projects/linux·blob v3.4-rc4 · refs

573 lines · 14.0 KiB

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 * Pid namespaces:
23 * (C) 2007 Pavel Emelyanov <[email protected]>, OpenVZ, SWsoft Inc.
24 * (C) 2007 Sukadev Bhattiprolu <[email protected]>, IBM
25 * Many thanks to Oleg Nesterov for comments and help
26 *
27 */
28
29#include <linux/mm.h>
30#include <linux/export.h>
31#include <linux/slab.h>
32#include <linux/init.h>
33#include <linux/rculist.h>
34#include <linux/bootmem.h>
35#include <linux/hash.h>
36#include <linux/pid_namespace.h>
37#include <linux/init_task.h>
38#include <linux/syscalls.h>
39
40#define pid_hashfn(nr, ns) \
41 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
42static struct hlist_head *pid_hash;
43static unsigned int pidhash_shift = 4;
44struct pid init_struct_pid = INIT_STRUCT_PID;
45
46int pid_max = PID_MAX_DEFAULT;
47
48#define RESERVED_PIDS 300
49
50int pid_max_min = RESERVED_PIDS + 1;
51int pid_max_max = PID_MAX_LIMIT;
52
53#define BITS_PER_PAGE (PAGE_SIZE*8)
54#define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1)
55
56static inline int mk_pid(struct pid_namespace *pid_ns,
57 struct pidmap *map, int off)
58{
59 return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
60}
61
62#define find_next_offset(map, off) \
63 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
64
65/*
66 * PID-map pages start out as NULL, they get allocated upon
67 * first use and are never deallocated. This way a low pid_max
68 * value does not cause lots of bitmaps to be allocated, but
69 * the scheme scales to up to 4 million PIDs, runtime.
70 */
71struct pid_namespace init_pid_ns = {
72 .kref = {
73 .refcount = ATOMIC_INIT(2),
74 },
75 .pidmap = {
76 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
77 },
78 .last_pid = 0,
79 .level = 0,
80 .child_reaper = &init_task,
81};
82EXPORT_SYMBOL_GPL(init_pid_ns);
83
84int is_container_init(struct task_struct *tsk)
85{
86 int ret = 0;
87 struct pid *pid;
88
89 rcu_read_lock();
90 pid = task_pid(tsk);
91 if (pid != NULL && pid->numbers[pid->level].nr == 1)
92 ret = 1;
93 rcu_read_unlock();
94
95 return ret;
96}
97EXPORT_SYMBOL(is_container_init);
98
99/*
100 * Note: disable interrupts while the pidmap_lock is held as an
101 * interrupt might come in and do read_lock(&tasklist_lock).
102 *
103 * If we don't disable interrupts there is a nasty deadlock between
104 * detach_pid()->free_pid() and another cpu that does
105 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
106 * read_lock(&tasklist_lock);
107 *
108 * After we clean up the tasklist_lock and know there are no
109 * irq handlers that take it we can leave the interrupts enabled.
110 * For now it is easier to be safe than to prove it can't happen.
111 */
112
113static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
114
115static void free_pidmap(struct upid *upid)
116{
117 int nr = upid->nr;
118 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
119 int offset = nr & BITS_PER_PAGE_MASK;
120
121 clear_bit(offset, map->page);
122 atomic_inc(&map->nr_free);
123}
124
125/*
126 * If we started walking pids at 'base', is 'a' seen before 'b'?
127 */
128static int pid_before(int base, int a, int b)
129{
130 /*
131 * This is the same as saying
132 *
133 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
134 * and that mapping orders 'a' and 'b' with respect to 'base'.
135 */
136 return (unsigned)(a - base) < (unsigned)(b - base);
137}
138
139/*
140 * We might be racing with someone else trying to set pid_ns->last_pid
141 * at the pid allocation time (there's also a sysctl for this, but racing
142 * with this one is OK, see comment in kernel/pid_namespace.c about it).
143 * We want the winner to have the "later" value, because if the
144 * "earlier" value prevails, then a pid may get reused immediately.
