[PATCH] sched: less aggressive idle balancing
[linux-2.6] / kernel / pid.c
1 /*
2  * Generic pidhash and scalable, time-bounded PID allocator
3  *
4  * (C) 2002-2003 William Irwin, IBM
5  * (C) 2004 William Irwin, Oracle
6  * (C) 2002-2004 Ingo Molnar, Red Hat
7  *
8  * pid-structures are backing objects for tasks sharing a given ID to chain
9  * against. There is very little to them aside from hashing them and
10  * parking tasks using given ID's on a list.
11  *
12  * The hash is always changed with the tasklist_lock write-acquired,
13  * and the hash is only accessed with the tasklist_lock at least
14  * read-acquired, so there's no additional SMP locking needed here.
15  *
16  * We have a list of bitmap pages, which bitmaps represent the PID space.
17  * Allocating and freeing PIDs is completely lockless. The worst-case
18  * allocation scenario when all but one out of 1 million PIDs possible are
19  * allocated already: the scanning of 32 list entries and at most PAGE_SIZE
20  * bytes. The typical fastpath is a single successful setbit. Freeing is O(1).
21  */
22
23 #include <linux/mm.h>
24 #include <linux/module.h>
25 #include <linux/slab.h>
26 #include <linux/init.h>
27 #include <linux/bootmem.h>
28 #include <linux/hash.h>
29
30 #define pid_hashfn(nr) hash_long((unsigned long)nr, pidhash_shift)
31 static struct hlist_head *pid_hash[PIDTYPE_MAX];
32 static int pidhash_shift;
33
34 int pid_max = PID_MAX_DEFAULT;
35 int last_pid;
36
37 #define RESERVED_PIDS           300
38
39 int pid_max_min = RESERVED_PIDS + 1;
40 int pid_max_max = PID_MAX_LIMIT;
41
42 #define PIDMAP_ENTRIES          ((PID_MAX_LIMIT + 8*PAGE_SIZE - 1)/PAGE_SIZE/8)
43 #define BITS_PER_PAGE           (PAGE_SIZE*8)
44 #define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
45 #define mk_pid(map, off)        (((map) - pidmap_array)*BITS_PER_PAGE + (off))
46 #define find_next_offset(map, off)                                      \
47                 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
48
49 /*
50  * PID-map pages start out as NULL, they get allocated upon
51  * first use and are never deallocated. This way a low pid_max
52  * value does not cause lots of bitmaps to be allocated, but
53  * the scheme scales to up to 4 million PIDs, runtime.
54  */
55 typedef struct pidmap {
56         atomic_t nr_free;
57         void *page;
58 } pidmap_t;
59
60 static pidmap_t pidmap_array[PIDMAP_ENTRIES] =
61          { [ 0 ... PIDMAP_ENTRIES-1 ] = { ATOMIC_INIT(BITS_PER_PAGE), NULL } };
62
63 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
64
65 fastcall void free_pidmap(int pid)
66 {
67         pidmap_t *map = pidmap_array + pid / BITS_PER_PAGE;
68         int offset = pid & BITS_PER_PAGE_MASK;
69
70         clear_bit(offset, map->page);
71         atomic_inc(&map->nr_free);
72 }
73
74 int alloc_pidmap(void)
75 {
76         int i, offset, max_scan, pid, last = last_pid;
77         pidmap_t *map;
78
79         pid = last + 1;
80         if (pid >= pid_max)
81                 pid = RESERVED_PIDS;
82         offset = pid & BITS_PER_PAGE_MASK;
83         map = &pidmap_array[pid/BITS_PER_PAGE];
84         max_scan = (pid_max + BITS_PER_PAGE - 1)/BITS_PER_PAGE - !offset;
85         for (i = 0; i <= max_scan; ++i) {
86                 if (unlikely(!map->page)) {
87                         unsigned long page = get_zeroed_page(GFP_KERNEL);
88                         /*
89                          * Free the page if someone raced with us
90                          * installing it:
91                          */
92                         spin_lock(&pidmap_lock);
93                         if (map->page)
94                                 free_page(page);
95                         else
96                                 map->page = (void *)page;
97                         spin_unlock(&pidmap_lock);
98                         if (unlikely(!map->page))
99                                 break;
100                 }
101                 if (likely(atomic_read(&map->nr_free))) {
102                         do {
103                                 if (!test_and_set_bit(offset, map->page)) {
104                                         atomic_dec(&map->nr_free);
105                                         last_pid = pid;
106                                         return pid;
107                                 }
108                                 offset = find_next_offset(map, offset);
109                                 pid = mk_pid(map, offset);
110                         /*
111                          * find_next_offset() found a bit, the pid from it
112                          * is in-bounds, and if we fell back to the last
113                          * bitmap block and the final block was the same
114                          * as the starting point, pid is before last_pid.
