
1200	struct io_worker *worker;
1201
1202	if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1203		return false;
1204	worker = container_of(cb, struct io_worker, create_work);
1205	return worker->wq == data;
1206}
1207
1208void io_wq_exit_start(struct io_wq *wq)
1209{
1210	set_bit(IO_WQ_BIT_EXIT, &wq->state);
1211}
1212
1213static void io_wq_cancel_tw_create(struct io_wq *wq)
1214{
1215	struct callback_head *cb;
1216
1217	while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1218		struct io_worker *worker;
1219
1220		worker = container_of(cb, struct io_worker, create_work);
1221		io_worker_cancel_cb(worker);
1222		/*
1223		 * Only the worker continuation helper has worker allocated and
1224		 * hence needs freeing.
1225		 */
1226		if (cb->func == create_worker_cont)
1227			kfree(worker);
1228	}
1229}
1230
1231static void io_wq_exit_workers(struct io_wq *wq)
1232{
1233	if (!wq->task)
1234		return;
1235
1236	io_wq_cancel_tw_create(wq);
1237
1238	rcu_read_lock();
1239	io_wq_for_each_worker(wq, io_wq_worker_wake, NULL);
1240	rcu_read_unlock();
1241	io_worker_ref_put(wq);
1242	wait_for_completion(&wq->worker_done);
1243
1244	spin_lock_irq(&wq->hash->wait.lock);
1245	list_del_init(&wq->wait.entry);
1246	spin_unlock_irq(&wq->hash->wait.lock);
1247
1248	put_task_struct(wq->task);
1249	wq->task = NULL;
1250}
1251
1252static void io_wq_destroy(struct io_wq *wq)
1253{
1254	struct io_cb_cancel_data match = {
1255		.fn		= io_wq_work_match_all,
1256		.cancel_all	= true,
1257	};
1258
1259	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1260	io_wq_cancel_pending_work(wq, &match);
1261	free_cpumask_var(wq->cpu_mask);
1262	io_wq_put_hash(wq->hash);
1263	kfree(wq);
1264}
1265
1266void io_wq_put_and_exit(struct io_wq *wq)
1267{
1268	WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1269
1270	io_wq_exit_workers(wq);
1271	io_wq_destroy(wq);
1272}
1273
1274struct online_data {
1275	unsigned int cpu;
1276	bool online;
1277};
1278
1279static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1280{
1281	struct online_data *od = data;
1282
1283	if (od->online)
1284		cpumask_set_cpu(od->cpu, worker->wq->cpu_mask);
1285	else
1286		cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask);
1287	return false;
1288}
1289
1290static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1291{
1292	struct online_data od = {
1293		.cpu = cpu,
1294		.online = online
1295	};
1296
1297	rcu_read_lock();
1298	io_wq_for_each_worker(wq, io_wq_worker_affinity, &od);
1299	rcu_read_unlock();
1300	return 0;
1301}
1302
1303static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1304{
1305	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1306
1307	return __io_wq_cpu_online(wq, cpu, true);
1308}
1309
1310static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1311{
1312	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1313
1314	return __io_wq_cpu_online(wq, cpu, false);
1315}
1316
1317int io_wq_cpu_affinity(struct io_uring_task *tctx, cpumask_var_t mask)
1318{
 
 
 
1319	if (!tctx || !tctx->io_wq)
1320		return -EINVAL;
1321
 
 
 
1322	rcu_read_lock();
1323	if (mask)
1324		cpumask_copy(tctx->io_wq->cpu_mask, mask);
1325	else
1326		cpumask_copy(tctx->io_wq->cpu_mask, cpu_possible_mask);
 
 
 
 
 
1327	rcu_read_unlock();
1328
1329	return 0;
 
1330}
1331
1332/*
1333 * Set max number of unbounded workers, returns old value. If new_count is 0,
1334 * then just return the old value.
1335 */
1336int io_wq_max_workers(struct io_wq *wq, int *new_count)
1337{
1338	struct io_wq_acct *acct;
1339	int prev[IO_WQ_ACCT_NR];
1340	int i;
1341
1342	BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND   != (int) IO_WQ_BOUND);
1343	BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1344	BUILD_BUG_ON((int) IO_WQ_ACCT_NR      != 2);
1345
1346	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1347		if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1348			new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1349	}
1350
1351	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1352		prev[i] = 0;
1353
1354	rcu_read_lock();
1355
1356	raw_spin_lock(&wq->lock);
1357	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1358		acct = &wq->acct[i];
1359		prev[i] = max_t(int, acct->max_workers, prev[i]);
1360		if (new_count[i])
1361			acct->max_workers = new_count[i];
1362	}
1363	raw_spin_unlock(&wq->lock);
1364	rcu_read_unlock();
1365
1366	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1367		new_count[i] = prev[i];
1368
1369	return 0;
1370}
1371
1372static __init int io_wq_init(void)
1373{
1374	int ret;
1375
1376	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1377					io_wq_cpu_online, io_wq_cpu_offline);
1378	if (ret < 0)
1379		return ret;
1380	io_wq_online = ret;
1381	return 0;
1382}
1383subsys_initcall(io_wq_init);
v6.13.7
   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/cpuset.h>
  17#include <linux/task_work.h>
  18#include <linux/audit.h>
  19#include <linux/mmu_context.h>
  20#include <uapi/linux/io_uring.h>
  21
  22#include "io-wq.h"
  23#include "slist.h"
  24#include "io_uring.h"
  25
  26#define WORKER_IDLE_TIMEOUT	(5 * HZ)
  27#define WORKER_INIT_LIMIT	3
  28
  29enum {
  30	IO_WORKER_F_UP		= 0,	/* up and active */
  31	IO_WORKER_F_RUNNING	= 1,	/* account as running */
  32	IO_WORKER_F_FREE	= 2,	/* worker on free list */
  33	IO_WORKER_F_BOUND	= 3,	/* is doing bounded work */
  34};
  35
  36enum {
  37	IO_WQ_BIT_EXIT		= 0,	/* wq exiting */
  38};
  39
  40enum {
  41	IO_ACCT_STALLED_BIT	= 0,	/* stalled on hash */
  42};
  43
  44/*
  45 * One for each thread in a wq pool
  46 */
  47struct io_worker {
  48	refcount_t ref;
  49	int create_index;
  50	unsigned long flags;
  51	struct hlist_nulls_node nulls_node;
  52	struct list_head all_list;
  53	struct task_struct *task;
  54	struct io_wq *wq;
  55
  56	struct io_wq_work *cur_work;
 
