]>
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 | ||
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 | #include <linux/pid_namespace.h> | |
30 | #include <linux/init_task.h> | |
31 | ||
32 | #define pid_hashfn(nr) hash_long((unsigned long)nr, pidhash_shift) | |
33 | static struct hlist_head *pid_hash; | |
34 | static int pidhash_shift; | |
35 | struct pid init_struct_pid = INIT_STRUCT_PID; | |
36 | ||
37 | int pid_max = PID_MAX_DEFAULT; | |
38 | ||
39 | #define RESERVED_PIDS 300 | |
40 | ||
41 | int pid_max_min = RESERVED_PIDS + 1; | |
42 | int pid_max_max = PID_MAX_LIMIT; | |
43 | ||
44 | #define BITS_PER_PAGE (PAGE_SIZE*8) | |
45 | #define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1) | |
46 | ||
47 | static inline int mk_pid(struct pid_namespace *pid_ns, | |
48 | struct pidmap *map, int off) | |
49 | { | |
50 | return (map - pid_ns->pidmap)*BITS_PER_PAGE + off; | |
51 | } | |
52 | ||
53 | #define find_next_offset(map, off) \ | |
54 | find_next_zero_bit((map)->page, BITS_PER_PAGE, off) | |
55 | ||
56 | /* | |
57 | * PID-map pages start out as NULL, they get allocated upon | |
58 | * first use and are never deallocated. This way a low pid_max | |
59 | * value does not cause lots of bitmaps to be allocated, but | |
60 | * the scheme scales to up to 4 million PIDs, runtime. | |
61 | */ | |
62 | struct pid_namespace init_pid_ns = { | |
63 | .kref = { | |
64 | .refcount = ATOMIC_INIT(2), | |
65 | }, | |
66 | .pidmap = { | |
67 | [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL } | |
68 | }, | |
69 | .last_pid = 0, | |
70 | .child_reaper = &init_task | |
71 | }; | |
72 | ||
73 | /* | |
74 | * Note: disable interrupts while the pidmap_lock is held as an | |
75 | * interrupt might come in and do read_lock(&tasklist_lock). | |
76 | * | |
77 | * If we don't disable interrupts there is a nasty deadlock between | |
78 | * detach_pid()->free_pid() and another cpu that does | |
79 | * spin_lock(&pidmap_lock) followed by an interrupt routine that does | |
80 | * read_lock(&tasklist_lock); | |
81 | * | |
82 | * After we clean up the tasklist_lock and know there are no | |
83 | * irq handlers that take it we can leave the interrupts enabled. | |
84 | * For now it is easier to be safe than to prove it can't happen. | |
85 | */ | |
86 | ||
87 | static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock); | |
88 | ||
89 | static fastcall void free_pidmap(struct pid_namespace *pid_ns, int pid) | |
90 | { | |
91 | struct pidmap *map = pid_ns->pidmap + pid / BITS_PER_PAGE; | |
92 | int offset = pid & BITS_PER_PAGE_MASK; | |
93 | ||
94 | clear_bit(offset, map->page); | |
95 | atomic_inc(&map->nr_free); | |
96 | } | |
97 | ||
98 | static int alloc_pidmap(struct pid_namespace *pid_ns) | |
99 | { | |
100 | int i, offset, max_scan, pid, last = pid_ns->last_pid; | |
101 | struct pidmap *map; | |
102 | ||
103 | pid = last + 1; | |
104 | if (pid >= pid_max) | |
105 | pid = RESERVED_PIDS; | |
106 | offset = pid & BITS_PER_PAGE_MASK; | |
107 | map = &pid_ns->pidmap[pid/BITS_PER_PAGE]; | |
108 | max_scan = (pid_max + BITS_PER_PAGE - 1)/BITS_PER_PAGE - !offset; | |
109 | for (i = 0; i <= max_scan; ++i) { | |
110 | if (unlikely(!map->page)) { | |
111 | void *page = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
112 | /* | |
113 | * Free the page if someone raced with us | |
114 | * installing it: | |
115 | */ | |
116 | spin_lock_irq(&pidmap_lock); | |
117 | if (map->page) | |
118 | kfree(page); | |
119 | else | |
120 | map->page = page; | |
121 | spin_unlock_irq(&pidmap_lock); | |
122 | if (unlikely(!map->page)) | |
123 | break; | |
124 | } | |
125 | if (likely(atomic_read(&map->nr_free))) { | |
