2 * Generic pidhash and scalable, time-bounded PID allocator
4 * (C) 2002 William Irwin, IBM
5 * (C) 2002 Ingo Molnar, Red Hat
7 * pid-structures are backing objects for tasks sharing a given ID to chain
8 * against. There is very little to them aside from hashing them and
9 * parking tasks using given ID's on a list.
11 * The hash is always changed with the tasklist_lock write-acquired,
12 * and the hash is only accessed with the tasklist_lock at least
13 * read-acquired, so there's no additional SMP locking needed here.
15 * We have a list of bitmap pages, which bitmaps represent the PID space.
16 * Allocating and freeing PIDs is completely lockless. The worst-case
17 * allocation scenario when all but one out of 1 million PIDs possible are
18 * allocated already: the scanning of 32 list entries and at most PAGE_SIZE
19 * bytes. The typical fastpath is a single successful setbit. Freeing is O(1).
23 #include <linux/module.h>
24 #include <linux/slab.h>
25 #include <linux/init.h>
26 #include <linux/bootmem.h>
27 #include <linux/hash.h>
29 #define pid_hashfn(nr) hash_long((unsigned long)nr, pidhash_shift)
30 static struct list_head *pid_hash[PIDTYPE_MAX];
31 static int pidhash_shift;
33 int pid_max = PID_MAX_DEFAULT;
36 #define RESERVED_PIDS 300
38 #define PIDMAP_ENTRIES (PID_MAX_LIMIT/PAGE_SIZE/8)
39 #define BITS_PER_PAGE (PAGE_SIZE*8)
40 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1)
43 * PID-map pages start out as NULL, they get allocated upon
44 * first use and are never deallocated. This way a low pid_max
45 * value does not cause lots of bitmaps to be allocated, but
46 * the scheme scales to up to 4 million PIDs, runtime.
48 typedef struct pidmap {
53 static pidmap_t pidmap_array[PIDMAP_ENTRIES] =
54 { [ 0 ... PIDMAP_ENTRIES-1 ] = { ATOMIC_INIT(BITS_PER_PAGE), NULL } };
56 static pidmap_t *map_limit = pidmap_array + PIDMAP_ENTRIES;
58 static spinlock_t pidmap_lock __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
60 fastcall void free_pidmap(int pid)
62 pidmap_t *map = pidmap_array + pid / BITS_PER_PAGE;
63 int offset = pid & BITS_PER_PAGE_MASK;
65 clear_bit(offset, map->page);
66 atomic_inc(&map->nr_free);
70 * Here we search for the next map that has free bits left.
71 * Normally the next map has free PIDs.
73 static inline pidmap_t *next_free_map(pidmap_t *map, int *max_steps)
75 while (--*max_steps) {
76 if (++map == map_limit)
78 if (unlikely(!map->page)) {
79 unsigned long page = get_zeroed_page(GFP_KERNEL);
81 * Free the page if someone raced with us
84 spin_lock(&pidmap_lock);
88 map->page = (void *)page;
89 spin_unlock(&pidmap_lock);
94 if (atomic_read(&map->nr_free))
100 int alloc_pidmap(void)
102 int pid, offset, max_steps = PIDMAP_ENTRIES + 1;
109 offset = pid & BITS_PER_PAGE_MASK;
110 map = pidmap_array + pid / BITS_PER_PAGE;
112 if (likely(map->page && !test_and_set_bit(offset, map->page))) {
114 * There is a small window for last_pid updates to race,
115 * but in that case the next allocation will go into the
116 * slowpath and that fixes things up.