145 *
146 * Since pids rollover, it is not sufficient to just pick the bigger
147 * value. We have to consider where we started counting from.
148 *
149 * 'base' is the value of pid_ns->last_pid that we observed when
150 * we started looking for a pid.
151 *
152 * 'pid' is the pid that we eventually found.
153 */
154static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
155{
156 int prev;
157 int last_write = base;
158 do {
159 prev = last_write;
160 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
161 } while ((prev != last_write) && (pid_before(base, last_write, pid)));
162}
163
164static int alloc_pidmap(struct pid_namespace *pid_ns)
165{
166 int i, offset, max_scan, pid, last = pid_ns->last_pid;
167 struct pidmap *map;
168
169 pid = last + 1;
170 if (pid >= pid_max)
171 pid = RESERVED_PIDS;
172 offset = pid & BITS_PER_PAGE_MASK;
173 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
174 /*
175 * If last_pid points into the middle of the map->page we
176 * want to scan this bitmap block twice, the second time
177 * we start with offset == 0 (or RESERVED_PIDS).
178 */
179 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
180 for (i = 0; i <= max_scan; ++i) {
181 if (unlikely(!map->page)) {
182 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
183 /*
184 * Free the page if someone raced with us
185 * installing it:
186 */
187 spin_lock_irq(&pidmap_lock);
188 if (!map->page) {
189 map->page = page;
190 page = NULL;
191 }
192 spin_unlock_irq(&pidmap_lock);
193 kfree(page);
194 if (unlikely(!map->page))
195 break;
196 }
197 if (likely(atomic_read(&map->nr_free))) {
198 do {
199 if (!test_and_set_bit(offset, map->page)) {
200 atomic_dec(&map->nr_free);
201 set_last_pid(pid_ns, last, pid);
202 return pid;
203 }
204 offset = find_next_offset(map, offset);
205 pid = mk_pid(pid_ns, map, offset);
206 } while (offset < BITS_PER_PAGE && pid < pid_max);
207 }
208 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
209 ++map;
210 offset = 0;
211 } else {
212 map = &pid_ns->pidmap[0];
213 offset = RESERVED_PIDS;
214 if (unlikely(last == offset))
215 break;
216 }
217 pid = mk_pid(pid_ns, map, offset);
218 }
219 return -1;
220}
221
222int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
223{
224 int offset;
225 struct pidmap *map, *end;
226
227 if (last >= PID_MAX_LIMIT)
228 return -1;
229
230 offset = (last + 1) & BITS_PER_PAGE_MASK;
231 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
232 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
233 for (; map < end; map++, offset = 0) {
234 if (unlikely(!map->page))
235 continue;
236 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
237 if (offset < BITS_PER_PAGE)
238 return mk_pid(pid_ns, map, offset);
239 }
240 return -1;
241}
242
243void put_pid(struct pid *pid)
244{
245 struct pid_namespace *ns;
246
247 if (!pid)
248 return;
249
250 ns = pid->numbers[pid->level].ns;
251 if ((atomic_read(&pid->count) == 1) ||
252 atomic_dec_and_test(&pid->count)) {
253 kmem_cache_free(ns->pid_cachep, pid);
254 put_pid_ns(ns);
255 }
256}
257EXPORT_SYMBOL_GPL(put_pid);
258
259static void delayed_put_pid(struct rcu_head *rhp)
260{
261 struct pid *pid = container_of(rhp, struct pid, rcu);
262 put_pid(pid);
263}
264
265void free_pid(struct pid *pid)
266{
267 /* We can be called with write_lock_irq(&tasklist_lock) held */
268 int i;
269 unsigned long flags;
270
271 spin_lock_irqsave(&pidmap_lock, flags);
272 for (i = 0; i <= pid->level; i++)
273 hlist_del_rcu(&pid->numbers[i].pid_chain);
