Version:  2.0.40 2.2.26 2.4.37 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10

Linux/kernel/cgroup.c

  1 /*
  2  *  Generic process-grouping system.
  3  *
  4  *  Based originally on the cpuset system, extracted by Paul Menage
  5  *  Copyright (C) 2006 Google, Inc
  6  *
  7  *  Notifications support
  8  *  Copyright (C) 2009 Nokia Corporation
  9  *  Author: Kirill A. Shutemov
 10  *
 11  *  Copyright notices from the original cpuset code:
 12  *  --------------------------------------------------
 13  *  Copyright (C) 2003 BULL SA.
 14  *  Copyright (C) 2004-2006 Silicon Graphics, Inc.
 15  *
 16  *  Portions derived from Patrick Mochel's sysfs code.
 17  *  sysfs is Copyright (c) 2001-3 Patrick Mochel
 18  *
 19  *  2003-10-10 Written by Simon Derr.
 20  *  2003-10-22 Updates by Stephen Hemminger.
 21  *  2004 May-July Rework by Paul Jackson.
 22  *  ---------------------------------------------------
 23  *
 24  *  This file is subject to the terms and conditions of the GNU General Public
 25  *  License.  See the file COPYING in the main directory of the Linux
 26  *  distribution for more details.
 27  */
 28 
 29 #include <linux/cgroup.h>
 30 #include <linux/cred.h>
 31 #include <linux/ctype.h>
 32 #include <linux/errno.h>
 33 #include <linux/init_task.h>
 34 #include <linux/kernel.h>
 35 #include <linux/list.h>
 36 #include <linux/mm.h>
 37 #include <linux/mutex.h>
 38 #include <linux/mount.h>
 39 #include <linux/pagemap.h>
 40 #include <linux/proc_fs.h>
 41 #include <linux/rcupdate.h>
 42 #include <linux/sched.h>
 43 #include <linux/backing-dev.h>
 44 #include <linux/slab.h>
 45 #include <linux/magic.h>
 46 #include <linux/spinlock.h>
 47 #include <linux/string.h>
 48 #include <linux/sort.h>
 49 #include <linux/kmod.h>
 50 #include <linux/module.h>
 51 #include <linux/delayacct.h>
 52 #include <linux/cgroupstats.h>
 53 #include <linux/hashtable.h>
 54 #include <linux/namei.h>
 55 #include <linux/pid_namespace.h>
 56 #include <linux/idr.h>
 57 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
 58 #include <linux/flex_array.h> /* used in cgroup_attach_task */
 59 #include <linux/kthread.h>
 60 
 61 #include <linux/atomic.h>
 62 
 63 /*
 64  * pidlists linger the following amount before being destroyed.  The goal
 65  * is avoiding frequent destruction in the middle of consecutive read calls
 66  * Expiring in the middle is a performance problem not a correctness one.
 67  * 1 sec should be enough.
 68  */
 69 #define CGROUP_PIDLIST_DESTROY_DELAY    HZ
 70 
 71 /*
 72  * cgroup_mutex is the master lock.  Any modification to cgroup or its
 73  * hierarchy must be performed while holding it.
 74  *
 75  * cgroup_root_mutex nests inside cgroup_mutex and should be held to modify
 76  * cgroupfs_root of any cgroup hierarchy - subsys list, flags,
 77  * release_agent_path and so on.  Modifying requires both cgroup_mutex and
 78  * cgroup_root_mutex.  Readers can acquire either of the two.  This is to
 79  * break the following locking order cycle.
 80  *
 81  *  A. cgroup_mutex -> cred_guard_mutex -> s_type->i_mutex_key -> namespace_sem
 82  *  B. namespace_sem -> cgroup_mutex
 83  *
 84  * B happens only through cgroup_show_options() and using cgroup_root_mutex
 85  * breaks it.
 86  */
 87 #ifdef CONFIG_PROVE_RCU
 88 DEFINE_MUTEX(cgroup_mutex);
 89 EXPORT_SYMBOL_GPL(cgroup_mutex);        /* only for lockdep */
 90 #else
 91 static DEFINE_MUTEX(cgroup_mutex);
 92 #endif
 93 
 94 static DEFINE_MUTEX(cgroup_root_mutex);
 95 
 96 #define cgroup_assert_mutex_or_rcu_locked()                             \
 97         rcu_lockdep_assert(rcu_read_lock_held() ||                      \
 98                            lockdep_is_held(&cgroup_mutex),              \
 99                            "cgroup_mutex or RCU read lock required");
100 
101 #ifdef CONFIG_LOCKDEP
102 #define cgroup_assert_mutex_or_root_locked()                            \
103         WARN_ON_ONCE(debug_locks && (!lockdep_is_held(&cgroup_mutex) && \
104                                      !lockdep_is_held(&cgroup_root_mutex)))
105 #else
106 #define cgroup_assert_mutex_or_root_locked()    do { } while (0)
107 #endif
108 
109 /*
110  * cgroup destruction makes heavy use of work items and there can be a lot
111  * of concurrent destructions.  Use a separate workqueue so that cgroup
112  * destruction work items don't end up filling up max_active of system_wq
113  * which may lead to deadlock.
114  */
115 static struct workqueue_struct *cgroup_destroy_wq;
116 
117 /*
118  * pidlist destructions need to be flushed on cgroup destruction.  Use a
119  * separate workqueue as flush domain.
120  */
121 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
122 
123 /*
124  * Generate an array of cgroup subsystem pointers. At boot time, this is
125  * populated with the built in subsystems, and modular subsystems are
126  * registered after that. The mutable section of this array is protected by
127  * cgroup_mutex.
128  */
129 #define SUBSYS(_x) [_x ## _subsys_id] = &_x ## _subsys,
130 #define IS_SUBSYS_ENABLED(option) IS_BUILTIN(option)
131 static struct cgroup_subsys *cgroup_subsys[CGROUP_SUBSYS_COUNT] = {
132 #include <linux/cgroup_subsys.h>
133 };
134 
135 /*
136  * The dummy hierarchy, reserved for the subsystems that are otherwise
137  * unattached - it never has more than a single cgroup, and all tasks are
138  * part of that cgroup.
139  */
140 static struct cgroupfs_root cgroup_dummy_root;
141 
142 /* dummy_top is a shorthand for the dummy hierarchy's top cgroup */
143 static struct cgroup * const cgroup_dummy_top = &cgroup_dummy_root.top_cgroup;
144 
145 /* The list of hierarchy roots */
146 
147 static LIST_HEAD(cgroup_roots);
148 static int cgroup_root_count;
149 
150 /*
151  * Hierarchy ID allocation and mapping.  It follows the same exclusion
152  * rules as other root ops - both cgroup_mutex and cgroup_root_mutex for
153  * writes, either for reads.
154  */
155 static DEFINE_IDR(cgroup_hierarchy_idr);
156 
157 static struct cgroup_name root_cgroup_name = { .name = "/" };
158 
159 /*
160  * Assign a monotonically increasing serial number to cgroups.  It
161  * guarantees cgroups with bigger numbers are newer than those with smaller
162  * numbers.  Also, as cgroups are always appended to the parent's
163  * ->children list, it guarantees that sibling cgroups are always sorted in
164  * the ascending serial number order on the list.  Protected by
165  * cgroup_mutex.
166  */
167 static u64 cgroup_serial_nr_next = 1;
168 
169 /* This flag indicates whether tasks in the fork and exit paths should
170  * check for fork/exit handlers to call. This avoids us having to do
171  * extra work in the fork/exit path if none of the subsystems need to
172  * be called.
173  */
174 static int need_forkexit_callback __read_mostly;
175 
176 static struct cftype cgroup_base_files[];
177 
178 static void cgroup_destroy_css_killed(struct cgroup *cgrp);
179 static int cgroup_destroy_locked(struct cgroup *cgrp);
180 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
181                               bool is_add);
182 static int cgroup_file_release(struct inode *inode, struct file *file);
183 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp);
184 
185 /**
186  * cgroup_css - obtain a cgroup's css for the specified subsystem
187  * @cgrp: the cgroup of interest
188  * @ss: the subsystem of interest (%NULL returns the dummy_css)
189  *
190  * Return @cgrp's css (cgroup_subsys_state) associated with @ss.  This
191  * function must be called either under cgroup_mutex or rcu_read_lock() and
192  * the caller is responsible for pinning the returned css if it wants to
193  * keep accessing it outside the said locks.  This function may return
194  * %NULL if @cgrp doesn't have @subsys_id enabled.
195  */
196 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
197                                               struct cgroup_subsys *ss)
198 {
199         if (ss)
200                 return rcu_dereference_check(cgrp->subsys[ss->subsys_id],
201                                              lockdep_is_held(&cgroup_mutex));
202         else
203                 return &cgrp->dummy_css;
204 }
205 
206 /* convenient tests for these bits */
207 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
208 {
209         return test_bit(CGRP_DEAD, &cgrp->flags);
210 }
211 
212 /**
213  * cgroup_is_descendant - test ancestry
214  * @cgrp: the cgroup to be tested
215  * @ancestor: possible ancestor of @cgrp
216  *
217  * Test whether @cgrp is a descendant of @ancestor.  It also returns %true
218  * if @cgrp == @ancestor.  This function is safe to call as long as @cgrp
219  * and @ancestor are accessible.
220  */
221 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
222 {
223         while (cgrp) {
224                 if (cgrp == ancestor)
225                         return true;
226                 cgrp = cgrp->parent;
227         }
228         return false;
229 }
230 EXPORT_SYMBOL_GPL(cgroup_is_descendant);
231 
232 static int cgroup_is_releasable(const struct cgroup *cgrp)
233 {
234         const int bits =
235                 (1 << CGRP_RELEASABLE) |
236                 (1 << CGRP_NOTIFY_ON_RELEASE);
237         return (cgrp->flags & bits) == bits;
238 }
239 
240 static int notify_on_release(const struct cgroup *cgrp)
241 {
242         return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
243 }
244 
245 /**
246  * for_each_css - iterate all css's of a cgroup
247  * @css: the iteration cursor
248  * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
249  * @cgrp: the target cgroup to iterate css's of
250  *
251  * Should be called under cgroup_mutex.
252  */
253 #define for_each_css(css, ssid, cgrp)                                   \
254         for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)        \
255                 if (!((css) = rcu_dereference_check(                    \
256                                 (cgrp)->subsys[(ssid)],                 \
257                                 lockdep_is_held(&cgroup_mutex)))) { }   \
258                 else
259 
260 /**
261  * for_each_subsys - iterate all loaded cgroup subsystems
262  * @ss: the iteration cursor
263  * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
264  *
265  * Iterates through all loaded subsystems.  Should be called under
266  * cgroup_mutex or cgroup_root_mutex.
267  */
268 #define for_each_subsys(ss, ssid)                                       \
269         for (({ cgroup_assert_mutex_or_root_locked(); (ssid) = 0; });   \
270              (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)                    \
271                 if (!((ss) = cgroup_subsys[(ssid)])) { }                \
272                 else
273 
274 /**
275  * for_each_builtin_subsys - iterate all built-in cgroup subsystems
276  * @ss: the iteration cursor
277  * @i: the index of @ss, CGROUP_BUILTIN_SUBSYS_COUNT after reaching the end
278  *
279  * Bulit-in subsystems are always present and iteration itself doesn't
280  * require any synchronization.
281  */
282 #define for_each_builtin_subsys(ss, i)                                  \
283         for ((i) = 0; (i) < CGROUP_BUILTIN_SUBSYS_COUNT &&              \
284              (((ss) = cgroup_subsys[i]) || true); (i)++)
285 
286 /* iterate across the active hierarchies */
287 #define for_each_active_root(root)                                      \
288         list_for_each_entry((root), &cgroup_roots, root_list)
289 
290 static inline struct cgroup *__d_cgrp(struct dentry *dentry)
291 {
292         return dentry->d_fsdata;
293 }
294 
295 static inline struct cfent *__d_cfe(struct dentry *dentry)
296 {
297         return dentry->d_fsdata;
298 }
299 
300 static inline struct cftype *__d_cft(struct dentry *dentry)
301 {
302         return __d_cfe(dentry)->type;
303 }
304 
305 /**
306  * cgroup_lock_live_group - take cgroup_mutex and check that cgrp is alive.
307  * @cgrp: the cgroup to be checked for liveness
308  *
309  * On success, returns true; the mutex should be later unlocked.  On
310  * failure returns false with no lock held.
311  */
312 static bool cgroup_lock_live_group(struct cgroup *cgrp)
313 {
314         mutex_lock(&cgroup_mutex);
315         if (cgroup_is_dead(cgrp)) {
316                 mutex_unlock(&cgroup_mutex);
317                 return false;
318         }
319         return true;
320 }
321 
322 /* the list of cgroups eligible for automatic release. Protected by
323  * release_list_lock */
324 static LIST_HEAD(release_list);
325 static DEFINE_RAW_SPINLOCK(release_list_lock);
326 static void cgroup_release_agent(struct work_struct *work);
327 static DECLARE_WORK(release_agent_work, cgroup_release_agent);
328 static void check_for_release(struct cgroup *cgrp);
329 
330 /*
331  * A cgroup can be associated with multiple css_sets as different tasks may
332  * belong to different cgroups on different hierarchies.  In the other
333  * direction, a css_set is naturally associated with multiple cgroups.
334  * This M:N relationship is represented by the following link structure
335  * which exists for each association and allows traversing the associations
336  * from both sides.
337  */
338 struct cgrp_cset_link {
339         /* the cgroup and css_set this link associates */
340         struct cgroup           *cgrp;
341         struct css_set          *cset;
342 
343         /* list of cgrp_cset_links anchored at cgrp->cset_links */
344         struct list_head        cset_link;
345 
346         /* list of cgrp_cset_links anchored at css_set->cgrp_links */
347         struct list_head        cgrp_link;
348 };
349 
350 /* The default css_set - used by init and its children prior to any
351  * hierarchies being mounted. It contains a pointer to the root state
352  * for each subsystem. Also used to anchor the list of css_sets. Not
353  * reference-counted, to improve performance when child cgroups
354  * haven't been created.
355  */
356 
357 static struct css_set init_css_set;
358 static struct cgrp_cset_link init_cgrp_cset_link;
359 
360 /*
361  * css_set_lock protects the list of css_set objects, and the chain of
362  * tasks off each css_set.  Nests outside task->alloc_lock due to
363  * css_task_iter_start().
364  */
365 static DEFINE_RWLOCK(css_set_lock);
366 static int css_set_count;
367 
368 /*
369  * hash table for cgroup groups. This improves the performance to find
370  * an existing css_set. This hash doesn't (currently) take into
371  * account cgroups in empty hierarchies.
372  */
373 #define CSS_SET_HASH_BITS       7
374 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
375 
376 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
377 {
378         unsigned long key = 0UL;
379         struct cgroup_subsys *ss;
380         int i;
381 
382         for_each_subsys(ss, i)
383                 key += (unsigned long)css[i];
384         key = (key >> 16) ^ key;
385 
386         return key;
387 }
388 
389 /*
390  * We don't maintain the lists running through each css_set to its task
391  * until after the first call to css_task_iter_start().  This reduces the
392  * fork()/exit() overhead for people who have cgroups compiled into their
393  * kernel but not actually in use.
394  */
395 static int use_task_css_set_links __read_mostly;
396 
397 static void __put_css_set(struct css_set *cset, int taskexit)
398 {
399         struct cgrp_cset_link *link, *tmp_link;
400 
401         /*
402          * Ensure that the refcount doesn't hit zero while any readers
403          * can see it. Similar to atomic_dec_and_lock(), but for an
404          * rwlock
405          */
406         if (atomic_add_unless(&cset->refcount, -1, 1))
407                 return;
408         write_lock(&css_set_lock);
409         if (!atomic_dec_and_test(&cset->refcount)) {
410                 write_unlock(&css_set_lock);
411                 return;
412         }
413 
414         /* This css_set is dead. unlink it and release cgroup refcounts */
415         hash_del(&cset->hlist);
416         css_set_count--;
417 
418         list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
419                 struct cgroup *cgrp = link->cgrp;
420 
421                 list_del(&link->cset_link);
422                 list_del(&link->cgrp_link);
423 
424                 /* @cgrp can't go away while we're holding css_set_lock */
425                 if (list_empty(&cgrp->cset_links) && notify_on_release(cgrp)) {
426                         if (taskexit)
427                                 set_bit(CGRP_RELEASABLE, &cgrp->flags);
428                         check_for_release(cgrp);
429                 }
430 
431                 kfree(link);
432         }
433 
434         write_unlock(&css_set_lock);
435         kfree_rcu(cset, rcu_head);
436 }
437 
438 /*
439  * refcounted get/put for css_set objects
440  */
441 static inline void get_css_set(struct css_set *cset)
442 {
443         atomic_inc(&cset->refcount);
444 }
445 
446 static inline void put_css_set(struct css_set *cset)
447 {
448         __put_css_set(cset, 0);
449 }
450 
451 static inline void put_css_set_taskexit(struct css_set *cset)
452 {
453         __put_css_set(cset, 1);
454 }
455 
456 /**
457  * compare_css_sets - helper function for find_existing_css_set().
458  * @cset: candidate css_set being tested
459  * @old_cset: existing css_set for a task
460  * @new_cgrp: cgroup that's being entered by the task
461  * @template: desired set of css pointers in css_set (pre-calculated)
462  *
463  * Returns true if "cset" matches "old_cset" except for the hierarchy
464  * which "new_cgrp" belongs to, for which it should match "new_cgrp".
465  */
466 static bool compare_css_sets(struct css_set *cset,
467                              struct css_set *old_cset,
468                              struct cgroup *new_cgrp,
469                              struct cgroup_subsys_state *template[])
470 {
471         struct list_head *l1, *l2;
472 
473         if (memcmp(template, cset->subsys, sizeof(cset->subsys))) {
474                 /* Not all subsystems matched */
475                 return false;
476         }
477 
478         /*
479          * Compare cgroup pointers in order to distinguish between
480          * different cgroups in heirarchies with no subsystems. We
481          * could get by with just this check alone (and skip the
482          * memcmp above) but on most setups the memcmp check will
483          * avoid the need for this more expensive check on almost all
484          * candidates.
485          */
486 
487         l1 = &cset->cgrp_links;
488         l2 = &old_cset->cgrp_links;
489         while (1) {
490                 struct cgrp_cset_link *link1, *link2;
491                 struct cgroup *cgrp1, *cgrp2;
492 
493                 l1 = l1->next;
494                 l2 = l2->next;
495                 /* See if we reached the end - both lists are equal length. */
496                 if (l1 == &cset->cgrp_links) {
497                         BUG_ON(l2 != &old_cset->cgrp_links);
498                         break;
499                 } else {
500                         BUG_ON(l2 == &old_cset->cgrp_links);
501                 }
502                 /* Locate the cgroups associated with these links. */
503                 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
504                 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
505                 cgrp1 = link1->cgrp;
506                 cgrp2 = link2->cgrp;
507                 /* Hierarchies should be linked in the same order. */
508                 BUG_ON(cgrp1->root != cgrp2->root);
509 
510                 /*
511                  * If this hierarchy is the hierarchy of the cgroup
512                  * that's changing, then we need to check that this
513                  * css_set points to the new cgroup; if it's any other
514                  * hierarchy, then this css_set should point to the
515                  * same cgroup as the old css_set.
516                  */
517                 if (cgrp1->root == new_cgrp->root) {
518                         if (cgrp1 != new_cgrp)
519                                 return false;
520                 } else {
521                         if (cgrp1 != cgrp2)
522                                 return false;
523                 }
524         }
525         return true;
526 }
527 
528 /**
529  * find_existing_css_set - init css array and find the matching css_set
530  * @old_cset: the css_set that we're using before the cgroup transition
531  * @cgrp: the cgroup that we're moving into
532  * @template: out param for the new set of csses, should be clear on entry
533  */
534 static struct css_set *find_existing_css_set(struct css_set *old_cset,
535                                         struct cgroup *cgrp,
536                                         struct cgroup_subsys_state *template[])
537 {
538         struct cgroupfs_root *root = cgrp->root;
539         struct cgroup_subsys *ss;
540         struct css_set *cset;
541         unsigned long key;
542         int i;
543 
544         /*
545          * Build the set of subsystem state objects that we want to see in the
546          * new css_set. while subsystems can change globally, the entries here
547          * won't change, so no need for locking.
548          */
549         for_each_subsys(ss, i) {
550                 if (root->subsys_mask & (1UL << i)) {
551                         /* Subsystem is in this hierarchy. So we want
552                          * the subsystem state from the new
553                          * cgroup */
554                         template[i] = cgroup_css(cgrp, ss);
555                 } else {
556                         /* Subsystem is not in this hierarchy, so we
557                          * don't want to change the subsystem state */
558                         template[i] = old_cset->subsys[i];
559                 }
560         }
561 
562         key = css_set_hash(template);
563         hash_for_each_possible(css_set_table, cset, hlist, key) {
564                 if (!compare_css_sets(cset, old_cset, cgrp, template))
565                         continue;
566 
567                 /* This css_set matches what we need */
568                 return cset;
569         }
570 
571         /* No existing cgroup group matched */
572         return NULL;
573 }
574 
575 static void free_cgrp_cset_links(struct list_head *links_to_free)
576 {
577         struct cgrp_cset_link *link, *tmp_link;
578 
579         list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
580                 list_del(&link->cset_link);
581                 kfree(link);
582         }
583 }
584 
585 /**
586  * allocate_cgrp_cset_links - allocate cgrp_cset_links
587  * @count: the number of links to allocate
588  * @tmp_links: list_head the allocated links are put on
589  *
590  * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
591  * through ->cset_link.  Returns 0 on success or -errno.
592  */
593 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
594 {
595         struct cgrp_cset_link *link;
596         int i;
597 
598         INIT_LIST_HEAD(tmp_links);
599 
600         for (i = 0; i < count; i++) {
601                 link = kzalloc(sizeof(*link), GFP_KERNEL);
602                 if (!link) {
603                         free_cgrp_cset_links(tmp_links);
604                         return -ENOMEM;
605                 }
606                 list_add(&link->cset_link, tmp_links);
607         }
608         return 0;
609 }
610 
611 /**
612  * link_css_set - a helper function to link a css_set to a cgroup
613  * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
614  * @cset: the css_set to be linked
615  * @cgrp: the destination cgroup
616  */
617 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
618                          struct cgroup *cgrp)
619 {
620         struct cgrp_cset_link *link;
621 
622         BUG_ON(list_empty(tmp_links));
623         link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
624         link->cset = cset;
625         link->cgrp = cgrp;
626         list_move(&link->cset_link, &cgrp->cset_links);
627         /*
628          * Always add links to the tail of the list so that the list
629          * is sorted by order of hierarchy creation
630          */
631         list_add_tail(&link->cgrp_link, &cset->cgrp_links);
632 }
633 
634 /**
635  * find_css_set - return a new css_set with one cgroup updated
636  * @old_cset: the baseline css_set
637  * @cgrp: the cgroup to be updated
638  *
639  * Return a new css_set that's equivalent to @old_cset, but with @cgrp
640  * substituted into the appropriate hierarchy.
