]>
Commit | Line | Data |
---|---|---|
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 <xemul@openvz.org>, OpenVZ, SWsoft Inc. | |
24 | * (C) 2007 Sukadev Bhattiprolu <sukadev@us.ibm.com>, IBM | |
25 | * Many thanks to Oleg Nesterov for comments and help | |
26 | * | |
27 | */ | |
28 | ||
29 | #include <linux/mm.h> | |
30 | #include <linux/module.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) | |
42 | static struct hlist_head *pid_hash; | |
43 | static unsigned int pidhash_shift = 4; | |
44 | struct pid init_struct_pid = INIT_STRUCT_PID; | |
45 | ||
46 | int pid_max = PID_MAX_DEFAULT; | |
47 | ||
48 | #define RESERVED_PIDS 300 | |
49 | ||
50 | int pid_max_min = RESERVED_PIDS + 1; | |
51 | int 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 | ||
56 | static 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 | */ | |
71 | struct 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 | }; | |
82 | EXPORT_SYMBOL_GPL(init_pid_ns); | |
83 | ||
84 | int 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 | } | |
97 | EXPORT_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 | ||
113 | static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock); | |
114 | ||
115 | static 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 | */ | |
128 | static 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 | * We want the winner to have the "later" value, because if the | |
142 | * "earlier" value prevails, then a pid may get reused immediately. | |
143 | * | |
144 | * Since pids rollover, it is not sufficient to just pick the bigger | |
145 | * value. We have to consider where we started counting from. | |
146 | * | |
147 | * 'base' is the value of pid_ns->last_pid that we observed when | |
148 | * we started looking for a pid. | |
149 | * | |
150 | * 'pid' is the pid that we eventually found. | |
151 | */ | |
152 | static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid) | |
153 | { | |
154 | int prev; | |
155 | int last_write = base; | |
156 | do { | |
157 | prev = last_write; | |
158 | last_write = cmpxchg(&pid_ns->last_pid, prev, pid); | |
159 | } while ((prev != last_write) && (pid_before(base, last_write, pid))); | |
160 | } | |
161 | ||
162 | static int alloc_pidmap(struct pid_namespace *pid_ns) | |
163 | { | |
164 | int i, offset, max_scan, pid, last = pid_ns->last_pid; | |
165 | struct pidmap *map; | |
166 | ||
167 | pid = last + 1; | |
168 | if (pid >= pid_max) | |
169 | pid = RESERVED_PIDS; | |
170 | offset = pid & BITS_PER_PAGE_MASK; | |
171 | map = &pid_ns->pidmap[pid/BITS_PER_PAGE]; | |
172 | /* | |
173 | * If last_pid points into the middle of the map->page we | |
174 | * want to scan this bitmap block twice, the second time | |
175 | * we start with offset == 0 (or RESERVED_PIDS). | |
176 | */ | |
177 | max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset; | |
178 | for (i = 0; i <= max_scan; ++i) { | |
179 | if (unlikely(!map->page)) { | |
180 | void *page = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
181 | /* | |
182 | * Free the page if someone raced with us | |
183 | * installing it: | |
184 | */ | |
185 | spin_lock_irq(&pidmap_lock); | |
186 | if (!map->page) { | |
187 | map->page = page; | |
188 | page = NULL; | |
189 | } | |
190 | spin_unlock_irq(&pidmap_lock); | |
191 | kfree(page); | |
192 | if (unlikely(!map->page)) | |
193 | break; | |
194 | } | |
195 | if (likely(atomic_read(&map->nr_free))) { | |
196 | do { | |
197 | if (!test_and_set_bit(offset, map->page)) { | |
198 | atomic_dec(&map->nr_free); | |
199 | set_last_pid(pid_ns, last, pid); | |
200 | return pid; | |
201 | } | |
202 | offset = find_next_offset(map, offset); | |
203 | pid = mk_pid(pid_ns, map, offset); | |
