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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 30 
 31 #include <linux/cgroup.h>
 32 #include <linux/cred.h>
 33 #include <linux/ctype.h>
 34 #include <linux/errno.h>
 35 #include <linux/init_task.h>
 36 #include <linux/kernel.h>
 37 #include <linux/list.h>
 38 #include <linux/magic.h>
 39 #include <linux/mm.h>
 40 #include <linux/mutex.h>
 41 #include <linux/mount.h>
 42 #include <linux/pagemap.h>
 43 #include <linux/proc_fs.h>
 44 #include <linux/rcupdate.h>
 45 #include <linux/sched.h>
 46 #include <linux/slab.h>
 47 #include <linux/spinlock.h>
 48 #include <linux/rwsem.h>
 49 #include <linux/string.h>
 50 #include <linux/sort.h>
 51 #include <linux/kmod.h>
 52 #include <linux/delayacct.h>
 53 #include <linux/cgroupstats.h>
 54 #include <linux/hashtable.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/kthread.h>
 59 #include <linux/delay.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 #define CGROUP_FILE_NAME_MAX            (MAX_CGROUP_TYPE_NAMELEN +      \
 72                                          MAX_CFTYPE_NAME + 2)
 73 
 74 /*
 75  * cgroup_mutex is the master lock.  Any modification to cgroup or its
 76  * hierarchy must be performed while holding it.
 77  *
 78  * css_set_rwsem protects task->cgroups pointer, the list of css_set
 79  * objects, and the chain of tasks off each css_set.
 80  *
 81  * These locks are exported if CONFIG_PROVE_RCU so that accessors in
 82  * cgroup.h can use them for lockdep annotations.
 83  */
 84 #ifdef CONFIG_PROVE_RCU
 85 DEFINE_MUTEX(cgroup_mutex);
 86 DECLARE_RWSEM(css_set_rwsem);
 87 EXPORT_SYMBOL_GPL(cgroup_mutex);
 88 EXPORT_SYMBOL_GPL(css_set_rwsem);
 89 #else
 90 static DEFINE_MUTEX(cgroup_mutex);
 91 static DECLARE_RWSEM(css_set_rwsem);
 92 #endif
 93 
 94 /*
 95  * Protects cgroup_idr and css_idr so that IDs can be released without
 96  * grabbing cgroup_mutex.
 97  */
 98 static DEFINE_SPINLOCK(cgroup_idr_lock);
 99 
100 /*
101  * Protects cgroup_subsys->release_agent_path.  Modifying it also requires
102  * cgroup_mutex.  Reading requires either cgroup_mutex or this spinlock.
103  */
104 static DEFINE_SPINLOCK(release_agent_path_lock);
105 
106 #define cgroup_assert_mutex_or_rcu_locked()                             \
107         rcu_lockdep_assert(rcu_read_lock_held() ||                      \
108                            lockdep_is_held(&cgroup_mutex),              \
109                            "cgroup_mutex or RCU read lock required");
110 
111 /*
112  * cgroup destruction makes heavy use of work items and there can be a lot
113  * of concurrent destructions.  Use a separate workqueue so that cgroup
114  * destruction work items don't end up filling up max_active of system_wq
115  * which may lead to deadlock.
116  */
117 static struct workqueue_struct *cgroup_destroy_wq;
118 
119 /*
120  * pidlist destructions need to be flushed on cgroup destruction.  Use a
121  * separate workqueue as flush domain.
122  */
123 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
124 
125 /* generate an array of cgroup subsystem pointers */
126 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
127 static struct cgroup_subsys *cgroup_subsys[] = {
128 #include <linux/cgroup_subsys.h>
129 };
130 #undef SUBSYS
131 
132 /* array of cgroup subsystem names */
133 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
134 static const char *cgroup_subsys_name[] = {
135 #include <linux/cgroup_subsys.h>
136 };
137 #undef SUBSYS
138 
139 /*
140  * The default hierarchy, reserved for the subsystems that are otherwise
141  * unattached - it never has more than a single cgroup, and all tasks are
142  * part of that cgroup.
143  */
144 struct cgroup_root cgrp_dfl_root;
145 
146 /*
147  * The default hierarchy always exists but is hidden until mounted for the
148  * first time.  This is for backward compatibility.
149  */
150 static bool cgrp_dfl_root_visible;
151 
152 /*
153  * Set by the boot param of the same name and makes subsystems with NULL
154  * ->dfl_files to use ->legacy_files on the default hierarchy.
155  */
156 static bool cgroup_legacy_files_on_dfl;
157 
158 /* some controllers are not supported in the default hierarchy */
159 static unsigned int cgrp_dfl_root_inhibit_ss_mask;
160 
161 /* The list of hierarchy roots */
162 
163 static LIST_HEAD(cgroup_roots);
164 static int cgroup_root_count;
165 
166 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
167 static DEFINE_IDR(cgroup_hierarchy_idr);
168 
169 /*
170  * Assign a monotonically increasing serial number to csses.  It guarantees
171  * cgroups with bigger numbers are newer than those with smaller numbers.
172  * Also, as csses are always appended to the parent's ->children list, it
173  * guarantees that sibling csses are always sorted in the ascending serial
174  * number order on the list.  Protected by cgroup_mutex.
175  */
176 static u64 css_serial_nr_next = 1;
177 
178 /* This flag indicates whether tasks in the fork and exit paths should
179  * check for fork/exit handlers to call. This avoids us having to do
180  * extra work in the fork/exit path if none of the subsystems need to
181  * be called.
182  */
183 static int need_forkexit_callback __read_mostly;
184 
185 static struct cftype cgroup_dfl_base_files[];
186 static struct cftype cgroup_legacy_base_files[];
187 
188 static int rebind_subsystems(struct cgroup_root *dst_root,
189                              unsigned int ss_mask);
190 static int cgroup_destroy_locked(struct cgroup *cgrp);
191 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
192                       bool visible);
193 static void css_release(struct percpu_ref *ref);
194 static void kill_css(struct cgroup_subsys_state *css);
195 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
196                               bool is_add);
197 
198 /* IDR wrappers which synchronize using cgroup_idr_lock */
199 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
200                             gfp_t gfp_mask)
201 {
202         int ret;
203 
204         idr_preload(gfp_mask);
205         spin_lock_bh(&cgroup_idr_lock);
206         ret = idr_alloc(idr, ptr, start, end, gfp_mask);
207         spin_unlock_bh(&cgroup_idr_lock);
208         idr_preload_end();
209         return ret;
210 }
211 
212 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
213 {
214         void *ret;
215 
216         spin_lock_bh(&cgroup_idr_lock);
217         ret = idr_replace(idr, ptr, id);
218         spin_unlock_bh(&cgroup_idr_lock);
219         return ret;
220 }
221 
222 static void cgroup_idr_remove(struct idr *idr, int id)
223 {
224         spin_lock_bh(&cgroup_idr_lock);
225         idr_remove(idr, id);
226         spin_unlock_bh(&cgroup_idr_lock);
227 }
228 
229 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
230 {
231         struct cgroup_subsys_state *parent_css = cgrp->self.parent;
232 
233         if (parent_css)
234                 return container_of(parent_css, struct cgroup, self);
235         return NULL;
236 }
237 
238 /**
239  * cgroup_css - obtain a cgroup's css for the specified subsystem
240  * @cgrp: the cgroup of interest
241  * @ss: the subsystem of interest (%NULL returns @cgrp->self)
242  *
243  * Return @cgrp's css (cgroup_subsys_state) associated with @ss.  This
244  * function must be called either under cgroup_mutex or rcu_read_lock() and
245  * the caller is responsible for pinning the returned css if it wants to
246  * keep accessing it outside the said locks.  This function may return
247  * %NULL if @cgrp doesn't have @subsys_id enabled.
248  */
249 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
250                                               struct cgroup_subsys *ss)
251 {
252         if (ss)
253                 return rcu_dereference_check(cgrp->subsys[ss->id],
254                                         lockdep_is_held(&cgroup_mutex));
255         else
256                 return &cgrp->self;
257 }
258 
259 /**
260  * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
261  * @cgrp: the cgroup of interest
262  * @ss: the subsystem of interest (%NULL returns @cgrp->self)
263  *
264  * Similar to cgroup_css() but returns the effctive css, which is defined
265  * as the matching css of the nearest ancestor including self which has @ss
266  * enabled.  If @ss is associated with the hierarchy @cgrp is on, this
267  * function is guaranteed to return non-NULL css.
268  */
269 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
270                                                 struct cgroup_subsys *ss)
271 {
272         lockdep_assert_held(&cgroup_mutex);
273 
274         if (!ss)
275                 return &cgrp->self;
276 
277         if (!(cgrp->root->subsys_mask & (1 << ss->id)))
278                 return NULL;
279 
280         /*
281          * This function is used while updating css associations and thus
282          * can't test the csses directly.  Use ->child_subsys_mask.
283          */
284         while (cgroup_parent(cgrp) &&
285                !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
286                 cgrp = cgroup_parent(cgrp);
287 
288         return cgroup_css(cgrp, ss);
289 }
290 
291 /**
292  * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
293  * @cgrp: the cgroup of interest
294  * @ss: the subsystem of interest
295  *
296  * Find and get the effective css of @cgrp for @ss.  The effective css is
297  * defined as the matching css of the nearest ancestor including self which
298  * has @ss enabled.  If @ss is not mounted on the hierarchy @cgrp is on,
299  * the root css is returned, so this function always returns a valid css.
300  * The returned css must be put using css_put().
301  */
302 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
303                                              struct cgroup_subsys *ss)
304 {
305         struct cgroup_subsys_state *css;
306 
307         rcu_read_lock();
308 
309         do {
310                 css = cgroup_css(cgrp, ss);
311 
312                 if (css && css_tryget_online(css))
313                         goto out_unlock;
314                 cgrp = cgroup_parent(cgrp);
315         } while (cgrp);
316 
317         css = init_css_set.subsys[ss->id];
318         css_get(css);
319 out_unlock:
320         rcu_read_unlock();
321         return css;
322 }
323 
324 /* convenient tests for these bits */
325 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
326 {
327         return !(cgrp->self.flags & CSS_ONLINE);
328 }
329 
330 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
331 {
332         struct cgroup *cgrp = of->kn->parent->priv;
333         struct cftype *cft = of_cft(of);
334 
335         /*
336          * This is open and unprotected implementation of cgroup_css().
337          * seq_css() is only called from a kernfs file operation which has
338          * an active reference on the file.  Because all the subsystem
339          * files are drained before a css is disassociated with a cgroup,
340          * the matching css from the cgroup's subsys table is guaranteed to
341          * be and stay valid until the enclosing operation is complete.
342          */
343         if (cft->ss)
344                 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
345         else
346                 return &cgrp->self;
347 }
348 EXPORT_SYMBOL_GPL(of_css);
349 
350 /**
351  * cgroup_is_descendant - test ancestry
352  * @cgrp: the cgroup to be tested
353  * @ancestor: possible ancestor of @cgrp
354  *
355  * Test whether @cgrp is a descendant of @ancestor.  It also returns %true
356  * if @cgrp == @ancestor.  This function is safe to call as long as @cgrp
357  * and @ancestor are accessible.
358  */
359 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
360 {
361         while (cgrp) {
362                 if (cgrp == ancestor)
363                         return true;
364                 cgrp = cgroup_parent(cgrp);
365         }
366         return false;
367 }
368 
369 static int notify_on_release(const struct cgroup *cgrp)
370 {
371         return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
372 }
373 
374 /**
375  * for_each_css - iterate all css's of a cgroup
376  * @css: the iteration cursor
377  * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
378  * @cgrp: the target cgroup to iterate css's of
379  *
380  * Should be called under cgroup_[tree_]mutex.
381  */
382 #define for_each_css(css, ssid, cgrp)                                   \
383         for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)        \
384                 if (!((css) = rcu_dereference_check(                    \
385                                 (cgrp)->subsys[(ssid)],                 \
386                                 lockdep_is_held(&cgroup_mutex)))) { }   \
387                 else
388 
389 /**
390  * for_each_e_css - iterate all effective css's of a cgroup
391  * @css: the iteration cursor
392  * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
393  * @cgrp: the target cgroup to iterate css's of
394  *
395  * Should be called under cgroup_[tree_]mutex.
396  */
397 #define for_each_e_css(css, ssid, cgrp)                                 \
398         for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)        \
399                 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
400                         ;                                               \
401                 else
402 
403 /**
404  * for_each_subsys - iterate all enabled cgroup subsystems
405  * @ss: the iteration cursor
406  * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
407  */
408 #define for_each_subsys(ss, ssid)                                       \
409         for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT &&                \
410              (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
411 
412 /* iterate across the hierarchies */
413 #define for_each_root(root)                                             \
414         list_for_each_entry((root), &cgroup_roots, root_list)
415 
416 /* iterate over child cgrps, lock should be held throughout iteration */
417 #define cgroup_for_each_live_child(child, cgrp)                         \
418         list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
419                 if (({ lockdep_assert_held(&cgroup_mutex);              \
420                        cgroup_is_dead(child); }))                       \
421                         ;                                               \
422                 else
423 
424 static void cgroup_release_agent(struct work_struct *work);
425 static void check_for_release(struct cgroup *cgrp);
426 
427 /*
428  * A cgroup can be associated with multiple css_sets as different tasks may
429  * belong to different cgroups on different hierarchies.  In the other
430  * direction, a css_set is naturally associated with multiple cgroups.
431  * This M:N relationship is represented by the following link structure
432  * which exists for each association and allows traversing the associations
433  * from both sides.
