xref: /openbmc/linux/io_uring/io-wq.c (revision e27a85c1)
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 <uapi/linux/io_uring.h>
19  
20  #include "io-wq.h"
21  #include "slist.h"
22  #include "io_uring.h"
23  
24  #define WORKER_IDLE_TIMEOUT	(5 * HZ)
25  
26  enum {
27  	IO_WORKER_F_UP		= 1,	/* up and active */
28  	IO_WORKER_F_RUNNING	= 2,	/* account as running */
29  	IO_WORKER_F_FREE	= 4,	/* worker on free list */
30  	IO_WORKER_F_BOUND	= 8,	/* is doing bounded work */
31  };
32  
33  enum {
34  	IO_WQ_BIT_EXIT		= 0,	/* wq exiting */
35  };
36  
37  enum {
38  	IO_ACCT_STALLED_BIT	= 0,	/* stalled on hash */
39  };
40  
41  /*
42   * One for each thread in a wqe pool
43   */
44  struct io_worker {
45  	refcount_t ref;
46  	unsigned flags;
47  	struct hlist_nulls_node nulls_node;
48  	struct list_head all_list;
49  	struct task_struct *task;
50  	struct io_wqe *wqe;
51  
52  	struct io_wq_work *cur_work;
53  	struct io_wq_work *next_work;
54  	raw_spinlock_t lock;
55  
56  	struct completion ref_done;
57  
58  	unsigned long create_state;
59  	struct callback_head create_work;
60  	int create_index;
61  
62  	union {
63  		struct rcu_head rcu;
64  		struct work_struct work;
65  	};
66  };
67  
68  #if BITS_PER_LONG == 64
69  #define IO_WQ_HASH_ORDER	6
70  #else
71  #define IO_WQ_HASH_ORDER	5
72  #endif
73  
74  #define IO_WQ_NR_HASH_BUCKETS	(1u << IO_WQ_HASH_ORDER)
75  
76  struct io_wqe_acct {
77  	unsigned nr_workers;
78  	unsigned max_workers;
79  	int index;
80  	atomic_t nr_running;
81  	raw_spinlock_t lock;
82  	struct io_wq_work_list work_list;
83  	unsigned long flags;
84  };
85  
86  enum {
87  	IO_WQ_ACCT_BOUND,
88  	IO_WQ_ACCT_UNBOUND,
89  	IO_WQ_ACCT_NR,
90  };
91  
92  /*
93   * Per-node worker thread pool
94   */
95  struct io_wqe {
96  	raw_spinlock_t lock;
97  	struct io_wqe_acct acct[IO_WQ_ACCT_NR];
98  
99  	int node;
100  
101  	struct hlist_nulls_head free_list;
102  	struct list_head all_list;
103  
104  	struct wait_queue_entry wait;
105  
106  	struct io_wq *wq;
107  	struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
108  
109  	cpumask_var_t cpu_mask;
110  };
111  
112  /*
113   * Per io_wq state
114    */
115  struct io_wq {
116  	unsigned long state;
117  
118  	free_work_fn *free_work;
119  	io_wq_work_fn *do_work;
120  
121  	struct io_wq_hash *hash;
122  
123  	atomic_t worker_refs;
124  	struct completion worker_done;
125  
126  	struct hlist_node cpuhp_node;
127  
128  	struct task_struct *task;
129  
130  	struct io_wqe *wqes[];
131  };
132  
133  static enum cpuhp_state io_wq_online;
134  
135  struct io_cb_cancel_data {
136  	work_cancel_fn *fn;
137  	void *data;
138  	int nr_running;
139  	int nr_pending;
140  	bool cancel_all;
141  };
142  
143  static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index);
144  static void io_wqe_dec_running(struct io_worker *worker);
145  static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
146  					struct io_wqe_acct *acct,
147  					struct io_cb_cancel_data *match);
148  static void create_worker_cb(struct callback_head *cb);
149  static void io_wq_cancel_tw_create(struct io_wq *wq);
150  
151  static bool io_worker_get(struct io_worker *worker)
152  {
153  	return refcount_inc_not_zero(&worker->ref);
154  }
155  
156  static void io_worker_release(struct io_worker *worker)
157  {
158  	if (refcount_dec_and_test(&worker->ref))
159  		complete(&worker->ref_done);
160  }
161  
162  static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
163  {
164  	return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
165  }
166  
167  static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
168  						   struct io_wq_work *work)
169  {
170  	return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
171  }
172  
173  static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
174  {
175  	return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
176  }
177  
178  static void io_worker_ref_put(struct io_wq *wq)
179  {
180  	if (atomic_dec_and_test(&wq->worker_refs))
181  		complete(&wq->worker_done);
182  }
183  
184  static void io_worker_cancel_cb(struct io_worker *worker)
185  {
186  	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
187  	struct io_wqe *wqe = worker->wqe;
188  	struct io_wq *wq = wqe->wq;
189  
190  	atomic_dec(&acct->nr_running);
191  	raw_spin_lock(&worker->wqe->lock);
192  	acct->nr_workers--;
193  	raw_spin_unlock(&worker->wqe->lock);
194  	io_worker_ref_put(wq);
195  	clear_bit_unlock(0, &worker->create_state);
196  	io_worker_release(worker);
197  }
198  
199  static bool io_task_worker_match(struct callback_head *cb, void *data)
200  {
201  	struct io_worker *worker;
202  
203  	if (cb->func != create_worker_cb)
204  		return false;