204 | } while (offset < BITS_PER_PAGE && pid < pid_max); | |
205 | } | |
206 | if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) { | |
207 | ++map; | |
208 | offset = 0; | |
209 | } else { | |
210 | map = &pid_ns->pidmap[0]; | |
211 | offset = RESERVED_PIDS; | |
212 | if (unlikely(last == offset)) | |
213 | break; | |
214 | } | |
215 | pid = mk_pid(pid_ns, map, offset); | |
216 | } | |
217 | return -1; | |
218 | } | |
219 | ||
220 | int next_pidmap(struct pid_namespace *pid_ns, int last) | |
221 | { | |
222 | int offset; | |
223 | struct pidmap *map, *end; | |
224 | ||
225 | offset = (last + 1) & BITS_PER_PAGE_MASK; | |
226 | map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE]; | |
227 | end = &pid_ns->pidmap[PIDMAP_ENTRIES]; | |
228 | for (; map < end; map++, offset = 0) { | |
229 | if (unlikely(!map->page)) | |
230 | continue; | |
231 | offset = find_next_bit((map)->page, BITS_PER_PAGE, offset); | |
232 | if (offset < BITS_PER_PAGE) | |
233 | return mk_pid(pid_ns, map, offset); | |
234 | } | |
235 | return -1; | |
236 | } | |
237 | ||
238 | void put_pid(struct pid *pid) | |
239 | { | |
240 | struct pid_namespace *ns; | |
241 | ||
242 | if (!pid) | |
243 | return; | |
244 | ||
245 | ns = pid->numbers[pid->level].ns; | |
246 | if ((atomic_read(&pid->count) == 1) || | |
247 | atomic_dec_and_test(&pid->count)) { | |
248 | kmem_cache_free(ns->pid_cachep, pid); | |
249 | put_pid_ns(ns); | |
250 | } | |
251 | } | |
252 | EXPORT_SYMBOL_GPL(put_pid); | |
253 | ||
254 | static void delayed_put_pid(struct rcu_head *rhp) | |
255 | { | |
256 | struct pid *pid = container_of(rhp, struct pid, rcu); | |
257 | put_pid(pid); | |
258 | } | |
259 | ||
260 | void free_pid(struct pid *pid) | |
261 | { | |
262 | /* We can be called with write_lock_irq(&tasklist_lock) held */ | |
263 | int i; | |
264 | unsigned long flags; | |
265 | ||
266 | spin_lock_irqsave(&pidmap_lock, flags); | |
267 | for (i = 0; i <= pid->level; i++) | |
268 | hlist_del_rcu(&pid->numbers[i].pid_chain); | |
269 | spin_unlock_irqrestore(&pidmap_lock, flags); | |
270 | ||
271 | for (i = 0; i <= pid->level; i++) | |
272 | free_pidmap(pid->numbers + i); | |
273 | ||
274 | call_rcu(&pid->rcu, delayed_put_pid); | |
275 | } | |
276 | ||
277 | struct pid *alloc_pid(struct pid_namespace *ns) | |
278 | { | |
279 | struct pid *pid; | |
280 | enum pid_type type; | |
281 | int i, nr; | |
282 | struct pid_namespace *tmp; | |
283 | struct upid *upid; | |
284 | ||
285 | pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL); | |
286 | if (!pid) | |
287 | goto out; | |
288 | ||
289 | tmp = ns; | |
290 | for (i = ns->level; i >= 0; i--) { | |
291 | nr = alloc_pidmap(tmp); | |
292 | if (nr < 0) | |
293 | goto out_free; | |
294 | ||
295 | pid->numbers[i].nr = nr; | |
296 | pid->numbers[i].ns = tmp; | |
297 | tmp = tmp->parent; | |
298 | } | |
299 | ||
300 | get_pid_ns(ns); | |
301 | pid->level = ns->level; | |
302 | atomic_set(&pid->count, 1); | |
303 | for (type = 0; type < PIDTYPE_MAX; ++type) | |
304 | INIT_HLIST_HEAD(&pid->tasks[type]); | |
305 | ||
306 | upid = pid->numbers + ns->level; | |
307 | spin_lock_irq(&pidmap_lock); | |
308 | for ( ; upid >= pid->numbers; --upid) | |
309 | hlist_add_head_rcu(&upid->pid_chain, | |
310 | &pid_hash[pid_hashfn(upid->nr, upid->ns)]); | |
311 | spin_unlock_irq(&pidmap_lock); | |
312 | ||
313 | out: | |
314 | return pid; | |
315 | ||
316 | out_free: | |
317 | while (++i <= ns->level) | |
318 | free_pidmap(pid->numbers + i); | |
319 | ||
320 | kmem_cache_free(ns->pid_cachep, pid); | |
321 | pid = NULL; | |
322 | goto out; | |
323 | } | |
324 | ||
325 | struct pid *find_pid_ns(int nr, struct pid_namespace *ns) | |
326 | { | |
327 | struct hlist_node *elem; | |