434  */
435 struct cgrp_cset_link {
436         /* the cgroup and css_set this link associates */
437         struct cgroup           *cgrp;
438         struct css_set          *cset;
439 
440         /* list of cgrp_cset_links anchored at cgrp->cset_links */
441         struct list_head        cset_link;
442 
443         /* list of cgrp_cset_links anchored at css_set->cgrp_links */
444         struct list_head        cgrp_link;
445 };
446 
447 /*
448  * The default css_set - used by init and its children prior to any
449  * hierarchies being mounted. It contains a pointer to the root state
450  * for each subsystem. Also used to anchor the list of css_sets. Not
451  * reference-counted, to improve performance when child cgroups
452  * haven't been created.
453  */
454 struct css_set init_css_set = {
455         .refcount               = ATOMIC_INIT(1),
456         .cgrp_links             = LIST_HEAD_INIT(init_css_set.cgrp_links),
457         .tasks                  = LIST_HEAD_INIT(init_css_set.tasks),
458         .mg_tasks               = LIST_HEAD_INIT(init_css_set.mg_tasks),
459         .mg_preload_node        = LIST_HEAD_INIT(init_css_set.mg_preload_node),
460         .mg_node                = LIST_HEAD_INIT(init_css_set.mg_node),
461 };
462 
463 static int css_set_count        = 1;    /* 1 for init_css_set */
464 
465 /**
466  * cgroup_update_populated - updated populated count of a cgroup
467  * @cgrp: the target cgroup
468  * @populated: inc or dec populated count
469  *
470  * @cgrp is either getting the first task (css_set) or losing the last.
471  * Update @cgrp->populated_cnt accordingly.  The count is propagated
472  * towards root so that a given cgroup's populated_cnt is zero iff the
473  * cgroup and all its descendants are empty.
474  *
475  * @cgrp's interface file "cgroup.populated" is zero if
476  * @cgrp->populated_cnt is zero and 1 otherwise.  When @cgrp->populated_cnt
477  * changes from or to zero, userland is notified that the content of the
478  * interface file has changed.  This can be used to detect when @cgrp and
479  * its descendants become populated or empty.
480  */
481 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
482 {
483         lockdep_assert_held(&css_set_rwsem);
484 
485         do {
486                 bool trigger;
487 
488                 if (populated)
489                         trigger = !cgrp->populated_cnt++;
490                 else
491                         trigger = !--cgrp->populated_cnt;
492 
493                 if (!trigger)
494                         break;
495 
496                 if (cgrp->populated_kn)
497                         kernfs_notify(cgrp->populated_kn);
498                 cgrp = cgroup_parent(cgrp);
499         } while (cgrp);
500 }
501 
502 /*
503  * hash table for cgroup groups. This improves the performance to find
504  * an existing css_set. This hash doesn't (currently) take into
505  * account cgroups in empty hierarchies.
506  */
507 #define CSS_SET_HASH_BITS       7
508 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
509 
510 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
511 {
512         unsigned long key = 0UL;
513         struct cgroup_subsys *ss;
514         int i;
515 
516         for_each_subsys(ss, i)
517                 key += (unsigned long)css[i];
518         key = (key >> 16) ^ key;
519 
520         return key;
521 }
522 
523 static void put_css_set_locked(struct css_set *cset)
524 {
525         struct cgrp_cset_link *link, *tmp_link;
526         struct cgroup_subsys *ss;
527         int ssid;
528 
529         lockdep_assert_held(&css_set_rwsem);
530 
531         if (!atomic_dec_and_test(&cset->refcount))
532                 return;
533 
534         /* This css_set is dead. unlink it and release cgroup refcounts */
535         for_each_subsys(ss, ssid)
536                 list_del(&cset->e_cset_node[ssid]);
537         hash_del(&cset->hlist);
538         css_set_count--;
539 
540         list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
541                 struct cgroup *cgrp = link->cgrp;
542 
543                 list_del(&link->cset_link);
544                 list_del(&link->cgrp_link);
545 
546                 /* @cgrp can't go away while we're holding css_set_rwsem */
547                 if (list_empty(&cgrp->cset_links)) {
548                         cgroup_update_populated(cgrp, false);
549                         check_for_release(cgrp);
550                 }
551 
552                 kfree(link);
553         }
554 
555         kfree_rcu(cset, rcu_head);
556 }
557 
558 static void put_css_set(struct css_set *cset)
559 {
560         /*
561          * Ensure that the refcount doesn't hit zero while any readers
562          * can see it. Similar to atomic_dec_and_lock(), but for an
563          * rwlock
564          */
565         if (atomic_add_unless(&cset->refcount, -1, 1))
566                 return;
567 
568         down_write(&css_set_rwsem);
569         put_css_set_locked(cset);
570         up_write(&css_set_rwsem);
571 }
572 
573 /*
574  * refcounted get/put for css_set objects
575  */
576 static inline void get_css_set(struct css_set *cset)
577 {
578         atomic_inc(&cset->refcount);
579 }
580 
581 /**
582  * compare_css_sets - helper function for find_existing_css_set().
583  * @cset: candidate css_set being tested
584  * @old_cset: existing css_set for a task
585  * @new_cgrp: cgroup that's being entered by the task
586  * @template: desired set of css pointers in css_set (pre-calculated)
587  *
588  * Returns true if "cset" matches "old_cset" except for the hierarchy
589  * which "new_cgrp" belongs to, for which it should match "new_cgrp".
590  */
591 static bool compare_css_sets(struct css_set *cset,
592                              struct css_set *old_cset,
593                              struct cgroup *new_cgrp,
594                              struct cgroup_subsys_state *template[])
595 {
596         struct list_head *l1, *l2;
597 
598         /*
599          * On the default hierarchy, there can be csets which are
600          * associated with the same set of cgroups but different csses.
601          * Let's first ensure that csses match.
602          */
603         if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
604                 return false;
605 
606         /*
607          * Compare cgroup pointers in order to distinguish between
608          * different cgroups in hierarchies.  As different cgroups may
609          * share the same effective css, this comparison is always
610          * necessary.
611          */
612         l1 = &cset->cgrp_links;
613         l2 = &old_cset->cgrp_links;
614         while (1) {
615                 struct cgrp_cset_link *link1, *link2;
616                 struct cgroup *cgrp1, *cgrp2;
617 
618                 l1 = l1->next;
619                 l2 = l2->next;
620                 /* See if we reached the end - both lists are equal length. */
621                 if (l1 == &cset->cgrp_links) {
622                         BUG_ON(l2 != &old_cset->cgrp_links);
623                         break;
624                 } else {
625                         BUG_ON(l2 == &old_cset->cgrp_links);
626                 }
627                 /* Locate the cgroups associated with these links. */
628                 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
629                 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
630                 cgrp1 = link1->cgrp;
631                 cgrp2 = link2->cgrp;
632                 /* Hierarchies should be linked in the same order. */
633                 BUG_ON(cgrp1->root != cgrp2->root);
634 
635                 /*
636                  * If this hierarchy is the hierarchy of the cgroup
637                  * that's changing, then we need to check that this
638                  * css_set points to the new cgroup; if it's any other
639                  * hierarchy, then this css_set should point to the
640                  * same cgroup as the old css_set.
641                  */
642                 if (cgrp1->root == new_cgrp->root) {
643                         if (cgrp1 != new_cgrp)
644                                 return false;
645                 } else {
646                         if (cgrp1 != cgrp2)
647                                 return false;
648                 }
649         }
650         return true;
651 }
652 
653 /**
654  * find_existing_css_set - init css array and find the matching css_set
655  * @old_cset: the css_set that we're using before the cgroup transition
656  * @cgrp: the cgroup that we're moving into
657  * @template: out param for the new set of csses, should be clear on entry
658  */
659 static struct css_set *find_existing_css_set(struct css_set *old_cset,
660                                         struct cgroup *cgrp,
661                                         struct cgroup_subsys_state *template[])
662 {
663         struct cgroup_root *root = cgrp->root;
664         struct cgroup_subsys *ss;
665         struct css_set *cset;
666         unsigned long key;
667         int i;
668 
669         /*
670          * Build the set of subsystem state objects that we want to see in the
671          * new css_set. while subsystems can change globally, the entries here
672          * won't change, so no need for locking.
673          */
674         for_each_subsys(ss, i) {
675                 if (root->subsys_mask & (1UL << i)) {
676                         /*
677                          * @ss is in this hierarchy, so we want the
678                          * effective css from @cgrp.
679                          */
680                         template[i] = cgroup_e_css(cgrp, ss);
681                 } else {
682                         /*
683                          * @ss is not in this hierarchy, so we don't want
684                          * to change the css.
685                          */
686                         template[i] = old_cset->subsys[i];
687                 }
688         }
689 
690         key = css_set_hash(template);
691         hash_for_each_possible(css_set_table, cset, hlist, key) {
692                 if (!compare_css_sets(cset, old_cset, cgrp, template))
693                         continue;
694 
695                 /* This css_set matches what we need */
696                 return cset;
697         }
698 
699         /* No existing cgroup group matched */
700         return NULL;
701 }
702 
703 static void free_cgrp_cset_links(struct list_head *links_to_free)
704 {
705         struct cgrp_cset_link *link, *tmp_link;
706 
707         list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
708                 list_del(&link->cset_link);
709                 kfree(link);
710         }
711 }
712 
713 /**
714  * allocate_cgrp_cset_links - allocate cgrp_cset_links
715  * @count: the number of links to allocate
716  * @tmp_links: list_head the allocated links are put on
717  *
718  * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
719  * through ->cset_link.  Returns 0 on success or -errno.
720  */
721 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
722 {
723         struct cgrp_cset_link *link;
724         int i;
725 
726         INIT_LIST_HEAD(tmp_links);
727 
728         for (i = 0; i < count; i++) {
729                 link = kzalloc(sizeof(*link), GFP_KERNEL);
730                 if (!link) {
731                         free_cgrp_cset_links(tmp_links);
732                         return -ENOMEM;
733                 }
734                 list_add(&link->cset_link, tmp_links);
735         }
736         return 0;
737 }
738 
739 /**
740  * link_css_set - a helper function to link a css_set to a cgroup
741  * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
742  * @cset: the css_set to be linked
743  * @cgrp: the destination cgroup
744  */
745 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
746                          struct cgroup *cgrp)
747 {
748         struct cgrp_cset_link *link;
749 
750         BUG_ON(list_empty(tmp_links));
751 
752         if (cgroup_on_dfl(cgrp))
753                 cset->dfl_cgrp = cgrp;
754 
755         link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
756         link->cset = cset;
757         link->cgrp = cgrp;
758 
759         if (list_empty(&cgrp->cset_links))
760                 cgroup_update_populated(cgrp, true);
761         list_move(&link->cset_link, &cgrp->cset_links);
762 
763         /*
764          * Always add links to the tail of the list so that the list
765          * is sorted by order of hierarchy creation
766          */
767         list_add_tail(&link->cgrp_link, &cset->cgrp_links);
768 }
769 
770 /**
771  * find_css_set - return a new css_set with one cgroup updated
772  * @old_cset: the baseline css_set
773  * @cgrp: the cgroup to be updated
774  *
775  * Return a new css_set that's equivalent to @old_cset, but with @cgrp
776  * substituted into the appropriate hierarchy.
777  */
778 static struct css_set *find_css_set(struct css_set *old_cset,
779                                     struct cgroup *cgrp)
780 {
781         struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
782         struct css_set *cset;
783         struct list_head tmp_links;
784         struct cgrp_cset_link *link;
785         struct cgroup_subsys *ss;
786         unsigned long key;
787         int ssid;
788 
789         lockdep_assert_held(&cgroup_mutex);
790 
791         /* First see if we already have a cgroup group that matches
792          * the desired set */
793         down_read(&css_set_rwsem);
794         cset = find_existing_css_set(old_cset, cgrp, template);
795         if (cset)
796                 get_css_set(cset);
797         up_read(&css_set_rwsem);
798 
799         if (cset)
800                 return cset;
801 
802         cset = kzalloc(sizeof(*cset), GFP_KERNEL);
803         if (!cset)
804                 return NULL;
805 
806         /* Allocate all the cgrp_cset_link objects that we'll need */
807         if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
808                 kfree(cset);
809                 return NULL;
810         }
811 
812         atomic_set(&cset->refcount, 1);
813         INIT_LIST_HEAD(&cset->cgrp_links);
814         INIT_LIST_HEAD(&cset->tasks);
815         INIT_LIST_HEAD(&cset->mg_tasks);
816         INIT_LIST_HEAD(&cset->mg_preload_node);
817         INIT_LIST_HEAD(&cset->mg_node);
818         INIT_HLIST_NODE(&cset->hlist);
819 
820         /* Copy the set of subsystem state objects generated in
821          * find_existing_css_set() */
822         memcpy(cset->subsys, template, sizeof(cset->subsys));
823 
824         down_write(&css_set_rwsem);
825         /* Add reference counts and links from the new css_set. */
826         list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
827                 struct cgroup *c = link->cgrp;
828 
829                 if (c->root == cgrp->root)
830                         c = cgrp;
831                 link_css_set(&tmp_links, cset, c);
832         }
833 
834         BUG_ON(!list_empty(&tmp_links));
835 
836         css_set_count++;
837 
838         /* Add @cset to the hash table */
839         key = css_set_hash(cset->subsys);
840         hash_add(css_set_table, &cset->hlist, key);
841 
842         for_each_subsys(ss, ssid)
843                 list_add_tail(&cset->e_cset_node[ssid],
844                               &cset->subsys[ssid]->cgroup->e_csets[ssid]);
845 
846         up_write(&css_set_rwsem);
847 
848         return cset;
849 }
850 
851 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
852 {
853         struct cgroup *root_cgrp = kf_root->kn->priv;
854 
855         return root_cgrp->root;
856 }
857 
858 static int cgroup_init_root_id(struct cgroup_root *root)
859 {
860         int id;
861 
862         lockdep_assert_held(&cgroup_mutex);
863 
864         id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
865         if (id < 0)
866                 return id;
867 
868         root->hierarchy_id = id;
869         return 0;
870 }
871 
872 static void cgroup_exit_root_id(struct cgroup_root *root)
873 {
874         lockdep_assert_held(&cgroup_mutex);
875 
876         if (root->hierarchy_id) {
877                 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
878                 root->hierarchy_id = 0;
879         }
880 }
881 
882 static void cgroup_free_root(struct cgroup_root *root)
883 {
884         if (root) {
885                 /* hierarhcy ID shoulid already have been released */
886                 WARN_ON_ONCE(root->hierarchy_id);
887 
888                 idr_destroy(&root->cgroup_idr);
889                 kfree(root);
890         }
891 }
892 
893 static void cgroup_destroy_root(struct cgroup_root *root)
894 {
895         struct cgroup *cgrp = &root->cgrp;
896         struct cgrp_cset_link *link, *tmp_link;
897 
898         mutex_lock(&cgroup_mutex);
899 
900         BUG_ON(atomic_read(&root->nr_cgrps));
901         BUG_ON(!list_empty(&cgrp->self.children));
902 
903         /* Rebind all subsystems back to the default hierarchy */
904         rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
905 
906         /*
907          * Release all the links from cset_links to this hierarchy's
908          * root cgroup
909          */
910         down_write(&css_set_rwsem);
911 
912         list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
913                 list_del(&link->cset_link);
914                 list_del(&link->cgrp_link);
915                 kfree(link);
916         }
917         up_write(&css_set_rwsem);
918 
919         if (!list_empty(&root->root_list)) {
920                 list_del(&root->root_list);
921                 cgroup_root_count--;
922         }
923 
924         cgroup_exit_root_id(root);
925 
926         mutex_unlock(&cgroup_mutex);
927 
928         kernfs_destroy_root(root->kf_root);
929         cgroup_free_root(root);
930 }
931 
932 /* look up cgroup associated with given css_set on the specified hierarchy */
933 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
934                                             struct cgroup_root *root)
935 {
936         struct cgroup *res = NULL;
937 
938         lockdep_assert_held(&cgroup_mutex);
939         lockdep_assert_held(&css_set_rwsem);
940 
941         if (cset == &init_css_set) {
942                 res = &root->cgrp;
943         } else {
944                 struct cgrp_cset_link *link;
945 
946                 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
947                         struct cgroup *c = link->cgrp;
948 
949                         if (c->root == root) {
950                                 res = c;
951                                 break;
952                         }
953                 }
954         }
955 
956         BUG_ON(!res);
957         return res;
958 }
959 
960 /*
961  * Return the cgroup for "task" from the given hierarchy. Must be
962  * called with cgroup_mutex and css_set_rwsem held.