126 | do { | |
127 | if (!test_and_set_bit(offset, map->page)) { | |
128 | atomic_dec(&map->nr_free); | |
129 | pid_ns->last_pid = pid; | |
130 | return pid; | |
131 | } | |
132 | offset = find_next_offset(map, offset); | |
133 | pid = mk_pid(pid_ns, map, offset); | |
134 | /* | |
135 | * find_next_offset() found a bit, the pid from it | |
136 | * is in-bounds, and if we fell back to the last | |
137 | * bitmap block and the final block was the same | |
138 | * as the starting point, pid is before last_pid. | |
139 | */ | |
140 | } while (offset < BITS_PER_PAGE && pid < pid_max && | |
141 | (i != max_scan || pid < last || | |
142 | !((last+1) & BITS_PER_PAGE_MASK))); | |
143 | } | |
144 | if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) { | |
145 | ++map; | |
146 | offset = 0; | |
147 | } else { | |
148 | map = &pid_ns->pidmap[0]; | |
149 | offset = RESERVED_PIDS; | |
150 | if (unlikely(last == offset)) | |
151 | break; | |
152 | } | |
153 | pid = mk_pid(pid_ns, map, offset); | |
154 | } | |
155 | return -1; | |
156 | } | |
157 | ||
158 | static int next_pidmap(struct pid_namespace *pid_ns, int last) | |
159 | { | |
160 | int offset; | |
161 | struct pidmap *map, *end; | |
162 | ||
163 | offset = (last + 1) & BITS_PER_PAGE_MASK; | |
164 | map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE]; | |
165 | end = &pid_ns->pidmap[PIDMAP_ENTRIES]; | |
166 | for (; map < end; map++, offset = 0) { | |
167 | if (unlikely(!map->page)) | |
168 | continue; | |
169 | offset = find_next_bit((map)->page, BITS_PER_PAGE, offset); | |
170 | if (offset < BITS_PER_PAGE) | |
171 | return mk_pid(pid_ns, map, offset); | |
172 | } | |
173 | return -1; | |
174 | } | |
175 | ||
176 | fastcall void put_pid(struct pid *pid) | |
177 | { | |
178 | struct pid_namespace *ns; | |
179 | ||
180 | if (!pid) | |
181 | return; | |
182 | ||
183 | /* FIXME - this must be the namespace this pid lives in */ | |
184 | ns = &init_pid_ns; | |
185 | if ((atomic_read(&pid->count) == 1) || | |
186 | atomic_dec_and_test(&pid->count)) | |
187 | kmem_cache_free(ns->pid_cachep, pid); | |
188 | } | |
189 | EXPORT_SYMBOL_GPL(put_pid); | |
190 | ||
191 | static void delayed_put_pid(struct rcu_head *rhp) | |
192 | { | |
193 | struct pid *pid = container_of(rhp, struct pid, rcu); | |
194 | put_pid(pid); | |
195 | } | |
196 | ||
197 | fastcall void free_pid(struct pid *pid) | |
198 | { | |
199 | /* We can be called with write_lock_irq(&tasklist_lock) held */ | |
200 | unsigned long flags; | |
201 | ||
202 | spin_lock_irqsave(&pidmap_lock, flags); | |
203 | hlist_del_rcu(&pid->pid_chain); | |
204 | spin_unlock_irqrestore(&pidmap_lock, flags); | |
205 | ||
206 | free_pidmap(&init_pid_ns, pid->nr); | |
207 | call_rcu(&pid->rcu, delayed_put_pid); | |
208 | } | |
209 | ||
210 | struct pid *alloc_pid(void) | |
211 | { | |
212 | struct pid *pid; | |
213 | enum pid_type type; | |
214 | int nr = -1; | |
215 | struct pid_namespace *ns; | |
216 | ||
217 | ns = task_active_pid_ns(current); | |
218 | pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL); | |
219 | if (!pid) | |
220 | goto out; | |
221 | ||
222 | nr = alloc_pidmap(ns); | |
223 | if (nr < 0) | |
224 | goto out_free; | |
225 | ||
226 | atomic_set(&pid->count, 1); | |
227 | pid->nr = nr; | |
228 | for (type = 0; type < PIDTYPE_MAX; ++type) | |
229 | INIT_HLIST_HEAD(&pid->tasks[type]); | |
230 | ||
231 | spin_lock_irq(&pidmap_lock); | |
232 | hlist_add_head_rcu(&pid->pid_chain, &pid_hash[pid_hashfn(pid->nr)]); | |
233 | spin_unlock_irq(&pidmap_lock); | |
234 | ||
235 | out: | |
236 | return pid; | |
237 | ||
238 | out_free: | |
