1 /* Worker thread pool for slow items, such as filesystem lookups or mkdirs
3 * Copyright (C) 2008 Red Hat, Inc. All Rights Reserved.
4 * Written by David Howells (dhowells@redhat.com)
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public Licence
8 * as published by the Free Software Foundation; either version
9 * 2 of the Licence, or (at your option) any later version.
11 * See Documentation/slow-work.txt
14 #include <linux/module.h>
15 #include <linux/slow-work.h>
16 #include <linux/kthread.h>
17 #include <linux/freezer.h>
18 #include <linux/wait.h>
20 #define SLOW_WORK_CULL_TIMEOUT (5 * HZ) /* cull threads 5s after running out of
22 #define SLOW_WORK_OOM_TIMEOUT (5 * HZ) /* can't start new threads for 5s after
25 static void slow_work_cull_timeout(unsigned long);
26 static void slow_work_oom_timeout(unsigned long);
29 static int slow_work_min_threads_sysctl(struct ctl_table *, int, struct file *,
30 void __user *, size_t *, loff_t *);
32 static int slow_work_max_threads_sysctl(struct ctl_table *, int , struct file *,
33 void __user *, size_t *, loff_t *);
37 * The pool of threads has at least min threads in it as long as someone is
38 * using the facility, and may have as many as max.
40 * A portion of the pool may be processing very slow operations.
42 static unsigned slow_work_min_threads = 2;
43 static unsigned slow_work_max_threads = 4;
44 static unsigned vslow_work_proportion = 50; /* % of threads that may process
48 static const int slow_work_min_min_threads = 2;
49 static int slow_work_max_max_threads = 255;
50 static const int slow_work_min_vslow = 1;
51 static const int slow_work_max_vslow = 99;
53 ctl_table slow_work_sysctls[] = {
55 .ctl_name = CTL_UNNUMBERED,
56 .procname = "min-threads",
57 .data = &slow_work_min_threads,
58 .maxlen = sizeof(unsigned),
60 .proc_handler = slow_work_min_threads_sysctl,
61 .extra1 = (void *) &slow_work_min_min_threads,
62 .extra2 = &slow_work_max_threads,
65 .ctl_name = CTL_UNNUMBERED,
66 .procname = "max-threads",
67 .data = &slow_work_max_threads,
68 .maxlen = sizeof(unsigned),
70 .proc_handler = slow_work_max_threads_sysctl,
71 .extra1 = &slow_work_min_threads,
72 .extra2 = (void *) &slow_work_max_max_threads,
75 .ctl_name = CTL_UNNUMBERED,
76 .procname = "vslow-percentage",
77 .data = &vslow_work_proportion,
78 .maxlen = sizeof(unsigned),
80 .proc_handler = &proc_dointvec_minmax,
81 .extra1 = (void *) &slow_work_min_vslow,
82 .extra2 = (void *) &slow_work_max_vslow,
89 * The active state of the thread pool
91 static atomic_t slow_work_thread_count;
92 static atomic_t vslow_work_executing_count;
94 static bool slow_work_may_not_start_new_thread;
95 static bool slow_work_cull; /* cull a thread due to lack of activity */
96 static DEFINE_TIMER(slow_work_cull_timer, slow_work_cull_timeout, 0, 0);
97 static DEFINE_TIMER(slow_work_oom_timer, slow_work_oom_timeout, 0, 0);
98 static struct slow_work slow_work_new_thread; /* new thread starter */
101 * The queues of work items and the lock governing access to them. These are
102 * shared between all the CPUs. It doesn't make sense to have per-CPU queues
103 * as the number of threads bears no relation to the number of CPUs.
105 * There are two queues of work items: one for slow work items, and one for
106 * very slow work items.
