vserver 1.9.5.x5
[linux-2.6.git] / mm / pdflush.c
1 /*
2  * mm/pdflush.c - worker threads for writing back filesystem data
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * 09Apr2002    akpm@zip.com.au
7  *              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>
20 #include <linux/fs.h>           // Needed by writeback.h
21 #include <linux/writeback.h>    // Prototypes pdflush_operation()
22 #include <linux/kthread.h>
23
24
25 /*
26  * Minimum and maximum number of pdflush instances
27  */
28 #define MIN_PDFLUSH_THREADS     2
29 #define MAX_PDFLUSH_THREADS     8
30
31 static void start_one_pdflush_thread(void);
32
33
34 /*
35  * The pdflush threads are worker threads for writing back dirty data.
36  * Ideally, we'd like one thread per active disk spindle.  But the disk
37  * topology is very hard to divine at this level.   Instead, we take
38  * care in various places to prevent more than one pdflush thread from
39  * performing writeback against a single filesystem.  pdflush threads
40  * have the PF_FLUSHER flag set in current->flags to aid in this.
41  */
42
43 /*
44  * All the pdflush threads.  Protected by pdflush_lock
45  */
46 static LIST_HEAD(pdflush_list);
47 static DEFINE_SPINLOCK(pdflush_lock);
48
49 /*
50  * The count of currently-running pdflush threads.  Protected
51  * by pdflush_lock.
52  *
53  * Readable by sysctl, but not writable.  Published to userspace at
54  * /proc/sys/vm/nr_pdflush_threads.
55  */
56 int nr_pdflush_threads = 0;
57
58 /*
59  * The time at which the pdflush thread pool last went empty
60  */
61 static unsigned long last_empty_jifs;
62
63 /*
64  * The pdflush thread.
65  *
66  * Thread pool management algorithm:
67  * 
68  * - The minimum and maximum number of pdflush instances are bound
69  *   by MIN_PDFLUSH_THREADS and MAX_PDFLUSH_THREADS.
70  * 
71  * - If there have been no idle pdflush instances for 1 second, create
72  *   a new one.
73  * 
74  * - If the least-recently-went-to-sleep pdflush thread has been asleep
75  *   for more than one second, terminate a thread.
76  */
77
78 /*
79  * A structure for passing work to a pdflush thread.  Also for passing
80  * state information between pdflush threads.  Protected by pdflush_lock.
81  */
82 struct pdflush_work {
83         struct task_struct *who;        /* The thread */
84         void (*fn)(unsigned long);      /* A callback function */
85         unsigned long arg0;             /* An argument to the callback */
86         struct list_head list;          /* On pdflush_list, when idle */
87         unsigned long when_i_went_to_sleep;
88 };
89
90 static int __pdflush(struct pdflush_work *my_work)
91 {
92         current->flags |= PF_FLUSHER;
93         my_work->fn = NULL;
94         my_work->who = current;
95         INIT_LIST_HEAD(&my_work->list);
96
97         spin_lock_irq(&pdflush_lock);
98         nr_pdflush_threads++;
99         for ( ; ; ) {
100                 struct pdflush_work *pdf;
101
102                 set_current_state(TASK_INTERRUPTIBLE);
103                 list_move(&my_work->list, &pdflush_list);
104                 my_work->when_i_went_to_sleep = jiffies;
105                 spin_unlock_irq(&pdflush_lock);
106
107                 schedule();
108                 if (try_to_freeze(PF_FREEZE)) {
109                         spin_lock_irq(&pdflush_lock);
110                         continue;
111                 }
112
113                 spin_lock_irq(&pdflush_lock);
114                 if (!list_empty(&my_work->list)) {
115                         printk("pdflush: bogus wakeup!\n");
116                         my_work->fn = NULL;
117                         continue;
118                 }
119                 if (my_work->fn == NULL) {
120                         printk("pdflush: NULL work function\n");
121                         continue;
122                 }
123                 spin_unlock_irq(&pdflush_lock);
124
125                 (*my_work->fn)(my_work->arg0);
126
127                 /*
128                  * Thread creation: For how long have there been zero
129                  * available threads?
130                  */
131                 if (jiffies - last_empty_jifs > 1 * HZ) {
132                         /* unlocked list_empty() test is OK here */
133                         if (list_empty(&pdflush_list)) {
134                                 /* unlocked test is OK here */
135                                 if (nr_pdflush_threads < MAX_PDFLUSH_THREADS)
136                                         start_one_pdflush_thread();
137                         }
138                 }
139
140                 spin_lock_irq(&pdflush_lock);
141                 my_work->fn = NULL;
142
143                 /*
144                  * Thread destruction: For how long has the sleepiest
145                  * thread slept?
146                  */
147                 if (list_empty(&pdflush_list))
148                         continue;
149                 if (nr_pdflush_threads <= MIN_PDFLUSH_THREADS)
150                         continue;
151                 pdf = list_entry(pdflush_list.prev, struct pdflush_work, list);
152                 if (jiffies - pdf->when_i_went_to_sleep > 1 * HZ) {
153                         /* Limit exit rate */
154                         pdf->when_i_went_to_sleep = jiffies;
155                         break;                                  /* exeunt */
156                 }
157         }
158         nr_pdflush_threads--;
159         spin_unlock_irq(&pdflush_lock);
160         return 0;
161 }
162
163 /*
164  * Of course, my_work wants to be just a local in __pdflush().  It is
165  * separated out in this manner to hopefully prevent the compiler from
166  * performing unfortunate optimisations against the auto variables.  Because
167  * these are visible to other tasks and CPUs.  (No problem has actually
168  * been observed.  This is just paranoia).
169  */
170 static int pdflush(void *dummy)
171 {
172         struct pdflush_work my_work;
173
174         /*
175          * pdflush can spend a lot of time doing encryption via dm-crypt.  We
176          * don't want to do that at keventd's priority.
177          */
178         set_user_nice(current, 0);
179         return __pdflush(&my_work);
180 }
181
182 /*
183  * Attempt to wake up a pdflush thread, and get it to do some work for you.
184  * Returns zero if it indeed managed to find a worker thread, and passed your
185  * payload to it.
186  */
187 int pdflush_operation(void (*fn)(unsigned long), unsigned long arg0)
188 {
189         unsigned long flags;
190         int ret = 0;
191
192         if (fn == NULL)
193                 BUG();          /* Hard to diagnose if it's deferred */
194
195         spin_lock_irqsave(&pdflush_lock, flags);
196         if (list_empty(&pdflush_list)) {
197                 spin_unlock_irqrestore(&pdflush_lock, flags);
198                 ret = -1;
199         } else {
200                 struct pdflush_work *pdf;
201
202                 pdf = list_entry(pdflush_list.next, struct pdflush_work, list);
203                 list_del_init(&pdf->list);
204                 if (list_empty(&pdflush_list))
205                         last_empty_jifs = jiffies;
206                 pdf->fn = fn;
207                 pdf->arg0 = arg0;
208                 wake_up_process(pdf->who);
209                 spin_unlock_irqrestore(&pdflush_lock, flags);
210         }
211         return ret;
212 }
213
214 static void start_one_pdflush_thread(void)
215 {
216         kthread_run(pdflush, NULL, "pdflush");
217 }
218
219 static int __init pdflush_init(void)
220 {
221         int i;
222
223         for (i = 0; i < MIN_PDFLUSH_THREADS; i++)
224                 start_one_pdflush_thread();
225         return 0;
226 }
227
228 module_init(pdflush_init);