328 | struct upid *pnr; | |
329 | ||
330 | hlist_for_each_entry_rcu(pnr, elem, | |
331 | &pid_hash[pid_hashfn(nr, ns)], pid_chain) | |
332 | if (pnr->nr == nr && pnr->ns == ns) | |
333 | return container_of(pnr, struct pid, | |
334 | numbers[ns->level]); | |
335 | ||
336 | return NULL; | |
337 | } | |
338 | EXPORT_SYMBOL_GPL(find_pid_ns); | |
339 | ||
340 | struct pid *find_vpid(int nr) | |
341 | { | |
342 | return find_pid_ns(nr, current->nsproxy->pid_ns); | |
343 | } | |
344 | EXPORT_SYMBOL_GPL(find_vpid); | |
345 | ||
346 | /* | |
347 | * attach_pid() must be called with the tasklist_lock write-held. | |
348 | */ | |
349 | void attach_pid(struct task_struct *task, enum pid_type type, | |
350 | struct pid *pid) | |
351 | { | |
352 | struct pid_link *link; | |
353 | ||
354 | link = &task->pids[type]; | |
355 | link->pid = pid; | |
356 | hlist_add_head_rcu(&link->node, &pid->tasks[type]); | |
357 | } | |
358 | ||
359 | static void __change_pid(struct task_struct *task, enum pid_type type, | |
360 | struct pid *new) | |
361 | { | |
362 | struct pid_link *link; | |
363 | struct pid *pid; | |
364 | int tmp; | |
365 | ||
366 | link = &task->pids[type]; | |
367 | pid = link->pid; | |
368 | ||
369 | hlist_del_rcu(&link->node); | |
370 | link->pid = new; | |
371 | ||
372 | for (tmp = PIDTYPE_MAX; --tmp >= 0; ) | |
373 | if (!hlist_empty(&pid->tasks[tmp])) | |
374 | return; | |
375 | ||
376 | free_pid(pid); | |
377 | } | |
378 | ||
379 | void detach_pid(struct task_struct *task, enum pid_type type) | |
380 | { | |
381 | __change_pid(task, type, NULL); | |
382 | } | |
383 | ||
384 | void change_pid(struct task_struct *task, enum pid_type type, | |
385 | struct pid *pid) | |
386 | { | |
387 | __change_pid(task, type, pid); | |
388 | attach_pid(task, type, pid); | |
389 | } | |
390 | ||
391 | /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */ | |
392 | void transfer_pid(struct task_struct *old, struct task_struct *new, | |
393 | enum pid_type type) | |
394 | { | |
395 | new->pids[type].pid = old->pids[type].pid; | |
396 | hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node); | |
397 | } | |
398 | ||
399 | struct task_struct *pid_task(struct pid *pid, enum pid_type type) | |
400 | { | |
401 | struct task_struct *result = NULL; | |
402 | if (pid) { | |
403 | struct hlist_node *first; | |
404 | first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]), | |
405 | rcu_read_lock_held() || | |
406 | lockdep_tasklist_lock_is_held()); | |
407 | if (first) | |
408 | result = hlist_entry(first, struct task_struct, pids[(type)].node); | |
409 | } | |
410 | return result; | |
411 | } | |
412 | EXPORT_SYMBOL(pid_task); | |
413 | ||
414 | /* | |
415 | * Must be called under rcu_read_lock(). | |
416 | */ | |
417 | struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns) | |
418 | { | |
419 | rcu_lockdep_assert(rcu_read_lock_held()); | |
420 | return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID); | |
421 | } | |
422 | ||
423 | struct task_struct *find_task_by_vpid(pid_t vnr) | |
424 | { | |
425 | return find_task_by_pid_ns(vnr, current->nsproxy->pid_ns); | |
426 | } | |
427 | ||
428 | struct pid *get_task_pid(struct task_struct *task, enum pid_type type) | |
429 | { | |
430 | struct pid *pid; | |
431 | rcu_read_lock(); | |
432 | if (type != PIDTYPE_PID) | |
433 | task = task->group_leader; | |
434 | pid = get_pid(task->pids[type].pid); | |
435 | rcu_read_unlock(); | |
436 | return pid; | |
437 | } | |
438 | ||
439 | struct task_struct *get_pid_task(struct pid *pid, enum pid_type type) | |
440 | { | |
441 | struct task_struct *result; | |
442 | rcu_read_lock(); | |
443 | result = pid_task(pid, type); | |
444 | if (result) | |
445 | get_task_struct(result); | |