641  */
642 static struct css_set *find_css_set(struct css_set *old_cset,
643                                     struct cgroup *cgrp)
644 {
645         struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
646         struct css_set *cset;
647         struct list_head tmp_links;
648         struct cgrp_cset_link *link;
649         unsigned long key;
650 
651         lockdep_assert_held(&cgroup_mutex);
652 
653         /* First see if we already have a cgroup group that matches
654          * the desired set */
655         read_lock(&css_set_lock);
656         cset = find_existing_css_set(old_cset, cgrp, template);
657         if (cset)
658                 get_css_set(cset);
659         read_unlock(&css_set_lock);
660 
661         if (cset)
662                 return cset;
663 
664         cset = kzalloc(sizeof(*cset), GFP_KERNEL);
665         if (!cset)
666                 return NULL;
667 
668         /* Allocate all the cgrp_cset_link objects that we'll need */
669         if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
670                 kfree(cset);
671                 return NULL;
672         }
673 
674         atomic_set(&cset->refcount, 1);
675         INIT_LIST_HEAD(&cset->cgrp_links);
676         INIT_LIST_HEAD(&cset->tasks);
677         INIT_HLIST_NODE(&cset->hlist);
678 
679         /* Copy the set of subsystem state objects generated in
680          * find_existing_css_set() */
681         memcpy(cset->subsys, template, sizeof(cset->subsys));
682 
683         write_lock(&css_set_lock);
684         /* Add reference counts and links from the new css_set. */
685         list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
686                 struct cgroup *c = link->cgrp;
687 
688                 if (c->root == cgrp->root)
689                         c = cgrp;
690                 link_css_set(&tmp_links, cset, c);
691         }
692 
693         BUG_ON(!list_empty(&tmp_links));
694 
695         css_set_count++;
696 
697         /* Add this cgroup group to the hash table */
698         key = css_set_hash(cset->subsys);
699         hash_add(css_set_table, &cset->hlist, key);
700 
701         write_unlock(&css_set_lock);
702 
703         return cset;
704 }
705 
706 /*
707  * Return the cgroup for "task" from the given hierarchy. Must be
708  * called with cgroup_mutex held.
709  */
710 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
711                                             struct cgroupfs_root *root)
712 {
713         struct css_set *cset;
714         struct cgroup *res = NULL;
715 
716         BUG_ON(!mutex_is_locked(&cgroup_mutex));
717         read_lock(&css_set_lock);
718         /*
719          * No need to lock the task - since we hold cgroup_mutex the
720          * task can't change groups, so the only thing that can happen
721          * is that it exits and its css is set back to init_css_set.
722          */
723         cset = task_css_set(task);
724         if (cset == &init_css_set) {
725                 res = &root->top_cgroup;
726         } else {
727                 struct cgrp_cset_link *link;
728 
729                 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
730                         struct cgroup *c = link->cgrp;
731 
732                         if (c->root == root) {
733                                 res = c;
734                                 break;
735                         }
736                 }
737         }
738         read_unlock(&css_set_lock);
739         BUG_ON(!res);
740         return res;
741 }
742 
743 /*
744  * There is one global cgroup mutex. We also require taking
745  * task_lock() when dereferencing a task's cgroup subsys pointers.
746  * See "The task_lock() exception", at the end of this comment.
747  *
748  * A task must hold cgroup_mutex to modify cgroups.
749  *
750  * Any task can increment and decrement the count field without lock.
751  * So in general, code holding cgroup_mutex can't rely on the count
752  * field not changing.  However, if the count goes to zero, then only
753  * cgroup_attach_task() can increment it again.  Because a count of zero
754  * means that no tasks are currently attached, therefore there is no
755  * way a task attached to that cgroup can fork (the other way to
756  * increment the count).  So code holding cgroup_mutex can safely
757  * assume that if the count is zero, it will stay zero. Similarly, if
758  * a task holds cgroup_mutex on a cgroup with zero count, it
759  * knows that the cgroup won't be removed, as cgroup_rmdir()
760  * needs that mutex.
761  *
762  * The fork and exit callbacks cgroup_fork() and cgroup_exit(), don't
763  * (usually) take cgroup_mutex.  These are the two most performance
764  * critical pieces of code here.  The exception occurs on cgroup_exit(),
765  * when a task in a notify_on_release cgroup exits.  Then cgroup_mutex
766  * is taken, and if the cgroup count is zero, a usermode call made
767  * to the release agent with the name of the cgroup (path relative to
768  * the root of cgroup file system) as the argument.
769  *
770  * A cgroup can only be deleted if both its 'count' of using tasks
771  * is zero, and its list of 'children' cgroups is empty.  Since all
772  * tasks in the system use _some_ cgroup, and since there is always at
773  * least one task in the system (init, pid == 1), therefore, top_cgroup
774  * always has either children cgroups and/or using tasks.  So we don't
775  * need a special hack to ensure that top_cgroup cannot be deleted.
776  *
777  *      The task_lock() exception
778  *
779  * The need for this exception arises from the action of
780  * cgroup_attach_task(), which overwrites one task's cgroup pointer with
781  * another.  It does so using cgroup_mutex, however there are
782  * several performance critical places that need to reference
783  * task->cgroup without the expense of grabbing a system global
784  * mutex.  Therefore except as noted below, when dereferencing or, as
785  * in cgroup_attach_task(), modifying a task's cgroup pointer we use
786  * task_lock(), which acts on a spinlock (task->alloc_lock) already in
787  * the task_struct routinely used for such matters.
788  *
789  * P.S.  One more locking exception.  RCU is used to guard the
790  * update of a tasks cgroup pointer by cgroup_attach_task()
791  */
792 
793 /*
794  * A couple of forward declarations required, due to cyclic reference loop:
795  * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
796  * cgroup_add_file -> cgroup_create_file -> cgroup_dir_inode_operations
797  * -> cgroup_mkdir.
798  */
799 
800 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode);
801 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
802 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
803 static const struct inode_operations cgroup_dir_inode_operations;
804 static const struct file_operations proc_cgroupstats_operations;
805 
806 static struct backing_dev_info cgroup_backing_dev_info = {
807         .name           = "cgroup",
808         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
809 };
810 
811 static struct inode *cgroup_new_inode(umode_t mode, struct super_block *sb)
812 {
813         struct inode *inode = new_inode(sb);
814 
815         if (inode) {
816                 inode->i_ino = get_next_ino();
817                 inode->i_mode = mode;
818                 inode->i_uid = current_fsuid();
819                 inode->i_gid = current_fsgid();
820                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
821                 inode->i_mapping->backing_dev_info = &cgroup_backing_dev_info;
822         }
823         return inode;
824 }
825 
826 static struct cgroup_name *cgroup_alloc_name(struct dentry *dentry)
827 {
828         struct cgroup_name *name;
829 
830         name = kmalloc(sizeof(*name) + dentry->d_name.len + 1, GFP_KERNEL);
831         if (!name)
832                 return NULL;
833         strcpy(name->name, dentry->d_name.name);
834         return name;
835 }
836 
837 static void cgroup_free_fn(struct work_struct *work)
838 {
839         struct cgroup *cgrp = container_of(work, struct cgroup, destroy_work);
840 
841         mutex_lock(&cgroup_mutex);
842         cgrp->root->number_of_cgroups--;
843         mutex_unlock(&cgroup_mutex);
844 
845         /*
846          * We get a ref to the parent's dentry, and put the ref when
847          * this cgroup is being freed, so it's guaranteed that the
848          * parent won't be destroyed before its children.
849          */
850         dput(cgrp->parent->dentry);
851 
852         /*
853          * Drop the active superblock reference that we took when we
854          * created the cgroup. This will free cgrp->root, if we are
855          * holding the last reference to @sb.
856          */
857         deactivate_super(cgrp->root->sb);
858 
859         cgroup_pidlist_destroy_all(cgrp);
860 
861         simple_xattrs_free(&cgrp->xattrs);
862 
863         kfree(rcu_dereference_raw(cgrp->name));
864         kfree(cgrp);
865 }
866 
867 static void cgroup_free_rcu(struct rcu_head *head)
868 {
869         struct cgroup *cgrp = container_of(head, struct cgroup, rcu_head);
870 
871         INIT_WORK(&cgrp->destroy_work, cgroup_free_fn);
872         queue_work(cgroup_destroy_wq, &cgrp->destroy_work);
873 }
874 
875 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
876 {
877         /* is dentry a directory ? if so, kfree() associated cgroup */
878         if (S_ISDIR(inode->i_mode)) {
879                 struct cgroup *cgrp = dentry->d_fsdata;
880 
881                 BUG_ON(!(cgroup_is_dead(cgrp)));
882 
883                 /*
884                  * XXX: cgrp->id is only used to look up css's.  As cgroup
885                  * and css's lifetimes will be decoupled, it should be made
886                  * per-subsystem and moved to css->id so that lookups are
887                  * successful until the target css is released.
888                  */
889                 mutex_lock(&cgroup_mutex);
890                 idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
891                 mutex_unlock(&cgroup_mutex);
892                 cgrp->id = -1;
893 
894                 call_rcu(&cgrp->rcu_head, cgroup_free_rcu);
895         } else {
896                 struct cfent *cfe = __d_cfe(dentry);
897                 struct cgroup *cgrp = dentry->d_parent->d_fsdata;
898 
899                 WARN_ONCE(!list_empty(&cfe->node) &&
900                           cgrp != &cgrp->root->top_cgroup,
901                           "cfe still linked for %s\n", cfe->type->name);
902                 simple_xattrs_free(&cfe->xattrs);
903                 kfree(cfe);
904         }
905         iput(inode);
906 }
907 
908 static void remove_dir(struct dentry *d)
909 {
910         struct dentry *parent = dget(d->d_parent);
911 
912         d_delete(d);
913         simple_rmdir(parent->d_inode, d);
914         dput(parent);
915 }
916 
917 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
918 {
919         struct cfent *cfe;
920 
921         lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
922         lockdep_assert_held(&cgroup_mutex);
923 
924         /*
925          * If we're doing cleanup due to failure of cgroup_create(),
926          * the corresponding @cfe may not exist.
927          */
928         list_for_each_entry(cfe, &cgrp->files, node) {
929                 struct dentry *d = cfe->dentry;
930 
931                 if (cft && cfe->type != cft)
932                         continue;
933 
934                 dget(d);
935                 d_delete(d);
936                 simple_unlink(cgrp->dentry->d_inode, d);
937                 list_del_init(&cfe->node);
938                 dput(d);
939 
940                 break;
941         }
942 }
943 
944 /**
945  * cgroup_clear_dir - remove subsys files in a cgroup directory
946  * @cgrp: target cgroup
947  * @subsys_mask: mask of the subsystem ids whose files should be removed
948  */
949 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
950 {
951         struct cgroup_subsys *ss;
952         int i;
953 
954         for_each_subsys(ss, i) {
955                 struct cftype_set *set;
956 
957                 if (!test_bit(i, &subsys_mask))
958                         continue;
959                 list_for_each_entry(set, &ss->cftsets, node)
960                         cgroup_addrm_files(cgrp, set->cfts, false);
961         }
962 }
963 
964 /*
965  * NOTE : the dentry must have been dget()'ed
966  */
967 static void cgroup_d_remove_dir(struct dentry *dentry)
968 {
969         struct dentry *parent;
970 
971         parent = dentry->d_parent;
972         spin_lock(&parent->d_lock);
973         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
974         list_del_init(&dentry->d_u.d_child);
975         spin_unlock(&dentry->d_lock);
976         spin_unlock(&parent->d_lock);
977         remove_dir(dentry);
978 }
979 
980 /*
981  * Call with cgroup_mutex held. Drops reference counts on modules, including
982  * any duplicate ones that parse_cgroupfs_options took. If this function
983  * returns an error, no reference counts are touched.
984  */
985 static int rebind_subsystems(struct cgroupfs_root *root,
986                              unsigned long added_mask, unsigned removed_mask)
987 {
988         struct cgroup *cgrp = &root->top_cgroup;
989         struct cgroup_subsys *ss;
990         unsigned long pinned = 0;
991         int i, ret;
992 
993         BUG_ON(!mutex_is_locked(&cgroup_mutex));
994         BUG_ON(!mutex_is_locked(&cgroup_root_mutex));
995 
996         /* Check that any added subsystems are currently free */
997         for_each_subsys(ss, i) {
998                 if (!(added_mask & (1 << i)))
999                         continue;
1000 
1001                 /* is the subsystem mounted elsewhere? */
1002                 if (ss->root != &cgroup_dummy_root) {
1003                         ret = -EBUSY;
1004                         goto out_put;
1005                 }
1006 
1007                 /* pin the module */
1008                 if (!try_module_get(ss->module)) {
1009                         ret = -ENOENT;
1010                         goto out_put;
1011                 }
1012                 pinned |= 1 << i;
1013         }
1014 
1015         /* subsys could be missing if unloaded between parsing and here */
1016         if (added_mask != pinned) {
1017                 ret = -ENOENT;
1018                 goto out_put;
1019         }
1020 
1021         ret = cgroup_populate_dir(cgrp, added_mask);
1022         if (ret)
1023                 goto out_put;
1024 
1025         /*
1026          * Nothing can fail from this point on.  Remove files for the
1027          * removed subsystems and rebind each subsystem.
1028          */
1029         cgroup_clear_dir(cgrp, removed_mask);
1030 
1031         for_each_subsys(ss, i) {
1032                 unsigned long bit = 1UL << i;
1033 
1034                 if (bit & added_mask) {
1035                         /* We're binding this subsystem to this hierarchy */
1036                         BUG_ON(cgroup_css(cgrp, ss));
1037                         BUG_ON(!cgroup_css(cgroup_dummy_top, ss));
1038                         BUG_ON(cgroup_css(cgroup_dummy_top, ss)->cgroup != cgroup_dummy_top);
1039 
1040                         rcu_assign_pointer(cgrp->subsys[i],
1041                                            cgroup_css(cgroup_dummy_top, ss));
1042                         cgroup_css(cgrp, ss)->cgroup = cgrp;
1043 
1044                         ss->root = root;
1045                         if (ss->bind)
1046                                 ss->bind(cgroup_css(cgrp, ss));
1047 
1048                         /* refcount was already taken, and we're keeping it */
1049                         root->subsys_mask |= bit;
1050                 } else if (bit & removed_mask) {
1051                         /* We're removing this subsystem */
1052                         BUG_ON(cgroup_css(cgrp, ss) != cgroup_css(cgroup_dummy_top, ss));
1053                         BUG_ON(cgroup_css(cgrp, ss)->cgroup != cgrp);
1054 
1055                         if (ss->bind)
1056                                 ss->bind(cgroup_css(cgroup_dummy_top, ss));
1057 
1058                         cgroup_css(cgroup_dummy_top, ss)->cgroup = cgroup_dummy_top;
1059                         RCU_INIT_POINTER(cgrp->subsys[i], NULL);
1060 
1061                         cgroup_subsys[i]->root = &cgroup_dummy_root;
1062 
1063                         /* subsystem is now free - drop reference on module */
1064                         module_put(ss->module);
1065                         root->subsys_mask &= ~bit;
1066                 }
1067         }
1068 
1069         /*
1070          * Mark @root has finished binding subsystems.  @root->subsys_mask
1071          * now matches the bound subsystems.
1072          */
1073         root->flags |= CGRP_ROOT_SUBSYS_BOUND;
1074 
1075         return 0;
1076 
1077 out_put:
1078         for_each_subsys(ss, i)
1079                 if (pinned & (1 << i))
1080                         module_put(ss->module);
1081         return ret;
1082 }
1083 
1084 static int cgroup_show_options(struct seq_file *seq, struct dentry *dentry)
1085 {
1086         struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
1087         struct cgroup_subsys *ss;
1088         int ssid;
1089 
1090         mutex_lock(&cgroup_root_mutex);
1091         for_each_subsys(ss, ssid)
1092                 if (root->subsys_mask & (1 << ssid))
1093                         seq_printf(seq, ",%s", ss->name);
1094         if (root->flags & CGRP_ROOT_SANE_BEHAVIOR)
1095                 seq_puts(seq, ",sane_behavior");
1096         if (root->flags & CGRP_ROOT_NOPREFIX)
1097                 seq_puts(seq, ",noprefix");
1098         if (root->flags & CGRP_ROOT_XATTR)
1099                 seq_puts(seq, ",xattr");
1100         if (strlen(root->release_agent_path))
1101                 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1102         if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags))
1103                 seq_puts(seq, ",clone_children");
1104         if (strlen(root->name))
1105                 seq_printf(seq, ",name=%s", root->name);
1106         mutex_unlock(&cgroup_root_mutex);
1107         return 0;
1108 }
1109 
1110 struct cgroup_sb_opts {
1111         unsigned long subsys_mask;
1112         unsigned long flags;
1113         char *release_agent;
1114         bool cpuset_clone_children;
1115         char *name;
1116         /* User explicitly requested empty subsystem */
1117         bool none;
1118 
1119         struct cgroupfs_root *new_root;
1120 
1121 };
1122 
1123 /*
1124  * Convert a hierarchy specifier into a bitmask of subsystems and
1125  * flags. Call with cgroup_mutex held to protect the cgroup_subsys[]
1126  * array. This function takes refcounts on subsystems to be used, unless it
1127  * returns error, in which case no refcounts are taken.
1128  */
1129 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1130 {
1131         char *token, *o = data;
1132         bool all_ss = false, one_ss = false;
1133         unsigned long mask = (unsigned long)-1;
1134         struct cgroup_subsys *ss;
1135         int i;
1136 
1137         BUG_ON(!mutex_is_locked(&cgroup_mutex));
1138 
1139 #ifdef CONFIG_CPUSETS
1140         mask = ~(1UL << cpuset_subsys_id);
1141 #endif
1142 
1143         memset(opts, 0, sizeof(*opts));
1144 
1145         while ((token = strsep(&o, ",")) != NULL) {
1146                 if (!*token)
1147                         return -EINVAL;
1148                 if (!strcmp(token, "none")) {
1149                         /* Explicitly have no subsystems */
1150                         opts->none = true;
1151                         continue;
1152                 }
1153                 if (!strcmp(token, "all")) {
1154                         /* Mutually exclusive option 'all' + subsystem name */
1155                         if (one_ss)
1156                                 return -EINVAL;
1157                         all_ss = true;
1158                         continue;
1159                 }
1160                 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1161                         opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1162                         continue;
1163                 }
1164                 if (!strcmp(token, "noprefix")) {
1165                         opts->flags |= CGRP_ROOT_NOPREFIX;
1166                         continue;
1167                 }
1168                 if (!strcmp(token, "clone_children")) {
1169                         opts->cpuset_clone_children = true;
1170                         continue;
1171                 }
1172                 if (!strcmp(token, "xattr")) {
1173                         opts->flags |= CGRP_ROOT_XATTR;
1174                         continue;
1175                 }
1176                 if (!strncmp(token, "release_agent=", 14)) {
1177                         /* Specifying two release agents is forbidden */
1178                         if (opts->release_agent)
1179                                 return -EINVAL;
1180                         opts->release_agent =
1181                                 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1182                         if (!opts->release_agent)
1183                                 return -ENOMEM;
1184                         continue;
1185                 }
1186                 if (!strncmp(token, "name=", 5)) {
1187                         const char *name = token + 5;
1188                         /* Can't specify an empty name */
1189                         if (!strlen(name))
1190                                 return -EINVAL;
1191                         /* Must match [\w.-]+ */
1192                         for (i = 0; i < strlen(name); i++) {
1193                                 char c = name[i];
1194                                 if (isalnum(c))
1195                                         continue;
1196                                 if ((c == '.') || (c == '-') || (c == '_'))
1197                                         continue;
1198                                 return -EINVAL;
1199                         }
1200                         /* Specifying two names is forbidden */
1201                         if (opts->name)
1202                                 return -EINVAL;
1203                         opts->name = kstrndup(name,
1204                                               MAX_CGROUP_ROOT_NAMELEN - 1,
1205                                               GFP_KERNEL);
1206                         if (!opts->name)
1207                                 return -ENOMEM;
1208 
1209                         continue;
1210                 }
1211 
1212                 for_each_subsys(ss, i) {
1213                         if (strcmp(token, ss->name))
1214                                 continue;
1215                         if (ss->disabled)
1216                                 continue;
1217 
1218                         /* Mutually exclusive option 'all' + subsystem name */
1219                         if (all_ss)
1220                                 return -EINVAL;
1221                         set_bit(i, &opts->subsys_mask);
1222                         one_ss = true;
1223 
1224                         break;
1225                 }
1226                 if (i == CGROUP_SUBSYS_COUNT)
1227                         return -ENOENT;
1228         }
1229 
1230         /*
1231          * If the 'all' option was specified select all the subsystems,
1232          * otherwise if 'none', 'name=' and a subsystem name options
1233          * were not specified, let's default to 'all'
1234          */
1235         if (all_ss || (!one_ss && !opts->none && !opts->name))
1236                 for_each_subsys(ss, i)
1237                         if (!ss->disabled)
1238                                 set_bit(i, &opts->subsys_mask);
1239 
1240         /* Consistency checks */
1241 
1242         if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1243                 pr_warning("cgroup: sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1244 
1245                 if (opts->flags & CGRP_ROOT_NOPREFIX) {
1246                         pr_err("cgroup: sane_behavior: noprefix is not allowed\n");
1247                         return -EINVAL;
1248                 }
1249 
1250                 if (opts->cpuset_clone_children) {
1251                         pr_err("cgroup: sane_behavior: clone_children is not allowed\n");
1252                         return -EINVAL;
1253                 }
1254         }
1255 
1256         /*
1257          * Option noprefix was introduced just for backward compatibility
1258          * with the old cpuset, so we allow noprefix only if mounting just
1259          * the cpuset subsystem.
1260          */
1261         if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1262                 return -EINVAL;
1263 
1264 
1265         /* Can't specify "none" and some subsystems */
1266         if (opts->subsys_mask && opts->none)
1267                 return -EINVAL;
1268 
1269         /*
1270          * We either have to specify by name or by subsystems. (So all
1271          * empty hierarchies must have a name).
1272          */
1273         if (!opts->subsys_mask && !opts->name)
1274                 return -EINVAL;
1275 
1276         return 0;
1277 }
1278 
1279 static int cgroup_remount(struct super_block *sb, int *flags, char *data)
1280 {
1281         int ret = 0;
1282         struct cgroupfs_root *root = sb->s_fs_info;
1283         struct cgroup *cgrp = &root->top_cgroup;
1284         struct cgroup_sb_opts opts;
1285         unsigned long added_mask, removed_mask;
1286 
1287         if (root->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1288                 pr_err("cgroup: sane_behavior: remount is not allowed\n");
1289                 return -EINVAL;
1290         }
1291 
1292         mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1293         mutex_lock(&cgroup_mutex);
1294         mutex_lock(&cgroup_root_mutex);
1295 
1296         /* See what subsystems are wanted */
1297         ret = parse_cgroupfs_options(data, &opts);
1298         if (ret)
1299                 goto out_unlock;
1300 
1301         if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1302                 pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
1303                            task_tgid_nr(current), current->comm);
1304 
1305         added_mask = opts.subsys_mask & ~root->subsys_mask;
1306         removed_mask = root->subsys_mask & ~opts.subsys_mask;
1307 
1308         /* Don't allow flags or name to change at remount */
1309         if (((opts.flags ^ root->flags) & CGRP_ROOT_OPTION_MASK) ||
1310             (opts.name && strcmp(opts.name, root->name))) {
1311                 pr_err("cgroup: option or name mismatch, new: 0x%lx \"%s\", old: 0x%lx \"%s\"\n",
1312                        opts.flags & CGRP_ROOT_OPTION_MASK, opts.name ?: "",
1313                        root->flags & CGRP_ROOT_OPTION_MASK, root->name);
1314                 ret = -EINVAL;
1315                 goto out_unlock;
1316         }
1317 
1318         /* remounting is not allowed for populated hierarchies */
1319         if (root->number_of_cgroups > 1) {
1320                 ret = -EBUSY;
1321                 goto out_unlock;
1322         }
1323 
1324         ret = rebind_subsystems(root, added_mask, removed_mask);
1325         if (ret)
1326                 goto out_unlock;
1327 
1328         if (opts.release_agent)
1329                 strcpy(root->release_agent_path, opts.release_agent);
1330  out_unlock:
1331         kfree(opts.release_agent);
1332         kfree(opts.name);
1333         mutex_unlock(&cgroup_root_mutex);
1334         mutex_unlock(&cgroup_mutex);
1335         mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1336         return ret;
1337 }
1338 
1339 static const struct super_operations cgroup_ops = {
1340         .statfs = simple_statfs,
1341         .drop_inode = generic_delete_inode,
1342         .show_options = cgroup_show_options,
1343         .remount_fs = cgroup_remount,
1344 };
1345 
1346 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1347 {
1348         INIT_LIST_HEAD(&cgrp->sibling);
1349         INIT_LIST_HEAD(&cgrp->children);
1350         INIT_LIST_HEAD(&cgrp->files);
1351         INIT_LIST_HEAD(&cgrp->cset_links);
1352         INIT_LIST_HEAD(&cgrp->release_list);
1353         INIT_LIST_HEAD(&cgrp->pidlists);
1354         mutex_init(&cgrp->pidlist_mutex);
1355         cgrp->dummy_css.cgroup = cgrp;
1356         simple_xattrs_init(&cgrp->xattrs);
1357 }
1358 
1359 static void init_cgroup_root(struct cgroupfs_root *root)
1360 {
1361         struct cgroup *cgrp = &root->top_cgroup;
1362 
1363         INIT_LIST_HEAD(&root->root_list);
1364         root->number_of_cgroups = 1;
1365         cgrp->root = root;
1366         RCU_INIT_POINTER(cgrp->name, &root_cgroup_name);
1367         init_cgroup_housekeeping(cgrp);
1368         idr_init(&root->cgroup_idr);
1369 }
1370 
1371 static int cgroup_init_root_id(struct cgroupfs_root *root, int start, int end)
1372 {
1373         int id;
1374 
1375         lockdep_assert_held(&cgroup_mutex);
1376         lockdep_assert_held(&cgroup_root_mutex);
1377 
1378         id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, start, end,
1379                               GFP_KERNEL);
1380         if (id < 0)
1381                 return id;
1382 
1383         root->hierarchy_id = id;
1384         return 0;
1385 }
1386 
1387 static void cgroup_exit_root_id(struct cgroupfs_root *root)
1388 {
1389         lockdep_assert_held(&cgroup_mutex);
1390         lockdep_assert_held(&cgroup_root_mutex);
1391 
1392         if (root->hierarchy_id) {
1393                 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1394                 root->hierarchy_id = 0;
1395         }
1396 }
1397 
1398 static int cgroup_test_super(struct super_block *sb, void *data)
1399 {
1400         struct cgroup_sb_opts *opts = data;
1401         struct cgroupfs_root *root = sb->s_fs_info;
1402 
1403         /* If we asked for a name then it must match */
1404         if (opts->name && strcmp(opts->name, root->name))
1405                 return 0;
1406 
1407         /*
1408          * If we asked for subsystems (or explicitly for no
1409          * subsystems) then they must match
1410          */
1411         if ((opts->subsys_mask || opts->none)
1412             && (opts->subsys_mask != root->subsys_mask))
1413                 return 0;
1414 
1415         return 1;
1416 }
1417 
1418 static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
1419 {
1420         struct cgroupfs_root *root;
1421 
1422         if (!opts->subsys_mask && !opts->none)
1423                 return NULL;
1424 
1425         root = kzalloc(sizeof(*root), GFP_KERNEL);
1426         if (!root)
1427                 return ERR_PTR(-ENOMEM);
1428 
1429         init_cgroup_root(root);
1430 
1431         /*
1432          * We need to set @root->subsys_mask now so that @root can be
1433          * matched by cgroup_test_super() before it finishes
1434          * initialization; otherwise, competing mounts with the same
1435          * options may try to bind the same subsystems instead of waiting
1436          * for the first one leading to unexpected mount errors.
