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,
30 void __user
*, size_t *, loff_t
*);
32 static int slow_work_max_threads_sysctl(struct ctl_table
*, int ,
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 .procname
= "min-threads",
56 .data
= &slow_work_min_threads
,
57 .maxlen
= sizeof(unsigned),
59 .proc_handler
= slow_work_min_threads_sysctl
,
60 .extra1
= (void *) &slow_work_min_min_threads
,
61 .extra2
= &slow_work_max_threads
,
64 .procname
= "max-threads",
65 .data
= &slow_work_max_threads
,
66 .maxlen
= sizeof(unsigned),
68 .proc_handler
= slow_work_max_threads_sysctl
,
69 .extra1
= &slow_work_min_threads
,
70 .extra2
= (void *) &slow_work_max_max_threads
,
73 .procname
= "vslow-percentage",
74 .data
= &vslow_work_proportion
,
75 .maxlen
= sizeof(unsigned),
77 .proc_handler
= &proc_dointvec_minmax
,
78 .extra1
= (void *) &slow_work_min_vslow
,
79 .extra2
= (void *) &slow_work_max_vslow
,
86 * The active state of the thread pool
88 static atomic_t slow_work_thread_count
;
89 static atomic_t vslow_work_executing_count
;
91 static bool slow_work_may_not_start_new_thread
;
92 static bool slow_work_cull
; /* cull a thread due to lack of activity */
93 static DEFINE_TIMER(slow_work_cull_timer
, slow_work_cull_timeout
, 0, 0);
94 static DEFINE_TIMER(slow_work_oom_timer
, slow_work_oom_timeout
, 0, 0);
95 static struct slow_work slow_work_new_thread
; /* new thread starter */
98 * The queues of work items and the lock governing access to them. These are
99 * shared between all the CPUs. It doesn't make sense to have per-CPU queues
100 * as the number of threads bears no relation to the number of CPUs.
102 * There are two queues of work items: one for slow work items, and one for
103 * very slow work items.
105 static LIST_HEAD(slow_work_queue
);
106 static LIST_HEAD(vslow_work_queue
);
107 static DEFINE_SPINLOCK(slow_work_queue_lock
);
110 * The thread controls. A variable used to signal to the threads that they
111 * should exit when the queue is empty, a waitqueue used by the threads to wait
112 * for signals, and a completion set by the last thread to exit.
114 static bool slow_work_threads_should_exit
;
115 static DECLARE_WAIT_QUEUE_HEAD(slow_work_thread_wq
);
116 static DECLARE_COMPLETION(slow_work_last_thread_exited
);
119 * The number of users of the thread pool and its lock. Whilst this is zero we
120 * have no threads hanging around, and when this reaches zero, we wait for all
121 * active or queued work items to complete and kill all the threads we do have.
123 static int slow_work_user_count
;
124 static DEFINE_MUTEX(slow_work_user_lock
);
127 * Calculate the maximum number of active threads in the pool that are
128 * permitted to process very slow work items.
130 * The answer is rounded up to at least 1, but may not equal or exceed the
131 * maximum number of the threads in the pool. This means we always have at
132 * least one thread that can process slow work items, and we always have at
133 * least one thread that won't get tied up doing so.
135 static unsigned slow_work_calc_vsmax(void)
139 vsmax
= atomic_read(&slow_work_thread_count
) * vslow_work_proportion
;
141 vsmax
= max(vsmax
, 1U);
142 return min(vsmax
, slow_work_max_threads
- 1);
146 * Attempt to execute stuff queued on a slow thread. Return true if we managed
147 * it, false if there was nothing to do.