205  	worker = container_of(cb, struct io_worker, create_work);
206  	return worker == data;
207  }
208  
209  static void io_worker_exit(struct io_worker *worker)
210  {
211  	struct io_wqe *wqe = worker->wqe;
212  	struct io_wq *wq = wqe->wq;
213  
214  	while (1) {
215  		struct callback_head *cb = task_work_cancel_match(wq->task,
216  						io_task_worker_match, worker);
217  
218  		if (!cb)
219  			break;
220  		io_worker_cancel_cb(worker);
221  	}
222  
223  	io_worker_release(worker);
224  	wait_for_completion(&worker->ref_done);
225  
226  	raw_spin_lock(&wqe->lock);
227  	if (worker->flags & IO_WORKER_F_FREE)
228  		hlist_nulls_del_rcu(&worker->nulls_node);
229  	list_del_rcu(&worker->all_list);
230  	raw_spin_unlock(&wqe->lock);
231  	io_wqe_dec_running(worker);
232  	worker->flags = 0;
233  	preempt_disable();
234  	current->flags &= ~PF_IO_WORKER;
235  	preempt_enable();
236  
237  	kfree_rcu(worker, rcu);
238  	io_worker_ref_put(wqe->wq);
239  	do_exit(0);
240  }
241  
242  static inline bool io_acct_run_queue(struct io_wqe_acct *acct)
243  {
244  	bool ret = false;
245  
246  	raw_spin_lock(&acct->lock);
247  	if (!wq_list_empty(&acct->work_list) &&
248  	    !test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
249  		ret = true;
250  	raw_spin_unlock(&acct->lock);
251  
252  	return ret;
253  }
254  
255  /*
256   * Check head of free list for an available worker. If one isn't available,
257   * caller must create one.
258   */
259  static bool io_wqe_activate_free_worker(struct io_wqe *wqe,
260  					struct io_wqe_acct *acct)
261  	__must_hold(RCU)
262  {
263  	struct hlist_nulls_node *n;
264  	struct io_worker *worker;
265  
266  	/*
267  	 * Iterate free_list and see if we can find an idle worker to
268  	 * activate. If a given worker is on the free_list but in the process
269  	 * of exiting, keep trying.
270  	 */
271  	hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
272  		if (!io_worker_get(worker))
273  			continue;
274  		if (io_wqe_get_acct(worker) != acct) {
275  			io_worker_release(worker);
276  			continue;
277  		}
278  		if (wake_up_process(worker->task)) {
279  			io_worker_release(worker);
280  			return true;
281  		}
282  		io_worker_release(worker);
283  	}
284  
285  	return false;
286  }
287  
288  /*
289   * We need a worker. If we find a free one, we're good. If not, and we're
290   * below the max number of workers, create one.
291   */
292  static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
293  {
294  	/*
295  	 * Most likely an attempt to queue unbounded work on an io_wq that
296  	 * wasn't setup with any unbounded workers.
297  	 */
298  	if (unlikely(!acct->max_workers))
299  		pr_warn_once("io-wq is not configured for unbound workers");
300  
301  	raw_spin_lock(&wqe->lock);
302  	if (acct->nr_workers >= acct->max_workers) {
303  		raw_spin_unlock(&wqe->lock);
304  		return true;
305  	}
306  	acct->nr_workers++;
307  	raw_spin_unlock(&wqe->lock);
308  	atomic_inc(&acct->nr_running);
309  	atomic_inc(&wqe->wq->worker_refs);
310  	return create_io_worker(wqe->wq, wqe, acct->index);
311  }
312  
313  static void io_wqe_inc_running(struct io_worker *worker)
314  {
315  	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
316  
317  	atomic_inc(&acct->nr_running);
318  }
319  
320  static void create_worker_cb(struct callback_head *cb)
321  {
322  	struct io_worker *worker;
323  	struct io_wq *wq;
324  	struct io_wqe *wqe;
325  	struct io_wqe_acct *acct;
326  	bool do_create = false;
327  
328  	worker = container_of(cb, struct io_worker, create_work);
329  	wqe = worker->wqe;
330  	wq = wqe->wq;
331  	acct = &wqe->acct[worker->create_index];
332  	raw_spin_lock(&wqe->lock);
333  	if (acct->nr_workers < acct->max_workers) {
334  		acct->nr_workers++;
335  		do_create = true;
336  	}
337  	raw_spin_unlock(&wqe->lock);
338  	if (do_create) {
339  		create_io_worker(wq, wqe, worker->create_index);
340  	} else {
341  		atomic_dec(&acct->nr_running);
342  		io_worker_ref_put(wq);
343  	}
344  	clear_bit_unlock(0, &worker->create_state);
345  	io_worker_release(worker);
346  }
347  
348  static bool io_queue_worker_create(struct io_worker *worker,
349  				   struct io_wqe_acct *acct,
350  				   task_work_func_t func)
351  {
352  	struct io_wqe *wqe = worker->wqe;
353  	struct io_wq *wq = wqe->wq;
354  
355  	/* raced with exit, just ignore create call */
356  	if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
357  		goto fail;
358  	if (!io_worker_get(worker))
359  		goto fail;
360  	/*
361  	 * create_state manages ownership of create_work/index. We should
362  	 * only need one entry per worker, as the worker going to sleep
363  	 * will trigger the condition, and waking will clear it once it
364  	 * runs the task_work.