1437          * SUBSYS_BOUND will be set once actual binding is complete.
1438          */
1439         root->subsys_mask = opts->subsys_mask;
1440         root->flags = opts->flags;
1441         if (opts->release_agent)
1442                 strcpy(root->release_agent_path, opts->release_agent);
1443         if (opts->name)
1444                 strcpy(root->name, opts->name);
1445         if (opts->cpuset_clone_children)
1446                 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->top_cgroup.flags);
1447         return root;
1448 }
1449 
1450 static void cgroup_free_root(struct cgroupfs_root *root)
1451 {
1452         if (root) {
1453                 /* hierarhcy ID shoulid already have been released */
1454                 WARN_ON_ONCE(root->hierarchy_id);
1455 
1456                 idr_destroy(&root->cgroup_idr);
1457                 kfree(root);
1458         }
1459 }
1460 
1461 static int cgroup_set_super(struct super_block *sb, void *data)
1462 {
1463         int ret;
1464         struct cgroup_sb_opts *opts = data;
1465 
1466         /* If we don't have a new root, we can't set up a new sb */
1467         if (!opts->new_root)
1468                 return -EINVAL;
1469 
1470         BUG_ON(!opts->subsys_mask && !opts->none);
1471 
1472         ret = set_anon_super(sb, NULL);
1473         if (ret)
1474                 return ret;
1475 
1476         sb->s_fs_info = opts->new_root;
1477         opts->new_root->sb = sb;
1478 
1479         sb->s_blocksize = PAGE_CACHE_SIZE;
1480         sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
1481         sb->s_magic = CGROUP_SUPER_MAGIC;
1482         sb->s_op = &cgroup_ops;
1483 
1484         return 0;
1485 }
1486 
1487 static int cgroup_get_rootdir(struct super_block *sb)
1488 {
1489         static const struct dentry_operations cgroup_dops = {
1490                 .d_iput = cgroup_diput,
1491                 .d_delete = always_delete_dentry,
1492         };
1493 
1494         struct inode *inode =
1495                 cgroup_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR, sb);
1496 
1497         if (!inode)
1498                 return -ENOMEM;
1499 
1500         inode->i_fop = &simple_dir_operations;
1501         inode->i_op = &cgroup_dir_inode_operations;
1502         /* directories start off with i_nlink == 2 (for "." entry) */
1503         inc_nlink(inode);
1504         sb->s_root = d_make_root(inode);
1505         if (!sb->s_root)
1506                 return -ENOMEM;
1507         /* for everything else we want ->d_op set */
1508         sb->s_d_op = &cgroup_dops;
1509         return 0;
1510 }
1511 
1512 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1513                          int flags, const char *unused_dev_name,
1514                          void *data)
1515 {
1516         struct cgroup_sb_opts opts;
1517         struct cgroupfs_root *root;
1518         int ret = 0;
1519         struct super_block *sb;
1520         struct cgroupfs_root *new_root;
1521         struct list_head tmp_links;
1522         struct inode *inode;
1523         const struct cred *cred;
1524 
1525         /* First find the desired set of subsystems */
1526         mutex_lock(&cgroup_mutex);
1527         ret = parse_cgroupfs_options(data, &opts);
1528         mutex_unlock(&cgroup_mutex);
1529         if (ret)
1530                 goto out_err;
1531 
1532         /*
1533          * Allocate a new cgroup root. We may not need it if we're
1534          * reusing an existing hierarchy.
1535          */
1536         new_root = cgroup_root_from_opts(&opts);
1537         if (IS_ERR(new_root)) {
1538                 ret = PTR_ERR(new_root);
1539                 goto out_err;
1540         }
1541         opts.new_root = new_root;
1542 
1543         /* Locate an existing or new sb for this hierarchy */
1544         sb = sget(fs_type, cgroup_test_super, cgroup_set_super, 0, &opts);
1545         if (IS_ERR(sb)) {
1546                 ret = PTR_ERR(sb);
1547                 cgroup_free_root(opts.new_root);
1548                 goto out_err;
1549         }
1550 
1551         root = sb->s_fs_info;
1552         BUG_ON(!root);
1553         if (root == opts.new_root) {
1554                 /* We used the new root structure, so this is a new hierarchy */
1555                 struct cgroup *root_cgrp = &root->top_cgroup;
1556                 struct cgroupfs_root *existing_root;
1557                 int i;
1558                 struct css_set *cset;
1559 
1560                 BUG_ON(sb->s_root != NULL);
1561 
1562                 ret = cgroup_get_rootdir(sb);
1563                 if (ret)
1564                         goto drop_new_super;
1565                 inode = sb->s_root->d_inode;
1566 
1567                 mutex_lock(&inode->i_mutex);
1568                 mutex_lock(&cgroup_mutex);
1569                 mutex_lock(&cgroup_root_mutex);
1570 
1571                 ret = idr_alloc(&root->cgroup_idr, root_cgrp, 0, 1, GFP_KERNEL);
1572                 if (ret < 0)
1573                         goto unlock_drop;
1574                 root_cgrp->id = ret;
1575 
1576                 /* Check for name clashes with existing mounts */
1577                 ret = -EBUSY;
1578                 if (strlen(root->name))
1579                         for_each_active_root(existing_root)
1580                                 if (!strcmp(existing_root->name, root->name))
1581                                         goto unlock_drop;
1582 
1583                 /*
1584                  * We're accessing css_set_count without locking
1585                  * css_set_lock here, but that's OK - it can only be
1586                  * increased by someone holding cgroup_lock, and
1587                  * that's us. The worst that can happen is that we
1588                  * have some link structures left over
1589                  */
1590                 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1591                 if (ret)
1592                         goto unlock_drop;
1593 
1594                 /* ID 0 is reserved for dummy root, 1 for unified hierarchy */
1595                 ret = cgroup_init_root_id(root, 2, 0);
1596                 if (ret)
1597                         goto unlock_drop;
1598 
1599                 sb->s_root->d_fsdata = root_cgrp;
1600                 root_cgrp->dentry = sb->s_root;
1601 
1602                 /*
1603                  * We're inside get_sb() and will call lookup_one_len() to
1604                  * create the root files, which doesn't work if SELinux is
1605                  * in use.  The following cred dancing somehow works around
1606                  * it.  See 2ce9738ba ("cgroupfs: use init_cred when
1607                  * populating new cgroupfs mount") for more details.
1608                  */
1609                 cred = override_creds(&init_cred);
1610 
1611                 ret = cgroup_addrm_files(root_cgrp, cgroup_base_files, true);
1612                 if (ret)
1613                         goto rm_base_files;
1614 
1615                 ret = rebind_subsystems(root, root->subsys_mask, 0);
1616                 if (ret)
1617                         goto rm_base_files;
1618 
1619                 revert_creds(cred);
1620 
1621                 /*
1622                  * There must be no failure case after here, since rebinding
1623                  * takes care of subsystems' refcounts, which are explicitly
1624                  * dropped in the failure exit path.
1625                  */
1626 
1627                 list_add(&root->root_list, &cgroup_roots);
1628                 cgroup_root_count++;
1629 
1630                 /* Link the top cgroup in this hierarchy into all
1631                  * the css_set objects */
1632                 write_lock(&css_set_lock);
1633                 hash_for_each(css_set_table, i, cset, hlist)
1634                         link_css_set(&tmp_links, cset, root_cgrp);
1635                 write_unlock(&css_set_lock);
1636 
1637                 free_cgrp_cset_links(&tmp_links);
1638 
1639                 BUG_ON(!list_empty(&root_cgrp->children));
1640                 BUG_ON(root->number_of_cgroups != 1);
1641 
1642                 mutex_unlock(&cgroup_root_mutex);
1643                 mutex_unlock(&cgroup_mutex);
1644                 mutex_unlock(&inode->i_mutex);
1645         } else {
1646                 /*
1647                  * We re-used an existing hierarchy - the new root (if
1648                  * any) is not needed
1649                  */
1650                 cgroup_free_root(opts.new_root);
1651 
1652                 if ((root->flags ^ opts.flags) & CGRP_ROOT_OPTION_MASK) {
1653                         if ((root->flags | opts.flags) & CGRP_ROOT_SANE_BEHAVIOR) {
1654                                 pr_err("cgroup: sane_behavior: new mount options should match the existing superblock\n");
1655                                 ret = -EINVAL;
1656                                 goto drop_new_super;
1657                         } else {
1658                                 pr_warning("cgroup: new mount options do not match the existing superblock, will be ignored\n");
1659                         }
1660                 }
1661         }
1662 
1663         kfree(opts.release_agent);
1664         kfree(opts.name);
1665         return dget(sb->s_root);
1666 
1667  rm_base_files:
1668         free_cgrp_cset_links(&tmp_links);
1669         cgroup_addrm_files(&root->top_cgroup, cgroup_base_files, false);
1670         revert_creds(cred);
1671  unlock_drop:
1672         cgroup_exit_root_id(root);
1673         mutex_unlock(&cgroup_root_mutex);
1674         mutex_unlock(&cgroup_mutex);
1675         mutex_unlock(&inode->i_mutex);
1676  drop_new_super:
1677         deactivate_locked_super(sb);
1678  out_err:
1679         kfree(opts.release_agent);
1680         kfree(opts.name);
1681         return ERR_PTR(ret);
1682 }
1683 
1684 static void cgroup_kill_sb(struct super_block *sb)
1685 {
1686         struct cgroupfs_root *root = sb->s_fs_info;
1687         struct cgroup *cgrp = &root->top_cgroup;
1688         struct cgrp_cset_link *link, *tmp_link;
1689         int ret;
1690 
1691         BUG_ON(!root);
1692 
1693         BUG_ON(root->number_of_cgroups != 1);
1694         BUG_ON(!list_empty(&cgrp->children));
1695 
1696         mutex_lock(&cgrp->dentry->d_inode->i_mutex);
1697         mutex_lock(&cgroup_mutex);
1698         mutex_lock(&cgroup_root_mutex);
1699 
1700         /* Rebind all subsystems back to the default hierarchy */
1701         if (root->flags & CGRP_ROOT_SUBSYS_BOUND) {
1702                 ret = rebind_subsystems(root, 0, root->subsys_mask);
1703                 /* Shouldn't be able to fail ... */
1704                 BUG_ON(ret);
1705         }
1706 
1707         /*
1708          * Release all the links from cset_links to this hierarchy's
1709          * root cgroup
1710          */
1711         write_lock(&css_set_lock);
1712 
1713         list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1714                 list_del(&link->cset_link);
1715                 list_del(&link->cgrp_link);
1716                 kfree(link);
1717         }
1718         write_unlock(&css_set_lock);
1719 
1720         if (!list_empty(&root->root_list)) {
1721                 list_del(&root->root_list);
1722                 cgroup_root_count--;
1723         }
1724 
1725         cgroup_exit_root_id(root);
1726 
1727         mutex_unlock(&cgroup_root_mutex);
1728         mutex_unlock(&cgroup_mutex);
1729         mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
1730 
1731         simple_xattrs_free(&cgrp->xattrs);
1732 
1733         kill_litter_super(sb);
1734         cgroup_free_root(root);
1735 }
1736 
1737 static struct file_system_type cgroup_fs_type = {
1738         .name = "cgroup",
1739         .mount = cgroup_mount,
1740         .kill_sb = cgroup_kill_sb,
1741 };
1742 
1743 static struct kobject *cgroup_kobj;
1744 
1745 /**
1746  * cgroup_path - generate the path of a cgroup
1747  * @cgrp: the cgroup in question
1748  * @buf: the buffer to write the path into
1749  * @buflen: the length of the buffer
1750  *
1751  * Writes path of cgroup into buf.  Returns 0 on success, -errno on error.
1752  *
1753  * We can't generate cgroup path using dentry->d_name, as accessing
1754  * dentry->name must be protected by irq-unsafe dentry->d_lock or parent
1755  * inode's i_mutex, while on the other hand cgroup_path() can be called
1756  * with some irq-safe spinlocks held.
1757  */
1758 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
1759 {
1760         int ret = -ENAMETOOLONG;
1761         char *start;
1762 
1763         if (!cgrp->parent) {
1764                 if (strlcpy(buf, "/", buflen) >= buflen)
1765                         return -ENAMETOOLONG;
1766                 return 0;
1767         }
1768 
1769         start = buf + buflen - 1;
1770         *start = '\0';
1771 
1772         rcu_read_lock();
1773         do {
1774                 const char *name = cgroup_name(cgrp);
1775                 int len;
1776 
1777                 len = strlen(name);
1778                 if ((start -= len) < buf)
1779                         goto out;
1780                 memcpy(start, name, len);
1781 
1782                 if (--start < buf)
1783                         goto out;
1784                 *start = '/';
1785 
1786                 cgrp = cgrp->parent;
1787         } while (cgrp->parent);
1788         ret = 0;
1789         memmove(buf, start, buf + buflen - start);
1790 out:
1791         rcu_read_unlock();
1792         return ret;
1793 }
1794 EXPORT_SYMBOL_GPL(cgroup_path);
1795 
1796 /**
1797  * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1798  * @task: target task
1799  * @buf: the buffer to write the path into
1800  * @buflen: the length of the buffer
1801  *
1802  * Determine @task's cgroup on the first (the one with the lowest non-zero
1803  * hierarchy_id) cgroup hierarchy and copy its path into @buf.  This
1804  * function grabs cgroup_mutex and shouldn't be used inside locks used by
1805  * cgroup controller callbacks.
1806  *
1807  * Returns 0 on success, fails with -%ENAMETOOLONG if @buflen is too short.
1808  */
1809 int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1810 {
1811         struct cgroupfs_root *root;
1812         struct cgroup *cgrp;
1813         int hierarchy_id = 1, ret = 0;
1814 
1815         if (buflen < 2)
1816                 return -ENAMETOOLONG;
1817 
1818         mutex_lock(&cgroup_mutex);
1819 
1820         root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1821 
1822         if (root) {
1823                 cgrp = task_cgroup_from_root(task, root);
1824                 ret = cgroup_path(cgrp, buf, buflen);
1825         } else {
1826                 /* if no hierarchy exists, everyone is in "/" */
1827                 memcpy(buf, "/", 2);
1828         }
1829 
1830         mutex_unlock(&cgroup_mutex);
1831         return ret;
1832 }
1833 EXPORT_SYMBOL_GPL(task_cgroup_path);
1834 
1835 /*
1836  * Control Group taskset
1837  */
1838 struct task_and_cgroup {
1839         struct task_struct      *task;
1840         struct cgroup           *cgrp;
1841         struct css_set          *cset;
1842 };
1843 
1844 struct cgroup_taskset {
1845         struct task_and_cgroup  single;
1846         struct flex_array       *tc_array;
1847         int                     tc_array_len;
1848         int                     idx;
1849         struct cgroup           *cur_cgrp;
1850 };
1851 
1852 /**
1853  * cgroup_taskset_first - reset taskset and return the first task
1854  * @tset: taskset of interest
1855  *
1856  * @tset iteration is initialized and the first task is returned.
1857  */
1858 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1859 {
1860         if (tset->tc_array) {
1861                 tset->idx = 0;
1862                 return cgroup_taskset_next(tset);
1863         } else {
1864                 tset->cur_cgrp = tset->single.cgrp;
1865                 return tset->single.task;
1866         }
1867 }
1868 EXPORT_SYMBOL_GPL(cgroup_taskset_first);
1869 
1870 /**
1871  * cgroup_taskset_next - iterate to the next task in taskset
1872  * @tset: taskset of interest
1873  *
1874  * Return the next task in @tset.  Iteration must have been initialized
1875  * with cgroup_taskset_first().
1876  */
1877 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
1878 {
1879         struct task_and_cgroup *tc;
1880 
1881         if (!tset->tc_array || tset->idx >= tset->tc_array_len)
1882                 return NULL;
1883 
1884         tc = flex_array_get(tset->tc_array, tset->idx++);
1885         tset->cur_cgrp = tc->cgrp;
1886         return tc->task;
1887 }
1888 EXPORT_SYMBOL_GPL(cgroup_taskset_next);
1889 
1890 /**
1891  * cgroup_taskset_cur_css - return the matching css for the current task
1892  * @tset: taskset of interest
1893  * @subsys_id: the ID of the target subsystem
1894  *
1895  * Return the css for the current (last returned) task of @tset for
1896  * subsystem specified by @subsys_id.  This function must be preceded by
1897  * either cgroup_taskset_first() or cgroup_taskset_next().
1898  */
1899 struct cgroup_subsys_state *cgroup_taskset_cur_css(struct cgroup_taskset *tset,
1900                                                    int subsys_id)
1901 {
1902         return cgroup_css(tset->cur_cgrp, cgroup_subsys[subsys_id]);
1903 }
1904 EXPORT_SYMBOL_GPL(cgroup_taskset_cur_css);
1905 
1906 /**
1907  * cgroup_taskset_size - return the number of tasks in taskset
1908  * @tset: taskset of interest
1909  */
1910 int cgroup_taskset_size(struct cgroup_taskset *tset)
1911 {
1912         return tset->tc_array ? tset->tc_array_len : 1;
1913 }
1914 EXPORT_SYMBOL_GPL(cgroup_taskset_size);
1915 
1916 
1917 /*
1918  * cgroup_task_migrate - move a task from one cgroup to another.
1919  *
1920  * Must be called with cgroup_mutex and threadgroup locked.
1921  */
1922 static void cgroup_task_migrate(struct cgroup *old_cgrp,
1923                                 struct task_struct *tsk,
1924                                 struct css_set *new_cset)
1925 {
1926         struct css_set *old_cset;
1927 
1928         /*
1929          * We are synchronized through threadgroup_lock() against PF_EXITING
1930          * setting such that we can't race against cgroup_exit() changing the
1931          * css_set to init_css_set and dropping the old one.
1932          */
1933         WARN_ON_ONCE(tsk->flags & PF_EXITING);
1934         old_cset = task_css_set(tsk);
1935 
1936         task_lock(tsk);
1937         rcu_assign_pointer(tsk->cgroups, new_cset);
1938         task_unlock(tsk);
1939 
1940         /* Update the css_set linked lists if we're using them */
1941         write_lock(&css_set_lock);
1942         if (!list_empty(&tsk->cg_list))
1943                 list_move(&tsk->cg_list, &new_cset->tasks);
1944         write_unlock(&css_set_lock);
1945 
1946         /*
1947          * We just gained a reference on old_cset by taking it from the
1948          * task. As trading it for new_cset is protected by cgroup_mutex,
1949          * we're safe to drop it here; it will be freed under RCU.
1950          */
1951         set_bit(CGRP_RELEASABLE, &old_cgrp->flags);
1952         put_css_set(old_cset);
1953 }
1954 
1955 /**
1956  * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
1957  * @cgrp: the cgroup to attach to
1958  * @tsk: the task or the leader of the threadgroup to be attached
1959  * @threadgroup: attach the whole threadgroup?
1960  *
1961  * Call holding cgroup_mutex and the group_rwsem of the leader. Will take
1962  * task_lock of @tsk or each thread in the threadgroup individually in turn.
1963  */
1964 static int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk,
1965                               bool threadgroup)
1966 {
1967         int retval, i, group_size;
1968         struct cgroupfs_root *root = cgrp->root;
1969         struct cgroup_subsys_state *css, *failed_css = NULL;
1970         /* threadgroup list cursor and array */
1971         struct task_struct *leader = tsk;
1972         struct task_and_cgroup *tc;
1973         struct flex_array *group;
1974         struct cgroup_taskset tset = { };
1975 
1976         /*
1977          * step 0: in order to do expensive, possibly blocking operations for
1978          * every thread, we cannot iterate the thread group list, since it needs
1979          * rcu or tasklist locked. instead, build an array of all threads in the
1980          * group - group_rwsem prevents new threads from appearing, and if
1981          * threads exit, this will just be an over-estimate.
1982          */
1983         if (threadgroup)
1984                 group_size = get_nr_threads(tsk);
1985         else
1986                 group_size = 1;
1987         /* flex_array supports very large thread-groups better than kmalloc. */
1988         group = flex_array_alloc(sizeof(*tc), group_size, GFP_KERNEL);
1989         if (!group)
1990                 return -ENOMEM;
1991         /* pre-allocate to guarantee space while iterating in rcu read-side. */
1992         retval = flex_array_prealloc(group, 0, group_size, GFP_KERNEL);
1993         if (retval)
1994                 goto out_free_group_list;
1995 
1996         i = 0;
1997         /*
1998          * Prevent freeing of tasks while we take a snapshot. Tasks that are
1999          * already PF_EXITING could be freed from underneath us unless we
2000          * take an rcu_read_lock.
2001          */
2002         rcu_read_lock();
2003         do {
2004                 struct task_and_cgroup ent;
2005 
2006                 /* @tsk either already exited or can't exit until the end */
2007                 if (tsk->flags & PF_EXITING)
2008                         goto next;
2009 
2010                 /* as per above, nr_threads may decrease, but not increase. */
2011                 BUG_ON(i >= group_size);
2012                 ent.task = tsk;
2013                 ent.cgrp = task_cgroup_from_root(tsk, root);
2014                 /* nothing to do if this task is already in the cgroup */
2015                 if (ent.cgrp == cgrp)
2016                         goto next;
2017                 /*
2018                  * saying GFP_ATOMIC has no effect here because we did prealloc
2019                  * earlier, but it's good form to communicate our expectations.