108 static LIST_HEAD(slow_work_queue);
109 static LIST_HEAD(vslow_work_queue);
110 static DEFINE_SPINLOCK(slow_work_queue_lock);
113 * The thread controls. A variable used to signal to the threads that they
114 * should exit when the queue is empty, a waitqueue used by the threads to wait
115 * for signals, and a completion set by the last thread to exit.
117 static bool slow_work_threads_should_exit;
118 static DECLARE_WAIT_QUEUE_HEAD(slow_work_thread_wq);
119 static DECLARE_COMPLETION(slow_work_last_thread_exited);
122 * The number of users of the thread pool and its lock. Whilst this is zero we
123 * have no threads hanging around, and when this reaches zero, we wait for all
124 * active or queued work items to complete and kill all the threads we do have.
126 static int slow_work_user_count;
127 static DEFINE_MUTEX(slow_work_user_lock);
130 * Calculate the maximum number of active threads in the pool that are
131 * permitted to process very slow work items.
133 * The answer is rounded up to at least 1, but may not equal or exceed the
134 * maximum number of the threads in the pool. This means we always have at
135 * least one thread that can process slow work items, and we always have at
136 * least one thread that won't get tied up doing so.
138 static unsigned slow_work_calc_vsmax(void)
142 vsmax = atomic_read(&slow_work_thread_count) * vslow_work_proportion;
144 vsmax = max(vsmax, 1U);
145 return min(vsmax, slow_work_max_threads - 1);
149 * Attempt to execute stuff queued on a slow thread. Return true if we managed
150 * it, false if there was nothing to do.
152 static bool slow_work_execute(void)
154 struct slow_work *work = NULL;
158 vsmax = slow_work_calc_vsmax();
160 /* see if we can schedule a new thread to be started if we're not
161 * keeping up with the work */
162 if (!waitqueue_active(&slow_work_thread_wq) &&
163 (!list_empty(&slow_work_queue) || !list_empty(&vslow_work_queue)) &&
164 atomic_read(&slow_work_thread_count) < slow_work_max_threads &&
165 !slow_work_may_not_start_new_thread)
166 slow_work_enqueue(&slow_work_new_thread);
168 /* find something to execute */
169 spin_lock_irq(&slow_work_queue_lock);
170 if (!list_empty(&vslow_work_queue) &&
171 atomic_read(&vslow_work_executing_count) < vsmax) {
172 work = list_entry(vslow_work_queue.next,
173 struct slow_work, link);
174 if (test_and_set_bit_lock(SLOW_WORK_EXECUTING, &work->flags))
176 list_del_init(&work->link);
177 atomic_inc(&vslow_work_executing_count);
179 } else if (!list_empty(&slow_work_queue)) {
180 work = list_entry(slow_work_queue.next,
181 struct slow_work, link);
182 if (test_and_set_bit_lock(SLOW_WORK_EXECUTING, &work->flags))
184 list_del_init(&work->link);
187 very_slow = false; /* avoid the compiler warning */
189 spin_unlock_irq(&slow_work_queue_lock);
194 if (!test_and_clear_bit(SLOW_WORK_PENDING, &work->flags))
197 work->ops->execute(work);
200 atomic_dec(&vslow_work_executing_count);
201 clear_bit_unlock(SLOW_WORK_EXECUTING, &work->flags);
203 /* if someone tried to enqueue the item whilst we were executing it,
204 * then it'll be left unenqueued to avoid multiple threads trying to
205 * execute it simultaneously
207 * there is, however, a race between us testing the pending flag and
208 * getting the spinlock, and between the enqueuer setting the pending
209 * flag and getting the spinlock, so we use a deferral bit to tell us
210 * if the enqueuer got there first
212 if (test_bit(SLOW_WORK_PENDING, &work->flags)) {
213 spin_lock_irq(&slow_work_queue_lock);
215 if (!test_bit(SLOW_WORK_EXECUTING, &work->flags) &&
216 test_and_clear_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags))
219 spin_unlock_irq(&slow_work_queue_lock);
222 work->ops->put_ref(work);
226 /* we must complete the enqueue operation
227 * - we transfer our ref on the item back to the appropriate queue
228 * - don't wake another thread up as we're awake already
230 if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags))
231 list_add_tail(&work->link, &vslow_work_queue);
233 list_add_tail(&work->link, &slow_work_queue);
234 spin_unlock_irq(&slow_work_queue_lock);
239 * slow_work_enqueue - Schedule a slow work item for processing
240 * @work: The work item to queue
242 * Schedule a slow work item for processing. If the item is already undergoing
243 * execution, this guarantees not to re-enter the execution routine until the
244 * first execution finishes.