365  	 */
366  	if (test_bit(0, &worker->create_state) ||
367  	    test_and_set_bit_lock(0, &worker->create_state))
368  		goto fail_release;
369  
370  	atomic_inc(&wq->worker_refs);
371  	init_task_work(&worker->create_work, func);
372  	worker->create_index = acct->index;
373  	if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
374  		/*
375  		 * EXIT may have been set after checking it above, check after
376  		 * adding the task_work and remove any creation item if it is
377  		 * now set. wq exit does that too, but we can have added this
378  		 * work item after we canceled in io_wq_exit_workers().
379  		 */
380  		if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
381  			io_wq_cancel_tw_create(wq);
382  		io_worker_ref_put(wq);
383  		return true;
384  	}
385  	io_worker_ref_put(wq);
386  	clear_bit_unlock(0, &worker->create_state);
387  fail_release:
388  	io_worker_release(worker);
389  fail:
390  	atomic_dec(&acct->nr_running);
391  	io_worker_ref_put(wq);
392  	return false;
393  }
394  
395  static void io_wqe_dec_running(struct io_worker *worker)
396  {
397  	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
398  	struct io_wqe *wqe = worker->wqe;
399  
400  	if (!(worker->flags & IO_WORKER_F_UP))
401  		return;
402  
403  	if (!atomic_dec_and_test(&acct->nr_running))
404  		return;
405  	if (!io_acct_run_queue(acct))
406  		return;
407  
408  	atomic_inc(&acct->nr_running);
409  	atomic_inc(&wqe->wq->worker_refs);
410  	io_queue_worker_create(worker, acct, create_worker_cb);
411  }
412  
413  /*
414   * Worker will start processing some work. Move it to the busy list, if
415   * it's currently on the freelist
416   */
417  static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker)
418  {
419  	if (worker->flags & IO_WORKER_F_FREE) {
420  		worker->flags &= ~IO_WORKER_F_FREE;
421  		raw_spin_lock(&wqe->lock);
422  		hlist_nulls_del_init_rcu(&worker->nulls_node);
423  		raw_spin_unlock(&wqe->lock);
424  	}
425  }
426  
427  /*
428   * No work, worker going to sleep. Move to freelist, and unuse mm if we
429   * have one attached. Dropping the mm may potentially sleep, so we drop
430   * the lock in that case and return success. Since the caller has to
431   * retry the loop in that case (we changed task state), we don't regrab
432   * the lock if we return success.
433   */
434  static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
435  	__must_hold(wqe->lock)
436  {
437  	if (!(worker->flags & IO_WORKER_F_FREE)) {
438  		worker->flags |= IO_WORKER_F_FREE;
439  		hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
440  	}
441  }
442  
443  static inline unsigned int io_get_work_hash(struct io_wq_work *work)
444  {
445  	return work->flags >> IO_WQ_HASH_SHIFT;
446  }
447  
448  static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
449  {
450  	struct io_wq *wq = wqe->wq;
451  	bool ret = false;
452  
453  	spin_lock_irq(&wq->hash->wait.lock);
454  	if (list_empty(&wqe->wait.entry)) {
455  		__add_wait_queue(&wq->hash->wait, &wqe->wait);
456  		if (!test_bit(hash, &wq->hash->map)) {
457  			__set_current_state(TASK_RUNNING);
458  			list_del_init(&wqe->wait.entry);
459  			ret = true;
460  		}
461  	}
462  	spin_unlock_irq(&wq->hash->wait.lock);
463  	return ret;
464  }
465  
466  static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
467  					   struct io_worker *worker)
468  	__must_hold(acct->lock)
469  {
470  	struct io_wq_work_node *node, *prev;
471  	struct io_wq_work *work, *tail;
472  	unsigned int stall_hash = -1U;
473  	struct io_wqe *wqe = worker->wqe;
474  
475  	wq_list_for_each(node, prev, &acct->work_list) {
476  		unsigned int hash;
477  
478  		work = container_of(node, struct io_wq_work, list);
479  
480  		/* not hashed, can run anytime */
481  		if (!io_wq_is_hashed(work)) {
482  			wq_list_del(&acct->work_list, node, prev);
483  			return work;
484  		}
485  
486  		hash = io_get_work_hash(work);
487  		/* all items with this hash lie in [work, tail] */
488  		tail = wqe->hash_tail[hash];
489  
490  		/* hashed, can run if not already running */
491  		if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
492  			wqe->hash_tail[hash] = NULL;
493  			wq_list_cut(&acct->work_list, &tail->list, prev);
494  			return work;
495  		}
496  		if (stall_hash == -1U)
497  			stall_hash = hash;
498  		/* fast forward to a next hash, for-each will fix up @prev */
499  		node = &tail->list;
500  	}
501  
502  	if (stall_hash != -1U) {
503  		bool unstalled;
504  
505  		/*
506  		 * Set this before dropping the lock to avoid racing with new
507  		 * work being added and clearing the stalled bit.