2020                  */
2021                 retval = flex_array_put(group, i, &ent, GFP_ATOMIC);
2022                 BUG_ON(retval != 0);
2023                 i++;
2024         next:
2025                 if (!threadgroup)
2026                         break;
2027         } while_each_thread(leader, tsk);
2028         rcu_read_unlock();
2029         /* remember the number of threads in the array for later. */
2030         group_size = i;
2031         tset.tc_array = group;
2032         tset.tc_array_len = group_size;
2033 
2034         /* methods shouldn't be called if no task is actually migrating */
2035         retval = 0;
2036         if (!group_size)
2037                 goto out_free_group_list;
2038 
2039         /*
2040          * step 1: check that we can legitimately attach to the cgroup.
2041          */
2042         for_each_css(css, i, cgrp) {
2043                 if (css->ss->can_attach) {
2044                         retval = css->ss->can_attach(css, &tset);
2045                         if (retval) {
2046                                 failed_css = css;
2047                                 goto out_cancel_attach;
2048                         }
2049                 }
2050         }
2051 
2052         /*
2053          * step 2: make sure css_sets exist for all threads to be migrated.
2054          * we use find_css_set, which allocates a new one if necessary.
2055          */
2056         for (i = 0; i < group_size; i++) {
2057                 struct css_set *old_cset;
2058 
2059                 tc = flex_array_get(group, i);
2060                 old_cset = task_css_set(tc->task);
2061                 tc->cset = find_css_set(old_cset, cgrp);
2062                 if (!tc->cset) {
2063                         retval = -ENOMEM;
2064                         goto out_put_css_set_refs;
2065                 }
2066         }
2067 
2068         /*
2069          * step 3: now that we're guaranteed success wrt the css_sets,
2070          * proceed to move all tasks to the new cgroup.  There are no
2071          * failure cases after here, so this is the commit point.
2072          */
2073         for (i = 0; i < group_size; i++) {
2074                 tc = flex_array_get(group, i);
2075                 cgroup_task_migrate(tc->cgrp, tc->task, tc->cset);
2076         }
2077         /* nothing is sensitive to fork() after this point. */
2078 
2079         /*
2080          * step 4: do subsystem attach callbacks.
2081          */
2082         for_each_css(css, i, cgrp)
2083                 if (css->ss->attach)
2084                         css->ss->attach(css, &tset);
2085 
2086         /*
2087          * step 5: success! and cleanup
2088          */
2089         retval = 0;
2090 out_put_css_set_refs:
2091         if (retval) {
2092                 for (i = 0; i < group_size; i++) {
2093                         tc = flex_array_get(group, i);
2094                         if (!tc->cset)
2095                                 break;
2096                         put_css_set(tc->cset);
2097                 }
2098         }
2099 out_cancel_attach:
2100         if (retval) {
2101                 for_each_css(css, i, cgrp) {
2102                         if (css == failed_css)
2103                                 break;
2104                         if (css->ss->cancel_attach)
2105                                 css->ss->cancel_attach(css, &tset);
2106                 }
2107         }
2108 out_free_group_list:
2109         flex_array_free(group);
2110         return retval;
2111 }
2112 
2113 /*
2114  * Find the task_struct of the task to attach by vpid and pass it along to the
2115  * function to attach either it or all tasks in its threadgroup. Will lock
2116  * cgroup_mutex and threadgroup; may take task_lock of task.
2117  */
2118 static int attach_task_by_pid(struct cgroup *cgrp, u64 pid, bool threadgroup)
2119 {
2120         struct task_struct *tsk;
2121         const struct cred *cred = current_cred(), *tcred;
2122         int ret;
2123 
2124         if (!cgroup_lock_live_group(cgrp))
2125                 return -ENODEV;
2126 
2127 retry_find_task:
2128         rcu_read_lock();
2129         if (pid) {
2130                 tsk = find_task_by_vpid(pid);
2131                 if (!tsk) {
2132                         rcu_read_unlock();
2133                         ret = -ESRCH;
2134                         goto out_unlock_cgroup;
2135                 }
2136                 /*
2137                  * even if we're attaching all tasks in the thread group, we
2138                  * only need to check permissions on one of them.
2139                  */
2140                 tcred = __task_cred(tsk);
2141                 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2142                     !uid_eq(cred->euid, tcred->uid) &&
2143                     !uid_eq(cred->euid, tcred->suid)) {
2144                         rcu_read_unlock();
2145                         ret = -EACCES;
2146                         goto out_unlock_cgroup;
2147                 }
2148         } else
2149                 tsk = current;
2150 
2151         if (threadgroup)
2152                 tsk = tsk->group_leader;
2153 
2154         /*
2155          * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2156          * trapped in a cpuset, or RT worker may be born in a cgroup
2157          * with no rt_runtime allocated.  Just say no.
2158          */
2159         if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2160                 ret = -EINVAL;
2161                 rcu_read_unlock();
2162                 goto out_unlock_cgroup;
2163         }
2164 
2165         get_task_struct(tsk);
2166         rcu_read_unlock();
2167 
2168         threadgroup_lock(tsk);
2169         if (threadgroup) {
2170                 if (!thread_group_leader(tsk)) {
2171                         /*
2172                          * a race with de_thread from another thread's exec()
2173                          * may strip us of our leadership, if this happens,
2174                          * there is no choice but to throw this task away and
2175                          * try again; this is
2176                          * "double-double-toil-and-trouble-check locking".
2177                          */
2178                         threadgroup_unlock(tsk);
2179                         put_task_struct(tsk);
2180                         goto retry_find_task;
2181                 }
2182         }
2183 
2184         ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2185 
2186         threadgroup_unlock(tsk);
2187 
2188         put_task_struct(tsk);
2189 out_unlock_cgroup:
2190         mutex_unlock(&cgroup_mutex);
2191         return ret;
2192 }
2193 
2194 /**
2195  * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2196  * @from: attach to all cgroups of a given task
2197  * @tsk: the task to be attached
2198  */
2199 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2200 {
2201         struct cgroupfs_root *root;
2202         int retval = 0;
2203 
2204         mutex_lock(&cgroup_mutex);
2205         for_each_active_root(root) {
2206                 struct cgroup *from_cgrp = task_cgroup_from_root(from, root);
2207 
2208                 retval = cgroup_attach_task(from_cgrp, tsk, false);
2209                 if (retval)
2210                         break;
2211         }
2212         mutex_unlock(&cgroup_mutex);
2213 
2214         return retval;
2215 }
2216 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2217 
2218 static int cgroup_tasks_write(struct cgroup_subsys_state *css,
2219                               struct cftype *cft, u64 pid)
2220 {
2221         return attach_task_by_pid(css->cgroup, pid, false);
2222 }
2223 
2224 static int cgroup_procs_write(struct cgroup_subsys_state *css,
2225                               struct cftype *cft, u64 tgid)
2226 {
2227         return attach_task_by_pid(css->cgroup, tgid, true);
2228 }
2229 
2230 static int cgroup_release_agent_write(struct cgroup_subsys_state *css,
2231                                       struct cftype *cft, const char *buffer)
2232 {
2233         BUILD_BUG_ON(sizeof(css->cgroup->root->release_agent_path) < PATH_MAX);
2234         if (strlen(buffer) >= PATH_MAX)
2235                 return -EINVAL;
2236         if (!cgroup_lock_live_group(css->cgroup))
2237                 return -ENODEV;
2238         mutex_lock(&cgroup_root_mutex);
2239         strcpy(css->cgroup->root->release_agent_path, buffer);
2240         mutex_unlock(&cgroup_root_mutex);
2241         mutex_unlock(&cgroup_mutex);
2242         return 0;
2243 }
2244 
2245 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2246 {
2247         struct cgroup *cgrp = seq_css(seq)->cgroup;
2248 
2249         if (!cgroup_lock_live_group(cgrp))
2250                 return -ENODEV;
2251         seq_puts(seq, cgrp->root->release_agent_path);
2252         seq_putc(seq, '\n');
2253         mutex_unlock(&cgroup_mutex);
2254         return 0;
2255 }
2256 
2257 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2258 {
2259         struct cgroup *cgrp = seq_css(seq)->cgroup;
2260 
2261         seq_printf(seq, "%d\n", cgroup_sane_behavior(cgrp));
2262         return 0;
2263 }
2264 
2265 /* A buffer size big enough for numbers or short strings */
2266 #define CGROUP_LOCAL_BUFFER_SIZE 64
2267 
2268 static ssize_t cgroup_file_write(struct file *file, const char __user *userbuf,
2269                                  size_t nbytes, loff_t *ppos)
2270 {
2271         struct cfent *cfe = __d_cfe(file->f_dentry);
2272         struct cftype *cft = __d_cft(file->f_dentry);
2273         struct cgroup_subsys_state *css = cfe->css;
2274         size_t max_bytes = cft->max_write_len ?: CGROUP_LOCAL_BUFFER_SIZE - 1;
2275         char *buf;
2276         int ret;
2277 
2278         if (nbytes >= max_bytes)
2279                 return -E2BIG;
2280 
2281         buf = kmalloc(nbytes + 1, GFP_KERNEL);
2282         if (!buf)
2283                 return -ENOMEM;
2284 
2285         if (copy_from_user(buf, userbuf, nbytes)) {
2286                 ret = -EFAULT;
2287                 goto out_free;
2288         }
2289 
2290         buf[nbytes] = '\0';
2291 
2292         if (cft->write_string) {
2293                 ret = cft->write_string(css, cft, strstrip(buf));
2294         } else if (cft->write_u64) {
2295                 unsigned long long v;
2296                 ret = kstrtoull(buf, 0, &v);
2297                 if (!ret)
2298                         ret = cft->write_u64(css, cft, v);
2299         } else if (cft->write_s64) {
2300                 long long v;
2301                 ret = kstrtoll(buf, 0, &v);
2302                 if (!ret)
2303                         ret = cft->write_s64(css, cft, v);
2304         } else if (cft->trigger) {
2305                 ret = cft->trigger(css, (unsigned int)cft->private);
2306         } else {
2307                 ret = -EINVAL;
2308         }
2309 out_free:
2310         kfree(buf);
2311         return ret ?: nbytes;
2312 }
2313 
2314 /*
2315  * seqfile ops/methods for returning structured data. Currently just
2316  * supports string->u64 maps, but can be extended in future.
2317  */
2318 
2319 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2320 {
2321         struct cftype *cft = seq_cft(seq);
2322 
2323         if (cft->seq_start) {
2324                 return cft->seq_start(seq, ppos);
2325         } else {
2326                 /*
2327                  * The same behavior and code as single_open().  Returns
2328                  * !NULL if pos is at the beginning; otherwise, NULL.
2329                  */
2330                 return NULL + !*ppos;
2331         }
2332 }
2333 
2334 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2335 {
2336         struct cftype *cft = seq_cft(seq);
2337 
2338         if (cft->seq_next) {
2339                 return cft->seq_next(seq, v, ppos);
2340         } else {
2341                 /*
2342                  * The same behavior and code as single_open(), always
2343                  * terminate after the initial read.
2344                  */
2345                 ++*ppos;
2346                 return NULL;
2347         }
2348 }
2349 
2350 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2351 {
2352         struct cftype *cft = seq_cft(seq);
2353 
2354         if (cft->seq_stop)
2355                 cft->seq_stop(seq, v);
2356 }
2357 
2358 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2359 {
2360         struct cftype *cft = seq_cft(m);
2361         struct cgroup_subsys_state *css = seq_css(m);
2362 
2363         if (cft->seq_show)
2364                 return cft->seq_show(m, arg);
2365 
2366         if (cft->read_u64)
2367                 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2368         else if (cft->read_s64)
2369                 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2370         else
2371                 return -EINVAL;
2372         return 0;
2373 }
2374 
2375 static struct seq_operations cgroup_seq_operations = {
2376         .start          = cgroup_seqfile_start,
2377         .next           = cgroup_seqfile_next,
2378         .stop           = cgroup_seqfile_stop,
2379         .show           = cgroup_seqfile_show,
2380 };
2381 
2382 static int cgroup_file_open(struct inode *inode, struct file *file)
2383 {
2384         struct cfent *cfe = __d_cfe(file->f_dentry);
2385         struct cftype *cft = __d_cft(file->f_dentry);
2386         struct cgroup *cgrp = __d_cgrp(cfe->dentry->d_parent);
2387         struct cgroup_subsys_state *css;
2388         struct cgroup_open_file *of;
2389         int err;
2390 
2391         err = generic_file_open(inode, file);
2392         if (err)
2393                 return err;
2394 
2395         /*
2396          * If the file belongs to a subsystem, pin the css.  Will be
2397          * unpinned either on open failure or release.  This ensures that
2398          * @css stays alive for all file operations.
2399          */
2400         rcu_read_lock();
2401         css = cgroup_css(cgrp, cft->ss);
2402         if (cft->ss && !css_tryget(css))
2403                 css = NULL;
2404         rcu_read_unlock();
2405 
2406         if (!css)
2407                 return -ENODEV;
2408 
2409         /*
2410          * @cfe->css is used by read/write/close to determine the
2411          * associated css.  @file->private_data would be a better place but
2412          * that's already used by seqfile.  Multiple accessors may use it
2413          * simultaneously which is okay as the association never changes.
2414          */
2415         WARN_ON_ONCE(cfe->css && cfe->css != css);
2416         cfe->css = css;
2417 
2418         of = __seq_open_private(file, &cgroup_seq_operations,
2419                                 sizeof(struct cgroup_open_file));
2420         if (of) {
2421                 of->cfe = cfe;
2422                 return 0;
2423         }
2424 
2425         if (css->ss)
2426                 css_put(css);
2427         return -ENOMEM;
2428 }
2429 
2430 static int cgroup_file_release(struct inode *inode, struct file *file)
2431 {
2432         struct cfent *cfe = __d_cfe(file->f_dentry);
2433         struct cgroup_subsys_state *css = cfe->css;
2434 
2435         if (css->ss)
2436                 css_put(css);
2437         return seq_release_private(inode, file);
2438 }
2439 
2440 /*
2441  * cgroup_rename - Only allow simple rename of directories in place.
2442  */
2443 static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
2444                             struct inode *new_dir, struct dentry *new_dentry)
2445 {
2446         int ret;
2447         struct cgroup_name *name, *old_name;
2448         struct cgroup *cgrp;
2449 
2450         /*
2451          * It's convinient to use parent dir's i_mutex to protected
2452          * cgrp->name.
2453          */
2454         lockdep_assert_held(&old_dir->i_mutex);
2455 
2456         if (!S_ISDIR(old_dentry->d_inode->i_mode))
2457                 return -ENOTDIR;
2458         if (new_dentry->d_inode)
2459                 return -EEXIST;
2460         if (old_dir != new_dir)
2461                 return -EIO;
2462 
2463         cgrp = __d_cgrp(old_dentry);
2464 
2465         /*
2466          * This isn't a proper migration and its usefulness is very
2467          * limited.  Disallow if sane_behavior.
2468          */
2469         if (cgroup_sane_behavior(cgrp))
2470                 return -EPERM;
2471 
2472         name = cgroup_alloc_name(new_dentry);
2473         if (!name)
2474                 return -ENOMEM;
2475 
2476         ret = simple_rename(old_dir, old_dentry, new_dir, new_dentry);
2477         if (ret) {
2478                 kfree(name);
2479                 return ret;
2480         }
2481 
2482         old_name = rcu_dereference_protected(cgrp->name, true);
2483         rcu_assign_pointer(cgrp->name, name);
2484 
2485         kfree_rcu(old_name, rcu_head);
2486         return 0;
2487 }
2488 
2489 static struct simple_xattrs *__d_xattrs(struct dentry *dentry)
2490 {
2491         if (S_ISDIR(dentry->d_inode->i_mode))
2492                 return &__d_cgrp(dentry)->xattrs;
2493         else
2494                 return &__d_cfe(dentry)->xattrs;
2495 }
2496 
2497 static inline int xattr_enabled(struct dentry *dentry)
2498 {
2499         struct cgroupfs_root *root = dentry->d_sb->s_fs_info;
2500         return root->flags & CGRP_ROOT_XATTR;
2501 }
2502 
2503 static bool is_valid_xattr(const char *name)
2504 {
2505         if (!strncmp(name, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) ||
2506             !strncmp(name, XATTR_SECURITY_PREFIX, XATTR_SECURITY_PREFIX_LEN))
2507                 return true;
2508         return false;
2509 }
2510 
2511 static int cgroup_setxattr(struct dentry *dentry, const char *name,
2512                            const void *val, size_t size, int flags)
2513 {
2514         if (!xattr_enabled(dentry))
2515                 return -EOPNOTSUPP;
2516         if (!is_valid_xattr(name))
2517                 return -EINVAL;
2518         return simple_xattr_set(__d_xattrs(dentry), name, val, size, flags);
2519 }
2520 
2521 static int cgroup_removexattr(struct dentry *dentry, const char *name)
2522 {
2523         if (!xattr_enabled(dentry))
2524                 return -EOPNOTSUPP;
2525         if (!is_valid_xattr(name))
2526                 return -EINVAL;
2527         return simple_xattr_remove(__d_xattrs(dentry), name);
2528 }
2529 
2530 static ssize_t cgroup_getxattr(struct dentry *dentry, const char *name,
2531                                void *buf, size_t size)
2532 {
2533         if (!xattr_enabled(dentry))
2534                 return -EOPNOTSUPP;
2535         if (!is_valid_xattr(name))
2536                 return -EINVAL;
2537         return simple_xattr_get(__d_xattrs(dentry), name, buf, size);
2538 }
2539 
2540 static ssize_t cgroup_listxattr(struct dentry *dentry, char *buf, size_t size)
2541 {
2542         if (!xattr_enabled(dentry))
2543                 return -EOPNOTSUPP;
2544         return simple_xattr_list(__d_xattrs(dentry), buf, size);
2545 }
2546 
2547 static const struct file_operations cgroup_file_operations = {
2548         .read = seq_read,
2549         .write = cgroup_file_write,
2550         .llseek = generic_file_llseek,
2551         .open = cgroup_file_open,
2552         .release = cgroup_file_release,
2553 };
2554 
2555 static const struct inode_operations cgroup_file_inode_operations = {
2556         .setxattr = cgroup_setxattr,
2557         .getxattr = cgroup_getxattr,
2558         .listxattr = cgroup_listxattr,
2559         .removexattr = cgroup_removexattr,
2560 };
2561 
2562 static const struct inode_operations cgroup_dir_inode_operations = {
2563         .lookup = simple_lookup,
2564         .mkdir = cgroup_mkdir,
2565         .rmdir = cgroup_rmdir,
2566         .rename = cgroup_rename,
2567         .setxattr = cgroup_setxattr,
2568         .getxattr = cgroup_getxattr,
2569         .listxattr = cgroup_listxattr,
2570         .removexattr = cgroup_removexattr,
2571 };
2572 
2573 static int cgroup_create_file(struct dentry *dentry, umode_t mode,
2574                                 struct super_block *sb)
2575 {
2576         struct inode *inode;
2577 
2578         if (!dentry)
2579                 return -ENOENT;
2580         if (dentry->d_inode)
2581                 return -EEXIST;
2582 
2583         inode = cgroup_new_inode(mode, sb);
2584         if (!inode)
2585                 return -ENOMEM;
2586 
2587         if (S_ISDIR(mode)) {
2588                 inode->i_op = &cgroup_dir_inode_operations;
2589                 inode->i_fop = &simple_dir_operations;
2590 
2591                 /* start off with i_nlink == 2 (for "." entry) */
2592                 inc_nlink(inode);
2593                 inc_nlink(dentry->d_parent->d_inode);
2594 
2595                 /*
2596                  * Control reaches here with cgroup_mutex held.
2597                  * @inode->i_mutex should nest outside cgroup_mutex but we
2598                  * want to populate it immediately without releasing
2599                  * cgroup_mutex.  As @inode isn't visible to anyone else
2600                  * yet, trylock will always succeed without affecting
2601                  * lockdep checks.
2602                  */
2603                 WARN_ON_ONCE(!mutex_trylock(&inode->i_mutex));
2604         } else if (S_ISREG(mode)) {
2605                 inode->i_size = 0;
2606                 inode->i_fop = &cgroup_file_operations;
2607                 inode->i_op = &cgroup_file_inode_operations;
2608         }
2609         d_instantiate(dentry, inode);
2610         dget(dentry);   /* Extra count - pin the dentry in core */
2611         return 0;
2612 }
2613 
2614 /**
2615  * cgroup_file_mode - deduce file mode of a control file
2616  * @cft: the control file in question
2617  *
2618  * returns cft->mode if ->mode is not 0
2619  * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
2620  * returns S_IRUGO if it has only a read handler
2621  * returns S_IWUSR if it has only a write hander
2622  */
2623 static umode_t cgroup_file_mode(const struct cftype *cft)
2624 {
2625         umode_t mode = 0;
2626 
2627         if (cft->mode)
2628                 return cft->mode;
2629 
2630         if (cft->read_u64 || cft->read_s64 || cft->seq_show)
2631                 mode |= S_IRUGO;
2632 
2633         if (cft->write_u64 || cft->write_s64 || cft->write_string ||
2634             cft->trigger)
2635                 mode |= S_IWUSR;
2636 
2637         return mode;
2638 }
2639 
2640 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
2641 {
2642         struct dentry *dir = cgrp->dentry;
2643         struct cgroup *parent = __d_cgrp(dir);
2644         struct dentry *dentry;
2645         struct cfent *cfe;
2646         int error;
2647         umode_t mode;
2648         char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
2649 
2650         if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
2651             !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
2652                 strcpy(name, cft->ss->name);
2653                 strcat(name, ".");
2654         }
2655         strcat(name, cft->name);
2656 
2657         BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
2658 
2659         cfe = kzalloc(sizeof(*cfe), GFP_KERNEL);
2660         if (!cfe)
2661                 return -ENOMEM;
2662 
2663         dentry = lookup_one_len(name, dir, strlen(name));
2664         if (IS_ERR(dentry)) {
2665                 error = PTR_ERR(dentry);
2666                 goto out;
2667         }
2668 
2669         cfe->type = (void *)cft;
2670         cfe->dentry = dentry;
2671         dentry->d_fsdata = cfe;
2672         simple_xattrs_init(&cfe->xattrs);
2673 
2674         mode = cgroup_file_mode(cft);
2675         error = cgroup_create_file(dentry, mode | S_IFREG, cgrp->root->sb);
2676         if (!error) {
2677                 list_add_tail(&cfe->node, &parent->files);
2678                 cfe = NULL;
2679         }
2680         dput(dentry);
2681 out:
2682         kfree(cfe);
2683         return error;
2684 }
2685 
2686 /**
2687  * cgroup_addrm_files - add or remove files to a cgroup directory
2688  * @cgrp: the target cgroup
2689  * @cfts: array of cftypes to be added
2690  * @is_add: whether to add or remove
2691  *
2692  * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
2693  * For removals, this function never fails.  If addition fails, this
2694  * function doesn't remove files already added.  The caller is responsible
2695  * for cleaning up.
2696  */
2697 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
2698                               bool is_add)
2699 {
2700         struct cftype *cft;
2701         int ret;
2702 
2703         lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
2704         lockdep_assert_held(&cgroup_mutex);
2705 
2706         for (cft = cfts; cft->name[0] != '\0'; cft++) {
2707                 /* does cft->flags tell us to skip this file on @cgrp? */
2708                 if ((cft->flags & CFTYPE_INSANE) && cgroup_sane_behavior(cgrp))
2709                         continue;
2710                 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgrp->parent)
2711                         continue;
2712                 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgrp->parent)
2713                         continue;
2714 
2715                 if (is_add) {
2716                         ret = cgroup_add_file(cgrp, cft);
2717                         if (ret) {
2718                                 pr_warn("cgroup_addrm_files: failed to add %s, err=%d\n",
2719                                         cft->name, ret);
2720                                 return ret;
2721                         }
2722                 } else {
2723                         cgroup_rm_file(cgrp, cft);
2724                 }
2725         }
2726         return 0;
2727 }
2728 
2729 static void cgroup_cfts_prepare(void)
2730         __acquires(&cgroup_mutex)
2731 {
2732         /*
2733          * Thanks to the entanglement with vfs inode locking, we can't walk
2734          * the existing cgroups under cgroup_mutex and create files.