119 atomic_dec(&map->nr_free);
124 if (!offset || !atomic_read(&map->nr_free)) {
126 map = next_free_map(map, &max_steps);
132 * Find the next zero bit:
135 offset = find_next_zero_bit(map->page, BITS_PER_PAGE, offset);
136 if (offset >= BITS_PER_PAGE)
138 if (test_and_set_bit(offset, map->page))
141 /* we got the PID: */
142 pid = (map - pidmap_array) * BITS_PER_PAGE + offset;
149 fastcall struct pid *find_pid(enum pid_type type, int nr)
151 struct list_head *elem, *bucket = &pid_hash[type][pid_hashfn(nr)];
154 __list_for_each(elem, bucket) {
155 pid = list_entry(elem, struct pid, hash_chain);
162 void fastcall link_pid(task_t *task, struct pid_link *link, struct pid *pid)
164 atomic_inc(&pid->count);
165 list_add_tail(&link->pid_chain, &pid->task_list);
169 int fastcall attach_pid(task_t *task, enum pid_type type, int nr)
171 struct pid *pid = find_pid(type, nr);
174 atomic_inc(&pid->count);
176 pid = &task->pids[type].pid;
178 atomic_set(&pid->count, 1);
179 INIT_LIST_HEAD(&pid->task_list);
181 get_task_struct(task);
182 list_add(&pid->hash_chain, &pid_hash[type][pid_hashfn(nr)]);
184 list_add_tail(&task->pids[type].pid_chain, &pid->task_list);
185 task->pids[type].pidptr = pid;
190 static inline int __detach_pid(task_t *task, enum pid_type type)
192 struct pid_link *link = task->pids + type;
193 struct pid *pid = link->pidptr;
196 list_del(&link->pid_chain);
197 if (!atomic_dec_and_test(&pid->count))
201 list_del(&pid->hash_chain);
202 put_task_struct(pid->task);
207 static void _detach_pid(task_t *task, enum pid_type type)
209 __detach_pid(task, type);
212 void fastcall detach_pid(task_t *task, enum pid_type type)
214 int nr = __detach_pid(task, type);
219 for (type = 0; type < PIDTYPE_MAX; ++type)
220 if (find_pid(type, nr))
225 task_t *find_task_by_pid(int nr)
227 struct pid *pid = find_pid(PIDTYPE_PID,
228 vx_rmap_tgid(current->vx_info, nr));
232 return pid_task(pid->task_list.next, PIDTYPE_PID);
235 EXPORT_SYMBOL(find_task_by_pid);
238 * This function switches the PIDs if a non-leader thread calls
239 * sys_execve() - this must be done without releasing the PID.
240 * (which a detach_pid() would eventually do.)
242 void switch_exec_pids(task_t *leader, task_t *thread)
244 _detach_pid(leader, PIDTYPE_PID);
245 _detach_pid(leader, PIDTYPE_TGID);
246 _detach_pid(leader, PIDTYPE_PGID);
247 _detach_pid(leader, PIDTYPE_SID);
249 _detach_pid(thread, PIDTYPE_PID);
250 _detach_pid(thread, PIDTYPE_TGID);
252 leader->pid = leader->tgid = thread->pid;
253 thread->pid = thread->tgid;
255 attach_pid(thread, PIDTYPE_PID, thread->pid);
256 attach_pid(thread, PIDTYPE_TGID, thread->tgid);
257 attach_pid(thread, PIDTYPE_PGID, thread->signal->pgrp);
258 attach_pid(thread, PIDTYPE_SID, thread->signal->session);
259 list_add_tail(&thread->tasks, &init_task.tasks);
261 attach_pid(leader, PIDTYPE_PID, leader->pid);
262 attach_pid(leader, PIDTYPE_TGID, leader->tgid);
263 attach_pid(leader, PIDTYPE_PGID, leader->signal->pgrp);
264 attach_pid(leader, PIDTYPE_SID, leader->signal->session);
268 * The pid hash table is scaled according to the amount of memory in the
269 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
272 void __init pidhash_init(void)
274 int i, j, pidhash_size;
275 unsigned long megabytes = max_pfn >> (20 - PAGE_SHIFT);
277 pidhash_shift = max(4, fls(megabytes * 4));
278 pidhash_shift = min(12, pidhash_shift);
279 pidhash_size = 1 << pidhash_shift;
281 printk("PID hash table entries: %d (order %d: %Zd bytes)\n",
282 pidhash_size, pidhash_shift,
283 pidhash_size * sizeof(struct list_head));
285 for (i = 0; i < PIDTYPE_MAX; i++) {
286 pid_hash[i] = alloc_bootmem(pidhash_size *
287 sizeof(struct list_head));
289 panic("Could not alloc pidhash!\n");
290 for (j = 0; j < pidhash_size; j++)
291 INIT_LIST_HEAD(&pid_hash[i][j]);
295 void __init pidmap_init(void)
299 pidmap_array->page = (void *)get_zeroed_page(GFP_KERNEL);
300 set_bit(0, pidmap_array->page);
301 atomic_dec(&pidmap_array->nr_free);
304 * Allocate PID 0, and hash it via all PID types:
307 for (i = 0; i < PIDTYPE_MAX; i++)
308 attach_pid(current, i, 0);