246 * The item is pinned by this function as it retains a reference to it, managed
247 * through the item operations. The item is unpinned once it has been
250 * An item may hog the thread that is running it for a relatively large amount
251 * of time, sufficient, for example, to perform several lookup, mkdir, create
252 * and setxattr operations. It may sleep on I/O and may sleep to obtain locks.
254 * Conversely, if a number of items are awaiting processing, it may take some
255 * time before any given item is given attention. The number of threads in the
256 * pool may be increased to deal with demand, but only up to a limit.
258 * If SLOW_WORK_VERY_SLOW is set on the work item, then it will be placed in
259 * the very slow queue, from which only a portion of the threads will be
260 * allowed to pick items to execute. This ensures that very slow items won't
261 * overly block ones that are just ordinarily slow.
263 * Returns 0 if successful, -EAGAIN if not.
265 int slow_work_enqueue(struct slow_work *work)
269 BUG_ON(slow_work_user_count <= 0);
272 BUG_ON(!work->ops->get_ref);
274 /* when honouring an enqueue request, we only promise that we will run
275 * the work function in the future; we do not promise to run it once
276 * per enqueue request
278 * we use the PENDING bit to merge together repeat requests without
279 * having to disable IRQs and take the spinlock, whilst still
280 * maintaining our promise
282 if (!test_and_set_bit_lock(SLOW_WORK_PENDING, &work->flags)) {
283 spin_lock_irqsave(&slow_work_queue_lock, flags);
285 /* we promise that we will not attempt to execute the work
286 * function in more than one thread simultaneously
288 * this, however, leaves us with a problem if we're asked to
289 * enqueue the work whilst someone is executing the work
290 * function as simply queueing the work immediately means that
291 * another thread may try executing it whilst it is already
294 * to deal with this, we set the ENQ_DEFERRED bit instead of
295 * enqueueing, and the thread currently executing the work
296 * function will enqueue the work item when the work function
297 * returns and it has cleared the EXECUTING bit
299 if (test_bit(SLOW_WORK_EXECUTING, &work->flags)) {
300 set_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags);
302 if (work->ops->get_ref(work) < 0)
304 if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags))
305 list_add_tail(&work->link, &vslow_work_queue);
307 list_add_tail(&work->link, &slow_work_queue);
308 wake_up(&slow_work_thread_wq);
311 spin_unlock_irqrestore(&slow_work_queue_lock, flags);
316 spin_unlock_irqrestore(&slow_work_queue_lock, flags);
319 EXPORT_SYMBOL(slow_work_enqueue);
322 * Worker thread culling algorithm
324 static bool slow_work_cull_thread(void)
327 bool do_cull = false;
329 spin_lock_irqsave(&slow_work_queue_lock, flags);
331 if (slow_work_cull) {
332 slow_work_cull = false;
334 if (list_empty(&slow_work_queue) &&
335 list_empty(&vslow_work_queue) &&
336 atomic_read(&slow_work_thread_count) >
337 slow_work_min_threads) {
338 mod_timer(&slow_work_cull_timer,
339 jiffies + SLOW_WORK_CULL_TIMEOUT);
344 spin_unlock_irqrestore(&slow_work_queue_lock, flags);
349 * Determine if there is slow work available for dispatch
351 static inline bool slow_work_available(int vsmax)