963  */
964 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
965                                             struct cgroup_root *root)
966 {
967         /*
968          * No need to lock the task - since we hold cgroup_mutex the
969          * task can't change groups, so the only thing that can happen
970          * is that it exits and its css is set back to init_css_set.
971          */
972         return cset_cgroup_from_root(task_css_set(task), root);
973 }
974 
975 /*
976  * A task must hold cgroup_mutex to modify cgroups.
977  *
978  * Any task can increment and decrement the count field without lock.
979  * So in general, code holding cgroup_mutex can't rely on the count
980  * field not changing.  However, if the count goes to zero, then only
981  * cgroup_attach_task() can increment it again.  Because a count of zero
982  * means that no tasks are currently attached, therefore there is no
983  * way a task attached to that cgroup can fork (the other way to
984  * increment the count).  So code holding cgroup_mutex can safely
985  * assume that if the count is zero, it will stay zero. Similarly, if
986  * a task holds cgroup_mutex on a cgroup with zero count, it
987  * knows that the cgroup won't be removed, as cgroup_rmdir()
988  * needs that mutex.
989  *
990  * A cgroup can only be deleted if both its 'count' of using tasks
991  * is zero, and its list of 'children' cgroups is empty.  Since all
992  * tasks in the system use _some_ cgroup, and since there is always at
993  * least one task in the system (init, pid == 1), therefore, root cgroup
994  * always has either children cgroups and/or using tasks.  So we don't
995  * need a special hack to ensure that root cgroup cannot be deleted.
996  *
997  * P.S.  One more locking exception.  RCU is used to guard the
998  * update of a tasks cgroup pointer by cgroup_attach_task()
999  */
1000 
1001 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask);
1002 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1003 static const struct file_operations proc_cgroupstats_operations;
1004 
1005 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1006                               char *buf)
1007 {
1008         if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1009             !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1010                 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1011                          cft->ss->name, cft->name);
1012         else
1013                 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1014         return buf;
1015 }
1016 
1017 /**
1018  * cgroup_file_mode - deduce file mode of a control file
1019  * @cft: the control file in question
1020  *
1021  * returns cft->mode if ->mode is not 0
1022  * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1023  * returns S_IRUGO if it has only a read handler
1024  * returns S_IWUSR if it has only a write hander
1025  */
1026 static umode_t cgroup_file_mode(const struct cftype *cft)
1027 {
1028         umode_t mode = 0;
1029 
1030         if (cft->mode)
1031                 return cft->mode;
1032 
1033         if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1034                 mode |= S_IRUGO;
1035 
1036         if (cft->write_u64 || cft->write_s64 || cft->write)
1037                 mode |= S_IWUSR;
1038 
1039         return mode;
1040 }
1041 
1042 static void cgroup_get(struct cgroup *cgrp)
1043 {
1044         WARN_ON_ONCE(cgroup_is_dead(cgrp));
1045         css_get(&cgrp->self);
1046 }
1047 
1048 static bool cgroup_tryget(struct cgroup *cgrp)
1049 {
1050         return css_tryget(&cgrp->self);
1051 }
1052 
1053 static void cgroup_put(struct cgroup *cgrp)
1054 {
1055         css_put(&cgrp->self);
1056 }
1057 
1058 /**
1059  * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1060  * @cgrp: the target cgroup
1061  * @subtree_control: the new subtree_control mask to consider
1062  *
1063  * On the default hierarchy, a subsystem may request other subsystems to be
1064  * enabled together through its ->depends_on mask.  In such cases, more
1065  * subsystems than specified in "cgroup.subtree_control" may be enabled.
1066  *
1067  * This function calculates which subsystems need to be enabled if
1068  * @subtree_control is to be applied to @cgrp.  The returned mask is always
1069  * a superset of @subtree_control and follows the usual hierarchy rules.
1070  */
1071 static unsigned int cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1072                                                   unsigned int subtree_control)
1073 {
1074         struct cgroup *parent = cgroup_parent(cgrp);
1075         unsigned int cur_ss_mask = subtree_control;
1076         struct cgroup_subsys *ss;
1077         int ssid;
1078 
1079         lockdep_assert_held(&cgroup_mutex);
1080 
1081         if (!cgroup_on_dfl(cgrp))
1082                 return cur_ss_mask;
1083 
1084         while (true) {
1085                 unsigned int new_ss_mask = cur_ss_mask;
1086 
1087                 for_each_subsys(ss, ssid)
1088                         if (cur_ss_mask & (1 << ssid))
1089                                 new_ss_mask |= ss->depends_on;
1090 
1091                 /*
1092                  * Mask out subsystems which aren't available.  This can
1093                  * happen only if some depended-upon subsystems were bound
1094                  * to non-default hierarchies.
1095                  */
1096                 if (parent)
1097                         new_ss_mask &= parent->child_subsys_mask;
1098                 else
1099                         new_ss_mask &= cgrp->root->subsys_mask;
1100 
1101                 if (new_ss_mask == cur_ss_mask)
1102                         break;
1103                 cur_ss_mask = new_ss_mask;
1104         }
1105 
1106         return cur_ss_mask;
1107 }
1108 
1109 /**
1110  * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1111  * @cgrp: the target cgroup
1112  *
1113  * Update @cgrp->child_subsys_mask according to the current
1114  * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1115  */
1116 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1117 {
1118         cgrp->child_subsys_mask =
1119                 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1120 }
1121 
1122 /**
1123  * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1124  * @kn: the kernfs_node being serviced
1125  *
1126  * This helper undoes cgroup_kn_lock_live() and should be invoked before
1127  * the method finishes if locking succeeded.  Note that once this function
1128  * returns the cgroup returned by cgroup_kn_lock_live() may become
1129  * inaccessible any time.  If the caller intends to continue to access the
1130  * cgroup, it should pin it before invoking this function.
1131  */
1132 static void cgroup_kn_unlock(struct kernfs_node *kn)
1133 {
1134         struct cgroup *cgrp;
1135 
1136         if (kernfs_type(kn) == KERNFS_DIR)
1137                 cgrp = kn->priv;
1138         else
1139                 cgrp = kn->parent->priv;
1140 
1141         mutex_unlock(&cgroup_mutex);
1142 
1143         kernfs_unbreak_active_protection(kn);
1144         cgroup_put(cgrp);
1145 }
1146 
1147 /**
1148  * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1149  * @kn: the kernfs_node being serviced
1150  *
1151  * This helper is to be used by a cgroup kernfs method currently servicing
1152  * @kn.  It breaks the active protection, performs cgroup locking and
1153  * verifies that the associated cgroup is alive.  Returns the cgroup if
1154  * alive; otherwise, %NULL.  A successful return should be undone by a
1155  * matching cgroup_kn_unlock() invocation.
1156  *
1157  * Any cgroup kernfs method implementation which requires locking the
1158  * associated cgroup should use this helper.  It avoids nesting cgroup
1159  * locking under kernfs active protection and allows all kernfs operations
1160  * including self-removal.
1161  */
1162 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1163 {
1164         struct cgroup *cgrp;
1165 
1166         if (kernfs_type(kn) == KERNFS_DIR)
1167                 cgrp = kn->priv;
1168         else
1169                 cgrp = kn->parent->priv;
1170 
1171         /*
1172          * We're gonna grab cgroup_mutex which nests outside kernfs
1173          * active_ref.  cgroup liveliness check alone provides enough
1174          * protection against removal.  Ensure @cgrp stays accessible and
1175          * break the active_ref protection.
1176          */
1177         if (!cgroup_tryget(cgrp))
1178                 return NULL;
1179         kernfs_break_active_protection(kn);
1180 
1181         mutex_lock(&cgroup_mutex);
1182 
1183         if (!cgroup_is_dead(cgrp))
1184                 return cgrp;
1185 
1186         cgroup_kn_unlock(kn);
1187         return NULL;
1188 }
1189 
1190 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1191 {
1192         char name[CGROUP_FILE_NAME_MAX];
1193 
1194         lockdep_assert_held(&cgroup_mutex);
1195         kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1196 }
1197 
1198 /**
1199  * cgroup_clear_dir - remove subsys files in a cgroup directory
1200  * @cgrp: target cgroup
1201  * @subsys_mask: mask of the subsystem ids whose files should be removed
1202  */
1203 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned int subsys_mask)
1204 {
1205         struct cgroup_subsys *ss;
1206         int i;
1207 
1208         for_each_subsys(ss, i) {
1209                 struct cftype *cfts;
1210 
1211                 if (!(subsys_mask & (1 << i)))
1212                         continue;
1213                 list_for_each_entry(cfts, &ss->cfts, node)
1214                         cgroup_addrm_files(cgrp, cfts, false);
1215         }
1216 }
1217 
1218 static int rebind_subsystems(struct cgroup_root *dst_root, unsigned int ss_mask)
1219 {
1220         struct cgroup_subsys *ss;
1221         unsigned int tmp_ss_mask;
1222         int ssid, i, ret;
1223 
1224         lockdep_assert_held(&cgroup_mutex);
1225 
1226         for_each_subsys(ss, ssid) {
1227                 if (!(ss_mask & (1 << ssid)))
1228                         continue;
1229 
1230                 /* if @ss has non-root csses attached to it, can't move */
1231                 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1232                         return -EBUSY;
1233 
1234                 /* can't move between two non-dummy roots either */
1235                 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1236                         return -EBUSY;
1237         }
1238 
1239         /* skip creating root files on dfl_root for inhibited subsystems */
1240         tmp_ss_mask = ss_mask;
1241         if (dst_root == &cgrp_dfl_root)
1242                 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1243 
1244         ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1245         if (ret) {
1246                 if (dst_root != &cgrp_dfl_root)
1247                         return ret;
1248 
1249                 /*
1250                  * Rebinding back to the default root is not allowed to
1251                  * fail.  Using both default and non-default roots should
1252                  * be rare.  Moving subsystems back and forth even more so.
1253                  * Just warn about it and continue.
1254                  */
1255                 if (cgrp_dfl_root_visible) {
1256                         pr_warn("failed to create files (%d) while rebinding 0x%x to default root\n",
1257                                 ret, ss_mask);
1258                         pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1259                 }
1260         }
1261 
1262         /*
1263          * Nothing can fail from this point on.  Remove files for the
1264          * removed subsystems and rebind each subsystem.