149 static bool slow_work_execute(void)
151 struct slow_work
*work
= NULL
;
155 vsmax
= slow_work_calc_vsmax();
157 /* see if we can schedule a new thread to be started if we're not
158 * keeping up with the work */
159 if (!waitqueue_active(&slow_work_thread_wq
) &&
160 (!list_empty(&slow_work_queue
) || !list_empty(&vslow_work_queue
)) &&
161 atomic_read(&slow_work_thread_count
) < slow_work_max_threads
&&
162 !slow_work_may_not_start_new_thread
)
163 slow_work_enqueue(&slow_work_new_thread
);
165 /* find something to execute */
166 spin_lock_irq(&slow_work_queue_lock
);
167 if (!list_empty(&vslow_work_queue
) &&
168 atomic_read(&vslow_work_executing_count
) < vsmax
) {
169 work
= list_entry(vslow_work_queue
.next
,
170 struct slow_work
, link
);
171 if (test_and_set_bit_lock(SLOW_WORK_EXECUTING
, &work
->flags
))
173 list_del_init(&work
->link
);
174 atomic_inc(&vslow_work_executing_count
);
176 } else if (!list_empty(&slow_work_queue
)) {
177 work
= list_entry(slow_work_queue
.next
,
178 struct slow_work
, link
);
179 if (test_and_set_bit_lock(SLOW_WORK_EXECUTING
, &work
->flags
))
181 list_del_init(&work
->link
);
184 very_slow
= false; /* avoid the compiler warning */
186 spin_unlock_irq(&slow_work_queue_lock
);
191 if (!test_and_clear_bit(SLOW_WORK_PENDING
, &work
->flags
))
194 work
->ops
->execute(work
);
197 atomic_dec(&vslow_work_executing_count
);
198 clear_bit_unlock(SLOW_WORK_EXECUTING
, &work
->flags
);
200 /* if someone tried to enqueue the item whilst we were executing it,
201 * then it'll be left unenqueued to avoid multiple threads trying to
202 * execute it simultaneously
204 * there is, however, a race between us testing the pending flag and
205 * getting the spinlock, and between the enqueuer setting the pending
206 * flag and getting the spinlock, so we use a deferral bit to tell us
207 * if the enqueuer got there first
209 if (test_bit(SLOW_WORK_PENDING
, &work
->flags
)) {
210 spin_lock_irq(&slow_work_queue_lock
);
212 if (!test_bit(SLOW_WORK_EXECUTING
, &work
->flags
) &&
213 test_and_clear_bit(SLOW_WORK_ENQ_DEFERRED
, &work
->flags
))
216 spin_unlock_irq(&slow_work_queue_lock
);
219 work
->ops
->put_ref(work
);
223 /* we must complete the enqueue operation
224 * - we transfer our ref on the item back to the appropriate queue
225 * - don't wake another thread up as we're awake already
227 if (test_bit(SLOW_WORK_VERY_SLOW
, &work
->flags
))
228 list_add_tail(&work
->link
, &vslow_work_queue
);
230 list_add_tail(&work
->link
, &slow_work_queue
);
231 spin_unlock_irq(&slow_work_queue_lock
);
236 * slow_work_enqueue - Schedule a slow work item for processing
237 * @work: The work item to queue
239 * Schedule a slow work item for processing. If the item is already undergoing
240 * execution, this guarantees not to re-enter the execution routine until the
241 * first execution finishes.
243 * The item is pinned by this function as it retains a reference to it, managed
244 * through the item operations. The item is unpinned once it has been
247 * An item may hog the thread that is running it for a relatively large amount
248 * of time, sufficient, for example, to perform several lookup, mkdir, create
249 * and setxattr operations. It may sleep on I/O and may sleep to obtain locks.
251 * Conversely, if a number of items are awaiting processing, it may take some
252 * time before any given item is given attention. The number of threads in the
253 * pool may be increased to deal with demand, but only up to a limit.
255 * If SLOW_WORK_VERY_SLOW is set on the work item, then it will be placed in
256 * the very slow queue, from which only a portion of the threads will be
257 * allowed to pick items to execute. This ensures that very slow items won't
258 * overly block ones that are just ordinarily slow.
260 * Returns 0 if successful, -EAGAIN if not.
262 int slow_work_enqueue(struct slow_work
*work
)
266 BUG_ON(slow_work_user_count
<= 0);
269 BUG_ON(!work
->ops
->get_ref
);
271 /* when honouring an enqueue request, we only promise that we will run
272 * the work function in the future; we do not promise to run it once
273 * per enqueue request
275 * we use the PENDING bit to merge together repeat requests without
276 * having to disable IRQs and take the spinlock, whilst still