353 return !list_empty(&slow_work_queue) ||
354 (!list_empty(&vslow_work_queue) &&
355 atomic_read(&vslow_work_executing_count) < vsmax);
359 * Worker thread dispatcher
361 static int slow_work_thread(void *_data)
368 set_user_nice(current, -5);
371 vsmax = vslow_work_proportion;
372 vsmax *= atomic_read(&slow_work_thread_count);
375 prepare_to_wait_exclusive(&slow_work_thread_wq, &wait,
377 if (!freezing(current) &&
378 !slow_work_threads_should_exit &&
379 !slow_work_available(vsmax) &&
382 finish_wait(&slow_work_thread_wq, &wait);
386 vsmax = vslow_work_proportion;
387 vsmax *= atomic_read(&slow_work_thread_count);
390 if (slow_work_available(vsmax) && slow_work_execute()) {
392 if (list_empty(&slow_work_queue) &&
393 list_empty(&vslow_work_queue) &&
394 atomic_read(&slow_work_thread_count) >
395 slow_work_min_threads)
396 mod_timer(&slow_work_cull_timer,
397 jiffies + SLOW_WORK_CULL_TIMEOUT);
401 if (slow_work_threads_should_exit)
404 if (slow_work_cull && slow_work_cull_thread())
408 if (atomic_dec_and_test(&slow_work_thread_count))
409 complete_and_exit(&slow_work_last_thread_exited, 0);
414 * Handle thread cull timer expiration
416 static void slow_work_cull_timeout(unsigned long data)
418 slow_work_cull = true;
419 wake_up(&slow_work_thread_wq);
423 * Get a reference on slow work thread starter
425 static int slow_work_new_thread_get_ref(struct slow_work *work)
431 * Drop a reference on slow work thread starter
433 static void slow_work_new_thread_put_ref(struct slow_work *work)
438 * Start a new slow work thread
440 static void slow_work_new_thread_execute(struct slow_work *work)
442 struct task_struct *p;
444 if (slow_work_threads_should_exit)
447 if (atomic_read(&slow_work_thread_count) >= slow_work_max_threads)
450 if (!mutex_trylock(&slow_work_user_lock))
453 slow_work_may_not_start_new_thread = true;
454 atomic_inc(&slow_work_thread_count);
455 p = kthread_run(slow_work_thread, NULL, "kslowd");
457 printk(KERN_DEBUG "Slow work thread pool: OOM\n");
458 if (atomic_dec_and_test(&slow_work_thread_count))
459 BUG(); /* we're running on a slow work thread... */
460 mod_timer(&slow_work_oom_timer,
461 jiffies + SLOW_WORK_OOM_TIMEOUT);
463 /* ratelimit the starting of new threads */
464 mod_timer(&slow_work_oom_timer, jiffies + 1);
467 mutex_unlock(&slow_work_user_lock);
470 static const struct slow_work_ops slow_work_new_thread_ops = {
471 .get_ref = slow_work_new_thread_get_ref,
472 .put_ref = slow_work_new_thread_put_ref,
473 .execute = slow_work_new_thread_execute,
477 * post-OOM new thread start suppression expiration
479 static void slow_work_oom_timeout(unsigned long data)
481 slow_work_may_not_start_new_thread = false;
486 * Handle adjustment of the minimum number of threads
488 static int slow_work_min_threads_sysctl(struct ctl_table *table, int write,
489 struct file *filp, void __user *buffer,
490 size_t *lenp, loff_t *ppos)
492 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
496 mutex_lock(&slow_work_user_lock);
497 if (slow_work_user_count > 0) {
498 /* see if we need to start or stop threads */
499 n = atomic_read(&slow_work_thread_count) -
500 slow_work_min_threads;
502 if (n < 0 && !slow_work_may_not_start_new_thread)
503 slow_work_enqueue(&slow_work_new_thread);
505 mod_timer(&slow_work_cull_timer,