1265          */
1266         for_each_subsys(ss, ssid)
1267                 if (ss_mask & (1 << ssid))
1268                         cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1269 
1270         for_each_subsys(ss, ssid) {
1271                 struct cgroup_root *src_root;
1272                 struct cgroup_subsys_state *css;
1273                 struct css_set *cset;
1274 
1275                 if (!(ss_mask & (1 << ssid)))
1276                         continue;
1277 
1278                 src_root = ss->root;
1279                 css = cgroup_css(&src_root->cgrp, ss);
1280 
1281                 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1282 
1283                 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1284                 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1285                 ss->root = dst_root;
1286                 css->cgroup = &dst_root->cgrp;
1287 
1288                 down_write(&css_set_rwsem);
1289                 hash_for_each(css_set_table, i, cset, hlist)
1290                         list_move_tail(&cset->e_cset_node[ss->id],
1291                                        &dst_root->cgrp.e_csets[ss->id]);
1292                 up_write(&css_set_rwsem);
1293 
1294                 src_root->subsys_mask &= ~(1 << ssid);
1295                 src_root->cgrp.subtree_control &= ~(1 << ssid);
1296                 cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1297 
1298                 /* default hierarchy doesn't enable controllers by default */
1299                 dst_root->subsys_mask |= 1 << ssid;
1300                 if (dst_root != &cgrp_dfl_root) {
1301                         dst_root->cgrp.subtree_control |= 1 << ssid;
1302                         cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1303                 }
1304 
1305                 if (ss->bind)
1306                         ss->bind(css);
1307         }
1308 
1309         kernfs_activate(dst_root->cgrp.kn);
1310         return 0;
1311 }
1312 
1313 static int cgroup_show_options(struct seq_file *seq,
1314                                struct kernfs_root *kf_root)
1315 {
1316         struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1317         struct cgroup_subsys *ss;
1318         int ssid;
1319 
1320         for_each_subsys(ss, ssid)
1321                 if (root->subsys_mask & (1 << ssid))
1322                         seq_printf(seq, ",%s", ss->name);
1323         if (root->flags & CGRP_ROOT_NOPREFIX)
1324                 seq_puts(seq, ",noprefix");
1325         if (root->flags & CGRP_ROOT_XATTR)
1326                 seq_puts(seq, ",xattr");
1327 
1328         spin_lock(&release_agent_path_lock);
1329         if (strlen(root->release_agent_path))
1330                 seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1331         spin_unlock(&release_agent_path_lock);
1332 
1333         if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1334                 seq_puts(seq, ",clone_children");
1335         if (strlen(root->name))
1336                 seq_printf(seq, ",name=%s", root->name);
1337         return 0;
1338 }
1339 
1340 struct cgroup_sb_opts {
1341         unsigned int subsys_mask;
1342         unsigned int flags;
1343         char *release_agent;
1344         bool cpuset_clone_children;
1345         char *name;
1346         /* User explicitly requested empty subsystem */
1347         bool none;
1348 };
1349 
1350 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1351 {
1352         char *token, *o = data;
1353         bool all_ss = false, one_ss = false;
1354         unsigned int mask = -1U;
1355         struct cgroup_subsys *ss;
1356         int nr_opts = 0;
1357         int i;
1358 
1359 #ifdef CONFIG_CPUSETS
1360         mask = ~(1U << cpuset_cgrp_id);
1361 #endif
1362 
1363         memset(opts, 0, sizeof(*opts));
1364 
1365         while ((token = strsep(&o, ",")) != NULL) {
1366                 nr_opts++;
1367 
1368                 if (!*token)
1369                         return -EINVAL;
1370                 if (!strcmp(token, "none")) {
1371                         /* Explicitly have no subsystems */
1372                         opts->none = true;
1373                         continue;
1374                 }
1375                 if (!strcmp(token, "all")) {
1376                         /* Mutually exclusive option 'all' + subsystem name */
1377                         if (one_ss)
1378                                 return -EINVAL;
1379                         all_ss = true;
1380                         continue;
1381                 }
1382                 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1383                         opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1384                         continue;
1385                 }
1386                 if (!strcmp(token, "noprefix")) {
1387                         opts->flags |= CGRP_ROOT_NOPREFIX;
1388                         continue;
1389                 }
1390                 if (!strcmp(token, "clone_children")) {
1391                         opts->cpuset_clone_children = true;
1392                         continue;
1393                 }
1394                 if (!strcmp(token, "xattr")) {
1395                         opts->flags |= CGRP_ROOT_XATTR;
1396                         continue;
1397                 }
1398                 if (!strncmp(token, "release_agent=", 14)) {
1399                         /* Specifying two release agents is forbidden */
1400                         if (opts->release_agent)
1401                                 return -EINVAL;
1402                         opts->release_agent =
1403                                 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1404                         if (!opts->release_agent)
1405                                 return -ENOMEM;
1406                         continue;
1407                 }
1408                 if (!strncmp(token, "name=", 5)) {
1409                         const char *name = token + 5;
1410                         /* Can't specify an empty name */
1411                         if (!strlen(name))
1412                                 return -EINVAL;
1413                         /* Must match [\w.-]+ */
1414                         for (i = 0; i < strlen(name); i++) {
1415                                 char c = name[i];
1416                                 if (isalnum(c))
1417                                         continue;
1418                                 if ((c == '.') || (c == '-') || (c == '_'))
1419                                         continue;
1420                                 return -EINVAL;
1421                         }
1422                         /* Specifying two names is forbidden */
1423                         if (opts->name)
1424                                 return -EINVAL;
1425                         opts->name = kstrndup(name,
1426                                               MAX_CGROUP_ROOT_NAMELEN - 1,
1427                                               GFP_KERNEL);
1428                         if (!opts->name)
1429                                 return -ENOMEM;
1430 
1431                         continue;
1432                 }
1433 
1434                 for_each_subsys(ss, i) {
1435                         if (strcmp(token, ss->name))
1436                                 continue;
1437                         if (ss->disabled)
1438                                 continue;
1439 
1440                         /* Mutually exclusive option 'all' + subsystem name */
1441                         if (all_ss)
1442                                 return -EINVAL;
1443                         opts->subsys_mask |= (1 << i);
1444                         one_ss = true;
1445 
1446                         break;
1447                 }
1448                 if (i == CGROUP_SUBSYS_COUNT)
1449                         return -ENOENT;
1450         }
1451 
1452         if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1453                 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1454                 if (nr_opts != 1) {
1455                         pr_err("sane_behavior: no other mount options allowed\n");
1456                         return -EINVAL;
1457                 }
1458                 return 0;
1459         }
1460 
1461         /*
1462          * If the 'all' option was specified select all the subsystems,
1463          * otherwise if 'none', 'name=' and a subsystem name options were
1464          * not specified, let's default to 'all'
1465          */
1466         if (all_ss || (!one_ss && !opts->none && !opts->name))
1467                 for_each_subsys(ss, i)
1468                         if (!ss->disabled)
1469                                 opts->subsys_mask |= (1 << i);
1470 
1471         /*
1472          * We either have to specify by name or by subsystems. (So all
1473          * empty hierarchies must have a name).
1474          */
1475         if (!opts->subsys_mask && !opts->name)
1476                 return -EINVAL;
1477 
1478         /*
1479          * Option noprefix was introduced just for backward compatibility
1480          * with the old cpuset, so we allow noprefix only if mounting just
1481          * the cpuset subsystem.
1482          */
1483         if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1484                 return -EINVAL;
1485 
1486         /* Can't specify "none" and some subsystems */
1487         if (opts->subsys_mask && opts->none)
1488                 return -EINVAL;
1489 
1490         return 0;
1491 }
1492 
1493 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1494 {
1495         int ret = 0;
1496         struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1497         struct cgroup_sb_opts opts;
1498         unsigned int added_mask, removed_mask;
1499 
1500         if (root == &cgrp_dfl_root) {
1501                 pr_err("remount is not allowed\n");
1502                 return -EINVAL;
1503         }
1504 
1505         mutex_lock(&cgroup_mutex);
1506 
1507         /* See what subsystems are wanted */
1508         ret = parse_cgroupfs_options(data, &opts);
1509         if (ret)
1510                 goto out_unlock;
1511 
1512         if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1513                 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1514                         task_tgid_nr(current), current->comm);
1515 
1516         added_mask = opts.subsys_mask & ~root->subsys_mask;
1517         removed_mask = root->subsys_mask & ~opts.subsys_mask;
1518 
1519         /* Don't allow flags or name to change at remount */
1520         if ((opts.flags ^ root->flags) ||
1521             (opts.name && strcmp(opts.name, root->name))) {
1522                 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1523                        opts.flags, opts.name ?: "", root->flags, root->name);
1524                 ret = -EINVAL;
1525                 goto out_unlock;
1526         }
1527 
1528         /* remounting is not allowed for populated hierarchies */
1529         if (!list_empty(&root->cgrp.self.children)) {
1530                 ret = -EBUSY;
1531                 goto out_unlock;
1532         }
1533 
1534         ret = rebind_subsystems(root, added_mask);
1535         if (ret)
1536                 goto out_unlock;
1537 
1538         rebind_subsystems(&cgrp_dfl_root, removed_mask);
1539 
1540         if (opts.release_agent) {
1541                 spin_lock(&release_agent_path_lock);
1542                 strcpy(root->release_agent_path, opts.release_agent);
1543                 spin_unlock(&release_agent_path_lock);
1544         }
1545  out_unlock:
1546         kfree(opts.release_agent);
1547         kfree(opts.name);
1548         mutex_unlock(&cgroup_mutex);
1549         return ret;
1550 }
1551 
1552 /*
1553  * To reduce the fork() overhead for systems that are not actually using
1554  * their cgroups capability, we don't maintain the lists running through
1555  * each css_set to its tasks until we see the list actually used - in other
1556  * words after the first mount.
1557  */
1558 static bool use_task_css_set_links __read_mostly;
1559 
1560 static void cgroup_enable_task_cg_lists(void)
1561 {
1562         struct task_struct *p, *g;
1563 
1564         down_write(&css_set_rwsem);
1565 
1566         if (use_task_css_set_links)
1567                 goto out_unlock;
1568 
1569         use_task_css_set_links = true;
1570 
1571         /*
1572          * We need tasklist_lock because RCU is not safe against
1573          * while_each_thread(). Besides, a forking task that has passed
1574          * cgroup_post_fork() without seeing use_task_css_set_links = 1
1575          * is not guaranteed to have its child immediately visible in the
1576          * tasklist if we walk through it with RCU.
1577          */
1578         read_lock(&tasklist_lock);
1579         do_each_thread(g, p) {
1580                 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1581                              task_css_set(p) != &init_css_set);
1582 
1583                 /*
1584                  * We should check if the process is exiting, otherwise
1585                  * it will race with cgroup_exit() in that the list
1586                  * entry won't be deleted though the process has exited.
1587                  * Do it while holding siglock so that we don't end up
1588                  * racing against cgroup_exit().
1589                  */
1590                 spin_lock_irq(&p->sighand->siglock);
1591                 if (!(p->flags & PF_EXITING)) {
1592                         struct css_set *cset = task_css_set(p);
1593 
1594                         list_add(&p->cg_list, &cset->tasks);
1595                         get_css_set(cset);
1596                 }
1597                 spin_unlock_irq(&p->sighand->siglock);
1598         } while_each_thread(g, p);
1599         read_unlock(&tasklist_lock);
1600 out_unlock:
1601         up_write(&css_set_rwsem);
1602 }
1603 
1604 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1605 {
1606         struct cgroup_subsys *ss;
1607         int ssid;
1608 
1609         INIT_LIST_HEAD(&cgrp->self.sibling);
1610         INIT_LIST_HEAD(&cgrp->self.children);
1611         INIT_LIST_HEAD(&cgrp->cset_links);
1612         INIT_LIST_HEAD(&cgrp->pidlists);
1613         mutex_init(&cgrp->pidlist_mutex);
1614         cgrp->self.cgroup = cgrp;
1615         cgrp->self.flags |= CSS_ONLINE;
1616 
1617         for_each_subsys(ss, ssid)
1618                 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1619 
1620         init_waitqueue_head(&cgrp->offline_waitq);
1621         INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1622 }
1623 
1624 static void init_cgroup_root(struct cgroup_root *root,
1625                              struct cgroup_sb_opts *opts)
1626 {
1627         struct cgroup *cgrp = &root->cgrp;
1628 
1629         INIT_LIST_HEAD(&root->root_list);
1630         atomic_set(&root->nr_cgrps, 1);
1631         cgrp->root = root;
1632         init_cgroup_housekeeping(cgrp);
1633         idr_init(&root->cgroup_idr);
1634 
1635         root->flags = opts->flags;
1636         if (opts->release_agent)
1637                 strcpy(root->release_agent_path, opts->release_agent);
1638         if (opts->name)
1639                 strcpy(root->name, opts->name);
1640         if (opts->cpuset_clone_children)
1641                 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1642 }
1643 
1644 static int cgroup_setup_root(struct cgroup_root *root, unsigned int ss_mask)
1645 {
1646         LIST_HEAD(tmp_links);
1647         struct cgroup *root_cgrp = &root->cgrp;
1648         struct cftype *base_files;
1649         struct css_set *cset;
1650         int i, ret;
1651 
1652         lockdep_assert_held(&cgroup_mutex);
1653 
1654         ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1655         if (ret < 0)
1656                 goto out;
1657         root_cgrp->id = ret;
1658 
1659         ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1660                               GFP_KERNEL);
1661         if (ret)
1662                 goto out;
1663 
1664         /*
1665          * We're accessing css_set_count without locking css_set_rwsem here,
1666          * but that's OK - it can only be increased by someone holding
1667          * cgroup_lock, and that's us. The worst that can happen is that we
1668          * have some link structures left over
1669          */
1670         ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1671         if (ret)
1672                 goto cancel_ref;
1673 
1674         ret = cgroup_init_root_id(root);
1675         if (ret)
1676                 goto cancel_ref;
1677 
1678         root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1679                                            KERNFS_ROOT_CREATE_DEACTIVATED,
1680                                            root_cgrp);
1681         if (IS_ERR(root->kf_root)) {
1682                 ret = PTR_ERR(root->kf_root);
1683                 goto exit_root_id;
1684         }
1685         root_cgrp->kn = root->kf_root->kn;
1686 
1687         if (root == &cgrp_dfl_root)
1688                 base_files = cgroup_dfl_base_files;
1689         else
1690                 base_files = cgroup_legacy_base_files;
1691 
1692         ret = cgroup_addrm_files(root_cgrp, base_files, true);
1693         if (ret)
1694                 goto destroy_root;
1695 
1696         ret = rebind_subsystems(root, ss_mask);
1697         if (ret)
1698                 goto destroy_root;
1699 
1700         /*
1701          * There must be no failure case after here, since rebinding takes
1702          * care of subsystems' refcounts, which are explicitly dropped in
1703          * the failure exit path.
