blob: 235ac440c44e186b916925de3bcaa6f085518963 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * mm/pdflush.c - worker threads for writing back filesystem data
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
Francois Camie1f8e872008-10-15 22:01:59 -07006 * 09Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -07007 * Initial version
8 * 29Feb2004 kaos@sgi.com
9 * Move worker thread creation to kthread to avoid chewing
10 * up stack space with nested calls to kernel_thread.
11 */
12
13#include <linux/sched.h>
14#include <linux/list.h>
15#include <linux/signal.h>
16#include <linux/spinlock.h>
17#include <linux/gfp.h>
18#include <linux/init.h>
19#include <linux/module.h>
Pavel Machekf5264482008-04-21 22:15:06 +000020#include <linux/fs.h> /* Needed by writeback.h */
21#include <linux/writeback.h> /* Prototypes pdflush_operation() */
Linus Torvalds1da177e2005-04-16 15:20:36 -070022#include <linux/kthread.h>
Paul Jackson28a42b92005-10-30 15:02:32 -080023#include <linux/cpuset.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080024#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025
26
27/*
28 * Minimum and maximum number of pdflush instances
29 */
30#define MIN_PDFLUSH_THREADS 2
31#define MAX_PDFLUSH_THREADS 8
32
33static void start_one_pdflush_thread(void);
34
35
36/*
37 * The pdflush threads are worker threads for writing back dirty data.
38 * Ideally, we'd like one thread per active disk spindle. But the disk
39 * topology is very hard to divine at this level. Instead, we take
40 * care in various places to prevent more than one pdflush thread from
41 * performing writeback against a single filesystem. pdflush threads
42 * have the PF_FLUSHER flag set in current->flags to aid in this.
43 */
44
45/*
46 * All the pdflush threads. Protected by pdflush_lock
47 */
48static LIST_HEAD(pdflush_list);
49static DEFINE_SPINLOCK(pdflush_lock);
50
51/*
52 * The count of currently-running pdflush threads. Protected
53 * by pdflush_lock.
54 *
55 * Readable by sysctl, but not writable. Published to userspace at
56 * /proc/sys/vm/nr_pdflush_threads.
57 */
58int nr_pdflush_threads = 0;
59
60/*
61 * The time at which the pdflush thread pool last went empty
62 */
63static unsigned long last_empty_jifs;
64
65/*
66 * The pdflush thread.
67 *
68 * Thread pool management algorithm:
69 *
70 * - The minimum and maximum number of pdflush instances are bound
Jens Axboecd17cbf2009-05-15 11:32:24 +020071 * by MIN_PDFLUSH_THREADS and MAX_PDFLUSH_THREADS.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 *
73 * - If there have been no idle pdflush instances for 1 second, create
74 * a new one.
75 *
76 * - If the least-recently-went-to-sleep pdflush thread has been asleep
77 * for more than one second, terminate a thread.
78 */
79
80/*
81 * A structure for passing work to a pdflush thread. Also for passing
82 * state information between pdflush threads. Protected by pdflush_lock.
83 */
84struct pdflush_work {
85 struct task_struct *who; /* The thread */
86 void (*fn)(unsigned long); /* A callback function */
87 unsigned long arg0; /* An argument to the callback */
88 struct list_head list; /* On pdflush_list, when idle */
89 unsigned long when_i_went_to_sleep;
90};
91
92static int __pdflush(struct pdflush_work *my_work)
93{
Christoph Lameter930d9152006-01-08 01:00:47 -080094 current->flags |= PF_FLUSHER | PF_SWAPWRITE;
Rafael J. Wysocki83144182007-07-17 04:03:35 -070095 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 my_work->fn = NULL;
97 my_work->who = current;
98 INIT_LIST_HEAD(&my_work->list);
99
100 spin_lock_irq(&pdflush_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101 for ( ; ; ) {
102 struct pdflush_work *pdf;
103
104 set_current_state(TASK_INTERRUPTIBLE);
105 list_move(&my_work->list, &pdflush_list);
106 my_work->when_i_went_to_sleep = jiffies;
107 spin_unlock_irq(&pdflush_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 schedule();
Andrew Mortond616e09a2006-06-25 05:47:46 -0700109 try_to_freeze();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110 spin_lock_irq(&pdflush_lock);
111 if (!list_empty(&my_work->list)) {
Andrew Mortond616e09a2006-06-25 05:47:46 -0700112 /*
113 * Someone woke us up, but without removing our control
114 * structure from the global list. swsusp will do this
115 * in try_to_freeze()->refrigerator(). Handle it.
116 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117 my_work->fn = NULL;
118 continue;
119 }
120 if (my_work->fn == NULL) {
Andrew Mortond616e09a2006-06-25 05:47:46 -0700121 printk("pdflush: bogus wakeup\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 continue;
123 }
124 spin_unlock_irq(&pdflush_lock);
125
126 (*my_work->fn)(my_work->arg0);
127
Peter W Morrealea56ed662009-04-06 19:00:28 -0700128 spin_lock_irq(&pdflush_lock);
129
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 /*
131 * Thread creation: For how long have there been zero
132 * available threads?