  57	raw_spinlock_t lock;
  58
  59	struct completion ref_done;
  60
  61	unsigned long create_state;
  62	struct callback_head create_work;
  63	int init_retries;
  64
  65	union {
  66		struct rcu_head rcu;
  67		struct work_struct work;
  68	};
  69};
  70
  71#if BITS_PER_LONG == 64
  72#define IO_WQ_HASH_ORDER	6
  73#else
  74#define IO_WQ_HASH_ORDER	5
  75#endif
  76
  77#define IO_WQ_NR_HASH_BUCKETS	(1u << IO_WQ_HASH_ORDER)
  78
  79struct io_wq_acct {
  80	unsigned nr_workers;
  81	unsigned max_workers;
  82	int index;
  83	atomic_t nr_running;
  84	raw_spinlock_t lock;
  85	struct io_wq_work_list work_list;
  86	unsigned long flags;
  87};
  88
  89enum {
  90	IO_WQ_ACCT_BOUND,
  91	IO_WQ_ACCT_UNBOUND,
  92	IO_WQ_ACCT_NR,
  93};
  94
  95/*
  96 * Per io_wq state
  97  */
  98struct io_wq {
  99	unsigned long state;
 100
 101	free_work_fn *free_work;
 102	io_wq_work_fn *do_work;
 103
 104	struct io_wq_hash *hash;
 105
 106	atomic_t worker_refs;
 107	struct completion worker_done;
 108
 109	struct hlist_node cpuhp_node;
 110
 111	struct task_struct *task;
 112
 113	struct io_wq_acct acct[IO_WQ_ACCT_NR];
 114
 115	/* lock protects access to elements below */
 116	raw_spinlock_t lock;
 117
 118	struct hlist_nulls_head free_list;
 119	struct list_head all_list;
 120
 121	struct wait_queue_entry wait;
 122
 123	struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
 124
 125	cpumask_var_t cpu_mask;
 126};
 127
 128static enum cpuhp_state io_wq_online;
 129
 130struct io_cb_cancel_data {
 131	work_cancel_fn *fn;
 132	void *data;
 133	int nr_running;
 134	int nr_pending;
 135	bool cancel_all;
 136};
 137
 138static bool create_io_worker(struct io_wq *wq, int index);
 139static void io_wq_dec_running(struct io_worker *worker);
 140static bool io_acct_cancel_pending_work(struct io_wq *wq,
 141					struct io_wq_acct *acct,
 142					struct io_cb_cancel_data *match);
 143static void create_worker_cb(struct callback_head *cb);
 144static void io_wq_cancel_tw_create(struct io_wq *wq);
 145
 146static bool io_worker_get(struct io_worker *worker)
 147{
 148	return refcount_inc_not_zero(&worker->ref);
 149}
 150
 151static void io_worker_release(struct io_worker *worker)
 152{
 153	if (refcount_dec_and_test(&worker->ref))
 154		complete(&worker->ref_done);
 155}
 156
 157static inline struct io_wq_acct *io_get_acct(struct io_wq *wq, bool bound)
 158{
 159	return &wq->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
 160}
 161
 162static inline struct io_wq_acct *io_work_get_acct(struct io_wq *wq,
 163						  struct io_wq_work *work)
 164{
 165	return io_get_acct(wq, !(atomic_read(&work->flags) & IO_WQ_WORK_UNBOUND));
 166}
 167
 168static inline struct io_wq_acct *io_wq_get_acct(struct io_worker *worker)
 169{
 170	return io_get_acct(worker->wq, test_bit(IO_WORKER_F_BOUND, &worker->flags));
 171}
 172
 173static void io_worker_ref_put(struct io_wq *wq)
 174{
 175	if (atomic_dec_and_test(&wq->worker_refs))
 176		complete(&wq->worker_done);
 177}
 178
 179bool io_wq_worker_stopped(void)
 180{
 181	struct io_worker *worker = current->worker_private;
 182
 183	if (WARN_ON_ONCE(!io_wq_current_is_worker()))
 184		return true;
 185
 186	return test_bit(IO_WQ_BIT_EXIT, &worker->wq->state);
 187}
 188
 189static void io_worker_cancel_cb(struct io_worker *worker)
 190{
 191	struct io_wq_acct *acct = io_wq_get_acct(worker);
 192	struct io_wq *wq = worker->wq;
 193
 194	atomic_dec(&acct->nr_running);
 195	raw_spin_lock(&wq->lock);
 196	acct->nr_workers--;
 197	raw_spin_unlock(&wq->lock);
 198	io_worker_ref_put(wq);
 199	clear_bit_unlock(0, &worker->create_state);
 200	io_worker_release(worker);
 201}
 202
 203static bool io_task_worker_match(struct callback_head *cb, void *data)
 204{
 205	struct io_worker *worker;
 206
 207	if (cb->func != create_worker_cb)
 208		return false;
 209	worker = container_of(cb, struct io_worker, create_work);
 210	return worker == data;
 211}
 212
 213static void io_worker_exit(struct io_worker *worker)
 214{
 215	struct io_wq *wq = worker->wq;
 216
 217	while (1) {
 218		struct callback_head *cb = task_work_cancel_match(wq->task,
 219						io_task_worker_match, worker);
 220
 221		if (!cb)
 222			break;
 223		io_worker_cancel_cb(worker);
 224	}
 225
 226	io_worker_release(worker);
 227	wait_for_completion(&worker->ref_done);
 228
 229	raw_spin_lock(&wq->lock);
 230	if (test_bit(IO_WORKER_F_FREE, &worker->flags))
 231		hlist_nulls_del_rcu(&worker->nulls_node);
 232	list_del_rcu(&worker->all_list);
 233	raw_spin_unlock(&wq->lock);
 234	io_wq_dec_running(worker);
 235	/*