1704          */
1705         list_add(&root->root_list, &cgroup_roots);
1706         cgroup_root_count++;
1707 
1708         /*
1709          * Link the root cgroup in this hierarchy into all the css_set
1710          * objects.
1711          */
1712         down_write(&css_set_rwsem);
1713         hash_for_each(css_set_table, i, cset, hlist)
1714                 link_css_set(&tmp_links, cset, root_cgrp);
1715         up_write(&css_set_rwsem);
1716 
1717         BUG_ON(!list_empty(&root_cgrp->self.children));
1718         BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1719 
1720         kernfs_activate(root_cgrp->kn);
1721         ret = 0;
1722         goto out;
1723 
1724 destroy_root:
1725         kernfs_destroy_root(root->kf_root);
1726         root->kf_root = NULL;
1727 exit_root_id:
1728         cgroup_exit_root_id(root);
1729 cancel_ref:
1730         percpu_ref_exit(&root_cgrp->self.refcnt);
1731 out:
1732         free_cgrp_cset_links(&tmp_links);
1733         return ret;
1734 }
1735 
1736 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1737                          int flags, const char *unused_dev_name,
1738                          void *data)
1739 {
1740         struct super_block *pinned_sb = NULL;
1741         struct cgroup_subsys *ss;
1742         struct cgroup_root *root;
1743         struct cgroup_sb_opts opts;
1744         struct dentry *dentry;
1745         int ret;
1746         int i;
1747         bool new_sb;
1748 
1749         /*
1750          * The first time anyone tries to mount a cgroup, enable the list
1751          * linking each css_set to its tasks and fix up all existing tasks.
1752          */
1753         if (!use_task_css_set_links)
1754                 cgroup_enable_task_cg_lists();
1755 
1756         mutex_lock(&cgroup_mutex);
1757 
1758         /* First find the desired set of subsystems */
1759         ret = parse_cgroupfs_options(data, &opts);
1760         if (ret)
1761                 goto out_unlock;
1762 
1763         /* look for a matching existing root */
1764         if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1765                 cgrp_dfl_root_visible = true;
1766                 root = &cgrp_dfl_root;
1767                 cgroup_get(&root->cgrp);
1768                 ret = 0;
1769                 goto out_unlock;
1770         }
1771 
1772         /*
1773          * Destruction of cgroup root is asynchronous, so subsystems may
1774          * still be dying after the previous unmount.  Let's drain the
1775          * dying subsystems.  We just need to ensure that the ones
1776          * unmounted previously finish dying and don't care about new ones
1777          * starting.  Testing ref liveliness is good enough.
1778          */
1779         for_each_subsys(ss, i) {
1780                 if (!(opts.subsys_mask & (1 << i)) ||
1781                     ss->root == &cgrp_dfl_root)
1782                         continue;
1783 
1784                 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1785                         mutex_unlock(&cgroup_mutex);
1786                         msleep(10);
1787                         ret = restart_syscall();
1788                         goto out_free;
1789                 }
1790                 cgroup_put(&ss->root->cgrp);
1791         }
1792 
1793         for_each_root(root) {
1794                 bool name_match = false;
1795 
1796                 if (root == &cgrp_dfl_root)
1797                         continue;
1798 
1799                 /*
1800                  * If we asked for a name then it must match.  Also, if
1801                  * name matches but sybsys_mask doesn't, we should fail.
1802                  * Remember whether name matched.
1803                  */
1804                 if (opts.name) {
1805                         if (strcmp(opts.name, root->name))
1806                                 continue;
1807                         name_match = true;
1808                 }
1809 
1810                 /*
1811                  * If we asked for subsystems (or explicitly for no
1812                  * subsystems) then they must match.
1813                  */
1814                 if ((opts.subsys_mask || opts.none) &&
1815                     (opts.subsys_mask != root->subsys_mask)) {
1816                         if (!name_match)
1817                                 continue;
1818                         ret = -EBUSY;
1819                         goto out_unlock;
1820                 }
1821 
1822                 if (root->flags ^ opts.flags)
1823                         pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1824 
1825                 /*
1826                  * We want to reuse @root whose lifetime is governed by its
1827                  * ->cgrp.  Let's check whether @root is alive and keep it
1828                  * that way.  As cgroup_kill_sb() can happen anytime, we
1829                  * want to block it by pinning the sb so that @root doesn't
1830                  * get killed before mount is complete.
1831                  *
1832                  * With the sb pinned, tryget_live can reliably indicate
1833                  * whether @root can be reused.  If it's being killed,
1834                  * drain it.  We can use wait_queue for the wait but this
1835                  * path is super cold.  Let's just sleep a bit and retry.
1836                  */
1837                 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
1838                 if (IS_ERR(pinned_sb) ||
1839                     !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1840                         mutex_unlock(&cgroup_mutex);
1841                         if (!IS_ERR_OR_NULL(pinned_sb))
1842                                 deactivate_super(pinned_sb);
1843                         msleep(10);
1844                         ret = restart_syscall();
1845                         goto out_free;
1846                 }
1847 
1848                 ret = 0;
1849                 goto out_unlock;
1850         }
1851 
1852         /*
1853          * No such thing, create a new one.  name= matching without subsys
1854          * specification is allowed for already existing hierarchies but we
1855          * can't create new one without subsys specification.
1856          */
1857         if (!opts.subsys_mask && !opts.none) {
1858                 ret = -EINVAL;
1859                 goto out_unlock;
1860         }
1861 
1862         root = kzalloc(sizeof(*root), GFP_KERNEL);
1863         if (!root) {
1864                 ret = -ENOMEM;
1865                 goto out_unlock;
1866         }
1867 
1868         init_cgroup_root(root, &opts);
1869 
1870         ret = cgroup_setup_root(root, opts.subsys_mask);
1871         if (ret)
1872                 cgroup_free_root(root);
1873 
1874 out_unlock:
1875         mutex_unlock(&cgroup_mutex);
1876 out_free:
1877         kfree(opts.release_agent);
1878         kfree(opts.name);
1879 
1880         if (ret)
1881                 return ERR_PTR(ret);
1882 
1883         dentry = kernfs_mount(fs_type, flags, root->kf_root,
1884                                 CGROUP_SUPER_MAGIC, &new_sb);
1885         if (IS_ERR(dentry) || !new_sb)
1886                 cgroup_put(&root->cgrp);
1887 
1888         /*
1889          * If @pinned_sb, we're reusing an existing root and holding an
1890          * extra ref on its sb.  Mount is complete.  Put the extra ref.
1891          */
1892         if (pinned_sb) {
1893                 WARN_ON(new_sb);
1894                 deactivate_super(pinned_sb);
1895         }
1896 
1897         return dentry;
1898 }
1899 
1900 static void cgroup_kill_sb(struct super_block *sb)
1901 {
1902         struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1903         struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1904 
1905         /*
1906          * If @root doesn't have any mounts or children, start killing it.
1907          * This prevents new mounts by disabling percpu_ref_tryget_live().
1908          * cgroup_mount() may wait for @root's release.
1909          *
1910          * And don't kill the default root.
1911          */
1912         if (!list_empty(&root->cgrp.self.children) ||
1913             root == &cgrp_dfl_root)
1914                 cgroup_put(&root->cgrp);
1915         else
1916                 percpu_ref_kill(&root->cgrp.self.refcnt);
1917 
1918         kernfs_kill_sb(sb);
1919 }
1920 
1921 static struct file_system_type cgroup_fs_type = {
1922         .name = "cgroup",
1923         .mount = cgroup_mount,
1924         .kill_sb = cgroup_kill_sb,
1925 };
1926 
1927 static struct kobject *cgroup_kobj;
1928 
1929 /**
1930  * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1931  * @task: target task
1932  * @buf: the buffer to write the path into
1933  * @buflen: the length of the buffer
1934  *
1935  * Determine @task's cgroup on the first (the one with the lowest non-zero
1936  * hierarchy_id) cgroup hierarchy and copy its path into @buf.  This
1937  * function grabs cgroup_mutex and shouldn't be used inside locks used by
1938  * cgroup controller callbacks.
1939  *
1940  * Return value is the same as kernfs_path().
1941  */
1942 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1943 {
1944         struct cgroup_root *root;
1945         struct cgroup *cgrp;
1946         int hierarchy_id = 1;
1947         char *path = NULL;
1948 
1949         mutex_lock(&cgroup_mutex);
1950         down_read(&css_set_rwsem);
1951 
1952         root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1953 
1954         if (root) {
1955                 cgrp = task_cgroup_from_root(task, root);
1956                 path = cgroup_path(cgrp, buf, buflen);
1957         } else {
1958                 /* if no hierarchy exists, everyone is in "/" */
1959                 if (strlcpy(buf, "/", buflen) < buflen)
1960                         path = buf;
1961         }
1962 
1963         up_read(&css_set_rwsem);
1964         mutex_unlock(&cgroup_mutex);
1965         return path;
1966 }
1967 EXPORT_SYMBOL_GPL(task_cgroup_path);
1968 
1969 /* used to track tasks and other necessary states during migration */
1970 struct cgroup_taskset {
1971         /* the src and dst cset list running through cset->mg_node */
1972         struct list_head        src_csets;
1973         struct list_head        dst_csets;
1974 
1975         /*
1976          * Fields for cgroup_taskset_*() iteration.
1977          *
1978          * Before migration is committed, the target migration tasks are on
1979          * ->mg_tasks of the csets on ->src_csets.  After, on ->mg_tasks of
1980          * the csets on ->dst_csets.  ->csets point to either ->src_csets
1981          * or ->dst_csets depending on whether migration is committed.
1982          *
1983          * ->cur_csets and ->cur_task point to the current task position
1984          * during iteration.
1985          */
1986         struct list_head        *csets;
1987         struct css_set          *cur_cset;
1988         struct task_struct      *cur_task;
1989 };
1990 
1991 /**
1992  * cgroup_taskset_first - reset taskset and return the first task
1993  * @tset: taskset of interest
1994  *
1995  * @tset iteration is initialized and the first task is returned.
1996  */
1997 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
1998 {
1999         tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2000         tset->cur_task = NULL;
2001 
2002         return cgroup_taskset_next(tset);
2003 }
2004 
2005 /**
2006  * cgroup_taskset_next - iterate to the next task in taskset
2007  * @tset: taskset of interest
2008  *
2009  * Return the next task in @tset.  Iteration must have been initialized
2010  * with cgroup_taskset_first().
2011  */
2012 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
2013 {
2014         struct css_set *cset = tset->cur_cset;
2015         struct task_struct *task = tset->cur_task;
2016 
2017         while (&cset->mg_node != tset->csets) {
2018                 if (!task)
2019                         task = list_first_entry(&cset->mg_tasks,
2020                                                 struct task_struct, cg_list);
2021                 else
2022                         task = list_next_entry(task, cg_list);
2023 
2024                 if (&task->cg_list != &cset->mg_tasks) {
2025                         tset->cur_cset = cset;
2026                         tset->cur_task = task;
2027                         return task;
2028                 }
2029 
2030                 cset = list_next_entry(cset, mg_node);
2031                 task = NULL;
2032         }
2033 
2034         return NULL;
2035 }
2036 
2037 /**
2038  * cgroup_task_migrate - move a task from one cgroup to another.
2039  * @old_cgrp: the cgroup @tsk is being migrated from
2040  * @tsk: the task being migrated
2041  * @new_cset: the new css_set @tsk is being attached to
2042  *
2043  * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2044  */
2045 static void cgroup_task_migrate(struct cgroup *old_cgrp,
2046                                 struct task_struct *tsk,
2047                                 struct css_set *new_cset)
2048 {
2049         struct css_set *old_cset;
2050 
2051         lockdep_assert_held(&cgroup_mutex);
2052         lockdep_assert_held(&css_set_rwsem);
2053 
2054         /*
2055          * We are synchronized through threadgroup_lock() against PF_EXITING
2056          * setting such that we can't race against cgroup_exit() changing the
2057          * css_set to init_css_set and dropping the old one.
2058          */
2059         WARN_ON_ONCE(tsk->flags & PF_EXITING);
2060         old_cset = task_css_set(tsk);
2061 
2062         get_css_set(new_cset);
2063         rcu_assign_pointer(tsk->cgroups, new_cset);
2064 
2065         /*
2066          * Use move_tail so that cgroup_taskset_first() still returns the
2067          * leader after migration.  This works because cgroup_migrate()
2068          * ensures that the dst_cset of the leader is the first on the
2069          * tset's dst_csets list.
2070          */
2071         list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2072 
2073         /*
2074          * We just gained a reference on old_cset by taking it from the
2075          * task. As trading it for new_cset is protected by cgroup_mutex,
2076          * we're safe to drop it here; it will be freed under RCU.
2077          */
2078         put_css_set_locked(old_cset);
2079 }
2080 
2081 /**
2082  * cgroup_migrate_finish - cleanup after attach
2083  * @preloaded_csets: list of preloaded css_sets
2084  *
2085  * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst().  See
2086  * those functions for details.
2087  */
2088 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2089 {
2090         struct css_set *cset, *tmp_cset;
2091 
2092         lockdep_assert_held(&cgroup_mutex);
2093 
2094         down_write(&css_set_rwsem);
2095         list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2096                 cset->mg_src_cgrp = NULL;
2097                 cset->mg_dst_cset = NULL;
2098                 list_del_init(&cset->mg_preload_node);
2099                 put_css_set_locked(cset);
2100         }
2101         up_write(&css_set_rwsem);
2102 }
2103 
2104 /**
2105  * cgroup_migrate_add_src - add a migration source css_set
2106  * @src_cset: the source css_set to add
2107  * @dst_cgrp: the destination cgroup
2108  * @preloaded_csets: list of preloaded css_sets
2109  *
2110  * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp.  Pin
2111  * @src_cset and add it to @preloaded_csets, which should later be cleaned
2112  * up by cgroup_migrate_finish().