2735          * Instead, we use css_for_each_descendant_pre() and drop RCU read
2736          * lock before calling cgroup_addrm_files().
2737          */
2738         mutex_lock(&cgroup_mutex);
2739 }
2740 
2741 static int cgroup_cfts_commit(struct cftype *cfts, bool is_add)
2742         __releases(&cgroup_mutex)
2743 {
2744         LIST_HEAD(pending);
2745         struct cgroup_subsys *ss = cfts[0].ss;
2746         struct cgroup *root = &ss->root->top_cgroup;
2747         struct super_block *sb = ss->root->sb;
2748         struct dentry *prev = NULL;
2749         struct inode *inode;
2750         struct cgroup_subsys_state *css;
2751         u64 update_before;
2752         int ret = 0;
2753 
2754         /* %NULL @cfts indicates abort and don't bother if @ss isn't attached */
2755         if (!cfts || ss->root == &cgroup_dummy_root ||
2756             !atomic_inc_not_zero(&sb->s_active)) {
2757                 mutex_unlock(&cgroup_mutex);
2758                 return 0;
2759         }
2760 
2761         /*
2762          * All cgroups which are created after we drop cgroup_mutex will
2763          * have the updated set of files, so we only need to update the
2764          * cgroups created before the current @cgroup_serial_nr_next.
2765          */
2766         update_before = cgroup_serial_nr_next;
2767 
2768         /* add/rm files for all cgroups created before */
2769         css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
2770                 struct cgroup *cgrp = css->cgroup;
2771 
2772                 if (cgroup_is_dead(cgrp))
2773                         continue;
2774 
2775                 inode = cgrp->dentry->d_inode;
2776                 dget(cgrp->dentry);
2777                 dput(prev);
2778                 prev = cgrp->dentry;
2779 
2780                 mutex_unlock(&cgroup_mutex);
2781                 mutex_lock(&inode->i_mutex);
2782                 mutex_lock(&cgroup_mutex);
2783                 if (cgrp->serial_nr < update_before && !cgroup_is_dead(cgrp))
2784                         ret = cgroup_addrm_files(cgrp, cfts, is_add);
2785                 mutex_unlock(&inode->i_mutex);
2786                 if (ret)
2787                         break;
2788         }
2789         mutex_unlock(&cgroup_mutex);
2790         dput(prev);
2791         deactivate_super(sb);
2792         return ret;
2793 }
2794 
2795 /**
2796  * cgroup_add_cftypes - add an array of cftypes to a subsystem
2797  * @ss: target cgroup subsystem
2798  * @cfts: zero-length name terminated array of cftypes
2799  *
2800  * Register @cfts to @ss.  Files described by @cfts are created for all
2801  * existing cgroups to which @ss is attached and all future cgroups will
2802  * have them too.  This function can be called anytime whether @ss is
2803  * attached or not.
2804  *
2805  * Returns 0 on successful registration, -errno on failure.  Note that this
2806  * function currently returns 0 as long as @cfts registration is successful
2807  * even if some file creation attempts on existing cgroups fail.
2808  */
2809 int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
2810 {
2811         struct cftype_set *set;
2812         struct cftype *cft;
2813         int ret;
2814 
2815         set = kzalloc(sizeof(*set), GFP_KERNEL);
2816         if (!set)
2817                 return -ENOMEM;
2818 
2819         for (cft = cfts; cft->name[0] != '\0'; cft++)
2820                 cft->ss = ss;
2821 
2822         cgroup_cfts_prepare();
2823         set->cfts = cfts;
2824         list_add_tail(&set->node, &ss->cftsets);
2825         ret = cgroup_cfts_commit(cfts, true);
2826         if (ret)
2827                 cgroup_rm_cftypes(cfts);
2828         return ret;
2829 }
2830 EXPORT_SYMBOL_GPL(cgroup_add_cftypes);
2831 
2832 /**
2833  * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
2834  * @cfts: zero-length name terminated array of cftypes
2835  *
2836  * Unregister @cfts.  Files described by @cfts are removed from all
2837  * existing cgroups and all future cgroups won't have them either.  This
2838  * function can be called anytime whether @cfts' subsys is attached or not.
2839  *
2840  * Returns 0 on successful unregistration, -ENOENT if @cfts is not
2841  * registered.
2842  */
2843 int cgroup_rm_cftypes(struct cftype *cfts)
2844 {
2845         struct cftype_set *set;
2846 
2847         if (!cfts || !cfts[0].ss)
2848                 return -ENOENT;
2849 
2850         cgroup_cfts_prepare();
2851 
2852         list_for_each_entry(set, &cfts[0].ss->cftsets, node) {
2853                 if (set->cfts == cfts) {
2854                         list_del(&set->node);
2855                         kfree(set);
2856                         cgroup_cfts_commit(cfts, false);
2857                         return 0;
2858                 }
2859         }
2860 
2861         cgroup_cfts_commit(NULL, false);
2862         return -ENOENT;
2863 }
2864 
2865 /**
2866  * cgroup_task_count - count the number of tasks in a cgroup.
2867  * @cgrp: the cgroup in question
2868  *
2869  * Return the number of tasks in the cgroup.
2870  */
2871 int cgroup_task_count(const struct cgroup *cgrp)
2872 {
2873         int count = 0;
2874         struct cgrp_cset_link *link;
2875 
2876         read_lock(&css_set_lock);
2877         list_for_each_entry(link, &cgrp->cset_links, cset_link)
2878                 count += atomic_read(&link->cset->refcount);
2879         read_unlock(&css_set_lock);
2880         return count;
2881 }
2882 
2883 /*
2884  * To reduce the fork() overhead for systems that are not actually using
2885  * their cgroups capability, we don't maintain the lists running through
2886  * each css_set to its tasks until we see the list actually used - in other
2887  * words after the first call to css_task_iter_start().
2888  */
2889 static void cgroup_enable_task_cg_lists(void)
2890 {
2891         struct task_struct *p, *g;
2892         write_lock(&css_set_lock);
2893         use_task_css_set_links = 1;
2894         /*
2895          * We need tasklist_lock because RCU is not safe against
2896          * while_each_thread(). Besides, a forking task that has passed
2897          * cgroup_post_fork() without seeing use_task_css_set_links = 1
2898          * is not guaranteed to have its child immediately visible in the
2899          * tasklist if we walk through it with RCU.
2900          */
2901         read_lock(&tasklist_lock);
2902         do_each_thread(g, p) {
2903                 task_lock(p);
2904                 /*
2905                  * We should check if the process is exiting, otherwise
2906                  * it will race with cgroup_exit() in that the list
2907                  * entry won't be deleted though the process has exited.
2908                  * Do it while holding siglock so that we don't end up
2909                  * racing against cgroup_exit().
2910                  */
2911                 spin_lock_irq(&p->sighand->siglock);
2912                 if (!(p->flags & PF_EXITING) && list_empty(&p->cg_list))
2913                         list_add(&p->cg_list, &task_css_set(p)->tasks);
2914                 spin_unlock_irq(&p->sighand->siglock);
2915 
2916                 task_unlock(p);
2917         } while_each_thread(g, p);
2918         read_unlock(&tasklist_lock);
2919         write_unlock(&css_set_lock);
2920 }
2921 
2922 /**
2923  * css_next_child - find the next child of a given css
2924  * @pos_css: the current position (%NULL to initiate traversal)
2925  * @parent_css: css whose children to walk
2926  *
2927  * This function returns the next child of @parent_css and should be called
2928  * under either cgroup_mutex or RCU read lock.  The only requirement is
2929  * that @parent_css and @pos_css are accessible.  The next sibling is
2930  * guaranteed to be returned regardless of their states.
2931  */
2932 struct cgroup_subsys_state *
2933 css_next_child(struct cgroup_subsys_state *pos_css,
2934                struct cgroup_subsys_state *parent_css)
2935 {
2936         struct cgroup *pos = pos_css ? pos_css->cgroup : NULL;
2937         struct cgroup *cgrp = parent_css->cgroup;
2938         struct cgroup *next;
2939 
2940         cgroup_assert_mutex_or_rcu_locked();
2941 
2942         /*
2943          * @pos could already have been removed.  Once a cgroup is removed,
2944          * its ->sibling.next is no longer updated when its next sibling
2945          * changes.  As CGRP_DEAD assertion is serialized and happens
2946          * before the cgroup is taken off the ->sibling list, if we see it
2947          * unasserted, it's guaranteed that the next sibling hasn't
2948          * finished its grace period even if it's already removed, and thus
2949          * safe to dereference from this RCU critical section.  If
2950          * ->sibling.next is inaccessible, cgroup_is_dead() is guaranteed
2951          * to be visible as %true here.
2952          *
2953          * If @pos is dead, its next pointer can't be dereferenced;
2954          * however, as each cgroup is given a monotonically increasing
2955          * unique serial number and always appended to the sibling list,
2956          * the next one can be found by walking the parent's children until
2957          * we see a cgroup with higher serial number than @pos's.  While
2958          * this path can be slower, it's taken only when either the current
2959          * cgroup is removed or iteration and removal race.
2960          */
2961         if (!pos) {
2962                 next = list_entry_rcu(cgrp->children.next, struct cgroup, sibling);
2963         } else if (likely(!cgroup_is_dead(pos))) {
2964                 next = list_entry_rcu(pos->sibling.next, struct cgroup, sibling);
2965         } else {
2966                 list_for_each_entry_rcu(next, &cgrp->children, sibling)
2967                         if (next->serial_nr > pos->serial_nr)
2968                                 break;
2969         }
2970 
2971         if (&next->sibling == &cgrp->children)
2972                 return NULL;
2973 
2974         return cgroup_css(next, parent_css->ss);
2975 }
2976 EXPORT_SYMBOL_GPL(css_next_child);
2977 
2978 /**
2979  * css_next_descendant_pre - find the next descendant for pre-order walk
2980  * @pos: the current position (%NULL to initiate traversal)
2981  * @root: css whose descendants to walk
2982  *
2983  * To be used by css_for_each_descendant_pre().  Find the next descendant
2984  * to visit for pre-order traversal of @root's descendants.  @root is
2985  * included in the iteration and the first node to be visited.
2986  *
2987  * While this function requires cgroup_mutex or RCU read locking, it
2988  * doesn't require the whole traversal to be contained in a single critical
2989  * section.  This function will return the correct next descendant as long
2990  * as both @pos and @root are accessible and @pos is a descendant of @root.
2991  */
2992 struct cgroup_subsys_state *
2993 css_next_descendant_pre(struct cgroup_subsys_state *pos,
2994                         struct cgroup_subsys_state *root)
2995 {
2996         struct cgroup_subsys_state *next;
2997 
2998         cgroup_assert_mutex_or_rcu_locked();
2999 
3000         /* if first iteration, visit @root */
3001         if (!pos)
3002                 return root;
3003 
3004         /* visit the first child if exists */
3005         next = css_next_child(NULL, pos);
3006         if (next)
3007                 return next;
3008 
3009         /* no child, visit my or the closest ancestor's next sibling */
3010         while (pos != root) {
3011                 next = css_next_child(pos, css_parent(pos));
3012                 if (next)
3013                         return next;
3014                 pos = css_parent(pos);
3015         }
3016 
3017         return NULL;
3018 }
3019 EXPORT_SYMBOL_GPL(css_next_descendant_pre);
3020 
3021 /**
3022  * css_rightmost_descendant - return the rightmost descendant of a css
3023  * @pos: css of interest
3024  *
3025  * Return the rightmost descendant of @pos.  If there's no descendant, @pos
3026  * is returned.  This can be used during pre-order traversal to skip
3027  * subtree of @pos.
3028  *
3029  * While this function requires cgroup_mutex or RCU read locking, it
3030  * doesn't require the whole traversal to be contained in a single critical
3031  * section.  This function will return the correct rightmost descendant as
3032  * long as @pos is accessible.
3033  */
3034 struct cgroup_subsys_state *
3035 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3036 {
3037         struct cgroup_subsys_state *last, *tmp;
3038 
3039         cgroup_assert_mutex_or_rcu_locked();
3040 
3041         do {
3042                 last = pos;
3043                 /* ->prev isn't RCU safe, walk ->next till the end */
3044                 pos = NULL;
3045                 css_for_each_child(tmp, last)
3046                         pos = tmp;
3047         } while (pos);
3048 
3049         return last;
3050 }
3051 EXPORT_SYMBOL_GPL(css_rightmost_descendant);
3052 
3053 static struct cgroup_subsys_state *
3054 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3055 {
3056         struct cgroup_subsys_state *last;
3057 
3058         do {
3059                 last = pos;
3060                 pos = css_next_child(NULL, pos);
3061         } while (pos);
3062 
3063         return last;
3064 }
3065 
3066 /**
3067  * css_next_descendant_post - find the next descendant for post-order walk
3068  * @pos: the current position (%NULL to initiate traversal)
3069  * @root: css whose descendants to walk
3070  *
3071  * To be used by css_for_each_descendant_post().  Find the next descendant
3072  * to visit for post-order traversal of @root's descendants.  @root is
3073  * included in the iteration and the last node to be visited.
3074  *
3075  * While this function requires cgroup_mutex or RCU read locking, it
3076  * doesn't require the whole traversal to be contained in a single critical
3077  * section.  This function will return the correct next descendant as long
3078  * as both @pos and @cgroup are accessible and @pos is a descendant of
3079  * @cgroup.
3080  */
3081 struct cgroup_subsys_state *
3082 css_next_descendant_post(struct cgroup_subsys_state *pos,
3083                          struct cgroup_subsys_state *root)
3084 {
3085         struct cgroup_subsys_state *next;
3086 
3087         cgroup_assert_mutex_or_rcu_locked();
3088 
3089         /* if first iteration, visit leftmost descendant which may be @root */
3090         if (!pos)
3091                 return css_leftmost_descendant(root);
3092 
3093         /* if we visited @root, we're done */
3094         if (pos == root)
3095                 return NULL;
3096 
3097         /* if there's an unvisited sibling, visit its leftmost descendant */
3098         next = css_next_child(pos, css_parent(pos));
3099         if (next)
3100                 return css_leftmost_descendant(next);
3101 
3102         /* no sibling left, visit parent */
3103         return css_parent(pos);
3104 }
3105 EXPORT_SYMBOL_GPL(css_next_descendant_post);
3106 
3107 /**
3108  * css_advance_task_iter - advance a task itererator to the next css_set
3109  * @it: the iterator to advance
3110  *
3111  * Advance @it to the next css_set to walk.
3112  */
3113 static void css_advance_task_iter(struct css_task_iter *it)
3114 {
3115         struct list_head *l = it->cset_link;
3116         struct cgrp_cset_link *link;
3117         struct css_set *cset;
3118 
3119         /* Advance to the next non-empty css_set */
3120         do {
3121                 l = l->next;
3122                 if (l == &it->origin_css->cgroup->cset_links) {
3123                         it->cset_link = NULL;
3124                         return;
3125                 }
3126                 link = list_entry(l, struct cgrp_cset_link, cset_link);
3127                 cset = link->cset;
3128         } while (list_empty(&cset->tasks));
3129         it->cset_link = l;
3130         it->task = cset->tasks.next;
3131 }
3132 
3133 /**
3134  * css_task_iter_start - initiate task iteration
3135  * @css: the css to walk tasks of
3136  * @it: the task iterator to use
3137  *
3138  * Initiate iteration through the tasks of @css.  The caller can call
3139  * css_task_iter_next() to walk through the tasks until the function
3140  * returns NULL.  On completion of iteration, css_task_iter_end() must be
3141  * called.
3142  *
3143  * Note that this function acquires a lock which is released when the
3144  * iteration finishes.  The caller can't sleep while iteration is in
3145  * progress.
3146  */
3147 void css_task_iter_start(struct cgroup_subsys_state *css,
3148                          struct css_task_iter *it)
3149         __acquires(css_set_lock)
3150 {
3151         /*
3152          * The first time anyone tries to iterate across a css, we need to
3153          * enable the list linking each css_set to its tasks, and fix up
3154          * all existing tasks.
3155          */
3156         if (!use_task_css_set_links)
3157                 cgroup_enable_task_cg_lists();
3158 
3159         read_lock(&css_set_lock);
3160 
3161         it->origin_css = css;
3162         it->cset_link = &css->cgroup->cset_links;
3163 
3164         css_advance_task_iter(it);
3165 }
3166 
3167 /**
3168  * css_task_iter_next - return the next task for the iterator
3169  * @it: the task iterator being iterated
3170  *
3171  * The "next" function for task iteration.  @it should have been
3172  * initialized via css_task_iter_start().  Returns NULL when the iteration
3173  * reaches the end.
3174  */
3175 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3176 {
3177         struct task_struct *res;
3178         struct list_head *l = it->task;
3179         struct cgrp_cset_link *link;
3180 
3181         /* If the iterator cg is NULL, we have no tasks */
3182         if (!it->cset_link)
3183                 return NULL;
3184         res = list_entry(l, struct task_struct, cg_list);
3185         /* Advance iterator to find next entry */
3186         l = l->next;
3187         link = list_entry(it->cset_link, struct cgrp_cset_link, cset_link);
3188         if (l == &link->cset->tasks) {
3189                 /*
3190                  * We reached the end of this task list - move on to the
3191                  * next cgrp_cset_link.
3192                  */
3193                 css_advance_task_iter(it);
3194         } else {
3195                 it->task = l;
3196         }
3197         return res;
3198 }
3199 
3200 /**
3201  * css_task_iter_end - finish task iteration
3202  * @it: the task iterator to finish
3203  *
3204  * Finish task iteration started by css_task_iter_start().
3205  */
3206 void css_task_iter_end(struct css_task_iter *it)
3207         __releases(css_set_lock)
3208 {
3209         read_unlock(&css_set_lock);
3210 }
3211 
3212 static inline int started_after_time(struct task_struct *t1,
3213                                      struct timespec *time,
3214                                      struct task_struct *t2)
3215 {
3216         int start_diff = timespec_compare(&t1->start_time, time);
3217         if (start_diff > 0) {
3218                 return 1;
3219         } else if (start_diff < 0) {
3220                 return 0;
3221         } else {
3222                 /*
3223                  * Arbitrarily, if two processes started at the same
3224                  * time, we'll say that the lower pointer value
3225                  * started first. Note that t2 may have exited by now
3226                  * so this may not be a valid pointer any longer, but
3227                  * that's fine - it still serves to distinguish
3228                  * between two tasks started (effectively) simultaneously.
3229                  */
3230                 return t1 > t2;
3231         }
3232 }
3233 
3234 /*
3235  * This function is a callback from heap_insert() and is used to order
3236  * the heap.
3237  * In this case we order the heap in descending task start time.
3238  */
3239 static inline int started_after(void *p1, void *p2)
3240 {
3241         struct task_struct *t1 = p1;
3242         struct task_struct *t2 = p2;
3243         return started_after_time(t1, &t2->start_time, t2);
3244 }
3245 
3246 /**
3247  * css_scan_tasks - iterate though all the tasks in a css
3248  * @css: the css to iterate tasks of
3249  * @test: optional test callback
3250  * @process: process callback
3251  * @data: data passed to @test and @process
3252  * @heap: optional pre-allocated heap used for task iteration
3253  *
3254  * Iterate through all the tasks in @css, calling @test for each, and if it
3255  * returns %true, call @process for it also.
3256  *
3257  * @test may be NULL, meaning always true (select all tasks), which
3258  * effectively duplicates css_task_iter_{start,next,end}() but does not
3259  * lock css_set_lock for the call to @process.
3260  *
3261  * It is guaranteed that @process will act on every task that is a member
3262  * of @css for the duration of this call.  This function may or may not
3263  * call @process for tasks that exit or move to a different css during the
3264  * call, or are forked or move into the css during the call.
3265  *
3266  * Note that @test may be called with locks held, and may in some
3267  * situations be called multiple times for the same task, so it should be
3268  * cheap.
3269  *
3270  * If @heap is non-NULL, a heap has been pre-allocated and will be used for
3271  * heap operations (and its "gt" member will be overwritten), else a
3272  * temporary heap will be used (allocation of which may cause this function
3273  * to fail).
3274  */
3275 int css_scan_tasks(struct cgroup_subsys_state *css,
3276                    bool (*test)(struct task_struct *, void *),
3277                    void (*process)(struct task_struct *, void *),
3278                    void *data, struct ptr_heap *heap)
3279 {
3280         int retval, i;
3281         struct css_task_iter it;
3282         struct task_struct *p, *dropped;
3283         /* Never dereference latest_task, since it's not refcounted */
3284         struct task_struct *latest_task = NULL;
3285         struct ptr_heap tmp_heap;
3286         struct timespec latest_time = { 0, 0 };
3287 
3288         if (heap) {
3289                 /* The caller supplied our heap and pre-allocated its memory */
3290                 heap->gt = &started_after;
3291         } else {
3292                 /* We need to allocate our own heap memory */
3293                 heap = &tmp_heap;
3294                 retval = heap_init(heap, PAGE_SIZE, GFP_KERNEL, &started_after);
3295                 if (retval)
3296                         /* cannot allocate the heap */
3297                         return retval;
3298         }
3299 
3300  again:
3301         /*
3302          * Scan tasks in the css, using the @test callback to determine
3303          * which are of interest, and invoking @process callback on the
3304          * ones which need an update.  Since we don't want to hold any
3305          * locks during the task updates, gather tasks to be processed in a
3306          * heap structure.  The heap is sorted by descending task start
3307          * time.  If the statically-sized heap fills up, we overflow tasks
3308          * that started later, and in future iterations only consider tasks
3309          * that started after the latest task in the previous pass. This
3310          * guarantees forward progress and that we don't miss any tasks.
3311          */
3312         heap->size = 0;
3313         css_task_iter_start(css, &it);
3314         while ((p = css_task_iter_next(&it))) {
3315                 /*
3316                  * Only affect tasks that qualify per the caller's callback,
3317                  * if he provided one
3318                  */
3319                 if (test && !test(p, data))
3320                         continue;
3321                 /*
3322                  * Only process tasks that started after the last task
3323                  * we processed
3324                  */
3325                 if (!started_after_time(p, &latest_time, latest_task))
3326                         continue;
3327                 dropped = heap_insert(heap, p);
3328                 if (dropped == NULL) {
3329                         /*
3330                          * The new task was inserted; the heap wasn't
3331                          * previously full
3332                          */
3333                         get_task_struct(p);
3334                 } else if (dropped != p) {
3335                         /*
3336                          * The new task was inserted, and pushed out a
3337                          * different task
3338                          */
3339                         get_task_struct(p);
3340                         put_task_struct(dropped);
3341                 }
3342                 /*
3343                  * Else the new task was newer than anything already in
3344                  * the heap and wasn't inserted
3345                  */
3346         }
3347         css_task_iter_end(&it);
3348 
3349         if (heap->size) {
3350                 for (i = 0; i < heap->size; i++) {
3351                         struct task_struct *q = heap->ptrs[i];
3352                         if (i == 0) {
3353                                 latest_time = q->start_time;
3354                                 latest_task = q;
3355                         }
3356                         /* Process the task per the caller's callback */
3357                         process(q, data);
3358                         put_task_struct(q);
3359                 }
3360                 /*
3361                  * If we had to process any tasks at all, scan again
3362                  * in case some of them were in the middle of forking
3363                  * children that didn't get processed.
3364                  * Not the most efficient way to do it, but it avoids
3365                  * having to take callback_mutex in the fork path
3366                  */
3367                 goto again;
3368         }
3369         if (heap == &tmp_heap)
3370                 heap_free(&tmp_heap);
3371         return 0;
3372 }
3373 
3374 static void cgroup_transfer_one_task(struct task_struct *task, void *data)
3375 {
3376         struct cgroup *new_cgroup = data;
3377 
3378         mutex_lock(&cgroup_mutex);
3379         cgroup_attach_task(new_cgroup, task, false);
3380         mutex_unlock(&cgroup_mutex);
3381 }
3382 
3383 /**
3384  * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3385  * @to: cgroup to which the tasks will be moved
3386  * @from: cgroup in which the tasks currently reside
3387  */
3388 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3389 {
3390         return css_scan_tasks(&from->dummy_css, NULL, cgroup_transfer_one_task,
3391                               to, NULL);
3392 }
3393 
3394 /*
3395  * Stuff for reading the 'tasks'/'procs' files.