 236	 * this worker is a goner, clear ->worker_private to avoid any
 237	 * inc/dec running calls that could happen as part of exit from
 238	 * touching 'worker'.
 239	 */
 240	current->worker_private = NULL;
 241
 242	kfree_rcu(worker, rcu);
 243	io_worker_ref_put(wq);
 244	do_exit(0);
 245}
 246
 247static inline bool __io_acct_run_queue(struct io_wq_acct *acct)
 248{
 249	return !test_bit(IO_ACCT_STALLED_BIT, &acct->flags) &&
 250		!wq_list_empty(&acct->work_list);
 251}
 252
 253/*
 254 * If there's work to do, returns true with acct->lock acquired. If not,
 255 * returns false with no lock held.
 256 */
 257static inline bool io_acct_run_queue(struct io_wq_acct *acct)
 258	__acquires(&acct->lock)
 259{
 260	raw_spin_lock(&acct->lock);
 261	if (__io_acct_run_queue(acct))
 262		return true;
 263
 264	raw_spin_unlock(&acct->lock);
 265	return false;
 266}
 267
 268/*
 269 * Check head of free list for an available worker. If one isn't available,
 270 * caller must create one.
 271 */
 272static bool io_wq_activate_free_worker(struct io_wq *wq,
 273					struct io_wq_acct *acct)
 274	__must_hold(RCU)
 275{
 276	struct hlist_nulls_node *n;
 277	struct io_worker *worker;
 278
 279	/*
 280	 * Iterate free_list and see if we can find an idle worker to
 281	 * activate. If a given worker is on the free_list but in the process
 282	 * of exiting, keep trying.
 283	 */
 284	hlist_nulls_for_each_entry_rcu(worker, n, &wq->free_list, nulls_node) {
 285		if (!io_worker_get(worker))
 286			continue;
 287		if (io_wq_get_acct(worker) != acct) {
 288			io_worker_release(worker);
 289			continue;
 290		}
 291		/*
 292		 * If the worker is already running, it's either already
 293		 * starting work or finishing work. In either case, if it does
 294		 * to go sleep, we'll kick off a new task for this work anyway.
 295		 */
 296		wake_up_process(worker->task);
 297		io_worker_release(worker);
 298		return true;
 299	}
 300
 301	return false;
 302}
 303
 304/*
 305 * We need a worker. If we find a free one, we're good. If not, and we're
 306 * below the max number of workers, create one.
 307 */
 308static bool io_wq_create_worker(struct io_wq *wq, struct io_wq_acct *acct)
 309{
 310	/*
 311	 * Most likely an attempt to queue unbounded work on an io_wq that
 312	 * wasn't setup with any unbounded workers.
 313	 */
 314	if (unlikely(!acct->max_workers))
 315		pr_warn_once("io-wq is not configured for unbound workers");
 316
 317	raw_spin_lock(&wq->lock);
 318	if (acct->nr_workers >= acct->max_workers) {
 319		raw_spin_unlock(&wq->lock);
 320		return true;
 321	}
 322	acct->nr_workers++;
 323	raw_spin_unlock(&wq->lock);
 324	atomic_inc(&acct->nr_running);
 325	atomic_inc(&wq->worker_refs);
 326	return create_io_worker(wq, acct->index);
 327}
 328
 329static void io_wq_inc_running(struct io_worker *worker)
 330{
 331	struct io_wq_acct *acct = io_wq_get_acct(worker);
 332
 333	atomic_inc(&acct->nr_running);
 334}
 335
 336static void create_worker_cb(struct callback_head *cb)
 337{
 338	struct io_worker *worker;
 339	struct io_wq *wq;
 340
 341	struct io_wq_acct *acct;
 342	bool do_create = false;
 343
 344	worker = container_of(cb, struct io_worker, create_work);
 345	wq = worker->wq;
 346	acct = &wq->acct[worker->create_index];
 347	raw_spin_lock(&wq->lock);
 348
 349	if (acct->nr_workers < acct->max_workers) {
 350		acct->nr_workers++;
 351		do_create = true;
 352	}
 353	raw_spin_unlock(&wq->lock);
 354	if (do_create) {
 355		create_io_worker(wq, worker->create_index);
 356	} else {
 357		atomic_dec(&acct->nr_running);
 358		io_worker_ref_put(wq);
 359	}
 360	clear_bit_unlock(0, &worker->create_state);
 361	io_worker_release(worker);
 362}
 363
 364static bool io_queue_worker_create(struct io_worker *worker,
 365				   struct io_wq_acct *acct,
 366				   task_work_func_t func)
 367{
 368	struct io_wq *wq = worker->wq;
 369
 370	/* raced with exit, just ignore create call */
 371	if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
 372		goto fail;
 373	if (!io_worker_get(worker))
 374		goto fail;
 375	/*
 376	 * create_state manages ownership of create_work/index. We should
 377	 * only need one entry per worker, as the worker going to sleep
 378	 * will trigger the condition, and waking will clear it once it
 379	 * runs the task_work.
 380	 */
 381	if (test_bit(0, &worker->create_state) ||
 382	    test_and_set_bit_lock(0, &worker->create_state))
 383		goto fail_release;
 384
 385	atomic_inc(&wq->worker_refs);
 386	init_task_work(&worker->create_work, func);
 387	worker->create_index = acct->index;
 388	if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