506 jiffies + SLOW_WORK_CULL_TIMEOUT);
508 mutex_unlock(&slow_work_user_lock);
515 * Handle adjustment of the maximum number of threads
517 static int slow_work_max_threads_sysctl(struct ctl_table *table, int write,
518 struct file *filp, void __user *buffer,
519 size_t *lenp, loff_t *ppos)
521 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
525 mutex_lock(&slow_work_user_lock);
526 if (slow_work_user_count > 0) {
527 /* see if we need to stop threads */
528 n = slow_work_max_threads -
529 atomic_read(&slow_work_thread_count);
532 mod_timer(&slow_work_cull_timer,
533 jiffies + SLOW_WORK_CULL_TIMEOUT);
535 mutex_unlock(&slow_work_user_lock);
540 #endif /* CONFIG_SYSCTL */
543 * slow_work_register_user - Register a user of the facility
545 * Register a user of the facility, starting up the initial threads if there
546 * aren't any other users at this point. This will return 0 if successful, or
549 int slow_work_register_user(void)
551 struct task_struct *p;
554 mutex_lock(&slow_work_user_lock);
556 if (slow_work_user_count == 0) {
557 printk(KERN_NOTICE "Slow work thread pool: Starting up\n");
558 init_completion(&slow_work_last_thread_exited);
560 slow_work_threads_should_exit = false;
561 slow_work_init(&slow_work_new_thread,
562 &slow_work_new_thread_ops);
563 slow_work_may_not_start_new_thread = false;
564 slow_work_cull = false;
566 /* start the minimum number of threads */
567 for (loop = 0; loop < slow_work_min_threads; loop++) {
568 atomic_inc(&slow_work_thread_count);
569 p = kthread_run(slow_work_thread, NULL, "kslowd");
573 printk(KERN_NOTICE "Slow work thread pool: Ready\n");
576 slow_work_user_count++;
577 mutex_unlock(&slow_work_user_lock);
581 if (atomic_dec_and_test(&slow_work_thread_count))
582 complete(&slow_work_last_thread_exited);
584 printk(KERN_ERR "Slow work thread pool:"
585 " Aborting startup on ENOMEM\n");
586 slow_work_threads_should_exit = true;
587 wake_up_all(&slow_work_thread_wq);
588 wait_for_completion(&slow_work_last_thread_exited);
589 printk(KERN_ERR "Slow work thread pool: Aborted\n");
591 mutex_unlock(&slow_work_user_lock);
594 EXPORT_SYMBOL(slow_work_register_user);
597 * slow_work_unregister_user - Unregister a user of the facility
599 * Unregister a user of the facility, killing all the threads if this was the
602 void slow_work_unregister_user(void)
604 mutex_lock(&slow_work_user_lock);
606 BUG_ON(slow_work_user_count <= 0);
608 slow_work_user_count--;
609 if (slow_work_user_count == 0) {
610 printk(KERN_NOTICE "Slow work thread pool: Shutting down\n");
611 slow_work_threads_should_exit = true;
612 del_timer_sync(&slow_work_cull_timer);
613 del_timer_sync(&slow_work_oom_timer);
614 wake_up_all(&slow_work_thread_wq);
615 wait_for_completion(&slow_work_last_thread_exited);
616 printk(KERN_NOTICE "Slow work thread pool:"
617 " Shut down complete\n");
620 mutex_unlock(&slow_work_user_lock);
622 EXPORT_SYMBOL(slow_work_unregister_user);
625 * Initialise the slow work facility
627 static int __init init_slow_work(void)
629 unsigned nr_cpus = num_possible_cpus();
631 if (slow_work_max_threads < nr_cpus)
632 slow_work_max_threads = nr_cpus;
634 if (slow_work_max_max_threads < nr_cpus * 2)
635 slow_work_max_max_threads = nr_cpus * 2;
640 subsys_initcall(init_slow_work);