3396  *
3397  * Reading this file can return large amounts of data if a cgroup has
3398  * *lots* of attached tasks. So it may need several calls to read(),
3399  * but we cannot guarantee that the information we produce is correct
3400  * unless we produce it entirely atomically.
3401  *
3402  */
3403 
3404 /* which pidlist file are we talking about? */
3405 enum cgroup_filetype {
3406         CGROUP_FILE_PROCS,
3407         CGROUP_FILE_TASKS,
3408 };
3409 
3410 /*
3411  * A pidlist is a list of pids that virtually represents the contents of one
3412  * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3413  * a pair (one each for procs, tasks) for each pid namespace that's relevant
3414  * to the cgroup.
3415  */
3416 struct cgroup_pidlist {
3417         /*
3418          * used to find which pidlist is wanted. doesn't change as long as
3419          * this particular list stays in the list.
3420         */
3421         struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3422         /* array of xids */
3423         pid_t *list;
3424         /* how many elements the above list has */
3425         int length;
3426         /* each of these stored in a list by its cgroup */
3427         struct list_head links;
3428         /* pointer to the cgroup we belong to, for list removal purposes */
3429         struct cgroup *owner;
3430         /* for delayed destruction */
3431         struct delayed_work destroy_dwork;
3432 };
3433 
3434 /*
3435  * The following two functions "fix" the issue where there are more pids
3436  * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3437  * TODO: replace with a kernel-wide solution to this problem
3438  */
3439 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3440 static void *pidlist_allocate(int count)
3441 {
3442         if (PIDLIST_TOO_LARGE(count))
3443                 return vmalloc(count * sizeof(pid_t));
3444         else
3445                 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3446 }
3447 
3448 static void pidlist_free(void *p)
3449 {
3450         if (is_vmalloc_addr(p))
3451                 vfree(p);
3452         else
3453                 kfree(p);
3454 }
3455 
3456 /*
3457  * Used to destroy all pidlists lingering waiting for destroy timer.  None
3458  * should be left afterwards.
3459  */
3460 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3461 {
3462         struct cgroup_pidlist *l, *tmp_l;
3463 
3464         mutex_lock(&cgrp->pidlist_mutex);
3465         list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3466                 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3467         mutex_unlock(&cgrp->pidlist_mutex);
3468 
3469         flush_workqueue(cgroup_pidlist_destroy_wq);
3470         BUG_ON(!list_empty(&cgrp->pidlists));
3471 }
3472 
3473 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3474 {
3475         struct delayed_work *dwork = to_delayed_work(work);
3476         struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3477                                                 destroy_dwork);
3478         struct cgroup_pidlist *tofree = NULL;
3479 
3480         mutex_lock(&l->owner->pidlist_mutex);
3481 
3482         /*
3483          * Destroy iff we didn't get queued again.  The state won't change
3484          * as destroy_dwork can only be queued while locked.
3485          */
3486         if (!delayed_work_pending(dwork)) {
3487                 list_del(&l->links);
3488                 pidlist_free(l->list);
3489                 put_pid_ns(l->key.ns);
3490                 tofree = l;
3491         }
3492 
3493         mutex_unlock(&l->owner->pidlist_mutex);
3494         kfree(tofree);
3495 }
3496 
3497 /*
3498  * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3499  * Returns the number of unique elements.
3500  */
3501 static int pidlist_uniq(pid_t *list, int length)
3502 {
3503         int src, dest = 1;
3504 
3505         /*
3506          * we presume the 0th element is unique, so i starts at 1. trivial
3507          * edge cases first; no work needs to be done for either
3508          */
3509         if (length == 0 || length == 1)
3510                 return length;
3511         /* src and dest walk down the list; dest counts unique elements */
3512         for (src = 1; src < length; src++) {
3513                 /* find next unique element */
3514                 while (list[src] == list[src-1]) {
3515                         src++;
3516                         if (src == length)
3517                                 goto after;
3518                 }
3519                 /* dest always points to where the next unique element goes */
3520                 list[dest] = list[src];
3521                 dest++;
3522         }
3523 after:
3524         return dest;
3525 }
3526 
3527 /*
3528  * The two pid files - task and cgroup.procs - guaranteed that the result
3529  * is sorted, which forced this whole pidlist fiasco.  As pid order is
3530  * different per namespace, each namespace needs differently sorted list,
3531  * making it impossible to use, for example, single rbtree of member tasks
3532  * sorted by task pointer.  As pidlists can be fairly large, allocating one
3533  * per open file is dangerous, so cgroup had to implement shared pool of
3534  * pidlists keyed by cgroup and namespace.
3535  *
3536  * All this extra complexity was caused by the original implementation
3537  * committing to an entirely unnecessary property.  In the long term, we
3538  * want to do away with it.  Explicitly scramble sort order if
3539  * sane_behavior so that no such expectation exists in the new interface.
3540  *
3541  * Scrambling is done by swapping every two consecutive bits, which is
3542  * non-identity one-to-one mapping which disturbs sort order sufficiently.
3543  */
3544 static pid_t pid_fry(pid_t pid)
3545 {
3546         unsigned a = pid & 0x55555555;
3547         unsigned b = pid & 0xAAAAAAAA;
3548 
3549         return (a << 1) | (b >> 1);
3550 }
3551 
3552 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3553 {
3554         if (cgroup_sane_behavior(cgrp))
3555                 return pid_fry(pid);
3556         else
3557                 return pid;
3558 }
3559 
3560 static int cmppid(const void *a, const void *b)
3561 {
3562         return *(pid_t *)a - *(pid_t *)b;
3563 }
3564 
3565 static int fried_cmppid(const void *a, const void *b)
3566 {
3567         return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3568 }
3569 
3570 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3571                                                   enum cgroup_filetype type)
3572 {
3573         struct cgroup_pidlist *l;
3574         /* don't need task_nsproxy() if we're looking at ourself */
3575         struct pid_namespace *ns = task_active_pid_ns(current);
3576 
3577         lockdep_assert_held(&cgrp->pidlist_mutex);
3578 
3579         list_for_each_entry(l, &cgrp->pidlists, links)
3580                 if (l->key.type == type && l->key.ns == ns)
3581                         return l;
3582         return NULL;
3583 }
3584 
3585 /*
3586  * find the appropriate pidlist for our purpose (given procs vs tasks)
3587  * returns with the lock on that pidlist already held, and takes care
3588  * of the use count, or returns NULL with no locks held if we're out of
3589  * memory.
3590  */
3591 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3592                                                 enum cgroup_filetype type)
3593 {
3594         struct cgroup_pidlist *l;
3595 
3596         lockdep_assert_held(&cgrp->pidlist_mutex);
3597 
3598         l = cgroup_pidlist_find(cgrp, type);
3599         if (l)
3600                 return l;
3601 
3602         /* entry not found; create a new one */
3603         l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3604         if (!l)
3605                 return l;
3606 
3607         INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3608         l->key.type = type;
3609         /* don't need task_nsproxy() if we're looking at ourself */
3610         l->key.ns = get_pid_ns(task_active_pid_ns(current));
3611         l->owner = cgrp;
3612         list_add(&l->links, &cgrp->pidlists);
3613         return l;
3614 }
3615 
3616 /*
3617  * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3618  */
3619 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3620                               struct cgroup_pidlist **lp)
3621 {
3622         pid_t *array;
3623         int length;
3624         int pid, n = 0; /* used for populating the array */
3625         struct css_task_iter it;
3626         struct task_struct *tsk;
3627         struct cgroup_pidlist *l;
3628 
3629         lockdep_assert_held(&cgrp->pidlist_mutex);
3630 
3631         /*
3632          * If cgroup gets more users after we read count, we won't have
3633          * enough space - tough.  This race is indistinguishable to the
3634          * caller from the case that the additional cgroup users didn't
3635          * show up until sometime later on.
3636          */
3637         length = cgroup_task_count(cgrp);
3638         array = pidlist_allocate(length);
3639         if (!array)
3640                 return -ENOMEM;
3641         /* now, populate the array */
3642         css_task_iter_start(&cgrp->dummy_css, &it);
3643         while ((tsk = css_task_iter_next(&it))) {
3644                 if (unlikely(n == length))
3645                         break;
3646                 /* get tgid or pid for procs or tasks file respectively */
3647                 if (type == CGROUP_FILE_PROCS)
3648                         pid = task_tgid_vnr(tsk);
3649                 else
3650                         pid = task_pid_vnr(tsk);
3651                 if (pid > 0) /* make sure to only use valid results */
3652                         array[n++] = pid;
3653         }
3654         css_task_iter_end(&it);
3655         length = n;
3656         /* now sort & (if procs) strip out duplicates */
3657         if (cgroup_sane_behavior(cgrp))
3658                 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
3659         else
3660                 sort(array, length, sizeof(pid_t), cmppid, NULL);
3661         if (type == CGROUP_FILE_PROCS)
3662                 length = pidlist_uniq(array, length);
3663 
3664         l = cgroup_pidlist_find_create(cgrp, type);
3665         if (!l) {
3666                 mutex_unlock(&cgrp->pidlist_mutex);
3667                 pidlist_free(array);
3668                 return -ENOMEM;
3669         }
3670 
3671         /* store array, freeing old if necessary */
3672         pidlist_free(l->list);
3673         l->list = array;
3674         l->length = length;
3675         *lp = l;
3676         return 0;
3677 }
3678 
3679 /**
3680  * cgroupstats_build - build and fill cgroupstats
3681  * @stats: cgroupstats to fill information into
3682  * @dentry: A dentry entry belonging to the cgroup for which stats have
3683  * been requested.
3684  *
3685  * Build and fill cgroupstats so that taskstats can export it to user
3686  * space.
3687  */
3688 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
3689 {
3690         int ret = -EINVAL;
3691         struct cgroup *cgrp;
3692         struct css_task_iter it;
3693         struct task_struct *tsk;
3694 
3695         /*
3696          * Validate dentry by checking the superblock operations,
3697          * and make sure it's a directory.
3698          */
3699         if (dentry->d_sb->s_op != &cgroup_ops ||
3700             !S_ISDIR(dentry->d_inode->i_mode))
3701                  goto err;
3702 
3703         ret = 0;
3704         cgrp = dentry->d_fsdata;
3705 
3706         css_task_iter_start(&cgrp->dummy_css, &it);
3707         while ((tsk = css_task_iter_next(&it))) {
3708                 switch (tsk->state) {
3709                 case TASK_RUNNING:
3710                         stats->nr_running++;
3711                         break;
3712                 case TASK_INTERRUPTIBLE:
3713                         stats->nr_sleeping++;
3714                         break;
3715                 case TASK_UNINTERRUPTIBLE:
3716                         stats->nr_uninterruptible++;
3717                         break;
3718                 case TASK_STOPPED:
3719                         stats->nr_stopped++;
3720                         break;
3721                 default:
3722                         if (delayacct_is_task_waiting_on_io(tsk))
3723                                 stats->nr_io_wait++;
3724                         break;
3725                 }
3726         }
3727         css_task_iter_end(&it);
3728 
3729 err:
3730         return ret;
3731 }
3732 
3733 
3734 /*
3735  * seq_file methods for the tasks/procs files. The seq_file position is the
3736  * next pid to display; the seq_file iterator is a pointer to the pid
3737  * in the cgroup->l->list array.
3738  */
3739 
3740 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
3741 {
3742         /*
3743          * Initially we receive a position value that corresponds to
3744          * one more than the last pid shown (or 0 on the first call or
3745          * after a seek to the start). Use a binary-search to find the
3746          * next pid to display, if any
3747          */
3748         struct cgroup_open_file *of = s->private;
3749         struct cgroup *cgrp = seq_css(s)->cgroup;
3750         struct cgroup_pidlist *l;
3751         enum cgroup_filetype type = seq_cft(s)->private;
3752         int index = 0, pid = *pos;
3753         int *iter, ret;
3754 
3755         mutex_lock(&cgrp->pidlist_mutex);
3756 
3757         /*
3758          * !NULL @of->priv indicates that this isn't the first start()
3759          * after open.  If the matching pidlist is around, we can use that.
3760          * Look for it.  Note that @of->priv can't be used directly.  It
3761          * could already have been destroyed.
3762          */
3763         if (of->priv)
3764                 of->priv = cgroup_pidlist_find(cgrp, type);
3765 
3766         /*
3767          * Either this is the first start() after open or the matching
3768          * pidlist has been destroyed inbetween.  Create a new one.
3769          */
3770         if (!of->priv) {
3771                 ret = pidlist_array_load(cgrp, type,
3772                                          (struct cgroup_pidlist **)&of->priv);
3773                 if (ret)
3774                         return ERR_PTR(ret);
3775         }
3776         l = of->priv;
3777 
3778         if (pid) {
3779                 int end = l->length;
3780 
3781                 while (index < end) {
3782                         int mid = (index + end) / 2;
3783                         if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
3784                                 index = mid;
3785                                 break;
3786                         } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
3787                                 index = mid + 1;
3788                         else
3789                                 end = mid;
3790                 }
3791         }
3792         /* If we're off the end of the array, we're done */
3793         if (index >= l->length)
3794                 return NULL;
3795         /* Update the abstract position to be the actual pid that we found */
3796         iter = l->list + index;
3797         *pos = cgroup_pid_fry(cgrp, *iter);
3798         return iter;
3799 }
3800 
3801 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
3802 {
3803         struct cgroup_open_file *of = s->private;
3804         struct cgroup_pidlist *l = of->priv;
3805 
3806         if (l)
3807                 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
3808                                  CGROUP_PIDLIST_DESTROY_DELAY);
3809         mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
3810 }
3811 
3812 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
3813 {
3814         struct cgroup_open_file *of = s->private;
3815         struct cgroup_pidlist *l = of->priv;
3816         pid_t *p = v;
3817         pid_t *end = l->list + l->length;
3818         /*
3819          * Advance to the next pid in the array. If this goes off the
3820          * end, we're done
3821          */
3822         p++;
3823         if (p >= end) {
3824                 return NULL;
3825         } else {
3826                 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
3827                 return p;
3828         }
3829 }
3830 
3831 static int cgroup_pidlist_show(struct seq_file *s, void *v)
3832 {
3833         return seq_printf(s, "%d\n", *(int *)v);
3834 }
3835 
3836 /*
3837  * seq_operations functions for iterating on pidlists through seq_file -
3838  * independent of whether it's tasks or procs
3839  */
3840 static const struct seq_operations cgroup_pidlist_seq_operations = {
3841         .start = cgroup_pidlist_start,
3842         .stop = cgroup_pidlist_stop,
3843         .next = cgroup_pidlist_next,
3844         .show = cgroup_pidlist_show,
3845 };
3846 
3847 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
3848                                          struct cftype *cft)
3849 {
3850         return notify_on_release(css->cgroup);
3851 }
3852 
3853 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
3854                                           struct cftype *cft, u64 val)
3855 {
3856         clear_bit(CGRP_RELEASABLE, &css->cgroup->flags);
3857         if (val)
3858                 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3859         else
3860                 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
3861         return 0;
3862 }
3863 
3864 /*
3865  * When dput() is called asynchronously, if umount has been done and
3866  * then deactivate_super() in cgroup_free_fn() kills the superblock,
3867  * there's a small window that vfs will see the root dentry with non-zero
3868  * refcnt and trigger BUG().
3869  *
3870  * That's why we hold a reference before dput() and drop it right after.
3871  */
3872 static void cgroup_dput(struct cgroup *cgrp)
3873 {
3874         struct super_block *sb = cgrp->root->sb;
3875 
3876         atomic_inc(&sb->s_active);
3877         dput(cgrp->dentry);
3878         deactivate_super(sb);
3879 }
3880 
3881 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
3882                                       struct cftype *cft)
3883 {
3884         return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3885 }
3886 
3887 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
3888                                        struct cftype *cft, u64 val)
3889 {
3890         if (val)
3891                 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3892         else
3893                 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
3894         return 0;
3895 }
3896 
3897 static struct cftype cgroup_base_files[] = {
3898         {
3899                 .name = "cgroup.procs",
3900                 .seq_start = cgroup_pidlist_start,
3901                 .seq_next = cgroup_pidlist_next,
3902                 .seq_stop = cgroup_pidlist_stop,
3903                 .seq_show = cgroup_pidlist_show,
3904                 .private = CGROUP_FILE_PROCS,
3905                 .write_u64 = cgroup_procs_write,
3906                 .mode = S_IRUGO | S_IWUSR,
3907         },
3908         {
3909                 .name = "cgroup.clone_children",
3910                 .flags = CFTYPE_INSANE,
3911                 .read_u64 = cgroup_clone_children_read,
3912                 .write_u64 = cgroup_clone_children_write,
3913         },
3914         {
3915                 .name = "cgroup.sane_behavior",
3916                 .flags = CFTYPE_ONLY_ON_ROOT,
3917                 .seq_show = cgroup_sane_behavior_show,
3918         },
3919 
3920         /*
3921          * Historical crazy stuff.  These don't have "cgroup."  prefix and
3922          * don't exist if sane_behavior.  If you're depending on these, be
3923          * prepared to be burned.
3924          */
3925         {
3926                 .name = "tasks",
3927                 .flags = CFTYPE_INSANE,         /* use "procs" instead */
3928                 .seq_start = cgroup_pidlist_start,
3929                 .seq_next = cgroup_pidlist_next,
3930                 .seq_stop = cgroup_pidlist_stop,
3931                 .seq_show = cgroup_pidlist_show,
3932                 .private = CGROUP_FILE_TASKS,
3933                 .write_u64 = cgroup_tasks_write,
3934                 .mode = S_IRUGO | S_IWUSR,
3935         },
3936         {
3937                 .name = "notify_on_release",
3938                 .flags = CFTYPE_INSANE,
3939                 .read_u64 = cgroup_read_notify_on_release,
3940                 .write_u64 = cgroup_write_notify_on_release,
3941         },
3942         {
3943                 .name = "release_agent",
3944                 .flags = CFTYPE_INSANE | CFTYPE_ONLY_ON_ROOT,
3945                 .seq_show = cgroup_release_agent_show,
3946                 .write_string = cgroup_release_agent_write,
3947                 .max_write_len = PATH_MAX,
3948         },
3949         { }     /* terminate */
3950 };
3951 
3952 /**
3953  * cgroup_populate_dir - create subsys files in a cgroup directory
3954  * @cgrp: target cgroup
3955  * @subsys_mask: mask of the subsystem ids whose files should be added
3956  *
3957  * On failure, no file is added.
3958  */
3959 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
3960 {
3961         struct cgroup_subsys *ss;
3962         int i, ret = 0;
3963 
3964         /* process cftsets of each subsystem */
3965         for_each_subsys(ss, i) {
3966                 struct cftype_set *set;
3967 
3968                 if (!test_bit(i, &subsys_mask))
3969                         continue;
3970 
3971                 list_for_each_entry(set, &ss->cftsets, node) {
3972                         ret = cgroup_addrm_files(cgrp, set->cfts, true);
3973                         if (ret < 0)
3974                                 goto err;
3975                 }
3976         }
3977         return 0;
3978 err:
3979         cgroup_clear_dir(cgrp, subsys_mask);
3980         return ret;
3981 }
3982 
3983 /*
3984  * css destruction is four-stage process.
3985  *
3986  * 1. Destruction starts.  Killing of the percpu_ref is initiated.
3987  *    Implemented in kill_css().
3988  *
3989  * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
3990  *    and thus css_tryget() is guaranteed to fail, the css can be offlined
3991  *    by invoking offline_css().  After offlining, the base ref is put.
3992  *    Implemented in css_killed_work_fn().
3993  *
3994  * 3. When the percpu_ref reaches zero, the only possible remaining
3995  *    accessors are inside RCU read sections.  css_release() schedules the
3996  *    RCU callback.
3997  *
3998  * 4. After the grace period, the css can be freed.  Implemented in
3999  *    css_free_work_fn().
4000  *
4001  * It is actually hairier because both step 2 and 4 require process context
4002  * and thus involve punting to css->destroy_work adding two additional
4003  * steps to the already complex sequence.
4004  */
4005 static void css_free_work_fn(struct work_struct *work)
4006 {
4007         struct cgroup_subsys_state *css =
4008                 container_of(work, struct cgroup_subsys_state, destroy_work);
4009         struct cgroup *cgrp = css->cgroup;
4010 
4011         if (css->parent)
4012                 css_put(css->parent);
4013 
4014         css->ss->css_free(css);
4015         cgroup_dput(cgrp);
4016 }
4017 
4018 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4019 {
4020         struct cgroup_subsys_state *css =
4021                 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4022 
4023         /*
4024          * css holds an extra ref to @cgrp->dentry which is put on the last
4025          * css_put().  dput() requires process context which we don't have.
4026          */
4027         INIT_WORK(&css->destroy_work, css_free_work_fn);
4028         queue_work(cgroup_destroy_wq, &css->destroy_work);
4029 }
4030 
4031 static void css_release(struct percpu_ref *ref)
4032 {
4033         struct cgroup_subsys_state *css =
4034                 container_of(ref, struct cgroup_subsys_state, refcnt);
4035 
4036         rcu_assign_pointer(css->cgroup->subsys[css->ss->subsys_id], NULL);
4037         call_rcu(&css->rcu_head, css_free_rcu_fn);
4038 }
4039 
4040 static void init_css(struct cgroup_subsys_state *css, struct cgroup_subsys *ss,
4041                      struct cgroup *cgrp)
4042 {
4043         css->cgroup = cgrp;
4044         css->ss = ss;
4045         css->flags = 0;
4046 
4047         if (cgrp->parent)
4048                 css->parent = cgroup_css(cgrp->parent, ss);
4049         else
4050                 css->flags |= CSS_ROOT;
4051 
4052         BUG_ON(cgroup_css(cgrp, ss));
4053 }
4054 
4055 /* invoke ->css_online() on a new CSS and mark it online if successful */
4056 static int online_css(struct cgroup_subsys_state *css)
4057 {
4058         struct cgroup_subsys *ss = css->ss;
4059         int ret = 0;
4060 
4061         lockdep_assert_held(&cgroup_mutex);
4062 
4063         if (ss->css_online)
4064                 ret = ss->css_online(css);
4065         if (!ret) {
4066                 css->flags |= CSS_ONLINE;
4067                 css->cgroup->nr_css++;
4068                 rcu_assign_pointer(css->cgroup->subsys[ss->subsys_id], css);
4069         }
4070         return ret;
4071 }
4072 
4073 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4074 static void offline_css(struct cgroup_subsys_state *css)
4075 {
4076         struct cgroup_subsys *ss = css->ss;
4077 
4078         lockdep_assert_held(&cgroup_mutex);
4079 
4080         if (!(css->flags & CSS_ONLINE))
4081                 return;
4082 
4083         if (ss->css_offline)
4084                 ss->css_offline(css);
4085 
4086         css->flags &= ~CSS_ONLINE;
4087         css->cgroup->nr_css--;
4088         RCU_INIT_POINTER(css->cgroup->subsys[ss->subsys_id], css);
4089 }
4090 
4091 /**
4092  * create_css - create a cgroup_subsys_state
4093  * @cgrp: the cgroup new css will be associated with
4094  * @ss: the subsys of new css
4095  *
4096  * Create a new css associated with @cgrp - @ss pair.  On success, the new
4097  * css is online and installed in @cgrp with all interface files created.
4098  * Returns 0 on success, -errno on failure.