508  		 */
509  		set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
510  		raw_spin_unlock(&acct->lock);
511  		unstalled = io_wait_on_hash(wqe, stall_hash);
512  		raw_spin_lock(&acct->lock);
513  		if (unstalled) {
514  			clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
515  			if (wq_has_sleeper(&wqe->wq->hash->wait))
516  				wake_up(&wqe->wq->hash->wait);
517  		}
518  	}
519  
520  	return NULL;
521  }
522  
523  static void io_assign_current_work(struct io_worker *worker,
524  				   struct io_wq_work *work)
525  {
526  	if (work) {
527  		io_run_task_work();
528  		cond_resched();
529  	}
530  
531  	raw_spin_lock(&worker->lock);
532  	worker->cur_work = work;
533  	worker->next_work = NULL;
534  	raw_spin_unlock(&worker->lock);
535  }
536  
537  static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
538  
539  static void io_worker_handle_work(struct io_worker *worker)
540  {
541  	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
542  	struct io_wqe *wqe = worker->wqe;
543  	struct io_wq *wq = wqe->wq;
544  	bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
545  
546  	do {
547  		struct io_wq_work *work;
548  
549  		/*
550  		 * If we got some work, mark us as busy. If we didn't, but
551  		 * the list isn't empty, it means we stalled on hashed work.
552  		 * Mark us stalled so we don't keep looking for work when we
553  		 * can't make progress, any work completion or insertion will
554  		 * clear the stalled flag.
555  		 */
556  		raw_spin_lock(&acct->lock);
557  		work = io_get_next_work(acct, worker);
558  		raw_spin_unlock(&acct->lock);
559  		if (work) {
560  			__io_worker_busy(wqe, worker);
561  
562  			/*
563  			 * Make sure cancelation can find this, even before
564  			 * it becomes the active work. That avoids a window
565  			 * where the work has been removed from our general
566  			 * work list, but isn't yet discoverable as the
567  			 * current work item for this worker.
568  			 */
569  			raw_spin_lock(&worker->lock);
570  			worker->next_work = work;
571  			raw_spin_unlock(&worker->lock);
572  		} else {
573  			break;
574  		}
575  		io_assign_current_work(worker, work);
576  		__set_current_state(TASK_RUNNING);
577  
578  		/* handle a whole dependent link */
579  		do {
580  			struct io_wq_work *next_hashed, *linked;
581  			unsigned int hash = io_get_work_hash(work);
582  
583  			next_hashed = wq_next_work(work);
584  
585  			if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
586  				work->flags |= IO_WQ_WORK_CANCEL;
587  			wq->do_work(work);
588  			io_assign_current_work(worker, NULL);
589  
590  			linked = wq->free_work(work);
591  			work = next_hashed;
592  			if (!work && linked && !io_wq_is_hashed(linked)) {
593  				work = linked;
594  				linked = NULL;
595  			}
596  			io_assign_current_work(worker, work);
597  			if (linked)
598  				io_wqe_enqueue(wqe, linked);
599  
600  			if (hash != -1U && !next_hashed) {
601  				/* serialize hash clear with wake_up() */
602  				spin_lock_irq(&wq->hash->wait.lock);
603  				clear_bit(hash, &wq->hash->map);
604  				clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
605  				spin_unlock_irq(&wq->hash->wait.lock);
606  				if (wq_has_sleeper(&wq->hash->wait))
607  					wake_up(&wq->hash->wait);
608  			}
609  		} while (work);
610  	} while (1);
611  }
612  
613  static int io_wqe_worker(void *data)
614  {
615  	struct io_worker *worker = data;
616  	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
617  	struct io_wqe *wqe = worker->wqe;
618  	struct io_wq *wq = wqe->wq;
619  	bool exit_mask = false, last_timeout = false;
620  	char buf[TASK_COMM_LEN];
621  
622  	worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
623  
624  	snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
625  	set_task_comm(current, buf);
626  
627  	while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
628  		long ret;
629  
630  		set_current_state(TASK_INTERRUPTIBLE);
631  		while (io_acct_run_queue(acct))
632  			io_worker_handle_work(worker);
633  
634  		raw_spin_lock(&wqe->lock);
635  		/*
636  		 * Last sleep timed out. Exit if we're not the last worker,
637  		 * or if someone modified our affinity.
638  		 */
639  		if (last_timeout && (exit_mask || acct->nr_workers > 1)) {
640  			acct->nr_workers--;
641  			raw_spin_unlock(&wqe->lock);
642  			__set_current_state(TASK_RUNNING);
643  			break;
644  		}
645  		last_timeout = false;
646  		__io_worker_idle(wqe, worker);
647  		raw_spin_unlock(&wqe->lock);
648  		if (io_run_task_work())
649  			continue;
650  		ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
651  		if (signal_pending(current)) {
652  			struct ksignal ksig;
653  
654  			if (!get_signal(&ksig))
655  				continue;
656  			break;
657  		}
658  		if (!ret) {
659  			last_timeout = true;
660  			exit_mask = !cpumask_test_cpu(raw_smp_processor_id(),
661  							wqe->cpu_mask);
662  		}
663  	}
664  
665  	if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
666  		io_worker_handle_work(worker);
667  
668  	io_worker_exit(worker);
669  	return 0;
670  }
671  
672  /*
673   * Called when a worker is scheduled in. Mark us as currently running.
674   */
675  void io_wq_worker_running(struct task_struct *tsk)
676  {
677  	struct io_worker *worker = tsk->worker_private;
678  
679  	if (!worker)
680  		return;
681  	if (!(worker->flags & IO_WORKER_F_UP))
682  		return;
683  	if (worker->flags & IO_WORKER_F_RUNNING)
684  		return;
685  	worker->flags |= IO_WORKER_F_RUNNING;
686  	io_wqe_inc_running(worker);
687  }
688  
689  /*
690   * Called when worker is going to sleep. If there are no workers currently
691   * running and we have work pending, wake up a free one or create a new one.