 389		/*
 390		 * EXIT may have been set after checking it above, check after
 391		 * adding the task_work and remove any creation item if it is
 392		 * now set. wq exit does that too, but we can have added this
 393		 * work item after we canceled in io_wq_exit_workers().
 394		 */
 395		if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
 396			io_wq_cancel_tw_create(wq);
 397		io_worker_ref_put(wq);
 398		return true;
 399	}
 400	io_worker_ref_put(wq);
 401	clear_bit_unlock(0, &worker->create_state);
 402fail_release:
 403	io_worker_release(worker);
 404fail:
 405	atomic_dec(&acct->nr_running);
 406	io_worker_ref_put(wq);
 407	return false;
 408}
 409
 410static void io_wq_dec_running(struct io_worker *worker)
 411{
 412	struct io_wq_acct *acct = io_wq_get_acct(worker);
 413	struct io_wq *wq = worker->wq;
 414
 415	if (!test_bit(IO_WORKER_F_UP, &worker->flags))
 416		return;
 417
 418	if (!atomic_dec_and_test(&acct->nr_running))
 419		return;
 420	if (!io_acct_run_queue(acct))
 421		return;
 422
 423	raw_spin_unlock(&acct->lock);
 424	atomic_inc(&acct->nr_running);
 425	atomic_inc(&wq->worker_refs);
 426	io_queue_worker_create(worker, acct, create_worker_cb);
 427}
 428
 429/*
 430 * Worker will start processing some work. Move it to the busy list, if
 431 * it's currently on the freelist
 432 */
 433static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker)
 434{
 435	if (test_bit(IO_WORKER_F_FREE, &worker->flags)) {
 436		clear_bit(IO_WORKER_F_FREE, &worker->flags);
 437		raw_spin_lock(&wq->lock);
 438		hlist_nulls_del_init_rcu(&worker->nulls_node);
 439		raw_spin_unlock(&wq->lock);
 440	}
 441}
 442
 443/*
 444 * No work, worker going to sleep. Move to freelist.
 445 */
 446static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker)
 447	__must_hold(wq->lock)
 448{
 449	if (!test_bit(IO_WORKER_F_FREE, &worker->flags)) {
 450		set_bit(IO_WORKER_F_FREE, &worker->flags);
 451		hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
 452	}
 453}
 454
 455static inline unsigned int io_get_work_hash(struct io_wq_work *work)
 456{
 457	return atomic_read(&work->flags) >> IO_WQ_HASH_SHIFT;
 458}
 459
 460static bool io_wait_on_hash(struct io_wq *wq, unsigned int hash)
 461{
 462	bool ret = false;
 463
 464	spin_lock_irq(&wq->hash->wait.lock);
 465	if (list_empty(&wq->wait.entry)) {
 466		__add_wait_queue(&wq->hash->wait, &wq->wait);
 467		if (!test_bit(hash, &wq->hash->map)) {
 468			__set_current_state(TASK_RUNNING);
 469			list_del_init(&wq->wait.entry);
 470			ret = true;
 471		}
 472	}
 473	spin_unlock_irq(&wq->hash->wait.lock);
 474	return ret;
 475}
 476
 477static struct io_wq_work *io_get_next_work(struct io_wq_acct *acct,
 478					   struct io_worker *worker)
 479	__must_hold(acct->lock)
 480{
 481	struct io_wq_work_node *node, *prev;
 482	struct io_wq_work *work, *tail;
 483	unsigned int stall_hash = -1U;
 484	struct io_wq *wq = worker->wq;
 485
 486	wq_list_for_each(node, prev, &acct->work_list) {
 487		unsigned int hash;
 488
 489		work = container_of(node, struct io_wq_work, list);
 490
 491		/* not hashed, can run anytime */
 492		if (!io_wq_is_hashed(work)) {
 493			wq_list_del(&acct->work_list, node, prev);
 494			return work;
 495		}
 496
 497		hash = io_get_work_hash(work);
 498		/* all items with this hash lie in [work, tail] */
 499		tail = wq->hash_tail[hash];
 500
 501		/* hashed, can run if not already running */
 502		if (!test_and_set_bit(hash, &wq->hash->map)) {
 503			wq->hash_tail[hash] = NULL;
 504			wq_list_cut(&acct->work_list, &tail->list, prev);
 505			return work;
 506		}
 507		if (stall_hash == -1U)
 508			stall_hash = hash;
 509		/* fast forward to a next hash, for-each will fix up @prev */
 510		node = &tail->list;
 511	}
 512
 513	if (stall_hash != -1U) {
 514		bool unstalled;
 515
 516		/*
 517		 * Set this before dropping the lock to avoid racing with new
 518		 * work being added and clearing the stalled bit.
 519		 */
 520		set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 521		raw_spin_unlock(&acct->lock);
 522		unstalled = io_wait_on_hash(wq, stall_hash);
 523		raw_spin_lock(&acct->lock);
 524		if (unstalled) {
 525			clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 526			if (wq_has_sleeper(&wq->hash->wait))
 527				wake_up(&wq->hash->wait);
 528		}
 529	}
 530
 531	return NULL;
 532}
 533
 534static void io_assign_current_work(struct io_worker *worker,
 535				   struct io_wq_work *work)
 536{
 537	if (work) {
 538		io_run_task_work();
 539		cond_resched();
 540	}
 541
 542	raw_spin_lock(&worker->lock);
 543	worker->cur_work = work;
 