277 * maintaining our promise
279 if (!test_and_set_bit_lock(SLOW_WORK_PENDING
, &work
->flags
)) {
280 spin_lock_irqsave(&slow_work_queue_lock
, flags
);
282 /* we promise that we will not attempt to execute the work
283 * function in more than one thread simultaneously
285 * this, however, leaves us with a problem if we're asked to
286 * enqueue the work whilst someone is executing the work
287 * function as simply queueing the work immediately means that
288 * another thread may try executing it whilst it is already
291 * to deal with this, we set the ENQ_DEFERRED bit instead of
292 * enqueueing, and the thread currently executing the work
293 * function will enqueue the work item when the work function
294 * returns and it has cleared the EXECUTING bit
296 if (test_bit(SLOW_WORK_EXECUTING
, &work
->flags
)) {
297 set_bit(SLOW_WORK_ENQ_DEFERRED
, &work
->flags
);
299 if (work
->ops
->get_ref(work
) < 0)
301 if (test_bit(SLOW_WORK_VERY_SLOW
, &work
->flags
))
302 list_add_tail(&work
->link
, &vslow_work_queue
);
304 list_add_tail(&work
->link
, &slow_work_queue
);
305 wake_up(&slow_work_thread_wq
);
308 spin_unlock_irqrestore(&slow_work_queue_lock
, flags
);
313 spin_unlock_irqrestore(&slow_work_queue_lock
, flags
);
316 EXPORT_SYMBOL(slow_work_enqueue
);
319 * Schedule a cull of the thread pool at some time in the near future
321 static void slow_work_schedule_cull(void)
323 mod_timer(&slow_work_cull_timer
,
324 round_jiffies(jiffies
+ SLOW_WORK_CULL_TIMEOUT
));
328 * Worker thread culling algorithm
330 static bool slow_work_cull_thread(void)
333 bool do_cull
= false;
335 spin_lock_irqsave(&slow_work_queue_lock
, flags
);
337 if (slow_work_cull
) {
338 slow_work_cull
= false;
340 if (list_empty(&slow_work_queue
) &&
341 list_empty(&vslow_work_queue
) &&
342 atomic_read(&slow_work_thread_count
) >
343 slow_work_min_threads
) {
344 slow_work_schedule_cull();
349 spin_unlock_irqrestore(&slow_work_queue_lock
, flags
);
354 * Determine if there is slow work available for dispatch
356 static inline bool slow_work_available(int vsmax
)
358 return !list_empty(&slow_work_queue
) ||
359 (!list_empty(&vslow_work_queue
) &&
360 atomic_read(&vslow_work_executing_count
) < vsmax
);
364 * Worker thread dispatcher
366 static int slow_work_thread(void *_data
)
373 set_user_nice(current
, -5);
376 vsmax
= vslow_work_proportion
;
377 vsmax
*= atomic_read(&slow_work_thread_count
);
380 prepare_to_wait_exclusive(&slow_work_thread_wq
, &wait
,
382 if (!freezing(current
) &&
383 !slow_work_threads_should_exit
&&
384 !slow_work_available(vsmax
) &&
387 finish_wait(&slow_work_thread_wq
, &wait
);
391 vsmax
= vslow_work_proportion
;
392 vsmax
*= atomic_read(&slow_work_thread_count
);
395 if (slow_work_available(vsmax
) && slow_work_execute()) {
397 if (list_empty(&slow_work_queue
) &&
398 list_empty(&vslow_work_queue
) &&
399 atomic_read(&slow_work_thread_count
) >
400 slow_work_min_threads
)
401 slow_work_schedule_cull();
405 if (slow_work_threads_should_exit
)
408 if (slow_work_cull
&& slow_work_cull_thread())
412 if (atomic_dec_and_test(&slow_work_thread_count
))
413 complete_and_exit(&slow_work_last_thread_exited
, 0);
418 * Handle thread cull timer expiration
420 static void slow_work_cull_timeout(unsigned long data
)
422 slow_work_cull
= true;
423 wake_up(&slow_work_thread_wq
);
427 * Get a reference on slow work thread starter
429 static int slow_work_new_thread_get_ref(struct slow_work
*work
)
435 * Drop a reference on slow work thread starter
437 static void slow_work_new_thread_put_ref(struct slow_work
*work
)
442 * Start a new slow work thread
444 static void slow_work_new_thread_execute(struct slow_work
*work
)
446 struct task_struct
*p
;
448 if (slow_work_threads_should_exit
)
451 if (atomic_read(&slow_work_thread_count
) >= slow_work_max_threads
)
454 if (!mutex_trylock(&slow_work_user_lock
))
457 slow_work_may_not_start_new_thread
= true;
458 atomic_inc(&slow_work_thread_count
);
459 p
= kthread_run(slow_work_thread
, NULL
, "kslowd");
461 printk(KERN_DEBUG
"Slow work thread pool: OOM\n");
462 if (atomic_dec_and_test(&slow_work_thread_count
))
463 BUG(); /* we're running on a slow work thread... */
464 mod_timer(&slow_work_oom_timer
,
465 round_jiffies(jiffies
+ SLOW_WORK_OOM_TIMEOUT