692   */
693  void io_wq_worker_sleeping(struct task_struct *tsk)
694  {
695  	struct io_worker *worker = tsk->worker_private;
696  
697  	if (!worker)
698  		return;
699  	if (!(worker->flags & IO_WORKER_F_UP))
700  		return;
701  	if (!(worker->flags & IO_WORKER_F_RUNNING))
702  		return;
703  
704  	worker->flags &= ~IO_WORKER_F_RUNNING;
705  	io_wqe_dec_running(worker);
706  }
707  
708  static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
709  			       struct task_struct *tsk)
710  {
711  	tsk->worker_private = worker;
712  	worker->task = tsk;
713  	set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
714  
715  	raw_spin_lock(&wqe->lock);
716  	hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
717  	list_add_tail_rcu(&worker->all_list, &wqe->all_list);
718  	worker->flags |= IO_WORKER_F_FREE;
719  	raw_spin_unlock(&wqe->lock);
720  	wake_up_new_task(tsk);
721  }
722  
723  static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
724  {
725  	return true;
726  }
727  
728  static inline bool io_should_retry_thread(long err)
729  {
730  	/*
731  	 * Prevent perpetual task_work retry, if the task (or its group) is
732  	 * exiting.
733  	 */
734  	if (fatal_signal_pending(current))
735  		return false;
736  
737  	switch (err) {
738  	case -EAGAIN:
739  	case -ERESTARTSYS:
740  	case -ERESTARTNOINTR:
741  	case -ERESTARTNOHAND:
742  		return true;
743  	default:
744  		return false;
745  	}
746  }
747  
748  static void create_worker_cont(struct callback_head *cb)
749  {
750  	struct io_worker *worker;
751  	struct task_struct *tsk;
752  	struct io_wqe *wqe;
753  
754  	worker = container_of(cb, struct io_worker, create_work);
755  	clear_bit_unlock(0, &worker->create_state);
756  	wqe = worker->wqe;
757  	tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
758  	if (!IS_ERR(tsk)) {
759  		io_init_new_worker(wqe, worker, tsk);
760  		io_worker_release(worker);
761  		return;
762  	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
763  		struct io_wqe_acct *acct = io_wqe_get_acct(worker);
764  
765  		atomic_dec(&acct->nr_running);
766  		raw_spin_lock(&wqe->lock);
767  		acct->nr_workers--;
768  		if (!acct->nr_workers) {
769  			struct io_cb_cancel_data match = {
770  				.fn		= io_wq_work_match_all,
771  				.cancel_all	= true,
772  			};
773  
774  			raw_spin_unlock(&wqe->lock);
775  			while (io_acct_cancel_pending_work(wqe, acct, &match))
776  				;
777  		} else {
778  			raw_spin_unlock(&wqe->lock);
779  		}
780  		io_worker_ref_put(wqe->wq);
781  		kfree(worker);
782  		return;
783  	}
784  
785  	/* re-create attempts grab a new worker ref, drop the existing one */
786  	io_worker_release(worker);
787  	schedule_work(&worker->work);
788  }
789  
790  static void io_workqueue_create(struct work_struct *work)
791  {
792  	struct io_worker *worker = container_of(work, struct io_worker, work);
793  	struct io_wqe_acct *acct = io_wqe_get_acct(worker);
794  
795  	if (!io_queue_worker_create(worker, acct, create_worker_cont))
796  		kfree(worker);
797  }
798  
799  static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
800  {
801  	struct io_wqe_acct *acct = &wqe->acct[index];
802  	struct io_worker *worker;
803  	struct task_struct *tsk;
804  
805  	__set_current_state(TASK_RUNNING);
806  
807  	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
808  	if (!worker) {
809  fail:
810  		atomic_dec(&acct->nr_running);
811  		raw_spin_lock(&wqe->lock);
812  		acct->nr_workers--;
813  		raw_spin_unlock(&wqe->lock);
814  		io_worker_ref_put(wq);
815  		return false;
816  	}
817  
818  	refcount_set(&worker->ref, 1);
819  	worker->wqe = wqe;
820  	raw_spin_lock_init(&worker->lock);
821  	init_completion(&worker->ref_done);
822  
823  	if (index == IO_WQ_ACCT_BOUND)
824  		worker->flags |= IO_WORKER_F_BOUND;
825  
826  	tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
827  	if (!IS_ERR(tsk)) {
828  		io_init_new_worker(wqe, worker, tsk);
829  	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
830  		kfree(worker);
831  		goto fail;
832  	} else {
833  		INIT_WORK(&worker->work, io_workqueue_create);
834  		schedule_work(&worker->work);
835  	}
836  
837  	return true;
838  }
839  
840  /*
841   * Iterate the passed in list and call the specific function for each
842   * worker that isn't exiting
843   */
844  static bool io_wq_for_each_worker(struct io_wqe *wqe,
845  				  bool (*func)(struct io_worker *, void *),
846  				  void *data)
847  {
848  	struct io_worker *worker;
849  	bool ret = false;
850  
851  	list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
852  		if (io_worker_get(worker)) {
853  			/* no task if node is/was offline */
854  			if (worker->task)
855  				ret = func(worker, data);
856  			io_worker_release(worker);
857  			if (ret)
858  				break;
859  		}
860  	}
861  
862  	return ret;
863  }
864  
865  static bool io_wq_worker_wake(struct io_worker *worker, void *data)
866  {
867  	__set_notify_signal(worker->task);
868  	wake_up_process(worker->task);
869  	return false;
870  }
871  
872  static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
873  {
874  	struct io_wq *wq = wqe->wq;
875  
876  	do {
877  		work->flags |= IO_WQ_WORK_CANCEL;
878  		wq->do_work(work);
879  		work = wq->free_work(work);
880  	} while (work);
881  }
882  
883  static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
884  {
885  	struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
886  	unsigned int hash;
887  	struct io_wq_work *tail;
888  
889  	if (!io_wq_is_hashed(work)) {
890  append:
891  		wq_list_add_tail(&work->list, &acct->work_list);
892  		return;
893  	}
894  
895  	hash = io_get_work_hash(work);
896  	tail = wqe->hash_tail[hash];
897  	wqe->hash_tail[hash] = work;
898  	if (!tail)
899  		goto append;
900  
901  	wq_list_add_after(&work->list, &tail->list, &acct->work_list);
902  }
903  
904  static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
905  {
906  	return work == data;
907  }
908  
909  static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
910  {
911  	struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
912  	struct io_cb_cancel_data match;
913  	unsigned work_flags = work->flags;
914  	bool do_create;
915  
916  	/*
917  	 * If io-wq is exiting for this task, or if the request has explicitly
918  	 * been marked as one that should not get executed, cancel it here.