 544	raw_spin_unlock(&worker->lock);
 545}
 546
 547/*
 548 * Called with acct->lock held, drops it before returning
 549 */
 550static void io_worker_handle_work(struct io_wq_acct *acct,
 551				  struct io_worker *worker)
 552	__releases(&acct->lock)
 553{
 554	struct io_wq *wq = worker->wq;
 555	bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
 556
 557	do {
 558		struct io_wq_work *work;
 559
 560		/*
 561		 * If we got some work, mark us as busy. If we didn't, but
 562		 * the list isn't empty, it means we stalled on hashed work.
 563		 * Mark us stalled so we don't keep looking for work when we
 564		 * can't make progress, any work completion or insertion will
 565		 * clear the stalled flag.
 566		 */
 567		work = io_get_next_work(acct, worker);
 
 568		if (work) {
 
 
 569			/*
 570			 * Make sure cancelation can find this, even before
 571			 * it becomes the active work. That avoids a window
 572			 * where the work has been removed from our general
 573			 * work list, but isn't yet discoverable as the
 574			 * current work item for this worker.
 575			 */
 576			raw_spin_lock(&worker->lock);
 577			worker->cur_work = work;
 578			raw_spin_unlock(&worker->lock);
 
 
 579		}
 580
 581		raw_spin_unlock(&acct->lock);
 582
 583		if (!work)
 584			break;
 585
 586		__io_worker_busy(wq, worker);
 587
 588		io_assign_current_work(worker, work);
 589		__set_current_state(TASK_RUNNING);
 590
 591		/* handle a whole dependent link */
 592		do {
 593			struct io_wq_work *next_hashed, *linked;
 594			unsigned int hash = io_get_work_hash(work);
 595
 596			next_hashed = wq_next_work(work);
 597
 598			if (do_kill &&
 599			    (atomic_read(&work->flags) & IO_WQ_WORK_UNBOUND))
 600				atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
 601			wq->do_work(work);
 602			io_assign_current_work(worker, NULL);
 603
 604			linked = wq->free_work(work);
 605			work = next_hashed;
 606			if (!work && linked && !io_wq_is_hashed(linked)) {
 607				work = linked;
 608				linked = NULL;
 609			}
 610			io_assign_current_work(worker, work);
 611			if (linked)
 612				io_wq_enqueue(wq, linked);
 613
 614			if (hash != -1U && !next_hashed) {
 615				/* serialize hash clear with wake_up() */
 616				spin_lock_irq(&wq->hash->wait.lock);
 617				clear_bit(hash, &wq->hash->map);
 618				clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 619				spin_unlock_irq(&wq->hash->wait.lock);
 620				if (wq_has_sleeper(&wq->hash->wait))
 621					wake_up(&wq->hash->wait);
 622			}
 623		} while (work);
 624
 625		if (!__io_acct_run_queue(acct))
 626			break;
 627		raw_spin_lock(&acct->lock);
 628	} while (1);
 629}
 630
 631static int io_wq_worker(void *data)
 632{
 633	struct io_worker *worker = data;
 634	struct io_wq_acct *acct = io_wq_get_acct(worker);
 635	struct io_wq *wq = worker->wq;
 636	bool exit_mask = false, last_timeout = false;
 637	char buf[TASK_COMM_LEN];
 638
 639	set_mask_bits(&worker->flags, 0,
 640		      BIT(IO_WORKER_F_UP) | BIT(IO_WORKER_F_RUNNING));
 641
 642	snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
 643	set_task_comm(current, buf);
 644
 645	while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 646		long ret;
 647
 648		set_current_state(TASK_INTERRUPTIBLE);
 649
 650		/*
 651		 * If we have work to do, io_acct_run_queue() returns with
 652		 * the acct->lock held. If not, it will drop it.
 653		 */
 654		while (io_acct_run_queue(acct))
 655			io_worker_handle_work(acct, worker);
 656
 657		raw_spin_lock(&wq->lock);
 658		/*
 659		 * Last sleep timed out. Exit if we're not the last worker,
 660		 * or if someone modified our affinity.
 661		 */
 662		if (last_timeout && (exit_mask || acct->nr_workers > 1)) {
 663			acct->nr_workers--;
 664			raw_spin_unlock(&wq->lock);
 665			__set_current_state(TASK_RUNNING);
 666			break;
 667		}
 668		last_timeout = false;
 669		__io_worker_idle(wq, worker);
 670		raw_spin_unlock(&wq->lock);
 671		if (io_run_task_work())
 672			continue;
 673		ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
 674		if (signal_pending(current)) {
 675			struct ksignal ksig;
 676
 677			if (!get_signal(&ksig))
 678				continue;
 679			break;
 680		}
 681		if (!ret) {
 682			last_timeout = true;
 683			exit_mask = !cpumask_test_cpu(raw_smp_processor_id(),
 684							wq->cpu_mask);
 685		}
 686	}
 687
 688	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && io_acct_run_queue(acct))
 689		io_worker_handle_work(acct, worker);
 690
 691	io_worker_exit(worker);
 692	return 0;
 693}
 694
 695/*
 696 * Called when a worker is scheduled in. Mark us as currently running.
 697 */
 698void io_wq_worker_running(struct task_struct *tsk)
 699{
 700	struct io_worker *worker = tsk->worker_private;
 701
 702	if (!worker)
 703		return;
 704	if (!test_bit(IO_WORKER_F_UP, &worker->flags))
 705		return;
 706	if (test_bit(IO_WORKER_F_RUNNING, &worker->flags))
 707		return;
 708	set_bit(IO_WORKER_F_RUNNING, &worker->flags);
 709	io_wq_inc_running(worker);
 710}
 711
 712/*
 713 * Called when worker is going to sleep. If there are no workers currently
 714 * running and we have work pending, wake up a free one or create a new one.
 715 */
 716void io_wq_worker_sleeping(struct task_struct *tsk)
 717{
 718	struct io_worker *worker = tsk->worker_private;
 719
 720	if (!worker)
 721		return;
 722	if (!test_bit(IO_WORKER_F_UP, &worker->flags))
 723		return;
 724	if (!test_bit(IO_WORKER_F_RUNNING, &worker->flags))
 725		return;
 726
 727	clear_bit(IO_WORKER_F_RUNNING, &worker->flags);
 728	io_wq_dec_running(worker);
 729}
 730
 731static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker,
 732			       struct task_struct *tsk)
 733{
 734	tsk->worker_private = worker;
 735	worker->task = tsk;
 736	set_cpus_allowed_ptr(tsk, wq->cpu_mask);
 737
 738	raw_spin_lock(&wq->lock);
 739	hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
 740	list_add_tail_rcu(&worker->all_list, &wq->all_list);
 741	set_bit(IO_WORKER_F_FREE, &worker->flags);
 742	raw_spin_unlock(&wq->lock);
 743	wake_up_new_task(tsk);
 744}
 745
 746static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
 747{
 748	return true;
 749}
 750
 751static inline bool io_should_retry_thread(struct io_worker *worker, long err)