115                          */
116                         } while (offset < BITS_PER_PAGE && pid < pid_max &&
117                                         (i != max_scan || pid < last ||
118                                             !((last+1) & BITS_PER_PAGE_MASK)));
119                 }
120                 if (map < &pidmap_array[(pid_max-1)/BITS_PER_PAGE]) {
121                         ++map;
122                         offset = 0;
123                 } else {
124                         map = &pidmap_array[0];
125                         offset = RESERVED_PIDS;
126                         if (unlikely(last == offset))
127                                 break;
128                 }
129                 pid = mk_pid(map, offset);
130         }
131         return -1;
132 }
133
134 struct pid * fastcall find_pid(enum pid_type type, int nr)
135 {
136         struct hlist_node *elem;
137         struct pid *pid;
138
139         hlist_for_each_entry(pid, elem,
140                         &pid_hash[type][pid_hashfn(nr)], pid_chain) {
141                 if (pid->nr == nr)
142                         return pid;
143         }
144         return NULL;
145 }
146
147 int fastcall attach_pid(task_t *task, enum pid_type type, int nr)
148 {
149         struct pid *pid, *task_pid;
150
151         task_pid = &task->pids[type];
152         pid = find_pid(type, nr);
153         if (pid == NULL) {
154                 hlist_add_head(&task_pid->pid_chain,
155                                 &pid_hash[type][pid_hashfn(nr)]);
156                 INIT_LIST_HEAD(&task_pid->pid_list);
157         } else {
158                 INIT_HLIST_NODE(&task_pid->pid_chain);
159                 list_add_tail(&task_pid->pid_list, &pid->pid_list);
160         }
161         task_pid->nr = nr;
162
163         return 0;
164 }
165
166 static fastcall int __detach_pid(task_t *task, enum pid_type type)
167 {
168         struct pid *pid, *pid_next;
169         int nr = 0;
170
171         pid = &task->pids[type];
172         if (!hlist_unhashed(&pid->pid_chain)) {
173                 hlist_del(&pid->pid_chain);
174
175                 if (list_empty(&pid->pid_list))
176                         nr = pid->nr;
177                 else {
178                         pid_next = list_entry(pid->pid_list.next,
179                                                 struct pid, pid_list);
180                         /* insert next pid from pid_list to hash */
181                         hlist_add_head(&pid_next->pid_chain,
182                                 &pid_hash[type][pid_hashfn(pid_next->nr)]);
183                 }
184         }
185
186         list_del(&pid->pid_list);
187         pid->nr = 0;
188
189         return nr;
190 }
191
192 void fastcall detach_pid(task_t *task, enum pid_type type)
193 {
194         int tmp, nr;
195
196         nr = __detach_pid(task, type);
197         if (!nr)
198                 return;
199
200         for (tmp = PIDTYPE_MAX; --tmp >= 0; )
201                 if (tmp != type && find_pid(tmp, nr))
202                         return;
203
204         free_pidmap(nr);
205 }
206
207 task_t *find_task_by_pid_type(int type, int nr)
208 {
209         struct pid *pid;
210
211         pid = find_pid(type, nr);
212         if (!pid)
213                 return NULL;
214
215         return pid_task(&pid->pid_list, type);
216 }
217
218 EXPORT_SYMBOL(find_task_by_pid_type);
219
220 /*
221  * This function switches the PIDs if a non-leader thread calls
222  * sys_execve() - this must be done without releasing the PID.
223  * (which a detach_pid() would eventually do.)
224  */
225 void switch_exec_pids(task_t *leader, task_t *thread)
226 {
227         __detach_pid(leader, PIDTYPE_PID);
228         __detach_pid(leader, PIDTYPE_TGID);
229         __detach_pid(leader, PIDTYPE_PGID);
230         __detach_pid(leader, PIDTYPE_SID);
231
232         __detach_pid(thread, PIDTYPE_PID);
233         __detach_pid(thread, PIDTYPE_TGID);
234
235         leader->pid = leader->tgid = thread->pid;
236         thread->pid = thread->tgid;
237
238         attach_pid(thread, PIDTYPE_PID, thread->pid);
239         attach_pid(thread, PIDTYPE_TGID, thread->tgid);
240         attach_pid(thread, PIDTYPE_PGID, thread->signal->pgrp);
241         attach_pid(thread, PIDTYPE_SID, thread->signal->session);
242         list_add_tail(&thread->tasks, &init_task.tasks);
243
244         attach_pid(leader, PIDTYPE_PID, leader->pid);
245         attach_pid(leader, PIDTYPE_TGID, leader->tgid);
246         attach_pid(leader, PIDTYPE_PGID, leader->signal->pgrp);
247         attach_pid(leader, PIDTYPE_SID, leader->signal->session);
248 }
249
250 /*
251  * The pid hash table is scaled according to the amount of memory in the
252  * machine.  From a minimum of 16 slots up to 4096 slots at one gigabyte or
253  * more.
254  */
255 void __init pidhash_init(void)
256 {
257         int i, j, pidhash_size;
258         unsigned long megabytes = nr_kernel_pages >> (20 - PAGE_SHIFT);
259
260         pidhash_shift = max(4, fls(megabytes * 4));
261         pidhash_shift = min(12, pidhash_shift);
262         pidhash_size = 1 << pidhash_shift;
263
264         printk("PID hash table entries: %d (order: %d, %Zd bytes)\n",
265                 pidhash_size, pidhash_shift,
266                 PIDTYPE_MAX * pidhash_size * sizeof(struct hlist_head));
267
268         for (i = 0; i < PIDTYPE_MAX; i++) {
269                 pid_hash[i] = alloc_bootmem(pidhash_size *
270                                         sizeof(*(pid_hash[i])));
271                 if (!pid_hash[i])
272                         panic("Could not alloc pidhash!\n");
273                 for (j = 0; j < pidhash_size; j++)
274                         INIT_HLIST_HEAD(&pid_hash[i][j]);
275         }
276 }
277
278 void __init pidmap_init(void)
279 {
280         int i;
281
282         pidmap_array->page = (void *)get_zeroed_page(GFP_KERNEL);
283         set_bit(0, pidmap_array->page);
284         atomic_dec(&pidmap_array->nr_free);
285
286         /*
287          * Allocate PID 0, and hash it via all PID types:
288          */
289
290         for (i = 0; i < PIDTYPE_MAX; i++)
291                 attach_pid(current, i, 0);
292 }