446 | rcu_read_unlock(); | |
447 | return result; | |
448 | } | |
449 | ||
450 | struct pid *find_get_pid(pid_t nr) | |
451 | { | |
452 | struct pid *pid; | |
453 | ||
454 | rcu_read_lock(); | |
455 | pid = get_pid(find_vpid(nr)); | |
456 | rcu_read_unlock(); | |
457 | ||
458 | return pid; | |
459 | } | |
460 | EXPORT_SYMBOL_GPL(find_get_pid); | |
461 | ||
462 | pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns) | |
463 | { | |
464 | struct upid *upid; | |
465 | pid_t nr = 0; | |
466 | ||
467 | if (pid && ns->level <= pid->level) { | |
468 | upid = &pid->numbers[ns->level]; | |
469 | if (upid->ns == ns) | |
470 | nr = upid->nr; | |
471 | } | |
472 | return nr; | |
473 | } | |
474 | ||
475 | pid_t pid_vnr(struct pid *pid) | |
476 | { | |
477 | return pid_nr_ns(pid, current->nsproxy->pid_ns); | |
478 | } | |
479 | EXPORT_SYMBOL_GPL(pid_vnr); | |
480 | ||
481 | pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type, | |
482 | struct pid_namespace *ns) | |
483 | { | |
484 | pid_t nr = 0; | |
485 | ||
486 | rcu_read_lock(); | |
487 | if (!ns) | |
488 | ns = current->nsproxy->pid_ns; | |
489 | if (likely(pid_alive(task))) { | |
490 | if (type != PIDTYPE_PID) | |
491 | task = task->group_leader; | |
492 | nr = pid_nr_ns(task->pids[type].pid, ns); | |
493 | } | |
494 | rcu_read_unlock(); | |
495 | ||
496 | return nr; | |
497 | } | |
498 | EXPORT_SYMBOL(__task_pid_nr_ns); | |
499 | ||
500 | pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns) | |
501 | { | |
502 | return pid_nr_ns(task_tgid(tsk), ns); | |
503 | } | |
504 | EXPORT_SYMBOL(task_tgid_nr_ns); | |
505 | ||
506 | struct pid_namespace *task_active_pid_ns(struct task_struct *tsk) | |
507 | { | |
508 | return ns_of_pid(task_pid(tsk)); | |
509 | } | |
510 | EXPORT_SYMBOL_GPL(task_active_pid_ns); | |
511 | ||
512 | /* | |
513 | * Used by proc to find the first pid that is greater than or equal to nr. | |
514 | * | |
515 | * If there is a pid at nr this function is exactly the same as find_pid_ns. | |
516 | */ | |
517 | struct pid *find_ge_pid(int nr, struct pid_namespace *ns) | |
518 | { | |
519 | struct pid *pid; | |
520 | ||
521 | do { | |
522 | pid = find_pid_ns(nr, ns); | |
523 | if (pid) | |
524 | break; | |
525 | nr = next_pidmap(ns, nr); | |
526 | } while (nr > 0); | |
527 | ||
528 | return pid; | |
529 | } | |
530 | ||
531 | /* | |
532 | * The pid hash table is scaled according to the amount of memory in the | |
533 | * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or | |
534 | * more. | |
535 | */ | |
536 | void __init pidhash_init(void) | |
537 | { | |
538 | int i, pidhash_size; | |
539 | ||
540 | pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18, | |
541 | HASH_EARLY | HASH_SMALL, | |
542 | &pidhash_shift, NULL, 4096); | |
543 | pidhash_size = 1 << pidhash_shift; | |
544 | ||
545 | for (i = 0; i < pidhash_size; i++) | |
546 | INIT_HLIST_HEAD(&pid_hash[i]); | |
547 | } | |
548 | ||
549 | void __init pidmap_init(void) | |
550 | { | |
551 | /* bump default and minimum pid_max based on number of cpus */ | |
552 | pid_max = min(pid_max_max, max_t(int, pid_max, | |
553 | PIDS_PER_CPU_DEFAULT * num_possible_cpus())); | |
554 | pid_max_min = max_t(int, pid_max_min, | |
555 | PIDS_PER_CPU_MIN * num_possible_cpus()); | |
556 | pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min); | |
557 | ||
558 | init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
559 | /* Reserve PID 0. We never call free_pidmap(0) */ | |
560 | set_bit(0, init_pid_ns.pidmap[0].page); | |
561 | atomic_dec(&init_pid_ns.pidmap[0].nr_free); | |
562 | ||
563 | init_pid_ns.pid_cachep = KMEM_CACHE(pid, | |
564 | SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
565 | } |