 752{
 753	/*
 754	 * Prevent perpetual task_work retry, if the task (or its group) is
 755	 * exiting.
 756	 */
 757	if (fatal_signal_pending(current))
 758		return false;
 759	if (worker->init_retries++ >= WORKER_INIT_LIMIT)
 760		return false;
 761
 762	switch (err) {
 763	case -EAGAIN:
 764	case -ERESTARTSYS:
 765	case -ERESTARTNOINTR:
 766	case -ERESTARTNOHAND:
 767		return true;
 768	default:
 769		return false;
 770	}
 771}
 772
 773static void create_worker_cont(struct callback_head *cb)
 774{
 775	struct io_worker *worker;
 776	struct task_struct *tsk;
 777	struct io_wq *wq;
 778
 779	worker = container_of(cb, struct io_worker, create_work);
 780	clear_bit_unlock(0, &worker->create_state);
 781	wq = worker->wq;
 782	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
 783	if (!IS_ERR(tsk)) {
 784		io_init_new_worker(wq, worker, tsk);
 785		io_worker_release(worker);
 786		return;
 787	} else if (!io_should_retry_thread(worker, PTR_ERR(tsk))) {
 788		struct io_wq_acct *acct = io_wq_get_acct(worker);
 789
 790		atomic_dec(&acct->nr_running);
 791		raw_spin_lock(&wq->lock);
 792		acct->nr_workers--;
 793		if (!acct->nr_workers) {
 794			struct io_cb_cancel_data match = {
 795				.fn		= io_wq_work_match_all,
 796				.cancel_all	= true,
 797			};
 798
 799			raw_spin_unlock(&wq->lock);
 800			while (io_acct_cancel_pending_work(wq, acct, &match))
 801				;
 802		} else {
 803			raw_spin_unlock(&wq->lock);
 804		}
 805		io_worker_ref_put(wq);
 806		kfree(worker);
 807		return;
 808	}
 809
 810	/* re-create attempts grab a new worker ref, drop the existing one */
 811	io_worker_release(worker);
 812	schedule_work(&worker->work);
 813}
 814
 815static void io_workqueue_create(struct work_struct *work)
 816{
 817	struct io_worker *worker = container_of(work, struct io_worker, work);
 818	struct io_wq_acct *acct = io_wq_get_acct(worker);
 819
 820	if (!io_queue_worker_create(worker, acct, create_worker_cont))
 821		kfree(worker);
 822}
 823
 824static bool create_io_worker(struct io_wq *wq, int index)
 825{
 826	struct io_wq_acct *acct = &wq->acct[index];
 827	struct io_worker *worker;
 828	struct task_struct *tsk;
 829
 830	__set_current_state(TASK_RUNNING);
 831
 832	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
 833	if (!worker) {
 834fail:
 835		atomic_dec(&acct->nr_running);
 836		raw_spin_lock(&wq->lock);
 837		acct->nr_workers--;
 838		raw_spin_unlock(&wq->lock);
 839		io_worker_ref_put(wq);
 840		return false;
 841	}
 842
 843	refcount_set(&worker->ref, 1);
 844	worker->wq = wq;
 845	raw_spin_lock_init(&worker->lock);
 846	init_completion(&worker->ref_done);
 847
 848	if (index == IO_WQ_ACCT_BOUND)
 849		set_bit(IO_WORKER_F_BOUND, &worker->flags);
 850
 851	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
 852	if (!IS_ERR(tsk)) {
 853		io_init_new_worker(wq, worker, tsk);
 854	} else if (!io_should_retry_thread(worker, PTR_ERR(tsk))) {
 855		kfree(worker);
 856		goto fail;
 857	} else {
 858		INIT_WORK(&worker->work, io_workqueue_create);
 859		schedule_work(&worker->work);
 860	}
 861
 862	return true;
 863}
 864
 865/*
 866 * Iterate the passed in list and call the specific function for each
 867 * worker that isn't exiting
 868 */
 869static bool io_wq_for_each_worker(struct io_wq *wq,
 870				  bool (*func)(struct io_worker *, void *),
 871				  void *data)
 872{
 873	struct io_worker *worker;
 874	bool ret = false;
 875
 876	list_for_each_entry_rcu(worker, &wq->all_list, all_list) {
 877		if (io_worker_get(worker)) {
 878			/* no task if node is/was offline */
 879			if (worker->task)
 880				ret = func(worker, data);
 881			io_worker_release(worker);
 882			if (ret)
 883				break;
 884		}
 885	}
 886
 887	return ret;
 888}
 889
 890static bool io_wq_worker_wake(struct io_worker *worker, void *data)
 891{
 892	__set_notify_signal(worker->task);
 893	wake_up_process(worker->task);
 894	return false;
 895}
 896
 897static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq)
 898{
 899	do {
 900		atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
 901		wq->do_work(work);
 902		work = wq->free_work(work);
 903	} while (work);
 904}
 905
 906static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work)
 907{
 908	struct io_wq_acct *acct = io_work_get_acct(wq, work);
 909	unsigned int hash;
 910	struct io_wq_work *tail;
 911
 912	if (!io_wq_is_hashed(work)) {
 913append:
 914		wq_list_add_tail(&work->list, &acct->work_list);
 915		return;
 916	}
 917
 918	hash = io_get_work_hash(work);
 919	tail = wq->hash_tail[hash];
 920	wq->hash_tail[hash] = work;
 921	if (!tail)
 922		goto append;
 923
 924	wq_list_add_after(&work->list, &tail->list, &acct->work_list);
 925}
 926
 927static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
 928{
 929	return work == data;
 930}
 931
 932void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 933{
 934	struct io_wq_acct *acct = io_work_get_acct(wq, work);
 935	unsigned int work_flags = atomic_read(&work->flags);
 936	struct io_cb_cancel_data match = {
 937		.fn		= io_wq_work_match_item,
 938		.data		= work,
 939		.cancel_all	= false,
 940	};
 941	bool do_create;
 942
 943	/*
 944	 * If io-wq is exiting for this task, or if the request has explicitly
 945	 * been marked as one that should not get executed, cancel it here.
 946	 */
 947	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 948	    (work_flags & IO_WQ_WORK_CANCEL)) {
 949		io_run_cancel(work, wq);
 950		return;
 951	}
 952
 953	raw_spin_lock(&acct->lock);
 954	io_wq_insert_work(wq, work);
 955	clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 956	raw_spin_unlock(&acct->lock);
 957
 958	rcu_read_lock();
 959	do_create = !io_wq_activate_free_worker(wq, acct);
 960	rcu_read_unlock();
 961
 962	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 963	    !atomic_read(&acct->nr_running))) {
 964		bool did_create;
 965
 966		did_create = io_wq_create_worker(wq, acct);
 967		if (likely(did_create))
 968			return;
 969
 970		raw_spin_lock(&wq->lock);
 971		if (acct->nr_workers) {
 972			raw_spin_unlock(&wq->lock);
 973			return;
 974		}
 975		raw_spin_unlock(&wq->lock);
 976
 977		/* fatal condition, failed to create the first worker */
 