Peter W Morrealea56ed662009-04-06 19:00:28 -0700133 *
134 * To throttle creation, we reset last_empty_jifs.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 */
OGAWA Hirofumi2b4bc462008-07-25 01:45:42 -0700136 if (time_after(jiffies, last_empty_jifs + 1 * HZ)) {
Jens Axboecd17cbf2009-05-15 11:32:24 +0200137 if (list_empty(&pdflush_list)) {
138 if (nr_pdflush_threads < MAX_PDFLUSH_THREADS) {
139 last_empty_jifs = jiffies;
140 nr_pdflush_threads++;
141 spin_unlock_irq(&pdflush_lock);
142 start_one_pdflush_thread();
143 spin_lock_irq(&pdflush_lock);
144 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 }
146 }
147
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 my_work->fn = NULL;
149
150 /*
151 * Thread destruction: For how long has the sleepiest
152 * thread slept?
153 */
154 if (list_empty(&pdflush_list))
155 continue;
Jens Axboecd17cbf2009-05-15 11:32:24 +0200156 if (nr_pdflush_threads <= MIN_PDFLUSH_THREADS)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157 continue;
158 pdf = list_entry(pdflush_list.prev, struct pdflush_work, list);
OGAWA Hirofumi2b4bc462008-07-25 01:45:42 -0700159 if (time_after(jiffies, pdf->when_i_went_to_sleep + 1 * HZ)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160 /* Limit exit rate */
161 pdf->when_i_went_to_sleep = jiffies;
162 break; /* exeunt */
163 }
164 }
165 nr_pdflush_threads--;
166 spin_unlock_irq(&pdflush_lock);
167 return 0;
168}
169
170/*
171 * Of course, my_work wants to be just a local in __pdflush(). It is
172 * separated out in this manner to hopefully prevent the compiler from
173 * performing unfortunate optimisations against the auto variables. Because
174 * these are visible to other tasks and CPUs. (No problem has actually
175 * been observed. This is just paranoia).
176 */
177static int pdflush(void *dummy)
178{
179 struct pdflush_work my_work;
Rusty Russell174596a2009-01-01 10:12:29 +1030180 cpumask_var_t cpus_allowed;
181
182 /*
183 * Since the caller doesn't even check kthread_run() worked, let's not
184 * freak out too much if this fails.
185 */
186 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
187 printk(KERN_WARNING "pdflush failed to allocate cpumask\n");
188 return 0;
189 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190
191 /*
192 * pdflush can spend a lot of time doing encryption via dm-crypt. We
193 * don't want to do that at keventd's priority.
194 */
195 set_user_nice(current, 0);
Paul Jackson28a42b92005-10-30 15:02:32 -0800196
197 /*
198 * Some configs put our parent kthread in a limited cpuset,
Rusty Russell1a2142a2009-03-30 22:05:10 -0600199 * which kthread() overrides, forcing cpus_allowed == cpu_all_mask.
Paul Jackson28a42b92005-10-30 15:02:32 -0800200 * Our needs are more modest - cut back to our cpusets cpus_allowed.
201 * This is needed as pdflush's are dynamically created and destroyed.
202 * The boottime pdflush's are easily placed w/o these 2 lines.
203 */
Rusty Russell174596a2009-01-01 10:12:29 +1030204 cpuset_cpus_allowed(current, cpus_allowed);
205 set_cpus_allowed_ptr(current, cpus_allowed);
206 free_cpumask_var(cpus_allowed);
Paul Jackson28a42b92005-10-30 15:02:32 -0800207
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208 return __pdflush(&my_work);
209}
210
211/*
212 * Attempt to wake up a pdflush thread, and get it to do some work for you.
213 * Returns zero if it indeed managed to find a worker thread, and passed your
214 * payload to it.
215 */
216int pdflush_operation(void (*fn)(unsigned long), unsigned long arg0)
217{
218 unsigned long flags;
219 int ret = 0;
220
Eric Sesterhenn125e1872006-06-23 02:06:06 -0700221 BUG_ON(fn == NULL); /* Hard to diagnose if it's deferred */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222
223 spin_lock_irqsave(&pdflush_lock, flags);
224 if (list_empty(&pdflush_list)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 ret = -1;
226 } else {
227 struct pdflush_work *pdf;
228
229 pdf = list_entry(pdflush_list.next, struct pdflush_work, list);
230 list_del_init(&pdf->list);
231 if (list_empty(&pdflush_list))
232 last_empty_jifs = jiffies;
233 pdf->fn = fn;
234 pdf->arg0 = arg0;
235 wake_up_process(pdf->who);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236 }
Denis Cheng5aecd552008-05-12 14:02:14 -0700237 spin_unlock_irqrestore(&pdflush_lock, flags);
238
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239 return ret;
240}
241
242static void start_one_pdflush_thread(void)
243{
Peter W Morrealea56ed662009-04-06 19:00:28 -0700244 struct task_struct *k;
245
246 k = kthread_run(pdflush, NULL, "pdflush");
247 if (unlikely(IS_ERR(k))) {
248 spin_lock_irq(&pdflush_lock);
249 nr_pdflush_threads--;
250 spin_unlock_irq(&pdflush_lock);
251 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252}
253
254static int __init pdflush_init(void)
255{
256 int i;
257
Peter W Morrealea56ed662009-04-06 19:00:28 -0700258 /*
259 * Pre-set nr_pdflush_threads... If we fail to create,
260 * the count will be decremented.
261 */
Jens Axboecd17cbf2009-05-15 11:32:24 +0200262 nr_pdflush_threads = MIN_PDFLUSH_THREADS;
Peter W Morrealea56ed662009-04-06 19:00:28 -0700263
Jens Axboecd17cbf2009-05-15 11:32:24 +0200264 for (i = 0; i < MIN_PDFLUSH_THREADS; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 start_one_pdflush_thread();
266 return 0;
267}
268
269module_init(pdflush_init);