2113  *
2114  * This function may be called without holding threadgroup_lock even if the
2115  * target is a process.  Threads may be created and destroyed but as long
2116  * as cgroup_mutex is not dropped, no new css_set can be put into play and
2117  * the preloaded css_sets are guaranteed to cover all migrations.
2118  */
2119 static void cgroup_migrate_add_src(struct css_set *src_cset,
2120                                    struct cgroup *dst_cgrp,
2121                                    struct list_head *preloaded_csets)
2122 {
2123         struct cgroup *src_cgrp;
2124 
2125         lockdep_assert_held(&cgroup_mutex);
2126         lockdep_assert_held(&css_set_rwsem);
2127 
2128         src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2129 
2130         if (!list_empty(&src_cset->mg_preload_node))
2131                 return;
2132 
2133         WARN_ON(src_cset->mg_src_cgrp);
2134         WARN_ON(!list_empty(&src_cset->mg_tasks));
2135         WARN_ON(!list_empty(&src_cset->mg_node));
2136 
2137         src_cset->mg_src_cgrp = src_cgrp;
2138         get_css_set(src_cset);
2139         list_add(&src_cset->mg_preload_node, preloaded_csets);
2140 }
2141 
2142 /**
2143  * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2144  * @dst_cgrp: the destination cgroup (may be %NULL)
2145  * @preloaded_csets: list of preloaded source css_sets
2146  *
2147  * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2148  * have been preloaded to @preloaded_csets.  This function looks up and
2149  * pins all destination css_sets, links each to its source, and append them
2150  * to @preloaded_csets.  If @dst_cgrp is %NULL, the destination of each
2151  * source css_set is assumed to be its cgroup on the default hierarchy.
2152  *
2153  * This function must be called after cgroup_migrate_add_src() has been
2154  * called on each migration source css_set.  After migration is performed
2155  * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2156  * @preloaded_csets.
2157  */
2158 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2159                                       struct list_head *preloaded_csets)
2160 {
2161         LIST_HEAD(csets);
2162         struct css_set *src_cset, *tmp_cset;
2163 
2164         lockdep_assert_held(&cgroup_mutex);
2165 
2166         /*
2167          * Except for the root, child_subsys_mask must be zero for a cgroup
2168          * with tasks so that child cgroups don't compete against tasks.
2169          */
2170         if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2171             dst_cgrp->child_subsys_mask)
2172                 return -EBUSY;
2173 
2174         /* look up the dst cset for each src cset and link it to src */
2175         list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2176                 struct css_set *dst_cset;
2177 
2178                 dst_cset = find_css_set(src_cset,
2179                                         dst_cgrp ?: src_cset->dfl_cgrp);
2180                 if (!dst_cset)
2181                         goto err;
2182 
2183                 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2184 
2185                 /*
2186                  * If src cset equals dst, it's noop.  Drop the src.
2187                  * cgroup_migrate() will skip the cset too.  Note that we
2188                  * can't handle src == dst as some nodes are used by both.
2189                  */
2190                 if (src_cset == dst_cset) {
2191                         src_cset->mg_src_cgrp = NULL;
2192                         list_del_init(&src_cset->mg_preload_node);
2193                         put_css_set(src_cset);
2194                         put_css_set(dst_cset);
2195                         continue;
2196                 }
2197 
2198                 src_cset->mg_dst_cset = dst_cset;
2199 
2200                 if (list_empty(&dst_cset->mg_preload_node))
2201                         list_add(&dst_cset->mg_preload_node, &csets);
2202                 else
2203                         put_css_set(dst_cset);
2204         }
2205 
2206         list_splice_tail(&csets, preloaded_csets);
2207         return 0;
2208 err:
2209         cgroup_migrate_finish(&csets);
2210         return -ENOMEM;
2211 }
2212 
2213 /**
2214  * cgroup_migrate - migrate a process or task to a cgroup
2215  * @cgrp: the destination cgroup
2216  * @leader: the leader of the process or the task to migrate
2217  * @threadgroup: whether @leader points to the whole process or a single task
2218  *
2219  * Migrate a process or task denoted by @leader to @cgrp.  If migrating a
2220  * process, the caller must be holding threadgroup_lock of @leader.  The
2221  * caller is also responsible for invoking cgroup_migrate_add_src() and
2222  * cgroup_migrate_prepare_dst() on the targets before invoking this
2223  * function and following up with cgroup_migrate_finish().
2224  *
2225  * As long as a controller's ->can_attach() doesn't fail, this function is
2226  * guaranteed to succeed.  This means that, excluding ->can_attach()
2227  * failure, when migrating multiple targets, the success or failure can be
2228  * decided for all targets by invoking group_migrate_prepare_dst() before
2229  * actually starting migrating.
2230  */
2231 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2232                           bool threadgroup)
2233 {
2234         struct cgroup_taskset tset = {
2235                 .src_csets      = LIST_HEAD_INIT(tset.src_csets),
2236                 .dst_csets      = LIST_HEAD_INIT(tset.dst_csets),
2237                 .csets          = &tset.src_csets,
2238         };
2239         struct cgroup_subsys_state *css, *failed_css = NULL;
2240         struct css_set *cset, *tmp_cset;
2241         struct task_struct *task, *tmp_task;
2242         int i, ret;
2243 
2244         /*
2245          * Prevent freeing of tasks while we take a snapshot. Tasks that are
2246          * already PF_EXITING could be freed from underneath us unless we
2247          * take an rcu_read_lock.
2248          */
2249         down_write(&css_set_rwsem);
2250         rcu_read_lock();
2251         task = leader;
2252         do {
2253                 /* @task either already exited or can't exit until the end */
2254                 if (task->flags & PF_EXITING)
2255                         goto next;
2256 
2257                 /* leave @task alone if post_fork() hasn't linked it yet */
2258                 if (list_empty(&task->cg_list))
2259                         goto next;
2260 
2261                 cset = task_css_set(task);
2262                 if (!cset->mg_src_cgrp)
2263                         goto next;
2264 
2265                 /*
2266                  * cgroup_taskset_first() must always return the leader.
2267                  * Take care to avoid disturbing the ordering.
2268                  */
2269                 list_move_tail(&task->cg_list, &cset->mg_tasks);
2270                 if (list_empty(&cset->mg_node))
2271                         list_add_tail(&cset->mg_node, &tset.src_csets);
2272                 if (list_empty(&cset->mg_dst_cset->mg_node))
2273                         list_move_tail(&cset->mg_dst_cset->mg_node,
2274                                        &tset.dst_csets);
2275         next:
2276                 if (!threadgroup)
2277                         break;
2278         } while_each_thread(leader, task);
2279         rcu_read_unlock();
2280         up_write(&css_set_rwsem);
2281 
2282         /* methods shouldn't be called if no task is actually migrating */
2283         if (list_empty(&tset.src_csets))
2284                 return 0;
2285 
2286         /* check that we can legitimately attach to the cgroup */
2287         for_each_e_css(css, i, cgrp) {
2288                 if (css->ss->can_attach) {
2289                         ret = css->ss->can_attach(css, &tset);
2290                         if (ret) {
2291                                 failed_css = css;
2292                                 goto out_cancel_attach;
2293                         }
2294                 }
2295         }
2296 
2297         /*
2298          * Now that we're guaranteed success, proceed to move all tasks to
2299          * the new cgroup.  There are no failure cases after here, so this
2300          * is the commit point.
2301          */
2302         down_write(&css_set_rwsem);
2303         list_for_each_entry(cset, &tset.src_csets, mg_node) {
2304                 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2305                         cgroup_task_migrate(cset->mg_src_cgrp, task,
2306                                             cset->mg_dst_cset);
2307         }
2308         up_write(&css_set_rwsem);
2309 
2310         /*
2311          * Migration is committed, all target tasks are now on dst_csets.
2312          * Nothing is sensitive to fork() after this point.  Notify
2313          * controllers that migration is complete.
2314          */
2315         tset.csets = &tset.dst_csets;
2316 
2317         for_each_e_css(css, i, cgrp)
2318                 if (css->ss->attach)
2319                         css->ss->attach(css, &tset);
2320 
2321         ret = 0;
2322         goto out_release_tset;
2323 
2324 out_cancel_attach:
2325         for_each_e_css(css, i, cgrp) {
2326                 if (css == failed_css)
2327                         break;
2328                 if (css->ss->cancel_attach)
2329                         css->ss->cancel_attach(css, &tset);
2330         }
2331 out_release_tset:
2332         down_write(&css_set_rwsem);
2333         list_splice_init(&tset.dst_csets, &tset.src_csets);
2334         list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2335                 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2336                 list_del_init(&cset->mg_node);
2337         }
2338         up_write(&css_set_rwsem);
2339         return ret;
2340 }
2341 
2342 /**
2343  * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2344  * @dst_cgrp: the cgroup to attach to
2345  * @leader: the task or the leader of the threadgroup to be attached
2346  * @threadgroup: attach the whole threadgroup?
2347  *
2348  * Call holding cgroup_mutex and threadgroup_lock of @leader.
2349  */
2350 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2351                               struct task_struct *leader, bool threadgroup)
2352 {
2353         LIST_HEAD(preloaded_csets);
2354         struct task_struct *task;
2355         int ret;
2356 
2357         /* look up all src csets */
2358         down_read(&css_set_rwsem);
2359         rcu_read_lock();
2360         task = leader;
2361         do {
2362                 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2363                                        &preloaded_csets);
2364                 if (!threadgroup)
2365                         break;
2366         } while_each_thread(leader, task);
2367         rcu_read_unlock();
2368         up_read(&css_set_rwsem);
2369 
2370         /* prepare dst csets and commit */
2371         ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2372         if (!ret)
2373                 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2374 
2375         cgroup_migrate_finish(&preloaded_csets);
2376         return ret;
2377 }
2378 
2379 /*
2380  * Find the task_struct of the task to attach by vpid and pass it along to the
2381  * function to attach either it or all tasks in its threadgroup. Will lock
2382  * cgroup_mutex and threadgroup.
2383  */
2384 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2385                                     size_t nbytes, loff_t off, bool threadgroup)
2386 {
2387         struct task_struct *tsk;
2388         const struct cred *cred = current_cred(), *tcred;
2389         struct cgroup *cgrp;
2390         pid_t pid;
2391         int ret;
2392 
2393         if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2394                 return -EINVAL;
2395 
2396         cgrp = cgroup_kn_lock_live(of->kn);
2397         if (!cgrp)
2398                 return -ENODEV;
2399 
2400 retry_find_task:
2401         rcu_read_lock();
2402         if (pid) {
2403                 tsk = find_task_by_vpid(pid);
2404                 if (!tsk) {
2405                         rcu_read_unlock();
2406                         ret = -ESRCH;
2407                         goto out_unlock_cgroup;
2408                 }
2409                 /*
2410                  * even if we're attaching all tasks in the thread group, we
2411                  * only need to check permissions on one of them.
2412                  */
2413                 tcred = __task_cred(tsk);
2414                 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2415                     !uid_eq(cred->euid, tcred->uid) &&
2416                     !uid_eq(cred->euid, tcred->suid)) {
2417                         rcu_read_unlock();
2418                         ret = -EACCES;
2419                         goto out_unlock_cgroup;
2420                 }
2421         } else
2422                 tsk = current;
2423 
2424         if (threadgroup)
2425                 tsk = tsk->group_leader;
2426 
2427         /*
2428          * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2429          * trapped in a cpuset, or RT worker may be born in a cgroup
2430          * with no rt_runtime allocated.  Just say no.
2431          */
2432         if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2433                 ret = -EINVAL;
2434                 rcu_read_unlock();
2435                 goto out_unlock_cgroup;
2436         }
2437 
2438         get_task_struct(tsk);
2439         rcu_read_unlock();
2440 
2441         threadgroup_lock(tsk);
2442         if (threadgroup) {
2443                 if (!thread_group_leader(tsk)) {
2444                         /*
2445                          * a race with de_thread from another thread's exec()
2446                          * may strip us of our leadership, if this happens,
2447                          * there is no choice but to throw this task away and
2448                          * try again; this is
2449                          * "double-double-toil-and-trouble-check locking".