));
467 /* ratelimit the starting of new threads */
468 mod_timer(&slow_work_oom_timer
, jiffies
+ 1);
471 mutex_unlock(&slow_work_user_lock
);
474 static const struct slow_work_ops slow_work_new_thread_ops
= {
475 .get_ref
= slow_work_new_thread_get_ref
,
476 .put_ref
= slow_work_new_thread_put_ref
,
477 .execute
= slow_work_new_thread_execute
,
481 * post-OOM new thread start suppression expiration
483 static void slow_work_oom_timeout(unsigned long data
)
485 slow_work_may_not_start_new_thread
= false;
490 * Handle adjustment of the minimum number of threads
492 static int slow_work_min_threads_sysctl(struct ctl_table
*table
, int write
,
494 size_t *lenp
, loff_t
*ppos
)
496 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
500 mutex_lock(&slow_work_user_lock
);
501 if (slow_work_user_count
> 0) {
502 /* see if we need to start or stop threads */
503 n
= atomic_read(&slow_work_thread_count
) -
504 slow_work_min_threads
;
506 if (n
< 0 && !slow_work_may_not_start_new_thread
)
507 slow_work_enqueue(&slow_work_new_thread
);
509 slow_work_schedule_cull();
511 mutex_unlock(&slow_work_user_lock
);
518 * Handle adjustment of the maximum number of threads
520 static int slow_work_max_threads_sysctl(struct ctl_table
*table
, int write
,
522 size_t *lenp
, loff_t
*ppos
)
524 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
528 mutex_lock(&slow_work_user_lock
);
529 if (slow_work_user_count
> 0) {
530 /* see if we need to stop threads */
531 n
= slow_work_max_threads
-
532 atomic_read(&slow_work_thread_count
);
535 slow_work_schedule_cull();
537 mutex_unlock(&slow_work_user_lock
);
542 #endif /* CONFIG_SYSCTL */
545 * slow_work_register_user - Register a user of the facility
547 * Register a user of the facility, starting up the initial threads if there
548 * aren't any other users at this point. This will return 0 if successful, or
551 int slow_work_register_user(void)
553 struct task_struct
*p
;
556 mutex_lock(&slow_work_user_lock
);
558 if (slow_work_user_count
== 0) {
559 printk(KERN_NOTICE
"Slow work thread pool: Starting up\n");
560 init_completion(&slow_work_last_thread_exited
);
562 slow_work_threads_should_exit
= false;
563 slow_work_init(&slow_work_new_thread
,
564 &slow_work_new_thread_ops
);
565 slow_work_may_not_start_new_thread
= false;
566 slow_work_cull
= false;
568 /* start the minimum number of threads */
569 for (loop
= 0; loop
< slow_work_min_threads
; loop
++) {
570 atomic_inc(&slow_work_thread_count
);
571 p
= kthread_run(slow_work_thread
, NULL
, "kslowd");
575 printk(KERN_NOTICE
"Slow work thread pool: Ready\n");
578 slow_work_user_count
++;
579 mutex_unlock(&slow_work_user_lock
);
583 if (atomic_dec_and_test(&slow_work_thread_count
))
584 complete(&slow_work_last_thread_exited
);
586 printk(KERN_ERR
"Slow work thread pool:"
587 " Aborting startup on ENOMEM\n");
588 slow_work_threads_should_exit
= true;
589 wake_up_all(&slow_work_thread_wq
);
590 wait_for_completion(&slow_work_last_thread_exited
);
591 printk(KERN_ERR
"Slow work thread pool: Aborted\n");
593 mutex_unlock(&slow_work_user_lock
);
596 EXPORT_SYMBOL(slow_work_register_user
);
599 * slow_work_unregister_user - Unregister a user of the facility
601 * Unregister a user of the facility, killing all the threads if this was the
604 void slow_work_unregister_user(void)
606 mutex_lock(&slow_work_user_lock
);
608 BUG_ON(slow_work_user_count
<= 0);
610 slow_work_user_count
--;
611 if (slow_work_user_count
== 0) {
612 printk(KERN_NOTICE
"Slow work thread pool: Shutting down\n");
613 slow_work_threads_should_exit
= true;
614 del_timer_sync(&slow_work_cull_timer
);
615 del_timer_sync(&slow_work_oom_timer
);
616 wake_up_all(&slow_work_thread_wq
);
617 wait_for_completion(&slow_work_last_thread_exited
);
618 printk(KERN_NOTICE
"Slow work thread pool:"
619 " Shut down complete\n");
622 mutex_unlock(&slow_work_user_lock
);
624 EXPORT_SYMBOL(slow_work_unregister_user
);
627 * Initialise the slow work facility
629 static int __init
init_slow_work(void)
631 unsigned nr_cpus
= num_possible_cpus();
633 if (slow_work_max_threads
< nr_cpus
)
634 slow_work_max_threads
= nr_cpus
;
636 if (slow_work_max_max_threads
< nr_cpus
* 2)
637 slow_work_max_max_threads
= nr_cpus
* 2;
642 subsys_initcall(init_slow_work
);