 
 
 
 978		io_acct_cancel_pending_work(wq, acct, &match);
 979	}
 980}
 981
 982/*
 983 * Work items that hash to the same value will not be done in parallel.
 984 * Used to limit concurrent writes, generally hashed by inode.
 985 */
 986void io_wq_hash_work(struct io_wq_work *work, void *val)
 987{
 988	unsigned int bit;
 989
 990	bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 991	atomic_or(IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT), &work->flags);
 992}
 993
 994static bool __io_wq_worker_cancel(struct io_worker *worker,
 995				  struct io_cb_cancel_data *match,
 996				  struct io_wq_work *work)
 997{
 998	if (work && match->fn(work, match->data)) {
 999		atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
1000		__set_notify_signal(worker->task);
1001		return true;
1002	}
1003
1004	return false;
1005}
1006
1007static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
1008{
1009	struct io_cb_cancel_data *match = data;
1010
1011	/*
1012	 * Hold the lock to avoid ->cur_work going out of scope, caller
1013	 * may dereference the passed in work.
1014	 */
1015	raw_spin_lock(&worker->lock);
1016	if (__io_wq_worker_cancel(worker, match, worker->cur_work))
 
1017		match->nr_running++;
1018	raw_spin_unlock(&worker->lock);
1019
1020	return match->nr_running && !match->cancel_all;
1021}
1022
1023static inline void io_wq_remove_pending(struct io_wq *wq,
1024					 struct io_wq_work *work,
1025					 struct io_wq_work_node *prev)
1026{
1027	struct io_wq_acct *acct = io_work_get_acct(wq, work);
1028	unsigned int hash = io_get_work_hash(work);
1029	struct io_wq_work *prev_work = NULL;
1030
1031	if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) {
1032		if (prev)
1033			prev_work = container_of(prev, struct io_wq_work, list);
1034		if (prev_work && io_get_work_hash(prev_work) == hash)
1035			wq->hash_tail[hash] = prev_work;
1036		else
1037			wq->hash_tail[hash] = NULL;
1038	}
1039	wq_list_del(&acct->work_list, &work->list, prev);
1040}
1041
1042static bool io_acct_cancel_pending_work(struct io_wq *wq,
1043					struct io_wq_acct *acct,
1044					struct io_cb_cancel_data *match)
1045{
1046	struct io_wq_work_node *node, *prev;
1047	struct io_wq_work *work;
1048
1049	raw_spin_lock(&acct->lock);
1050	wq_list_for_each(node, prev, &acct->work_list) {
1051		work = container_of(node, struct io_wq_work, list);
1052		if (!match->fn(work, match->data))
1053			continue;
1054		io_wq_remove_pending(wq, work, prev);
1055		raw_spin_unlock(&acct->lock);
1056		io_run_cancel(work, wq);
1057		match->nr_pending++;
1058		/* not safe to continue after unlock */
1059		return true;
1060	}
1061	raw_spin_unlock(&acct->lock);
1062
1063	return false;
1064}
1065
1066static void io_wq_cancel_pending_work(struct io_wq *wq,
1067				      struct io_cb_cancel_data *match)
1068{
1069	int i;
1070retry:
1071	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1072		struct io_wq_acct *acct = io_get_acct(wq, i == 0);
1073
1074		if (io_acct_cancel_pending_work(wq, acct, match)) {
1075			if (match->cancel_all)
1076				goto retry;
1077			break;
1078		}
1079	}
1080}
1081
1082static void io_wq_cancel_running_work(struct io_wq *wq,
1083				       struct io_cb_cancel_data *match)
1084{
1085	rcu_read_lock();
1086	io_wq_for_each_worker(wq, io_wq_worker_cancel, match);
1087	rcu_read_unlock();
1088}
1089
1090enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1091				  void *data, bool cancel_all)
1092{
1093	struct io_cb_cancel_data match = {
1094		.fn		= cancel,
1095		.data		= data,
1096		.cancel_all	= cancel_all,
1097	};
1098
1099	/*
1100	 * First check pending list, if we're lucky we can just remove it
1101	 * from there. CANCEL_OK means that the work is returned as-new,
1102	 * no completion will be posted for it.
1103	 *
1104	 * Then check if a free (going busy) or busy worker has the work
1105	 * currently running. If we find it there, we'll return CANCEL_RUNNING
1106	 * as an indication that we attempt to signal cancellation. The
1107	 * completion will run normally in this case.
1108	 *
1109	 * Do both of these while holding the wq->lock, to ensure that
1110	 * we'll find a work item regardless of state.
1111	 */
1112	io_wq_cancel_pending_work(wq, &match);
1113	if (match.nr_pending && !match.cancel_all)
1114		return IO_WQ_CANCEL_OK;
1115
1116	raw_spin_lock(&wq->lock);
1117	io_wq_cancel_running_work(wq, &match);
1118	raw_spin_unlock(&wq->lock);
1119	if (match.nr_running && !match.cancel_all)
1120		return IO_WQ_CANCEL_RUNNING;
1121
1122	if (match.nr_running)
1123		return IO_WQ_CANCEL_RUNNING;
1124	if (match.nr_pending)
1125		return IO_WQ_CANCEL_OK;
1126	return IO_WQ_CANCEL_NOTFOUND;
1127}
1128
1129static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1130			    int sync, void *key)
1131{
1132	struct io_wq *wq = container_of(wait, struct io_wq, wait);
1133	int i;
1134
1135	list_del_init(&wait->entry);
1136
1137	rcu_read_lock();
1138	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1139		struct io_wq_acct *acct = &wq->acct[i];
1140
1141		if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1142			io_wq_activate_free_worker(wq, acct);
1143	}
1144	rcu_read_unlock();
1145	return 1;
1146}
1147
1148struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1149{
1150	int ret, i;
1151	struct io_wq *wq;
1152
1153	if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1154		return ERR_PTR(-EINVAL);
1155	if (WARN_ON_ONCE(!bounded))
1156		return ERR_PTR(-EINVAL);
1157
1158	wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL);
1159	if (!wq)
1160		return ERR_PTR(-ENOMEM);
1161
1162	refcount_inc(&data->hash->refs);
1163	wq->hash = data->hash;
1164	wq->free_work = data->free_work;
1165	wq->do_work = data->do_work;
1166
1167	ret = -ENOMEM;
1168
1169	if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL))
1170		goto err;
1171	cpuset_cpus_allowed(data->task, wq->cpu_mask);
1172	wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1173	wq->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1174				task_rlimit(current, RLIMIT_NPROC);