919  	 */
920  	if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
921  	    (work->flags & IO_WQ_WORK_CANCEL)) {
922  		io_run_cancel(work, wqe);
923  		return;
924  	}
925  
926  	raw_spin_lock(&acct->lock);
927  	io_wqe_insert_work(wqe, work);
928  	clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
929  	raw_spin_unlock(&acct->lock);
930  
931  	raw_spin_lock(&wqe->lock);
932  	rcu_read_lock();
933  	do_create = !io_wqe_activate_free_worker(wqe, acct);
934  	rcu_read_unlock();
935  
936  	raw_spin_unlock(&wqe->lock);
937  
938  	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
939  	    !atomic_read(&acct->nr_running))) {
940  		bool did_create;
941  
942  		did_create = io_wqe_create_worker(wqe, acct);
943  		if (likely(did_create))
944  			return;
945  
946  		raw_spin_lock(&wqe->lock);
947  		if (acct->nr_workers) {
948  			raw_spin_unlock(&wqe->lock);
949  			return;
950  		}
951  		raw_spin_unlock(&wqe->lock);
952  
953  		/* fatal condition, failed to create the first worker */
954  		match.fn		= io_wq_work_match_item,
955  		match.data		= work,
956  		match.cancel_all	= false,
957  
958  		io_acct_cancel_pending_work(wqe, acct, &match);
959  	}
960  }
961  
962  void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
963  {
964  	struct io_wqe *wqe = wq->wqes[numa_node_id()];
965  
966  	io_wqe_enqueue(wqe, work);
967  }
968  
969  /*
970   * Work items that hash to the same value will not be done in parallel.
971   * Used to limit concurrent writes, generally hashed by inode.
972   */
973  void io_wq_hash_work(struct io_wq_work *work, void *val)
974  {
975  	unsigned int bit;
976  
977  	bit = hash_ptr(val, IO_WQ_HASH_ORDER);
978  	work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
979  }
980  
981  static bool __io_wq_worker_cancel(struct io_worker *worker,
982  				  struct io_cb_cancel_data *match,
983  				  struct io_wq_work *work)
984  {
985  	if (work && match->fn(work, match->data)) {
986  		work->flags |= IO_WQ_WORK_CANCEL;
987  		__set_notify_signal(worker->task);
988  		return true;
989  	}
990  
991  	return false;
992  }
993  
994  static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
995  {
996  	struct io_cb_cancel_data *match = data;
997  
998  	/*
999  	 * Hold the lock to avoid ->cur_work going out of scope, caller
1000  	 * may dereference the passed in work.
1001  	 */
1002  	raw_spin_lock(&worker->lock);
1003  	if (__io_wq_worker_cancel(worker, match, worker->cur_work) ||
1004  	    __io_wq_worker_cancel(worker, match, worker->next_work))
1005  		match->nr_running++;
1006  	raw_spin_unlock(&worker->lock);
1007  
1008  	return match->nr_running && !match->cancel_all;
1009  }
1010  
1011  static inline void io_wqe_remove_pending(struct io_wqe *wqe,
1012  					 struct io_wq_work *work,
1013  					 struct io_wq_work_node *prev)
1014  {
1015  	struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
1016  	unsigned int hash = io_get_work_hash(work);
1017  	struct io_wq_work *prev_work = NULL;
1018  
1019  	if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
1020  		if (prev)
1021  			prev_work = container_of(prev, struct io_wq_work, list);
1022  		if (prev_work && io_get_work_hash(prev_work) == hash)
1023  			wqe->hash_tail[hash] = prev_work;
1024  		else
1025  			wqe->hash_tail[hash] = NULL;
1026  	}
1027  	wq_list_del(&acct->work_list, &work->list, prev);
1028  }
1029  
1030  static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
1031  					struct io_wqe_acct *acct,
1032  					struct io_cb_cancel_data *match)
1033  {
1034  	struct io_wq_work_node *node, *prev;
1035  	struct io_wq_work *work;
1036  
1037  	raw_spin_lock(&acct->lock);
1038  	wq_list_for_each(node, prev, &acct->work_list) {
1039  		work = container_of(node, struct io_wq_work, list);
1040  		if (!match->fn(work, match->data))
1041  			continue;
1042  		io_wqe_remove_pending(wqe, work, prev);
1043  		raw_spin_unlock(&acct->lock);
1044  		io_run_cancel(work, wqe);
1045  		match->nr_pending++;
1046  		/* not safe to continue after unlock */
1047  		return true;
1048  	}
1049  	raw_spin_unlock(&acct->lock);
1050  
1051  	return false;
1052  }
1053  
1054  static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1055  				       struct io_cb_cancel_data *match)
1056  {
1057  	int i;
1058  retry:
1059  	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1060  		struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1061  
1062  		if (io_acct_cancel_pending_work(wqe, acct, match)) {
1063  			if (match->cancel_all)
1064  				goto retry;
1065  			break;
1066  		}
1067  	}
1068  }
1069  
1070  static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1071  				       struct io_cb_cancel_data *match)
1072  {
1073  	rcu_read_lock();
1074  	io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1075  	rcu_read_unlock();
1076  }
1077  
1078  enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1079  				  void *data, bool cancel_all)
1080  {
1081  	struct io_cb_cancel_data match = {
1082  		.fn		= cancel,
1083  		.data		= data,
1084  		.cancel_all	= cancel_all,
1085  	};
1086  	int node;
1087  
1088  	/*
1089  	 * First check pending list, if we're lucky we can just remove it
1090  	 * from there. CANCEL_OK means that the work is returned as-new,
1091  	 * no completion will be posted for it.