2450                          */
2451                         threadgroup_unlock(tsk);
2452                         put_task_struct(tsk);
2453                         goto retry_find_task;
2454                 }
2455         }
2456 
2457         ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2458 
2459         threadgroup_unlock(tsk);
2460 
2461         put_task_struct(tsk);
2462 out_unlock_cgroup:
2463         cgroup_kn_unlock(of->kn);
2464         return ret ?: nbytes;
2465 }
2466 
2467 /**
2468  * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2469  * @from: attach to all cgroups of a given task
2470  * @tsk: the task to be attached
2471  */
2472 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2473 {
2474         struct cgroup_root *root;
2475         int retval = 0;
2476 
2477         mutex_lock(&cgroup_mutex);
2478         for_each_root(root) {
2479                 struct cgroup *from_cgrp;
2480 
2481                 if (root == &cgrp_dfl_root)
2482                         continue;
2483 
2484                 down_read(&css_set_rwsem);
2485                 from_cgrp = task_cgroup_from_root(from, root);
2486                 up_read(&css_set_rwsem);
2487 
2488                 retval = cgroup_attach_task(from_cgrp, tsk, false);
2489                 if (retval)
2490                         break;
2491         }
2492         mutex_unlock(&cgroup_mutex);
2493 
2494         return retval;
2495 }
2496 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2497 
2498 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2499                                   char *buf, size_t nbytes, loff_t off)
2500 {
2501         return __cgroup_procs_write(of, buf, nbytes, off, false);
2502 }
2503 
2504 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2505                                   char *buf, size_t nbytes, loff_t off)
2506 {
2507         return __cgroup_procs_write(of, buf, nbytes, off, true);
2508 }
2509 
2510 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2511                                           char *buf, size_t nbytes, loff_t off)
2512 {
2513         struct cgroup *cgrp;
2514 
2515         BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2516 
2517         cgrp = cgroup_kn_lock_live(of->kn);
2518         if (!cgrp)
2519                 return -ENODEV;
2520         spin_lock(&release_agent_path_lock);
2521         strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2522                 sizeof(cgrp->root->release_agent_path));
2523         spin_unlock(&release_agent_path_lock);
2524         cgroup_kn_unlock(of->kn);
2525         return nbytes;
2526 }
2527 
2528 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2529 {
2530         struct cgroup *cgrp = seq_css(seq)->cgroup;
2531 
2532         spin_lock(&release_agent_path_lock);
2533         seq_puts(seq, cgrp->root->release_agent_path);
2534         spin_unlock(&release_agent_path_lock);
2535         seq_putc(seq, '\n');
2536         return 0;
2537 }
2538 
2539 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2540 {
2541         seq_puts(seq, "\n");
2542         return 0;
2543 }
2544 
2545 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned int ss_mask)
2546 {
2547         struct cgroup_subsys *ss;
2548         bool printed = false;
2549         int ssid;
2550 
2551         for_each_subsys(ss, ssid) {
2552                 if (ss_mask & (1 << ssid)) {
2553                         if (printed)
2554                                 seq_putc(seq, ' ');
2555                         seq_printf(seq, "%s", ss->name);
2556                         printed = true;
2557                 }
2558         }
2559         if (printed)
2560                 seq_putc(seq, '\n');
2561 }
2562 
2563 /* show controllers which are currently attached to the default hierarchy */
2564 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2565 {
2566         struct cgroup *cgrp = seq_css(seq)->cgroup;
2567 
2568         cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2569                              ~cgrp_dfl_root_inhibit_ss_mask);
2570         return 0;
2571 }
2572 
2573 /* show controllers which are enabled from the parent */
2574 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2575 {
2576         struct cgroup *cgrp = seq_css(seq)->cgroup;
2577 
2578         cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2579         return 0;
2580 }
2581 
2582 /* show controllers which are enabled for a given cgroup's children */
2583 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2584 {
2585         struct cgroup *cgrp = seq_css(seq)->cgroup;
2586 
2587         cgroup_print_ss_mask(seq, cgrp->subtree_control);
2588         return 0;
2589 }
2590 
2591 /**
2592  * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2593  * @cgrp: root of the subtree to update csses for
2594  *
2595  * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2596  * css associations need to be updated accordingly.  This function looks up
2597  * all css_sets which are attached to the subtree, creates the matching
2598  * updated css_sets and migrates the tasks to the new ones.
2599  */
2600 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2601 {
2602         LIST_HEAD(preloaded_csets);
2603         struct cgroup_subsys_state *css;
2604         struct css_set *src_cset;
2605         int ret;
2606 
2607         lockdep_assert_held(&cgroup_mutex);
2608 
2609         /* look up all csses currently attached to @cgrp's subtree */
2610         down_read(&css_set_rwsem);
2611         css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2612                 struct cgrp_cset_link *link;
2613 
2614                 /* self is not affected by child_subsys_mask change */
2615                 if (css->cgroup == cgrp)
2616                         continue;
2617 
2618                 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2619                         cgroup_migrate_add_src(link->cset, cgrp,
2620                                                &preloaded_csets);
2621         }
2622         up_read(&css_set_rwsem);
2623 
2624         /* NULL dst indicates self on default hierarchy */
2625         ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2626         if (ret)
2627                 goto out_finish;
2628 
2629         list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2630                 struct task_struct *last_task = NULL, *task;
2631 
2632                 /* src_csets precede dst_csets, break on the first dst_cset */
2633                 if (!src_cset->mg_src_cgrp)
2634                         break;
2635 
2636                 /*
2637                  * All tasks in src_cset need to be migrated to the
2638                  * matching dst_cset.  Empty it process by process.  We
2639                  * walk tasks but migrate processes.  The leader might even
2640                  * belong to a different cset but such src_cset would also
2641                  * be among the target src_csets because the default
2642                  * hierarchy enforces per-process membership.
2643                  */
2644                 while (true) {
2645                         down_read(&css_set_rwsem);
2646                         task = list_first_entry_or_null(&src_cset->tasks,
2647                                                 struct task_struct, cg_list);
2648                         if (task) {
2649                                 task = task->group_leader;
2650                                 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2651                                 get_task_struct(task);
2652                         }
2653                         up_read(&css_set_rwsem);
2654 
2655                         if (!task)
2656                                 break;
2657 
2658                         /* guard against possible infinite loop */
2659                         if (WARN(last_task == task,
2660                                  "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2661                                 goto out_finish;
2662                         last_task = task;
2663 
2664                         threadgroup_lock(task);
2665                         /* raced against de_thread() from another thread? */
2666                         if (!thread_group_leader(task)) {
2667                                 threadgroup_unlock(task);
2668                                 put_task_struct(task);
2669                                 continue;
2670                         }
2671 
2672                         ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2673 
2674                         threadgroup_unlock(task);
2675                         put_task_struct(task);
2676 
2677                         if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2678                                 goto out_finish;
2679                 }
2680         }
2681 
2682 out_finish:
2683         cgroup_migrate_finish(&preloaded_csets);
2684         return ret;
2685 }
2686 
2687 /* change the enabled child controllers for a cgroup in the default hierarchy */
2688 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2689                                             char *buf, size_t nbytes,
2690                                             loff_t off)
2691 {
2692         unsigned int enable = 0, disable = 0;
2693         unsigned int css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2694         struct cgroup *cgrp, *child;
2695         struct cgroup_subsys *ss;
2696         char *tok;
2697         int ssid, ret;
2698 
2699         /*
2700          * Parse input - space separated list of subsystem names prefixed
2701          * with either + or -.
2702          */
2703         buf = strstrip(buf);
2704         while ((tok = strsep(&buf, " "))) {
2705                 if (tok[0] == '\0')
2706                         continue;
2707                 for_each_subsys(ss, ssid) {
2708                         if (ss->disabled || strcmp(tok + 1, ss->name) ||
2709                             ((1 << ss->id) & cgrp_dfl_root_inhibit_ss_mask))
2710                                 continue;
2711 
2712                         if (*tok == '+') {
2713                                 enable |= 1 << ssid;
2714                                 disable &= ~(1 << ssid);
2715                         } else if (*tok == '-') {
2716                                 disable |= 1 << ssid;
2717                                 enable &= ~(1 << ssid);
2718                         } else {
2719                                 return -EINVAL;
2720                         }
2721                         break;
2722                 }
2723                 if (ssid == CGROUP_SUBSYS_COUNT)
2724                         return -EINVAL;
2725         }
2726 
2727         cgrp = cgroup_kn_lock_live(of->kn);
2728         if (!cgrp)
2729                 return -ENODEV;
2730 
2731         for_each_subsys(ss, ssid) {
2732                 if (enable & (1 << ssid)) {
2733                         if (cgrp->subtree_control & (1 << ssid)) {
2734                                 enable &= ~(1 << ssid);
2735                                 continue;
2736                         }
2737 
2738                         /* unavailable or not enabled on the parent? */
2739                         if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2740                             (cgroup_parent(cgrp) &&
2741                              !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2742                                 ret = -ENOENT;
2743                                 goto out_unlock;
2744                         }
2745                 } else if (disable & (1 << ssid)) {
2746                         if (!(cgrp->subtree_control & (1 << ssid))) {
2747                                 disable &= ~(1 << ssid);
2748                                 continue;
2749                         }
2750 
2751                         /* a child has it enabled? */
2752                         cgroup_for_each_live_child(child, cgrp) {
2753                                 if (child->subtree_control & (1 << ssid)) {
2754                                         ret = -EBUSY;
2755                                         goto out_unlock;
2756                                 }
2757                         }
2758                 }
2759         }
2760 
2761         if (!enable && !disable) {
2762                 ret = 0;
2763                 goto out_unlock;
2764         }
2765 
2766         /*
2767          * Except for the root, subtree_control must be zero for a cgroup
2768          * with tasks so that child cgroups don't compete against tasks.
2769          */
2770         if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2771                 ret = -EBUSY;
2772                 goto out_unlock;
2773         }
2774 
2775         /*
2776          * Update subsys masks and calculate what needs to be done.  More
2777          * subsystems than specified may need to be enabled or disabled
2778          * depending on subsystem dependencies.
2779          */
2780         old_sc = cgrp->subtree_control;
2781         old_ss = cgrp->child_subsys_mask;
2782         new_sc = (old_sc | enable) & ~disable;
2783         new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
2784 
2785         css_enable = ~old_ss & new_ss;
2786         css_disable = old_ss & ~new_ss;
2787         enable |= css_enable;
2788         disable |= css_disable;
2789 
2790         /*
2791          * Because css offlining is asynchronous, userland might try to
2792          * re-enable the same controller while the previous instance is
2793          * still around.  In such cases, wait till it's gone using
2794          * offline_waitq.
2795          */
2796         for_each_subsys(ss, ssid) {
2797                 if (!(css_enable & (1 << ssid)))
2798                         continue;
2799 
2800                 cgroup_for_each_live_child(child, cgrp) {
2801                         DEFINE_WAIT(wait);
2802 
2803                         if (!cgroup_css(child, ss))
2804                                 continue;
2805 
2806                         cgroup_get(child);
2807                         prepare_to_wait(&child->offline_waitq, &wait,
2808                                         TASK_UNINTERRUPTIBLE);
2809                         cgroup_kn_unlock(of->kn);
2810                         schedule();
2811                         finish_wait(&child->offline_waitq, &wait);
2812                         cgroup_put(child);
2813 
2814                         return restart_syscall();
2815                 }
2816         }
2817 
2818         cgrp->subtree_control = new_sc;
2819         cgrp->child_subsys_mask = new_ss;
2820 
2821         /*
2822          * Create new csses or make the existing ones visible.  A css is
2823          * created invisible if it's being implicitly enabled through
2824          * dependency.  An invisible css is made visible when the userland
2825          * explicitly enables it.
2826          */
2827         for_each_subsys(ss, ssid) {
2828                 if (!(enable & (1 << ssid)))
2829                         continue;
2830 
2831                 cgroup_for_each_live_child(child, cgrp) {
2832                         if (css_enable & (1 << ssid))
2833                                 ret = create_css(child, ss,
2834                                         cgrp->subtree_control & (1 << ssid));
2835                         else
2836                                 ret = cgroup_populate_dir(child, 1 << ssid);
2837                         if (ret)
2838                                 goto err_undo_css;
2839                 }
2840         }
2841 
2842         /*
2843          * At this point, cgroup_e_css() results reflect the new csses
2844          * making the following cgroup_update_dfl_csses() properly update
2845          * css associations of all tasks in the subtree.
2846          */
2847         ret = cgroup_update_dfl_csses(cgrp);
2848         if (ret)
2849                 goto err_undo_css;
2850 
2851         /*
2852          * All tasks are migrated out of disabled csses.  Kill or hide
2853          * them.  A css is hidden when the userland requests it to be
2854          * disabled while other subsystems are still depending on it.  The
2855          * css must not actively control resources and be in the vanilla
2856          * state if it's made visible again later.  Controllers which may
2857          * be depended upon should provide ->css_reset() for this purpose.
2858          */
2859         for_each_subsys(ss, ssid) {
2860                 if (!(disable & (1 << ssid)))
2861                         continue;
2862 
2863                 cgroup_for_each_live_child(child, cgrp) {
2864                         struct cgroup_subsys_state *css = cgroup_css(child, ss);
2865 
2866                         if (css_disable & (1 << ssid)) {
2867                                 kill_css(css);
2868                         } else {
2869                                 cgroup_clear_dir(child, 1 << ssid);
2870                                 if (ss->css_reset)
2871                                         ss->css_reset(css);
2872                         }
2873                 }
2874         }
2875 
2876         /*
2877          * The effective csses of all the descendants (excluding @cgrp) may
2878          * have changed.  Subsystems can optionally subscribe to this event
2879          * by implementing ->css_e_css_changed() which is invoked if any of
2880          * the effective csses seen from the css's cgroup may have changed.
2881          */
2882         for_each_subsys(ss, ssid) {
2883                 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
2884                 struct cgroup_subsys_state *css;
2885 
2886                 if (!ss->css_e_css_changed || !this_css)
2887                         continue;
2888 
2889                 css_for_each_descendant_pre(css, this_css)
2890                         if (css != this_css)
2891                                 ss->css_e_css_changed(css);
2892         }
2893 
2894         kernfs_activate(cgrp->kn);
2895         ret = 0;
2896 out_unlock:
2897         cgroup_kn_unlock(of->kn);
2898         return ret ?: nbytes;
2899 
2900 err_undo_css:
2901         cgrp->subtree_control = old_sc;
2902         cgrp->child_subsys_mask = old_ss;
2903 
2904         for_each_subsys(ss, ssid) {
2905                 if (!(enable & (1 << ssid)))
2906                         continue;
2907 
2908                 cgroup_for_each_live_child(child, cgrp) {
2909                         struct cgroup_subsys_state *css = cgroup_css(child, ss);
2910 
2911                         if (!css)
2912                                 continue;
2913 
2914                         if (css_enable & (1 << ssid))
2915                                 kill_css(css);
2916                         else
2917                                 cgroup_clear_dir(child, 1 << ssid);
2918                 }
2919         }
2920         goto out_unlock;
2921 }
2922 
2923 static int cgroup_populated_show(struct seq_file *seq, void *v)
2924 {
2925         seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2926         return 0;
2927 }
2928 
2929 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2930                                  size_t nbytes, loff_t off)
2931 {
2932         struct cgroup *cgrp = of->kn->parent->priv;
2933         struct cftype *cft = of->kn->priv;
2934         struct cgroup_subsys_state *css;
2935         int ret;
2936 
2937         if (cft->write)
2938                 return cft->write(of, buf, nbytes, off);
2939 
2940         /*
2941          * kernfs guarantees that a file isn't deleted with operations in
2942          * flight, which means that the matching css is and stays alive and
2943          * doesn't need to be pinned.  The RCU locking is not necessary
2944          * either.  It's just for the convenience of using cgroup_css().