1175	INIT_LIST_HEAD(&wq->wait.entry);
1176	wq->wait.func = io_wq_hash_wake;
1177	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1178		struct io_wq_acct *acct = &wq->acct[i];
1179
1180		acct->index = i;
1181		atomic_set(&acct->nr_running, 0);
1182		INIT_WQ_LIST(&acct->work_list);
1183		raw_spin_lock_init(&acct->lock);
1184	}
1185
1186	raw_spin_lock_init(&wq->lock);
1187	INIT_HLIST_NULLS_HEAD(&wq->free_list, 0);
1188	INIT_LIST_HEAD(&wq->all_list);
1189
1190	wq->task = get_task_struct(data->task);
1191	atomic_set(&wq->worker_refs, 1);
1192	init_completion(&wq->worker_done);
1193	ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1194	if (ret)
1195		goto err;
1196
1197	return wq;
1198err:
1199	io_wq_put_hash(data->hash);
1200	free_cpumask_var(wq->cpu_mask);
1201	kfree(wq);
1202	return ERR_PTR(ret);
1203}
1204
1205static bool io_task_work_match(struct callback_head *cb, void *data)
1206{
1207	struct io_worker *worker;
1208
1209	if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1210		return false;
1211	worker = container_of(cb, struct io_worker, create_work);
1212	return worker->wq == data;
1213}
1214
1215void io_wq_exit_start(struct io_wq *wq)
1216{
1217	set_bit(IO_WQ_BIT_EXIT, &wq->state);
1218}
1219
1220static void io_wq_cancel_tw_create(struct io_wq *wq)
1221{
1222	struct callback_head *cb;
1223
1224	while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1225		struct io_worker *worker;
1226
1227		worker = container_of(cb, struct io_worker, create_work);
1228		io_worker_cancel_cb(worker);
1229		/*
1230		 * Only the worker continuation helper has worker allocated and
1231		 * hence needs freeing.
1232		 */
1233		if (cb->func == create_worker_cont)
1234			kfree(worker);
1235	}
1236}
1237
1238static void io_wq_exit_workers(struct io_wq *wq)
1239{
1240	if (!wq->task)
1241		return;
1242
1243	io_wq_cancel_tw_create(wq);
1244
1245	rcu_read_lock();
1246	io_wq_for_each_worker(wq, io_wq_worker_wake, NULL);
1247	rcu_read_unlock();
1248	io_worker_ref_put(wq);
1249	wait_for_completion(&wq->worker_done);
1250
1251	spin_lock_irq(&wq->hash->wait.lock);
1252	list_del_init(&wq->wait.entry);
1253	spin_unlock_irq(&wq->hash->wait.lock);
1254
1255	put_task_struct(wq->task);
1256	wq->task = NULL;
1257}
1258
1259static void io_wq_destroy(struct io_wq *wq)
1260{
1261	struct io_cb_cancel_data match = {
1262		.fn		= io_wq_work_match_all,
1263		.cancel_all	= true,
1264	};
1265
1266	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1267	io_wq_cancel_pending_work(wq, &match);
1268	free_cpumask_var(wq->cpu_mask);
1269	io_wq_put_hash(wq->hash);
1270	kfree(wq);
1271}
1272
1273void io_wq_put_and_exit(struct io_wq *wq)
1274{
1275	WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1276
1277	io_wq_exit_workers(wq);
1278	io_wq_destroy(wq);
1279}
1280
1281struct online_data {
1282	unsigned int cpu;
1283	bool online;
1284};
1285
1286static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1287{
1288	struct online_data *od = data;
1289
1290	if (od->online)
1291		cpumask_set_cpu(od->cpu, worker->wq->cpu_mask);
1292	else
1293		cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask);
1294	return false;
1295}
1296
1297static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1298{
1299	struct online_data od = {
1300		.cpu = cpu,
1301		.online = online
1302	};
1303
1304	rcu_read_lock();
1305	io_wq_for_each_worker(wq, io_wq_worker_affinity, &od);
1306	rcu_read_unlock();
1307	return 0;
1308}
1309
1310static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1311{
1312	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1313
1314	return __io_wq_cpu_online(wq, cpu, true);
1315}
1316
1317static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1318{
1319	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1320
1321	return __io_wq_cpu_online(wq, cpu, false);
1322}
1323
1324int io_wq_cpu_affinity(struct io_uring_task *tctx, cpumask_var_t mask)
1325{
1326	cpumask_var_t allowed_mask;
1327	int ret = 0;
1328
1329	if (!tctx || !tctx->io_wq)
1330		return -EINVAL;
1331
1332	if (!alloc_cpumask_var(&allowed_mask, GFP_KERNEL))
1333		return -ENOMEM;
1334
1335	rcu_read_lock();
1336	cpuset_cpus_allowed(tctx->io_wq->task, allowed_mask);
1337	if (mask) {
1338		if (cpumask_subset(mask, allowed_mask))
1339			cpumask_copy(tctx->io_wq->cpu_mask, mask);
1340		else
1341			ret = -EINVAL;
1342	} else {
1343		cpumask_copy(tctx->io_wq->cpu_mask, allowed_mask);
1344	}
1345	rcu_read_unlock();
1346
1347	free_cpumask_var(allowed_mask);
1348	return ret;
1349}
1350
1351/*
1352 * Set max number of unbounded workers, returns old value. If new_count is 0,
1353 * then just return the old value.
1354 */
1355int io_wq_max_workers(struct io_wq *wq, int *new_count)
1356{
1357	struct io_wq_acct *acct;
1358	int prev[IO_WQ_ACCT_NR];
1359	int i;
1360
1361	BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND   != (int) IO_WQ_BOUND);
1362	BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1363	BUILD_BUG_ON((int) IO_WQ_ACCT_NR      != 2);
1364
1365	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1366		if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1367			new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1368	}
1369
1370	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1371		prev[i] = 0;
1372
1373	rcu_read_lock();
1374
1375	raw_spin_lock(&wq->lock);
1376	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1377		acct = &wq->acct[i];
1378		prev[i] = max_t(int, acct->max_workers, prev[i]);
1379		if (new_count[i])
1380			acct->max_workers = new_count[i];
1381	}
1382	raw_spin_unlock(&wq->lock);
1383	rcu_read_unlock();
1384
1385	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1386		new_count[i] = prev[i];
1387
1388	return 0;
1389}
1390
1391static __init int io_wq_init(void)
1392{
1393	int ret;
1394
1395	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1396					io_wq_cpu_online, io_wq_cpu_offline);
1397	if (ret < 0)
1398		return ret;
1399	io_wq_online = ret;
1400	return 0;
1401}
1402subsys_initcall(io_wq_init);