1092  	 *
1093  	 * Then check if a free (going busy) or busy worker has the work
1094  	 * currently running. If we find it there, we'll return CANCEL_RUNNING
1095  	 * as an indication that we attempt to signal cancellation. The
1096  	 * completion will run normally in this case.
1097  	 *
1098  	 * Do both of these while holding the wqe->lock, to ensure that
1099  	 * we'll find a work item regardless of state.
1100  	 */
1101  	for_each_node(node) {
1102  		struct io_wqe *wqe = wq->wqes[node];
1103  
1104  		io_wqe_cancel_pending_work(wqe, &match);
1105  		if (match.nr_pending && !match.cancel_all)
1106  			return IO_WQ_CANCEL_OK;
1107  
1108  		raw_spin_lock(&wqe->lock);
1109  		io_wqe_cancel_running_work(wqe, &match);
1110  		raw_spin_unlock(&wqe->lock);
1111  		if (match.nr_running && !match.cancel_all)
1112  			return IO_WQ_CANCEL_RUNNING;
1113  	}
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  
1122  static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1123  			    int sync, void *key)
1124  {
1125  	struct io_wqe *wqe = container_of(wait, struct io_wqe, 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_wqe_acct *acct = &wqe->acct[i];
1133  
1134  		if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1135  			io_wqe_activate_free_worker(wqe, acct);
1136  	}
1137  	rcu_read_unlock();
1138  	return 1;
1139  }
1140  
1141  struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1142  {
1143  	int ret, node, 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(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1152  	if (!wq)
1153  		return ERR_PTR(-ENOMEM);
1154  	ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1155  	if (ret)
1156  		goto err_wq;
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  	for_each_node(node) {
1165  		struct io_wqe *wqe;
1166  		int alloc_node = node;
1167  
1168  		if (!node_online(alloc_node))
1169  			alloc_node = NUMA_NO_NODE;
1170  		wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1171  		if (!wqe)
1172  			goto err;
1173  		wq->wqes[node] = wqe;
1174  		if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1175  			goto err;
1176  		cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1177  		wqe->node = alloc_node;
1178  		wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1179  		wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1180  					task_rlimit(current, RLIMIT_NPROC);
1181  		INIT_LIST_HEAD(&wqe->wait.entry);
1182  		wqe->wait.func = io_wqe_hash_wake;
1183  		for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1184  			struct io_wqe_acct *acct = &wqe->acct[i];
1185  
1186  			acct->index = i;
1187  			atomic_set(&acct->nr_running, 0);
1188  			INIT_WQ_LIST(&acct->work_list);
1189  			raw_spin_lock_init(&acct->lock);
1190  		}
1191  		wqe->wq = wq;
1192  		raw_spin_lock_init(&wqe->lock);
1193  		INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1194  		INIT_LIST_HEAD(&wqe->all_list);
1195  	}
1196  
1197  	wq->task = get_task_struct(data->task);
1198  	atomic_set(&wq->worker_refs, 1);
1199  	init_completion(&wq->worker_done);
1200  	return wq;
1201  err:
1202  	io_wq_put_hash(data->hash);
1203  	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1204  	for_each_node(node) {
1205  		if (!wq->wqes[node])
1206  			continue;
1207  		free_cpumask_var(wq->wqes[node]->cpu_mask);
1208  		kfree(wq->wqes[node]);
1209  	}
1210  err_wq:
1211  	kfree(wq);
1212  	return ERR_PTR(ret);
1213  }
1214  
1215  static bool io_task_work_match(struct callback_head *cb, void *data)
1216  {
1217  	struct io_worker *worker;
1218  
1219  	if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1220  		return false;
1221  	worker = container_of(cb, struct io_worker, create_work);
1222  	return worker->wqe->wq == data;
1223  }
1224  
1225  void io_wq_exit_start(struct io_wq *wq)
1226  {
1227  	set_bit(IO_WQ_BIT_EXIT, &wq->state);
1228  }
1229  
1230  static void io_wq_cancel_tw_create(struct io_wq *wq)
1231  {
1232  	struct callback_head *cb;
1233  
1234  	while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1235  		struct io_worker *worker;
1236  
1237  		worker = container_of(cb, struct io_worker, create_work);
1238  		io_worker_cancel_cb(worker);
1239  		/*
1240  		 * Only the worker continuation helper has worker allocated and
1241  		 * hence needs freeing.