274 spin_unlock_irqrestore(&pidmap_lock, flags);
275
276 for (i = 0; i <= pid->level; i++)
277 free_pidmap(pid->numbers + i);
278
279 call_rcu(&pid->rcu, delayed_put_pid);
280}
281
282struct pid *alloc_pid(struct pid_namespace *ns)
283{
284 struct pid *pid;
285 enum pid_type type;
286 int i, nr;
287 struct pid_namespace *tmp;
288 struct upid *upid;
289
290 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
291 if (!pid)
292 goto out;
293
294 tmp = ns;
295 for (i = ns->level; i >= 0; i--) {
296 nr = alloc_pidmap(tmp);
297 if (nr < 0)
298 goto out_free;
299
300 pid->numbers[i].nr = nr;
301 pid->numbers[i].ns = tmp;
302 tmp = tmp->parent;
303 }
304
305 get_pid_ns(ns);
306 pid->level = ns->level;
307 atomic_set(&pid->count, 1);
308 for (type = 0; type < PIDTYPE_MAX; ++type)
309 INIT_HLIST_HEAD(&pid->tasks[type]);
310
311 upid = pid->numbers + ns->level;
312 spin_lock_irq(&pidmap_lock);
313 for ( ; upid >= pid->numbers; --upid)
314 hlist_add_head_rcu(&upid->pid_chain,
315 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
316 spin_unlock_irq(&pidmap_lock);
317
318out:
319 return pid;
320
321out_free:
322 while (++i <= ns->level)
323 free_pidmap(pid->numbers + i);
324
325 kmem_cache_free(ns->pid_cachep, pid);
326 pid = NULL;
327 goto out;
328}
329
330struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
331{
332 struct hlist_node *elem;
333 struct upid *pnr;
334
335 hlist_for_each_entry_rcu(pnr, elem,
336 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
337 if (pnr->nr == nr && pnr->ns == ns)
338 return container_of(pnr, struct pid,
339 numbers[ns->level]);
340
341 return NULL;
342}
343EXPORT_SYMBOL_GPL(find_pid_ns);
344
345struct pid *find_vpid(int nr)
346{
347 return find_pid_ns(nr, current->nsproxy->pid_ns);
348}
349EXPORT_SYMBOL_GPL(find_vpid);
350
351/*
352 * attach_pid() must be called with the tasklist_lock write-held.
353 */
354void attach_pid(struct task_struct *task, enum pid_type type,
355 struct pid *pid)
356{
357 struct pid_link *link;
358
359 link = &task->pids[type];
360 link->pid = pid;
361 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
362}
363
364static void __change_pid(struct task_struct *task, enum pid_type type,
365 struct pid *new)
366{
367 struct pid_link *link;
368 struct pid *pid;
369 int tmp;
370
371 link = &task->pids[type];
372 pid = link->pid;
373
374 hlist_del_rcu(&link->node);
375 link->pid = new;
376
377 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
378 if (!hlist_empty(&pid->tasks[tmp]))
379 return;
380
381 free_pid(pid);
382}
383
384void detach_pid(struct task_struct *task, enum pid_type type)
385{
386 __change_pid(task, type, NULL);
387}
388
389void change_pid(struct task_struct *task, enum pid_type type,
390 struct pid *pid)
391{
392 __change_pid(task, type, pid);
393 attach_pid(task, type, pid);
394}
395
396/* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
397void transfer_pid(struct task_struct *old, struct task_struct *new,
398 enum pid_type type)
399{
400 new->pids[type].pid = old->pids[type].pid;
401 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
402}
403
404struct task_struct *pid_task(struct pid *pid, enum pid_type type)
405{
406 struct task_struct *result = NULL;
407 if (pid) {
408 struct hlist_node *first;
409 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
410 lockdep_tasklist_lock_is_held());
411 if (first)
412 result = hlist_entry(first, struct task_struct, pids[(type)].node);
413 }
414 return result;
415}
416EXPORT_SYMBOL(pid_task);
417
418/*
419 * Must be called under rcu_read_lock().