239 | kmem_cache_free(ns->pid_cachep, pid); | |
240 | pid = NULL; | |
241 | goto out; | |
242 | } | |
243 | ||
244 | struct pid * fastcall find_pid(int nr) | |
245 | { | |
246 | struct hlist_node *elem; | |
247 | struct pid *pid; | |
248 | ||
249 | hlist_for_each_entry_rcu(pid, elem, | |
250 | &pid_hash[pid_hashfn(nr)], pid_chain) { | |
251 | if (pid->nr == nr) | |
252 | return pid; | |
253 | } | |
254 | return NULL; | |
255 | } | |
256 | EXPORT_SYMBOL_GPL(find_pid); | |
257 | ||
258 | /* | |
259 | * attach_pid() must be called with the tasklist_lock write-held. | |
260 | */ | |
261 | int fastcall attach_pid(struct task_struct *task, enum pid_type type, | |
262 | struct pid *pid) | |
263 | { | |
264 | struct pid_link *link; | |
265 | ||
266 | link = &task->pids[type]; | |
267 | link->pid = pid; | |
268 | hlist_add_head_rcu(&link->node, &pid->tasks[type]); | |
269 | ||
270 | return 0; | |
271 | } | |
272 | ||
273 | void fastcall detach_pid(struct task_struct *task, enum pid_type type) | |
274 | { | |
275 | struct pid_link *link; | |
276 | struct pid *pid; | |
277 | int tmp; | |
278 | ||
279 | link = &task->pids[type]; | |
280 | pid = link->pid; | |
281 | ||
282 | hlist_del_rcu(&link->node); | |
283 | link->pid = NULL; | |
284 | ||
285 | for (tmp = PIDTYPE_MAX; --tmp >= 0; ) | |
286 | if (!hlist_empty(&pid->tasks[tmp])) | |
287 | return; | |
288 | ||
289 | free_pid(pid); | |
290 | } | |
291 | ||
292 | /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */ | |
293 | void fastcall transfer_pid(struct task_struct *old, struct task_struct *new, | |
294 | enum pid_type type) | |
295 | { | |
296 | new->pids[type].pid = old->pids[type].pid; | |
297 | hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node); | |
298 | old->pids[type].pid = NULL; | |
299 | } | |
300 | ||
301 | struct task_struct * fastcall pid_task(struct pid *pid, enum pid_type type) | |
302 | { | |
303 | struct task_struct *result = NULL; | |
304 | if (pid) { | |
305 | struct hlist_node *first; | |
306 | first = rcu_dereference(pid->tasks[type].first); | |
307 | if (first) | |
308 | result = hlist_entry(first, struct task_struct, pids[(type)].node); | |
309 | } | |
310 | return result; | |
311 | } | |
312 | ||
313 | /* | |
314 | * Must be called under rcu_read_lock() or with tasklist_lock read-held. | |
315 | */ | |
316 | struct task_struct *find_task_by_pid_type(int type, int nr) | |
317 | { | |
318 | return pid_task(find_pid(nr), type); | |
319 | } | |
320 | ||
321 | EXPORT_SYMBOL(find_task_by_pid_type); | |
322 | ||
323 | struct pid *get_task_pid(struct task_struct *task, enum pid_type type) | |
324 | { | |
325 | struct pid *pid; | |
326 | rcu_read_lock(); | |
327 | pid = get_pid(task->pids[type].pid); | |
328 | rcu_read_unlock(); | |
329 | return pid; | |
330 | } | |
331 | ||
332 | struct task_struct *fastcall get_pid_task(struct pid *pid, enum pid_type type) | |
333 | { | |
334 | struct task_struct *result; | |
335 | rcu_read_lock(); | |
336 | result = pid_task(pid, type); | |
337 | if (result) | |
338 | get_task_struct(result); | |
339 | rcu_read_unlock(); | |
340 | return result; | |
341 | } | |
342 | ||
343 | struct pid *find_get_pid(pid_t nr) | |
344 | { | |
345 | struct pid *pid; | |
346 | ||
347 | rcu_read_lock(); | |
348 | pid = get_pid(find_pid(nr)); | |
349 | rcu_read_unlock(); | |
350 | ||
351 | return pid; | |
352 | } | |
353 | ||
354 | /* | |
355 | * Used by proc to find the first pid that is greater then or equal to nr. | |
356 | * | |
357 | * If there is a pid at nr this function is exactly the same as find_pid. | |
358 | */ | |
359 | struct pid *find_ge_pid(int nr) | |
360 | { | |
361 | struct pid *pid; | |
362 | ||