1242  		 */
1243  		if (cb->func == create_worker_cont)
1244  			kfree(worker);
1245  	}
1246  }
1247  
1248  static void io_wq_exit_workers(struct io_wq *wq)
1249  {
1250  	int node;
1251  
1252  	if (!wq->task)
1253  		return;
1254  
1255  	io_wq_cancel_tw_create(wq);
1256  
1257  	rcu_read_lock();
1258  	for_each_node(node) {
1259  		struct io_wqe *wqe = wq->wqes[node];
1260  
1261  		io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1262  	}
1263  	rcu_read_unlock();
1264  	io_worker_ref_put(wq);
1265  	wait_for_completion(&wq->worker_done);
1266  
1267  	for_each_node(node) {
1268  		spin_lock_irq(&wq->hash->wait.lock);
1269  		list_del_init(&wq->wqes[node]->wait.entry);
1270  		spin_unlock_irq(&wq->hash->wait.lock);
1271  	}
1272  	put_task_struct(wq->task);
1273  	wq->task = NULL;
1274  }
1275  
1276  static void io_wq_destroy(struct io_wq *wq)
1277  {
1278  	int node;
1279  
1280  	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1281  
1282  	for_each_node(node) {
1283  		struct io_wqe *wqe = wq->wqes[node];
1284  		struct io_cb_cancel_data match = {
1285  			.fn		= io_wq_work_match_all,
1286  			.cancel_all	= true,
1287  		};
1288  		io_wqe_cancel_pending_work(wqe, &match);
1289  		free_cpumask_var(wqe->cpu_mask);
1290  		kfree(wqe);
1291  	}
1292  	io_wq_put_hash(wq->hash);
1293  	kfree(wq);
1294  }
1295  
1296  void io_wq_put_and_exit(struct io_wq *wq)
1297  {
1298  	WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1299  
1300  	io_wq_exit_workers(wq);
1301  	io_wq_destroy(wq);
1302  }
1303  
1304  struct online_data {
1305  	unsigned int cpu;
1306  	bool online;
1307  };
1308  
1309  static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1310  {
1311  	struct online_data *od = data;
1312  
1313  	if (od->online)
1314  		cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1315  	else
1316  		cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1317  	return false;
1318  }
1319  
1320  static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1321  {
1322  	struct online_data od = {
1323  		.cpu = cpu,
1324  		.online = online
1325  	};
1326  	int i;
1327  
1328  	rcu_read_lock();
1329  	for_each_node(i)
1330  		io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1331  	rcu_read_unlock();
1332  	return 0;
1333  }
1334  
1335  static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1336  {
1337  	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1338  
1339  	return __io_wq_cpu_online(wq, cpu, true);
1340  }
1341  
1342  static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1343  {
1344  	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1345  
1346  	return __io_wq_cpu_online(wq, cpu, false);
1347  }
1348  
1349  int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1350  {
1351  	int i;
1352  
1353  	rcu_read_lock();
1354  	for_each_node(i) {
1355  		struct io_wqe *wqe = wq->wqes[i];
1356  
1357  		if (mask)
1358  			cpumask_copy(wqe->cpu_mask, mask);
1359  		else
1360  			cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1361  	}
1362  	rcu_read_unlock();
1363  	return 0;
1364  }
1365  
1366  /*
1367   * Set max number of unbounded workers, returns old value. If new_count is 0,
1368   * then just return the old value.
1369   */
1370  int io_wq_max_workers(struct io_wq *wq, int *new_count)
1371  {
1372  	int prev[IO_WQ_ACCT_NR];
1373  	bool first_node = true;
1374  	int i, node;
1375  
1376  	BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND   != (int) IO_WQ_BOUND);
1377  	BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1378  	BUILD_BUG_ON((int) IO_WQ_ACCT_NR      != 2);
1379  
1380  	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1381  		if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1382  			new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1383  	}
1384  
1385  	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1386  		prev[i] = 0;
1387  
1388  	rcu_read_lock();
1389  	for_each_node(node) {
1390  		struct io_wqe *wqe = wq->wqes[node];
1391  		struct io_wqe_acct *acct;
1392  
1393  		raw_spin_lock(&wqe->lock);
1394  		for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1395  			acct = &wqe->acct[i];
1396  			if (first_node)
1397  				prev[i] = max_t(int, acct->max_workers, prev[i]);
1398  			if (new_count[i])
1399  				acct->max_workers = new_count[i];
1400  		}
1401  		raw_spin_unlock(&wqe->lock);
1402  		first_node = false;
1403  	}
1404  	rcu_read_unlock();
1405  
1406  	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1407  		new_count[i] = prev[i];
1408  
1409  	return 0;
1410  }
1411  
1412  static __init int io_wq_init(void)
1413  {
1414  	int ret;
1415  
1416  	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1417  					io_wq_cpu_online, io_wq_cpu_offline);
1418  	if (ret < 0)
1419  		return ret;
1420  	io_wq_online = ret;
1421  	return 0;
1422  }
1423  subsys_initcall(io_wq_init);
1424