4099  */
4100 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss)
4101 {
4102         struct cgroup *parent = cgrp->parent;
4103         struct cgroup_subsys_state *css;
4104         int err;
4105 
4106         lockdep_assert_held(&cgrp->dentry->d_inode->i_mutex);
4107         lockdep_assert_held(&cgroup_mutex);
4108 
4109         css = ss->css_alloc(cgroup_css(parent, ss));
4110         if (IS_ERR(css))
4111                 return PTR_ERR(css);
4112 
4113         err = percpu_ref_init(&css->refcnt, css_release);
4114         if (err)
4115                 goto err_free_css;
4116 
4117         init_css(css, ss, cgrp);
4118 
4119         err = cgroup_populate_dir(cgrp, 1 << ss->subsys_id);
4120         if (err)
4121                 goto err_free_percpu_ref;
4122 
4123         err = online_css(css);
4124         if (err)
4125                 goto err_clear_dir;
4126 
4127         dget(cgrp->dentry);
4128         css_get(css->parent);
4129 
4130         if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4131             parent->parent) {
4132                 pr_warning("cgroup: %s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4133                            current->comm, current->pid, ss->name);
4134                 if (!strcmp(ss->name, "memory"))
4135                         pr_warning("cgroup: \"memory\" requires setting use_hierarchy to 1 on the root.\n");
4136                 ss->warned_broken_hierarchy = true;
4137         }
4138 
4139         return 0;
4140 
4141 err_clear_dir:
4142         cgroup_clear_dir(css->cgroup, 1 << css->ss->subsys_id);
4143 err_free_percpu_ref:
4144         percpu_ref_cancel_init(&css->refcnt);
4145 err_free_css:
4146         ss->css_free(css);
4147         return err;
4148 }
4149 
4150 /*
4151  * cgroup_create - create a cgroup
4152  * @parent: cgroup that will be parent of the new cgroup
4153  * @dentry: dentry of the new cgroup
4154  * @mode: mode to set on new inode
4155  *
4156  * Must be called with the mutex on the parent inode held
4157  */
4158 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
4159                              umode_t mode)
4160 {
4161         struct cgroup *cgrp;
4162         struct cgroup_name *name;
4163         struct cgroupfs_root *root = parent->root;
4164         int ssid, err;
4165         struct cgroup_subsys *ss;
4166         struct super_block *sb = root->sb;
4167 
4168         /* allocate the cgroup and its ID, 0 is reserved for the root */
4169         cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4170         if (!cgrp)
4171                 return -ENOMEM;
4172 
4173         name = cgroup_alloc_name(dentry);
4174         if (!name) {
4175                 err = -ENOMEM;
4176                 goto err_free_cgrp;
4177         }
4178         rcu_assign_pointer(cgrp->name, name);
4179 
4180         /*
4181          * Only live parents can have children.  Note that the liveliness
4182          * check isn't strictly necessary because cgroup_mkdir() and
4183          * cgroup_rmdir() are fully synchronized by i_mutex; however, do it
4184          * anyway so that locking is contained inside cgroup proper and we
4185          * don't get nasty surprises if we ever grow another caller.
4186          */
4187         if (!cgroup_lock_live_group(parent)) {
4188                 err = -ENODEV;
4189                 goto err_free_name;
4190         }
4191 
4192         /*
4193          * Temporarily set the pointer to NULL, so idr_find() won't return
4194          * a half-baked cgroup.
4195          */
4196         cgrp->id = idr_alloc(&root->cgroup_idr, NULL, 1, 0, GFP_KERNEL);
4197         if (cgrp->id < 0) {
4198                 err = -ENOMEM;
4199                 goto err_unlock;
4200         }
4201 
4202         /* Grab a reference on the superblock so the hierarchy doesn't
4203          * get deleted on unmount if there are child cgroups.  This
4204          * can be done outside cgroup_mutex, since the sb can't
4205          * disappear while someone has an open control file on the
4206          * fs */
4207         atomic_inc(&sb->s_active);
4208 
4209         init_cgroup_housekeeping(cgrp);
4210 
4211         dentry->d_fsdata = cgrp;
4212         cgrp->dentry = dentry;
4213 
4214         cgrp->parent = parent;
4215         cgrp->dummy_css.parent = &parent->dummy_css;
4216         cgrp->root = parent->root;
4217 
4218         if (notify_on_release(parent))
4219                 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4220 
4221         if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4222                 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4223 
4224         /*
4225          * Create directory.  cgroup_create_file() returns with the new
4226          * directory locked on success so that it can be populated without
4227          * dropping cgroup_mutex.
4228          */
4229         err = cgroup_create_file(dentry, S_IFDIR | mode, sb);
4230         if (err < 0)
4231                 goto err_free_id;
4232         lockdep_assert_held(&dentry->d_inode->i_mutex);
4233 
4234         cgrp->serial_nr = cgroup_serial_nr_next++;
4235 
4236         /* allocation complete, commit to creation */
4237         list_add_tail_rcu(&cgrp->sibling, &cgrp->parent->children);
4238         root->number_of_cgroups++;
4239 
4240         /* hold a ref to the parent's dentry */
4241         dget(parent->dentry);
4242 
4243         /*
4244          * @cgrp is now fully operational.  If something fails after this
4245          * point, it'll be released via the normal destruction path.
4246          */
4247         idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4248 
4249         err = cgroup_addrm_files(cgrp, cgroup_base_files, true);
4250         if (err)
4251                 goto err_destroy;
4252 
4253         /* let's create and online css's */
4254         for_each_subsys(ss, ssid) {
4255                 if (root->subsys_mask & (1 << ssid)) {
4256                         err = create_css(cgrp, ss);
4257                         if (err)
4258                                 goto err_destroy;
4259                 }
4260         }
4261 
4262         mutex_unlock(&cgroup_mutex);
4263         mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
4264 
4265         return 0;
4266 
4267 err_free_id:
4268         idr_remove(&root->cgroup_idr, cgrp->id);
4269         /* Release the reference count that we took on the superblock */
4270         deactivate_super(sb);
4271 err_unlock:
4272         mutex_unlock(&cgroup_mutex);
4273 err_free_name:
4274         kfree(rcu_dereference_raw(cgrp->name));
4275 err_free_cgrp:
4276         kfree(cgrp);
4277         return err;
4278 
4279 err_destroy:
4280         cgroup_destroy_locked(cgrp);
4281         mutex_unlock(&cgroup_mutex);
4282         mutex_unlock(&dentry->d_inode->i_mutex);
4283         return err;
4284 }
4285 
4286 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
4287 {
4288         struct cgroup *c_parent = dentry->d_parent->d_fsdata;
4289 
4290         /* the vfs holds inode->i_mutex already */
4291         return cgroup_create(c_parent, dentry, mode | S_IFDIR);
4292 }
4293 
4294 /*
4295  * This is called when the refcnt of a css is confirmed to be killed.
4296  * css_tryget() is now guaranteed to fail.
4297  */
4298 static void css_killed_work_fn(struct work_struct *work)
4299 {
4300         struct cgroup_subsys_state *css =
4301                 container_of(work, struct cgroup_subsys_state, destroy_work);
4302         struct cgroup *cgrp = css->cgroup;
4303 
4304         mutex_lock(&cgroup_mutex);
4305 
4306         /*
4307          * css_tryget() is guaranteed to fail now.  Tell subsystems to
4308          * initate destruction.
4309          */
4310         offline_css(css);
4311 
4312         /*
4313          * If @cgrp is marked dead, it's waiting for refs of all css's to
4314          * be disabled before proceeding to the second phase of cgroup
4315          * destruction.  If we are the last one, kick it off.
4316          */
4317         if (!cgrp->nr_css && cgroup_is_dead(cgrp))
4318                 cgroup_destroy_css_killed(cgrp);
4319 
4320         mutex_unlock(&cgroup_mutex);
4321 
4322         /*
4323          * Put the css refs from kill_css().  Each css holds an extra
4324          * reference to the cgroup's dentry and cgroup removal proceeds
4325          * regardless of css refs.  On the last put of each css, whenever
4326          * that may be, the extra dentry ref is put so that dentry
4327          * destruction happens only after all css's are released.
4328          */
4329         css_put(css);
4330 }
4331 
4332 /* css kill confirmation processing requires process context, bounce */
4333 static void css_killed_ref_fn(struct percpu_ref *ref)
4334 {
4335         struct cgroup_subsys_state *css =
4336                 container_of(ref, struct cgroup_subsys_state, refcnt);
4337 
4338         INIT_WORK(&css->destroy_work, css_killed_work_fn);
4339         queue_work(cgroup_destroy_wq, &css->destroy_work);
4340 }
4341 
4342 /**
4343  * kill_css - destroy a css
4344  * @css: css to destroy
4345  *
4346  * This function initiates destruction of @css by removing cgroup interface
4347  * files and putting its base reference.  ->css_offline() will be invoked
4348  * asynchronously once css_tryget() is guaranteed to fail and when the
4349  * reference count reaches zero, @css will be released.
4350  */
4351 static void kill_css(struct cgroup_subsys_state *css)
4352 {
4353         cgroup_clear_dir(css->cgroup, 1 << css->ss->subsys_id);
4354 
4355         /*
4356          * Killing would put the base ref, but we need to keep it alive
4357          * until after ->css_offline().
4358          */
4359         css_get(css);
4360 
4361         /*
4362          * cgroup core guarantees that, by the time ->css_offline() is
4363          * invoked, no new css reference will be given out via
4364          * css_tryget().  We can't simply call percpu_ref_kill() and
4365          * proceed to offlining css's because percpu_ref_kill() doesn't
4366          * guarantee that the ref is seen as killed on all CPUs on return.
4367          *
4368          * Use percpu_ref_kill_and_confirm() to get notifications as each
4369          * css is confirmed to be seen as killed on all CPUs.
4370          */
4371         percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4372 }
4373 
4374 /**
4375  * cgroup_destroy_locked - the first stage of cgroup destruction
4376  * @cgrp: cgroup to be destroyed
4377  *
4378  * css's make use of percpu refcnts whose killing latency shouldn't be
4379  * exposed to userland and are RCU protected.  Also, cgroup core needs to
4380  * guarantee that css_tryget() won't succeed by the time ->css_offline() is
4381  * invoked.  To satisfy all the requirements, destruction is implemented in
4382  * the following two steps.
4383  *
4384  * s1. Verify @cgrp can be destroyed and mark it dying.  Remove all
4385  *     userland visible parts and start killing the percpu refcnts of
4386  *     css's.  Set up so that the next stage will be kicked off once all
4387  *     the percpu refcnts are confirmed to be killed.
4388  *
4389  * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4390  *     rest of destruction.  Once all cgroup references are gone, the
4391  *     cgroup is RCU-freed.
4392  *
4393  * This function implements s1.  After this step, @cgrp is gone as far as
4394  * the userland is concerned and a new cgroup with the same name may be
4395  * created.  As cgroup doesn't care about the names internally, this
4396  * doesn't cause any problem.
4397  */
4398 static int cgroup_destroy_locked(struct cgroup *cgrp)
4399         __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4400 {
4401         struct dentry *d = cgrp->dentry;
4402         struct cgroup_subsys_state *css;
4403         struct cgroup *child;
4404         bool empty;
4405         int ssid;
4406 
4407         lockdep_assert_held(&d->d_inode->i_mutex);
4408         lockdep_assert_held(&cgroup_mutex);
4409 
4410         /*
4411          * css_set_lock synchronizes access to ->cset_links and prevents
4412          * @cgrp from being removed while __put_css_set() is in progress.
4413          */
4414         read_lock(&css_set_lock);
4415         empty = list_empty(&cgrp->cset_links);
4416         read_unlock(&css_set_lock);
4417         if (!empty)
4418                 return -EBUSY;
4419 
4420         /*
4421          * Make sure there's no live children.  We can't test ->children
4422          * emptiness as dead children linger on it while being destroyed;
4423          * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
4424          */
4425         empty = true;
4426         rcu_read_lock();
4427         list_for_each_entry_rcu(child, &cgrp->children, sibling) {
4428                 empty = cgroup_is_dead(child);
4429                 if (!empty)
4430                         break;
4431         }
4432         rcu_read_unlock();
4433         if (!empty)
4434                 return -EBUSY;
4435 
4436         /*
4437          * Initiate massacre of all css's.  cgroup_destroy_css_killed()
4438          * will be invoked to perform the rest of destruction once the
4439          * percpu refs of all css's are confirmed to be killed.
4440          */
4441         for_each_css(css, ssid, cgrp)
4442                 kill_css(css);
4443 
4444         /*
4445          * Mark @cgrp dead.  This prevents further task migration and child
4446          * creation by disabling cgroup_lock_live_group().  Note that
4447          * CGRP_DEAD assertion is depended upon by css_next_child() to
4448          * resume iteration after dropping RCU read lock.  See
4449          * css_next_child() for details.
4450          */
4451         set_bit(CGRP_DEAD, &cgrp->flags);
4452 
4453         /* CGRP_DEAD is set, remove from ->release_list for the last time */
4454         raw_spin_lock(&release_list_lock);
4455         if (!list_empty(&cgrp->release_list))
4456                 list_del_init(&cgrp->release_list);
4457         raw_spin_unlock(&release_list_lock);
4458 
4459         /*
4460          * If @cgrp has css's attached, the second stage of cgroup
4461          * destruction is kicked off from css_killed_work_fn() after the
4462          * refs of all attached css's are killed.  If @cgrp doesn't have
4463          * any css, we kick it off here.
4464          */
4465         if (!cgrp->nr_css)
4466                 cgroup_destroy_css_killed(cgrp);
4467 
4468         /*
4469          * Clear the base files and remove @cgrp directory.  The removal
4470          * puts the base ref but we aren't quite done with @cgrp yet, so
4471          * hold onto it.
4472          */
4473         cgroup_addrm_files(cgrp, cgroup_base_files, false);
4474         dget(d);
4475         cgroup_d_remove_dir(d);
4476 
4477         return 0;
4478 };
4479 
4480 /**
4481  * cgroup_destroy_css_killed - the second step of cgroup destruction
4482  * @work: cgroup->destroy_free_work
4483  *
4484  * This function is invoked from a work item for a cgroup which is being
4485  * destroyed after all css's are offlined and performs the rest of
4486  * destruction.  This is the second step of destruction described in the
4487  * comment above cgroup_destroy_locked().
4488  */
4489 static void cgroup_destroy_css_killed(struct cgroup *cgrp)
4490 {
4491         struct cgroup *parent = cgrp->parent;
4492         struct dentry *d = cgrp->dentry;
4493 
4494         lockdep_assert_held(&cgroup_mutex);
4495 
4496         /* delete this cgroup from parent->children */
4497         list_del_rcu(&cgrp->sibling);
4498 
4499         dput(d);
4500 
4501         set_bit(CGRP_RELEASABLE, &parent->flags);
4502         check_for_release(parent);
4503 }
4504 
4505 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
4506 {
4507         int ret;
4508 
4509         mutex_lock(&cgroup_mutex);
4510         ret = cgroup_destroy_locked(dentry->d_fsdata);
4511         mutex_unlock(&cgroup_mutex);
4512 
4513         return ret;
4514 }
4515 
4516 static void __init_or_module cgroup_init_cftsets(struct cgroup_subsys *ss)
4517 {
4518         INIT_LIST_HEAD(&ss->cftsets);
4519 
4520         /*
4521          * base_cftset is embedded in subsys itself, no need to worry about
4522          * deregistration.
4523          */
4524         if (ss->base_cftypes) {
4525                 struct cftype *cft;
4526 
4527                 for (cft = ss->base_cftypes; cft->name[0] != '\0'; cft++)
4528                         cft->ss = ss;
4529 
4530                 ss->base_cftset.cfts = ss->base_cftypes;
4531                 list_add_tail(&ss->base_cftset.node, &ss->cftsets);
4532         }
4533 }
4534 
4535 static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
4536 {
4537         struct cgroup_subsys_state *css;
4538 
4539         printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4540 
4541         mutex_lock(&cgroup_mutex);
4542 
4543         /* init base cftset */
4544         cgroup_init_cftsets(ss);
4545 
4546         /* Create the top cgroup state for this subsystem */
4547         ss->root = &cgroup_dummy_root;
4548         css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4549         /* We don't handle early failures gracefully */
4550         BUG_ON(IS_ERR(css));
4551         init_css(css, ss, cgroup_dummy_top);
4552 
4553         /* Update the init_css_set to contain a subsys
4554          * pointer to this state - since the subsystem is
4555          * newly registered, all tasks and hence the
4556          * init_css_set is in the subsystem's top cgroup. */
4557         init_css_set.subsys[ss->subsys_id] = css;
4558 
4559         need_forkexit_callback |= ss->fork || ss->exit;
4560 
4561         /* At system boot, before all subsystems have been
4562          * registered, no tasks have been forked, so we don't
4563          * need to invoke fork callbacks here. */
4564         BUG_ON(!list_empty(&init_task.tasks));
4565 
4566         BUG_ON(online_css(css));
4567 
4568         mutex_unlock(&cgroup_mutex);
4569 
4570         /* this function shouldn't be used with modular subsystems, since they
4571          * need to register a subsys_id, among other things */
4572         BUG_ON(ss->module);
4573 }
4574 
4575 /**
4576  * cgroup_load_subsys: load and register a modular subsystem at runtime
4577  * @ss: the subsystem to load
4578  *
4579  * This function should be called in a modular subsystem's initcall. If the
4580  * subsystem is built as a module, it will be assigned a new subsys_id and set
4581  * up for use. If the subsystem is built-in anyway, work is delegated to the
4582  * simpler cgroup_init_subsys.
4583  */
4584 int __init_or_module cgroup_load_subsys(struct cgroup_subsys *ss)
4585 {
4586         struct cgroup_subsys_state *css;
4587         int i, ret;
4588         struct hlist_node *tmp;
4589         struct css_set *cset;
4590         unsigned long key;
4591 
4592         /* check name and function validity */
4593         if (ss->name == NULL || strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN ||
4594             ss->css_alloc == NULL || ss->css_free == NULL)
4595                 return -EINVAL;
4596 
4597         /*
4598          * we don't support callbacks in modular subsystems. this check is
4599          * before the ss->module check for consistency; a subsystem that could
4600          * be a module should still have no callbacks even if the user isn't
4601          * compiling it as one.
4602          */
4603         if (ss->fork || ss->exit)
4604                 return -EINVAL;
4605 
4606         /*
4607          * an optionally modular subsystem is built-in: we want to do nothing,
4608          * since cgroup_init_subsys will have already taken care of it.
4609          */
4610         if (ss->module == NULL) {
4611                 /* a sanity check */
4612                 BUG_ON(cgroup_subsys[ss->subsys_id] != ss);
4613                 return 0;
4614         }
4615 
4616         /* init base cftset */
4617         cgroup_init_cftsets(ss);
4618 
4619         mutex_lock(&cgroup_mutex);
4620         mutex_lock(&cgroup_root_mutex);
4621         cgroup_subsys[ss->subsys_id] = ss;
4622 
4623         /*
4624          * no ss->css_alloc seems to need anything important in the ss
4625          * struct, so this can happen first (i.e. before the dummy root
4626          * attachment).
4627          */
4628         css = ss->css_alloc(cgroup_css(cgroup_dummy_top, ss));
4629         if (IS_ERR(css)) {
4630                 /* failure case - need to deassign the cgroup_subsys[] slot. */
4631                 cgroup_subsys[ss->subsys_id] = NULL;
4632                 mutex_unlock(&cgroup_root_mutex);
4633                 mutex_unlock(&cgroup_mutex);
4634                 return PTR_ERR(css);
4635         }
4636 
4637         ss->root = &cgroup_dummy_root;
4638 
4639         /* our new subsystem will be attached to the dummy hierarchy. */
4640         init_css(css, ss, cgroup_dummy_top);
4641 
4642         /*
4643          * Now we need to entangle the css into the existing css_sets. unlike
4644          * in cgroup_init_subsys, there are now multiple css_sets, so each one
4645          * will need a new pointer to it; done by iterating the css_set_table.
4646          * furthermore, modifying the existing css_sets will corrupt the hash
4647          * table state, so each changed css_set will need its hash recomputed.
4648          * this is all done under the css_set_lock.
4649          */
4650         write_lock(&css_set_lock);
4651         hash_for_each_safe(css_set_table, i, tmp, cset, hlist) {
4652                 /* skip entries that we already rehashed */
4653                 if (cset->subsys[ss->subsys_id])
4654                         continue;
4655                 /* remove existing entry */
4656                 hash_del(&cset->hlist);
4657                 /* set new value */
4658                 cset->subsys[ss->subsys_id] = css;
4659                 /* recompute hash and restore entry */
4660                 key = css_set_hash(cset->subsys);
4661                 hash_add(css_set_table, &cset->hlist, key);
4662         }
4663         write_unlock(&css_set_lock);
4664 
4665         ret = online_css(css);
4666         if (ret) {
4667                 ss->css_free(css);
4668                 goto err_unload;
4669         }
4670 
4671         /* success! */
4672         mutex_unlock(&cgroup_root_mutex);
4673         mutex_unlock(&cgroup_mutex);
4674         return 0;
4675 
4676 err_unload:
4677         mutex_unlock(&cgroup_root_mutex);
4678         mutex_unlock(&cgroup_mutex);
4679         /* @ss can't be mounted here as try_module_get() would fail */
4680         cgroup_unload_subsys(ss);
4681         return ret;
4682 }
4683 EXPORT_SYMBOL_GPL(cgroup_load_subsys);
4684 
4685 /**
4686  * cgroup_unload_subsys: unload a modular subsystem
4687  * @ss: the subsystem to unload
4688  *
4689  * This function should be called in a modular subsystem's exitcall. When this
4690  * function is invoked, the refcount on the subsystem's module will be 0, so
4691  * the subsystem will not be attached to any hierarchy.
4692  */
4693 void cgroup_unload_subsys(struct cgroup_subsys *ss)
4694 {
4695         struct cgrp_cset_link *link;
4696         struct cgroup_subsys_state *css;
4697 
4698         BUG_ON(ss->module == NULL);
4699 
4700         /*
4701          * we shouldn't be called if the subsystem is in use, and the use of
4702          * try_module_get() in rebind_subsystems() should ensure that it
4703          * doesn't start being used while we're killing it off.
4704          */
4705         BUG_ON(ss->root != &cgroup_dummy_root);
4706 
4707         mutex_lock(&cgroup_mutex);
4708         mutex_lock(&cgroup_root_mutex);
4709 
4710         css = cgroup_css(cgroup_dummy_top, ss);
4711         if (css)
4712                 offline_css(css);
4713 
4714         /* deassign the subsys_id */
4715         cgroup_subsys[ss->subsys_id] = NULL;
4716 
4717         /*
4718          * disentangle the css from all css_sets attached to the dummy
4719          * top. as in loading, we need to pay our respects to the hashtable
4720          * gods.
4721          */
4722         write_lock(&css_set_lock);
4723         list_for_each_entry(link, &cgroup_dummy_top->cset_links, cset_link) {
4724                 struct css_set *cset = link->cset;
4725                 unsigned long key;
4726 
4727                 hash_del(&cset->hlist);
4728                 cset->subsys[ss->subsys_id] = NULL;
4729                 key = css_set_hash(cset->subsys);
4730                 hash_add(css_set_table, &cset->hlist, key);
4731         }
4732         write_unlock(&css_set_lock);
4733 
4734         /*
4735          * remove subsystem's css from the cgroup_dummy_top and free it -
4736          * need to free before marking as null because ss->css_free needs
4737          * the cgrp->subsys pointer to find their state.
4738          */
4739         if (css)
4740                 ss->css_free(css);
4741         RCU_INIT_POINTER(cgroup_dummy_top->subsys[ss->subsys_id], NULL);
4742 
4743         mutex_unlock(&cgroup_root_mutex);
4744         mutex_unlock(&cgroup_mutex);
4745 }
4746 EXPORT_SYMBOL_GPL(cgroup_unload_subsys);
4747 
4748 /**
4749  * cgroup_init_early - cgroup initialization at system boot
4750  *
4751  * Initialize cgroups at system boot, and initialize any
4752  * subsystems that request early init.