363 | do { | |
364 | pid = find_pid(nr); | |
365 | if (pid) | |
366 | break; | |
367 | nr = next_pidmap(task_active_pid_ns(current), nr); | |
368 | } while (nr > 0); | |
369 | ||
370 | return pid; | |
371 | } | |
372 | EXPORT_SYMBOL_GPL(find_get_pid); | |
373 | ||
374 | struct pid_cache { | |
375 | int nr_ids; | |
376 | char name[16]; | |
377 | struct kmem_cache *cachep; | |
378 | struct list_head list; | |
379 | }; | |
380 | ||
381 | static LIST_HEAD(pid_caches_lh); | |
382 | static DEFINE_MUTEX(pid_caches_mutex); | |
383 | ||
384 | /* | |
385 | * creates the kmem cache to allocate pids from. | |
386 | * @nr_ids: the number of numerical ids this pid will have to carry | |
387 | */ | |
388 | ||
389 | static struct kmem_cache *create_pid_cachep(int nr_ids) | |
390 | { | |
391 | struct pid_cache *pcache; | |
392 | struct kmem_cache *cachep; | |
393 | ||
394 | mutex_lock(&pid_caches_mutex); | |
395 | list_for_each_entry (pcache, &pid_caches_lh, list) | |
396 | if (pcache->nr_ids == nr_ids) | |
397 | goto out; | |
398 | ||
399 | pcache = kmalloc(sizeof(struct pid_cache), GFP_KERNEL); | |
400 | if (pcache == NULL) | |
401 | goto err_alloc; | |
402 | ||
403 | snprintf(pcache->name, sizeof(pcache->name), "pid_%d", nr_ids); | |
404 | cachep = kmem_cache_create(pcache->name, | |
405 | /* FIXME add numerical ids here */ | |
406 | sizeof(struct pid), 0, SLAB_HWCACHE_ALIGN, NULL); | |
407 | if (cachep == NULL) | |
408 | goto err_cachep; | |
409 | ||
410 | pcache->nr_ids = nr_ids; | |
411 | pcache->cachep = cachep; | |
412 | list_add(&pcache->list, &pid_caches_lh); | |
413 | out: | |
414 | mutex_unlock(&pid_caches_mutex); | |
415 | return pcache->cachep; | |
416 | ||
417 | err_cachep: | |
418 | kfree(pcache); | |
419 | err_alloc: | |
420 | mutex_unlock(&pid_caches_mutex); | |
421 | return NULL; | |
422 | } | |
423 | ||
424 | struct pid_namespace *copy_pid_ns(unsigned long flags, struct pid_namespace *old_ns) | |
425 | { | |
426 | BUG_ON(!old_ns); | |
427 | get_pid_ns(old_ns); | |
428 | return old_ns; | |
429 | } | |
430 | ||
431 | void free_pid_ns(struct kref *kref) | |
432 | { | |
433 | struct pid_namespace *ns; | |
434 | ||
435 | ns = container_of(kref, struct pid_namespace, kref); | |
436 | kfree(ns); | |
437 | } | |
438 | ||
439 | /* | |
440 | * The pid hash table is scaled according to the amount of memory in the | |
441 | * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or | |
442 | * more. | |
443 | */ | |
444 | void __init pidhash_init(void) | |
445 | { | |
446 | int i, pidhash_size; | |
447 | unsigned long megabytes = nr_kernel_pages >> (20 - PAGE_SHIFT); | |
448 | ||
449 | pidhash_shift = max(4, fls(megabytes * 4)); | |
450 | pidhash_shift = min(12, pidhash_shift); | |
451 | pidhash_size = 1 << pidhash_shift; | |
452 | ||
453 | printk("PID hash table entries: %d (order: %d, %Zd bytes)\n", | |
454 | pidhash_size, pidhash_shift, | |
455 | pidhash_size * sizeof(struct hlist_head)); | |
456 | ||
457 | pid_hash = alloc_bootmem(pidhash_size * sizeof(*(pid_hash))); | |
458 | if (!pid_hash) | |
459 | panic("Could not alloc pidhash!\n"); | |
460 | for (i = 0; i < pidhash_size; i++) | |
461 | INIT_HLIST_HEAD(&pid_hash[i]); | |
462 | } | |
463 | ||
464 | void __init pidmap_init(void) | |
465 | { | |
466 | init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
467 | /* Reserve PID 0. We never call free_pidmap(0) */ | |
468 | set_bit(0, init_pid_ns.pidmap[0].page); | |
469 | atomic_dec(&init_pid_ns.pidmap[0].nr_free); | |
470 | ||
471 | init_pid_ns.pid_cachep = create_pid_cachep(1); | |
472 | if (init_pid_ns.pid_cachep == NULL) | |
473 | panic("Can't create pid_1 cachep\n"); | |
474 | } |