420 */
421struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
422{
423 rcu_lockdep_assert(rcu_read_lock_held(),
424 "find_task_by_pid_ns() needs rcu_read_lock()"
425 " protection");
426 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
427}
428
429struct task_struct *find_task_by_vpid(pid_t vnr)
430{
431 return find_task_by_pid_ns(vnr, current->nsproxy->pid_ns);
432}
433
434struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
435{
436 struct pid *pid;
437 rcu_read_lock();
438 if (type != PIDTYPE_PID)
439 task = task->group_leader;
440 pid = get_pid(task->pids[type].pid);
441 rcu_read_unlock();
442 return pid;
443}
444EXPORT_SYMBOL_GPL(get_task_pid);
445
446struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
447{
448 struct task_struct *result;
449 rcu_read_lock();
450 result = pid_task(pid, type);
451 if (result)
452 get_task_struct(result);
453 rcu_read_unlock();
454 return result;
455}
456EXPORT_SYMBOL_GPL(get_pid_task);
457
458struct pid *find_get_pid(pid_t nr)
459{
460 struct pid *pid;
461
462 rcu_read_lock();
463 pid = get_pid(find_vpid(nr));
464 rcu_read_unlock();
465
466 return pid;
467}
468EXPORT_SYMBOL_GPL(find_get_pid);
469
470pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
471{
472 struct upid *upid;
473 pid_t nr = 0;
474
475 if (pid && ns->level <= pid->level) {
476 upid = &pid->numbers[ns->level];
477 if (upid->ns == ns)
478 nr = upid->nr;
479 }
480 return nr;
481}
482
483pid_t pid_vnr(struct pid *pid)
484{
485 return pid_nr_ns(pid, current->nsproxy->pid_ns);
486}
487EXPORT_SYMBOL_GPL(pid_vnr);
488
489pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
490 struct pid_namespace *ns)
491{
492 pid_t nr = 0;
493
494 rcu_read_lock();
495 if (!ns)
496 ns = current->nsproxy->pid_ns;
497 if (likely(pid_alive(task))) {
498 if (type != PIDTYPE_PID)
499 task = task->group_leader;
500 nr = pid_nr_ns(task->pids[type].pid, ns);
501 }
502 rcu_read_unlock();
503
504 return nr;
505}
506EXPORT_SYMBOL(__task_pid_nr_ns);
507
508pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
509{
510 return pid_nr_ns(task_tgid(tsk), ns);
511}
512EXPORT_SYMBOL(task_tgid_nr_ns);
513
514struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
515{
516 return ns_of_pid(task_pid(tsk));
517}
518EXPORT_SYMBOL_GPL(task_active_pid_ns);
519
520/*
521 * Used by proc to find the first pid that is greater than or equal to nr.
522 *
523 * If there is a pid at nr this function is exactly the same as find_pid_ns.
524 */
525struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
526{
527 struct pid *pid;
528
529 do {
530 pid = find_pid_ns(nr, ns);
531 if (pid)
532 break;
533 nr = next_pidmap(ns, nr);
534 } while (nr > 0);
535
536 return pid;
537}
538
539/*
540 * The pid hash table is scaled according to the amount of memory in the
541 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
542 * more.
543 */
544void __init pidhash_init(void)
545{
546 unsigned int i, pidhash_size;
547
548 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
549 HASH_EARLY | HASH_SMALL,
550 &pidhash_shift, NULL, 4096);
551 pidhash_size = 1U << pidhash_shift;
552
553 for (i = 0; i < pidhash_size; i++)
554 INIT_HLIST_HEAD(&pid_hash[i]);
555}
556
557void __init pidmap_init(void)
558{
559 /* bump default and minimum pid_max based on number of cpus */
560 pid_max = min(pid_max_max, max_t(int, pid_max,
561 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
562 pid_max_min = max_t(int, pid_max_min,
563 PIDS_PER_CPU_MIN * num_possible_cpus());
564 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
565
566 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
567 /* Reserve PID 0. We never call free_pidmap(0) */
568 set_bit(0, init_pid_ns.pidmap[0].page);
569 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
570
571 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
572 SLAB_HWCACHE_ALIGN | SLAB_PANIC);
573}