4753  */
4754 int __init cgroup_init_early(void)
4755 {
4756         struct cgroup_subsys *ss;
4757         int i;
4758 
4759         atomic_set(&init_css_set.refcount, 1);
4760         INIT_LIST_HEAD(&init_css_set.cgrp_links);
4761         INIT_LIST_HEAD(&init_css_set.tasks);
4762         INIT_HLIST_NODE(&init_css_set.hlist);
4763         css_set_count = 1;
4764         init_cgroup_root(&cgroup_dummy_root);
4765         cgroup_root_count = 1;
4766         RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4767 
4768         init_cgrp_cset_link.cset = &init_css_set;
4769         init_cgrp_cset_link.cgrp = cgroup_dummy_top;
4770         list_add(&init_cgrp_cset_link.cset_link, &cgroup_dummy_top->cset_links);
4771         list_add(&init_cgrp_cset_link.cgrp_link, &init_css_set.cgrp_links);
4772 
4773         /* at bootup time, we don't worry about modular subsystems */
4774         for_each_builtin_subsys(ss, i) {
4775                 BUG_ON(!ss->name);
4776                 BUG_ON(strlen(ss->name) > MAX_CGROUP_TYPE_NAMELEN);
4777                 BUG_ON(!ss->css_alloc);
4778                 BUG_ON(!ss->css_free);
4779                 if (ss->subsys_id != i) {
4780                         printk(KERN_ERR "cgroup: Subsys %s id == %d\n",
4781                                ss->name, ss->subsys_id);
4782                         BUG();
4783                 }
4784 
4785                 if (ss->early_init)
4786                         cgroup_init_subsys(ss);
4787         }
4788         return 0;
4789 }
4790 
4791 /**
4792  * cgroup_init - cgroup initialization
4793  *
4794  * Register cgroup filesystem and /proc file, and initialize
4795  * any subsystems that didn't request early init.
4796  */
4797 int __init cgroup_init(void)
4798 {
4799         struct cgroup_subsys *ss;
4800         unsigned long key;
4801         int i, err;
4802 
4803         err = bdi_init(&cgroup_backing_dev_info);
4804         if (err)
4805                 return err;
4806 
4807         for_each_builtin_subsys(ss, i) {
4808                 if (!ss->early_init)
4809                         cgroup_init_subsys(ss);
4810         }
4811 
4812         /* allocate id for the dummy hierarchy */
4813         mutex_lock(&cgroup_mutex);
4814         mutex_lock(&cgroup_root_mutex);
4815 
4816         /* Add init_css_set to the hash table */
4817         key = css_set_hash(init_css_set.subsys);
4818         hash_add(css_set_table, &init_css_set.hlist, key);
4819 
4820         BUG_ON(cgroup_init_root_id(&cgroup_dummy_root, 0, 1));
4821 
4822         err = idr_alloc(&cgroup_dummy_root.cgroup_idr, cgroup_dummy_top,
4823                         0, 1, GFP_KERNEL);
4824         BUG_ON(err < 0);
4825 
4826         mutex_unlock(&cgroup_root_mutex);
4827         mutex_unlock(&cgroup_mutex);
4828 
4829         cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
4830         if (!cgroup_kobj) {
4831                 err = -ENOMEM;
4832                 goto out;
4833         }
4834 
4835         err = register_filesystem(&cgroup_fs_type);
4836         if (err < 0) {
4837                 kobject_put(cgroup_kobj);
4838                 goto out;
4839         }
4840 
4841         proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
4842 
4843 out:
4844         if (err)
4845                 bdi_destroy(&cgroup_backing_dev_info);
4846 
4847         return err;
4848 }
4849 
4850 static int __init cgroup_wq_init(void)
4851 {
4852         /*
4853          * There isn't much point in executing destruction path in
4854          * parallel.  Good chunk is serialized with cgroup_mutex anyway.
4855          * Use 1 for @max_active.
4856          *
4857          * We would prefer to do this in cgroup_init() above, but that
4858          * is called before init_workqueues(): so leave this until after.
4859          */
4860         cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
4861         BUG_ON(!cgroup_destroy_wq);
4862 
4863         /*
4864          * Used to destroy pidlists and separate to serve as flush domain.
4865          * Cap @max_active to 1 too.
4866          */
4867         cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
4868                                                     0, 1);
4869         BUG_ON(!cgroup_pidlist_destroy_wq);
4870 
4871         return 0;
4872 }
4873 core_initcall(cgroup_wq_init);
4874 
4875 /*
4876  * proc_cgroup_show()
4877  *  - Print task's cgroup paths into seq_file, one line for each hierarchy
4878  *  - Used for /proc/<pid>/cgroup.
4879  *  - No need to task_lock(tsk) on this tsk->cgroup reference, as it
4880  *    doesn't really matter if tsk->cgroup changes after we read it,
4881  *    and we take cgroup_mutex, keeping cgroup_attach_task() from changing it
4882  *    anyway.  No need to check that tsk->cgroup != NULL, thanks to
4883  *    the_top_cgroup_hack in cgroup_exit(), which sets an exiting tasks
4884  *    cgroup to top_cgroup.
4885  */
4886 
4887 /* TODO: Use a proper seq_file iterator */
4888 int proc_cgroup_show(struct seq_file *m, void *v)
4889 {
4890         struct pid *pid;
4891         struct task_struct *tsk;
4892         char *buf;
4893         int retval;
4894         struct cgroupfs_root *root;
4895 
4896         retval = -ENOMEM;
4897         buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
4898         if (!buf)
4899                 goto out;
4900 
4901         retval = -ESRCH;
4902         pid = m->private;
4903         tsk = get_pid_task(pid, PIDTYPE_PID);
4904         if (!tsk)
4905                 goto out_free;
4906 
4907         retval = 0;
4908 
4909         mutex_lock(&cgroup_mutex);
4910 
4911         for_each_active_root(root) {
4912                 struct cgroup_subsys *ss;
4913                 struct cgroup *cgrp;
4914                 int ssid, count = 0;
4915 
4916                 seq_printf(m, "%d:", root->hierarchy_id);
4917                 for_each_subsys(ss, ssid)
4918                         if (root->subsys_mask & (1 << ssid))
4919                                 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
4920                 if (strlen(root->name))
4921                         seq_printf(m, "%sname=%s", count ? "," : "",
4922                                    root->name);
4923                 seq_putc(m, ':');
4924                 cgrp = task_cgroup_from_root(tsk, root);
4925                 retval = cgroup_path(cgrp, buf, PAGE_SIZE);
4926                 if (retval < 0)
4927                         goto out_unlock;
4928                 seq_puts(m, buf);
4929                 seq_putc(m, '\n');
4930         }
4931 
4932 out_unlock:
4933         mutex_unlock(&cgroup_mutex);
4934         put_task_struct(tsk);
4935 out_free:
4936         kfree(buf);
4937 out:
4938         return retval;
4939 }
4940 
4941 /* Display information about each subsystem and each hierarchy */
4942 static int proc_cgroupstats_show(struct seq_file *m, void *v)
4943 {
4944         struct cgroup_subsys *ss;
4945         int i;
4946 
4947         seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
4948         /*
4949          * ideally we don't want subsystems moving around while we do this.
4950          * cgroup_mutex is also necessary to guarantee an atomic snapshot of
4951          * subsys/hierarchy state.
4952          */
4953         mutex_lock(&cgroup_mutex);
4954 
4955         for_each_subsys(ss, i)
4956                 seq_printf(m, "%s\t%d\t%d\t%d\n",
4957                            ss->name, ss->root->hierarchy_id,
4958                            ss->root->number_of_cgroups, !ss->disabled);
4959 
4960         mutex_unlock(&cgroup_mutex);
4961         return 0;
4962 }
4963 
4964 static int cgroupstats_open(struct inode *inode, struct file *file)
4965 {
4966         return single_open(file, proc_cgroupstats_show, NULL);
4967 }
4968 
4969 static const struct file_operations proc_cgroupstats_operations = {
4970         .open = cgroupstats_open,
4971         .read = seq_read,
4972         .llseek = seq_lseek,
4973         .release = single_release,
4974 };
4975 
4976 /**
4977  * cgroup_fork - attach newly forked task to its parents cgroup.
4978  * @child: pointer to task_struct of forking parent process.
4979  *
4980  * Description: A task inherits its parent's cgroup at fork().
4981  *
4982  * A pointer to the shared css_set was automatically copied in
4983  * fork.c by dup_task_struct().  However, we ignore that copy, since
4984  * it was not made under the protection of RCU or cgroup_mutex, so
4985  * might no longer be a valid cgroup pointer.  cgroup_attach_task() might
4986  * have already changed current->cgroups, allowing the previously
4987  * referenced cgroup group to be removed and freed.
4988  *
4989  * At the point that cgroup_fork() is called, 'current' is the parent
4990  * task, and the passed argument 'child' points to the child task.
4991  */
4992 void cgroup_fork(struct task_struct *child)
4993 {
4994         task_lock(current);
4995         get_css_set(task_css_set(current));
4996         child->cgroups = current->cgroups;
4997         task_unlock(current);
4998         INIT_LIST_HEAD(&child->cg_list);
4999 }
5000 
5001 /**
5002  * cgroup_post_fork - called on a new task after adding it to the task list
5003  * @child: the task in question
5004  *
5005  * Adds the task to the list running through its css_set if necessary and
5006  * call the subsystem fork() callbacks.  Has to be after the task is
5007  * visible on the task list in case we race with the first call to
5008  * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5009  * list.
5010  */
5011 void cgroup_post_fork(struct task_struct *child)
5012 {
5013         struct cgroup_subsys *ss;
5014         int i;
5015 
5016         /*
5017          * use_task_css_set_links is set to 1 before we walk the tasklist
5018          * under the tasklist_lock and we read it here after we added the child
5019          * to the tasklist under the tasklist_lock as well. If the child wasn't
5020          * yet in the tasklist when we walked through it from
5021          * cgroup_enable_task_cg_lists(), then use_task_css_set_links value
5022          * should be visible now due to the paired locking and barriers implied
5023          * by LOCK/UNLOCK: it is written before the tasklist_lock unlock
5024          * in cgroup_enable_task_cg_lists() and read here after the tasklist_lock
5025          * lock on fork.
5026          */
5027         if (use_task_css_set_links) {
5028                 write_lock(&css_set_lock);
5029                 task_lock(child);
5030                 if (list_empty(&child->cg_list))
5031                         list_add(&child->cg_list, &task_css_set(child)->tasks);
5032                 task_unlock(child);
5033                 write_unlock(&css_set_lock);
5034         }
5035 
5036         /*
5037          * Call ss->fork().  This must happen after @child is linked on
5038          * css_set; otherwise, @child might change state between ->fork()
5039          * and addition to css_set.
5040          */
5041         if (need_forkexit_callback) {
5042                 /*
5043                  * fork/exit callbacks are supported only for builtin
5044                  * subsystems, and the builtin section of the subsys
5045                  * array is immutable, so we don't need to lock the
5046                  * subsys array here. On the other hand, modular section
5047                  * of the array can be freed at module unload, so we
5048                  * can't touch that.
5049                  */
5050                 for_each_builtin_subsys(ss, i)
5051                         if (ss->fork)
5052                                 ss->fork(child);
5053         }
5054 }
5055 
5056 /**
5057  * cgroup_exit - detach cgroup from exiting task
5058  * @tsk: pointer to task_struct of exiting process
5059  * @run_callback: run exit callbacks?
5060  *
5061  * Description: Detach cgroup from @tsk and release it.
5062  *
5063  * Note that cgroups marked notify_on_release force every task in
5064  * them to take the global cgroup_mutex mutex when exiting.
5065  * This could impact scaling on very large systems.  Be reluctant to
5066  * use notify_on_release cgroups where very high task exit scaling
5067  * is required on large systems.
5068  *
5069  * the_top_cgroup_hack:
5070  *
5071  *    Set the exiting tasks cgroup to the root cgroup (top_cgroup).
5072  *
5073  *    We call cgroup_exit() while the task is still competent to
5074  *    handle notify_on_release(), then leave the task attached to the
5075  *    root cgroup in each hierarchy for the remainder of its exit.
5076  *
5077  *    To do this properly, we would increment the reference count on
5078  *    top_cgroup, and near the very end of the kernel/exit.c do_exit()
5079  *    code we would add a second cgroup function call, to drop that
5080  *    reference.  This would just create an unnecessary hot spot on
5081  *    the top_cgroup reference count, to no avail.
5082  *
5083  *    Normally, holding a reference to a cgroup without bumping its
5084  *    count is unsafe.   The cgroup could go away, or someone could
5085  *    attach us to a different cgroup, decrementing the count on
5086  *    the first cgroup that we never incremented.  But in this case,
5087  *    top_cgroup isn't going away, and either task has PF_EXITING set,
5088  *    which wards off any cgroup_attach_task() attempts, or task is a failed
5089  *    fork, never visible to cgroup_attach_task.
5090  */
5091 void cgroup_exit(struct task_struct *tsk, int run_callbacks)
5092 {
5093         struct cgroup_subsys *ss;
5094         struct css_set *cset;
5095         int i;
5096 
5097         /*
5098          * Unlink from the css_set task list if necessary.
5099          * Optimistically check cg_list before taking
5100          * css_set_lock
5101          */
5102         if (!list_empty(&tsk->cg_list)) {
5103                 write_lock(&css_set_lock);
5104                 if (!list_empty(&tsk->cg_list))
5105                         list_del_init(&tsk->cg_list);
5106                 write_unlock(&css_set_lock);
5107         }
5108 
5109         /* Reassign the task to the init_css_set. */
5110         task_lock(tsk);
5111         cset = task_css_set(tsk);
5112         RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5113 
5114         if (run_callbacks && need_forkexit_callback) {
5115                 /*
5116                  * fork/exit callbacks are supported only for builtin
5117                  * subsystems, see cgroup_post_fork() for details.
5118                  */
5119                 for_each_builtin_subsys(ss, i) {
5120                         if (ss->exit) {
5121                                 struct cgroup_subsys_state *old_css = cset->subsys[i];
5122                                 struct cgroup_subsys_state *css = task_css(tsk, i);
5123 
5124                                 ss->exit(css, old_css, tsk);
5125                         }
5126                 }
5127         }
5128         task_unlock(tsk);
5129 
5130         put_css_set_taskexit(cset);
5131 }
5132 
5133 static void check_for_release(struct cgroup *cgrp)
5134 {
5135         if (cgroup_is_releasable(cgrp) &&
5136             list_empty(&cgrp->cset_links) && list_empty(&cgrp->children)) {
5137                 /*
5138                  * Control Group is currently removeable. If it's not
5139                  * already queued for a userspace notification, queue
5140                  * it now
5141                  */
5142                 int need_schedule_work = 0;
5143 
5144                 raw_spin_lock(&release_list_lock);
5145                 if (!cgroup_is_dead(cgrp) &&
5146                     list_empty(&cgrp->release_list)) {
5147                         list_add(&cgrp->release_list, &release_list);
5148                         need_schedule_work = 1;
5149                 }
5150                 raw_spin_unlock(&release_list_lock);
5151                 if (need_schedule_work)
5152                         schedule_work(&release_agent_work);
5153         }
5154 }
5155 
5156 /*
5157  * Notify userspace when a cgroup is released, by running the
5158  * configured release agent with the name of the cgroup (path
5159  * relative to the root of cgroup file system) as the argument.
5160  *
5161  * Most likely, this user command will try to rmdir this cgroup.
5162  *
5163  * This races with the possibility that some other task will be
5164  * attached to this cgroup before it is removed, or that some other
5165  * user task will 'mkdir' a child cgroup of this cgroup.  That's ok.
5166  * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5167  * unused, and this cgroup will be reprieved from its death sentence,
5168  * to continue to serve a useful existence.  Next time it's released,
5169  * we will get notified again, if it still has 'notify_on_release' set.
5170  *
5171  * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5172  * means only wait until the task is successfully execve()'d.  The
5173  * separate release agent task is forked by call_usermodehelper(),
5174  * then control in this thread returns here, without waiting for the
5175  * release agent task.  We don't bother to wait because the caller of
5176  * this routine has no use for the exit status of the release agent
5177  * task, so no sense holding our caller up for that.
5178  */
5179 static void cgroup_release_agent(struct work_struct *work)
5180 {
5181         BUG_ON(work != &release_agent_work);
5182         mutex_lock(&cgroup_mutex);
5183         raw_spin_lock(&release_list_lock);
5184         while (!list_empty(&release_list)) {
5185                 char *argv[3], *envp[3];
5186                 int i;
5187                 char *pathbuf = NULL, *agentbuf = NULL;
5188                 struct cgroup *cgrp = list_entry(release_list.next,
5189                                                     struct cgroup,
5190                                                     release_list);
5191                 list_del_init(&cgrp->release_list);
5192                 raw_spin_unlock(&release_list_lock);
5193                 pathbuf = kmalloc(PAGE_SIZE, GFP_KERNEL);
5194                 if (!pathbuf)
5195                         goto continue_free;
5196                 if (cgroup_path(cgrp, pathbuf, PAGE_SIZE) < 0)
5197                         goto continue_free;
5198                 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5199                 if (!agentbuf)
5200                         goto continue_free;
5201 
5202                 i = 0;
5203                 argv[i++] = agentbuf;
5204                 argv[i++] = pathbuf;
5205                 argv[i] = NULL;
5206 
5207                 i = 0;
5208                 /* minimal command environment */
5209                 envp[i++] = "HOME=/";
5210                 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5211                 envp[i] = NULL;
5212 
5213                 /* Drop the lock while we invoke the usermode helper,
5214                  * since the exec could involve hitting disk and hence
5215                  * be a slow process */
5216                 mutex_unlock(&cgroup_mutex);
5217                 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5218                 mutex_lock(&cgroup_mutex);
5219  continue_free:
5220                 kfree(pathbuf);
5221                 kfree(agentbuf);
5222                 raw_spin_lock(&release_list_lock);
5223         }
5224         raw_spin_unlock(&release_list_lock);
5225         mutex_unlock(&cgroup_mutex);
5226 }
5227 
5228 static int __init cgroup_disable(char *str)
5229 {
5230         struct cgroup_subsys *ss;
5231         char *token;
5232         int i;
5233 
5234         while ((token = strsep(&str, ",")) != NULL) {
5235                 if (!*token)
5236                         continue;
5237 
5238                 /*
5239                  * cgroup_disable, being at boot time, can't know about
5240                  * module subsystems, so we don't worry about them.
5241                  */
5242                 for_each_builtin_subsys(ss, i) {
5243                         if (!strcmp(token, ss->name)) {
5244                                 ss->disabled = 1;
5245                                 printk(KERN_INFO "Disabling %s control group"
5246                                         " subsystem\n", ss->name);
5247                                 break;
5248                         }
5249                 }
5250         }
5251         return 1;
5252 }
5253 __setup("cgroup_disable=", cgroup_disable);
5254 
5255 /**
5256  * css_from_dir - get corresponding css from the dentry of a cgroup dir
5257  * @dentry: directory dentry of interest
5258  * @ss: subsystem of interest
5259  *
5260  * Must be called under cgroup_mutex or RCU read lock.  The caller is
5261  * responsible for pinning the returned css if it needs to be accessed
5262  * outside the critical section.
5263  */
5264 struct cgroup_subsys_state *css_from_dir(struct dentry *dentry,
5265                                          struct cgroup_subsys *ss)
5266 {
5267         struct cgroup *cgrp;
5268 
5269         cgroup_assert_mutex_or_rcu_locked();
5270 
5271         /* is @dentry a cgroup dir? */
5272         if (!dentry->d_inode ||
5273             dentry->d_inode->i_op != &cgroup_dir_inode_operations)
5274                 return ERR_PTR(-EBADF);
5275 
5276         cgrp = __d_cgrp(dentry);
5277         return cgroup_css(cgrp, ss) ?: ERR_PTR(-ENOENT);
5278 }
5279 
5280 /**
5281  * css_from_id - lookup css by id
5282  * @id: the cgroup id
5283  * @ss: cgroup subsys to be looked into
5284  *
5285  * Returns the css if there's valid one with @id, otherwise returns NULL.
5286  * Should be called under rcu_read_lock().
5287  */
5288 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5289 {
5290         struct cgroup *cgrp;
5291 
5292         cgroup_assert_mutex_or_rcu_locked();
5293 
5294         cgrp = idr_find(&ss->root->cgroup_idr, id);
5295         if (cgrp)
5296                 return cgroup_css(cgrp, ss);
5297         return NULL;
5298 }
5299 
5300 #ifdef CONFIG_CGROUP_DEBUG
5301 static struct cgroup_subsys_state *
5302 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5303 {
5304         struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5305 
5306         if (!css)
5307                 return ERR_PTR(-ENOMEM);
5308 
5309         return css;
5310 }
5311 
5312 static void debug_css_free(struct cgroup_subsys_state *css)
5313 {
5314         kfree(css);
5315 }
5316 
5317 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5318                                 struct cftype *cft)
5319 {
5320         return cgroup_task_count(css->cgroup);
5321 }
5322 
5323 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5324                                 struct cftype *cft)
5325 {
5326         return (u64)(unsigned long)current->cgroups;
5327 }
5328 
5329 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5330                                          struct cftype *cft)
5331 {
5332         u64 count;
5333 
5334         rcu_read_lock();
5335         count = atomic_read(&task_css_set(current)->refcount);
5336         rcu_read_unlock();
5337         return count;
5338 }
5339 
5340 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5341 {
5342         struct cgrp_cset_link *link;
5343         struct css_set *cset;
5344 
5345         read_lock(&css_set_lock);
5346         rcu_read_lock();
5347         cset = rcu_dereference(current->cgroups);
5348         list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5349                 struct cgroup *c = link->cgrp;
5350                 const char *name;
5351 
5352                 if (c->dentry)
5353                         name = c->dentry->d_name.name;
5354                 else
5355                         name = "?";
5356                 seq_printf(seq, "Root %d group %s\n",
5357                            c->root->hierarchy_id, name);
5358         }
5359         rcu_read_unlock();
5360         read_unlock(&css_set_lock);
5361         return 0;
5362 }
5363 
5364 #define MAX_TASKS_SHOWN_PER_CSS 25
5365 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5366 {
5367         struct cgroup_subsys_state *css = seq_css(seq);
5368         struct cgrp_cset_link *link;
5369 
5370         read_lock(&css_set_lock);
5371         list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5372                 struct css_set *cset = link->cset;
5373                 struct task_struct *task;
5374                 int count = 0;
5375                 seq_printf(seq, "css_set %p\n", cset);
5376                 list_for_each_entry(task, &cset->tasks, cg_list) {
5377                         if (count++ > MAX_TASKS_SHOWN_PER_CSS) {
5378                                 seq_puts(seq, "  ...\n");
5379                                 break;
5380                         } else {
5381                                 seq_printf(seq, "  task %d\n",
5382                                            task_pid_vnr(task));
5383                         }
5384                 }
5385         }
5386         read_unlock(&css_set_lock);
5387         return 0;
5388 }
5389 
5390 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5391 {
5392         return test_bit(CGRP_RELEASABLE, &css->cgroup->flags);
5393 }
5394 
5395 static struct cftype debug_files[] =  {
5396         {
5397                 .name = "taskcount",
5398                 .read_u64 = debug_taskcount_read,
5399         },
5400 
5401         {
5402                 .name = "current_css_set",
5403                 .read_u64 = current_css_set_read,
5404         },
5405 
5406         {
5407                 .name = "current_css_set_refcount",
5408                 .read_u64 = current_css_set_refcount_read,
5409         },
5410 
5411         {
5412                 .name = "current_css_set_cg_links",
5413                 .seq_show = current_css_set_cg_links_read,
5414         },
5415 
5416         {
5417                 .name = "cgroup_css_links",
5418                 .seq_show = cgroup_css_links_read,
5419         },
5420 
5421         {
5422                 .name = "releasable",
5423                 .read_u64 = releasable_read,
5424         },
5425 
5426         { }     /* terminate */
5427 };
5428 
5429 struct cgroup_subsys debug_subsys = {
5430         .name = "debug",
5431         .css_alloc = debug_css_alloc,
5432         .css_free = debug_css_free,
5433         .subsys_id = debug_subsys_id,
5434         .base_cftypes = debug_files,
5435 };
5436 #endif /* CONFIG_CGROUP_DEBUG */
5437 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us