2945          */
2946         rcu_read_lock();
2947         css = cgroup_css(cgrp, cft->ss);
2948         rcu_read_unlock();
2949 
2950         if (cft->write_u64) {
2951                 unsigned long long v;
2952                 ret = kstrtoull(buf, 0, &v);
2953                 if (!ret)
2954                         ret = cft->write_u64(css, cft, v);
2955         } else if (cft->write_s64) {
2956                 long long v;
2957                 ret = kstrtoll(buf, 0, &v);
2958                 if (!ret)
2959                         ret = cft->write_s64(css, cft, v);
2960         } else {
2961                 ret = -EINVAL;
2962         }
2963 
2964         return ret ?: nbytes;
2965 }
2966 
2967 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
2968 {
2969         return seq_cft(seq)->seq_start(seq, ppos);
2970 }
2971 
2972 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
2973 {
2974         return seq_cft(seq)->seq_next(seq, v, ppos);
2975 }
2976 
2977 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
2978 {
2979         seq_cft(seq)->seq_stop(seq, v);
2980 }
2981 
2982 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
2983 {
2984         struct cftype *cft = seq_cft(m);
2985         struct cgroup_subsys_state *css = seq_css(m);
2986 
2987         if (cft->seq_show)
2988                 return cft->seq_show(m, arg);
2989 
2990         if (cft->read_u64)
2991                 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
2992         else if (cft->read_s64)
2993                 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
2994         else
2995                 return -EINVAL;
2996         return 0;
2997 }
2998 
2999 static struct kernfs_ops cgroup_kf_single_ops = {
3000         .atomic_write_len       = PAGE_SIZE,
3001         .write                  = cgroup_file_write,
3002         .seq_show               = cgroup_seqfile_show,
3003 };
3004 
3005 static struct kernfs_ops cgroup_kf_ops = {
3006         .atomic_write_len       = PAGE_SIZE,
3007         .write                  = cgroup_file_write,
3008         .seq_start              = cgroup_seqfile_start,
3009         .seq_next               = cgroup_seqfile_next,
3010         .seq_stop               = cgroup_seqfile_stop,
3011         .seq_show               = cgroup_seqfile_show,
3012 };
3013 
3014 /*
3015  * cgroup_rename - Only allow simple rename of directories in place.
3016  */
3017 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3018                          const char *new_name_str)
3019 {
3020         struct cgroup *cgrp = kn->priv;
3021         int ret;
3022 
3023         if (kernfs_type(kn) != KERNFS_DIR)
3024                 return -ENOTDIR;
3025         if (kn->parent != new_parent)
3026                 return -EIO;
3027 
3028         /*
3029          * This isn't a proper migration and its usefulness is very
3030          * limited.  Disallow on the default hierarchy.
3031          */
3032         if (cgroup_on_dfl(cgrp))
3033                 return -EPERM;
3034 
3035         /*
3036          * We're gonna grab cgroup_mutex which nests outside kernfs
3037          * active_ref.  kernfs_rename() doesn't require active_ref
3038          * protection.  Break them before grabbing cgroup_mutex.
3039          */
3040         kernfs_break_active_protection(new_parent);
3041         kernfs_break_active_protection(kn);
3042 
3043         mutex_lock(&cgroup_mutex);
3044 
3045         ret = kernfs_rename(kn, new_parent, new_name_str);
3046 
3047         mutex_unlock(&cgroup_mutex);
3048 
3049         kernfs_unbreak_active_protection(kn);
3050         kernfs_unbreak_active_protection(new_parent);
3051         return ret;
3052 }
3053 
3054 /* set uid and gid of cgroup dirs and files to that of the creator */
3055 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3056 {
3057         struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3058                                .ia_uid = current_fsuid(),
3059                                .ia_gid = current_fsgid(), };
3060 
3061         if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3062             gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3063                 return 0;
3064 
3065         return kernfs_setattr(kn, &iattr);
3066 }
3067 
3068 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
3069 {
3070         char name[CGROUP_FILE_NAME_MAX];
3071         struct kernfs_node *kn;
3072         struct lock_class_key *key = NULL;
3073         int ret;
3074 
3075 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3076         key = &cft->lockdep_key;
3077 #endif
3078         kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3079                                   cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3080                                   NULL, key);
3081         if (IS_ERR(kn))
3082                 return PTR_ERR(kn);
3083 
3084         ret = cgroup_kn_set_ugid(kn);
3085         if (ret) {
3086                 kernfs_remove(kn);
3087                 return ret;
3088         }
3089 
3090         if (cft->seq_show == cgroup_populated_show)
3091                 cgrp->populated_kn = kn;
3092         return 0;
3093 }
3094 
3095 /**
3096  * cgroup_addrm_files - add or remove files to a cgroup directory
3097  * @cgrp: the target cgroup
3098  * @cfts: array of cftypes to be added
3099  * @is_add: whether to add or remove
3100  *
3101  * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3102  * For removals, this function never fails.  If addition fails, this
3103  * function doesn't remove files already added.  The caller is responsible
3104  * for cleaning up.
3105  */
3106 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
3107                               bool is_add)
3108 {
3109         struct cftype *cft;
3110         int ret;
3111 
3112         lockdep_assert_held(&cgroup_mutex);
3113 
3114         for (cft = cfts; cft->name[0] != '\0'; cft++) {
3115                 /* does cft->flags tell us to skip this file on @cgrp? */
3116                 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3117                         continue;
3118                 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3119                         continue;
3120                 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3121                         continue;
3122                 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3123                         continue;
3124 
3125                 if (is_add) {
3126                         ret = cgroup_add_file(cgrp, cft);
3127                         if (ret) {
3128                                 pr_warn("%s: failed to add %s, err=%d\n",
3129                                         __func__, cft->name, ret);
3130                                 return ret;
3131                         }
3132                 } else {
3133                         cgroup_rm_file(cgrp, cft);
3134                 }
3135         }
3136         return 0;
3137 }
3138 
3139 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3140 {
3141         LIST_HEAD(pending);
3142         struct cgroup_subsys *ss = cfts[0].ss;
3143         struct cgroup *root = &ss->root->cgrp;
3144         struct cgroup_subsys_state *css;
3145         int ret = 0;
3146 
3147         lockdep_assert_held(&cgroup_mutex);
3148 
3149         /* add/rm files for all cgroups created before */
3150         css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3151                 struct cgroup *cgrp = css->cgroup;
3152 
3153                 if (cgroup_is_dead(cgrp))
3154                         continue;
3155 
3156                 ret = cgroup_addrm_files(cgrp, cfts, is_add);
3157                 if (ret)
3158                         break;
3159         }
3160 
3161         if (is_add && !ret)
3162                 kernfs_activate(root->kn);
3163         return ret;
3164 }
3165 
3166 static void cgroup_exit_cftypes(struct cftype *cfts)
3167 {
3168         struct cftype *cft;
3169 
3170         for (cft = cfts; cft->name[0] != '\0'; cft++) {
3171                 /* free copy for custom atomic_write_len, see init_cftypes() */
3172                 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3173                         kfree(cft->kf_ops);
3174                 cft->kf_ops = NULL;
3175                 cft->ss = NULL;
3176 
3177                 /* revert flags set by cgroup core while adding @cfts */
3178                 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3179         }
3180 }
3181 
3182 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3183 {
3184         struct cftype *cft;
3185 
3186         for (cft = cfts; cft->name[0] != '\0'; cft++) {
3187                 struct kernfs_ops *kf_ops;
3188 
3189                 WARN_ON(cft->ss || cft->kf_ops);
3190 
3191                 if (cft->seq_start)
3192                         kf_ops = &cgroup_kf_ops;
3193                 else
3194                         kf_ops = &cgroup_kf_single_ops;
3195 
3196                 /*
3197                  * Ugh... if @cft wants a custom max_write_len, we need to
3198                  * make a copy of kf_ops to set its atomic_write_len.
3199                  */
3200                 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3201                         kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3202                         if (!kf_ops) {
3203                                 cgroup_exit_cftypes(cfts);
3204                                 return -ENOMEM;
3205                         }
3206                         kf_ops->atomic_write_len = cft->max_write_len;
3207                 }
3208 
3209                 cft->kf_ops = kf_ops;
3210                 cft->ss = ss;
3211         }
3212 
3213         return 0;
3214 }
3215 
3216 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3217 {
3218         lockdep_assert_held(&cgroup_mutex);
3219 
3220         if (!cfts || !cfts[0].ss)
3221                 return -ENOENT;
3222 
3223         list_del(&cfts->node);
3224         cgroup_apply_cftypes(cfts, false);
3225         cgroup_exit_cftypes(cfts);
3226         return 0;
3227 }
3228 
3229 /**
3230  * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3231  * @cfts: zero-length name terminated array of cftypes
3232  *
3233  * Unregister @cfts.  Files described by @cfts are removed from all
3234  * existing cgroups and all future cgroups won't have them either.  This
3235  * function can be called anytime whether @cfts' subsys is attached or not.
3236  *
3237  * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3238  * registered.
3239  */
3240 int cgroup_rm_cftypes(struct cftype *cfts)
3241 {
3242         int ret;
3243 
3244         mutex_lock(&cgroup_mutex);
3245         ret = cgroup_rm_cftypes_locked(cfts);
3246         mutex_unlock(&cgroup_mutex);
3247         return ret;
3248 }
3249 
3250 /**
3251  * cgroup_add_cftypes - add an array of cftypes to a subsystem
3252  * @ss: target cgroup subsystem
3253  * @cfts: zero-length name terminated array of cftypes
3254  *
3255  * Register @cfts to @ss.  Files described by @cfts are created for all
3256  * existing cgroups to which @ss is attached and all future cgroups will
3257  * have them too.  This function can be called anytime whether @ss is
3258  * attached or not.
3259  *
3260  * Returns 0 on successful registration, -errno on failure.  Note that this
3261  * function currently returns 0 as long as @cfts registration is successful
3262  * even if some file creation attempts on existing cgroups fail.
3263  */
3264 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3265 {
3266         int ret;
3267 
3268         if (ss->disabled)
3269                 return 0;
3270 
3271         if (!cfts || cfts[0].name[0] == '\0')
3272                 return 0;
3273 
3274         ret = cgroup_init_cftypes(ss, cfts);
3275         if (ret)
3276                 return ret;
3277 
3278         mutex_lock(&cgroup_mutex);
3279 
3280         list_add_tail(&cfts->node, &ss->cfts);
3281         ret = cgroup_apply_cftypes(cfts, true);
3282         if (ret)
3283                 cgroup_rm_cftypes_locked(cfts);
3284 
3285         mutex_unlock(&cgroup_mutex);
3286         return ret;
3287 }
3288 
3289 /**
3290  * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3291  * @ss: target cgroup subsystem
3292  * @cfts: zero-length name terminated array of cftypes
3293  *
3294  * Similar to cgroup_add_cftypes() but the added files are only used for
3295  * the default hierarchy.
3296  */
3297 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3298 {
3299         struct cftype *cft;
3300 
3301         for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3302                 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3303         return cgroup_add_cftypes(ss, cfts);
3304 }
3305 
3306 /**
3307  * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3308  * @ss: target cgroup subsystem
3309  * @cfts: zero-length name terminated array of cftypes
3310  *
3311  * Similar to cgroup_add_cftypes() but the added files are only used for
3312  * the legacy hierarchies.
3313  */
3314 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3315 {
3316         struct cftype *cft;
3317 
3318         /*
3319          * If legacy_flies_on_dfl, we want to show the legacy files on the
3320          * dfl hierarchy but iff the target subsystem hasn't been updated
3321          * for the dfl hierarchy yet.
3322          */
3323         if (!cgroup_legacy_files_on_dfl ||
3324             ss->dfl_cftypes != ss->legacy_cftypes) {
3325                 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3326                         cft->flags |= __CFTYPE_NOT_ON_DFL;
3327         }
3328 
3329         return cgroup_add_cftypes(ss, cfts);
3330 }
3331 
3332 /**
3333  * cgroup_task_count - count the number of tasks in a cgroup.
3334  * @cgrp: the cgroup in question
3335  *
3336  * Return the number of tasks in the cgroup.
3337  */
3338 static int cgroup_task_count(const struct cgroup *cgrp)
3339 {
3340         int count = 0;
3341         struct cgrp_cset_link *link;
3342 
3343         down_read(&css_set_rwsem);
3344         list_for_each_entry(link, &cgrp->cset_links, cset_link)
3345                 count += atomic_read(&link->cset->refcount);
3346         up_read(&css_set_rwsem);
3347         return count;
3348 }
3349 
3350 /**
3351  * css_next_child - find the next child of a given css
3352  * @pos: the current position (%NULL to initiate traversal)
3353  * @parent: css whose children to walk
3354  *
3355  * This function returns the next child of @parent and should be called
3356  * under either cgroup_mutex or RCU read lock.  The only requirement is
3357  * that @parent and @pos are accessible.  The next sibling is guaranteed to
3358  * be returned regardless of their states.
3359  *
3360  * If a subsystem synchronizes ->css_online() and the start of iteration, a
3361  * css which finished ->css_online() is guaranteed to be visible in the
3362  * future iterations and will stay visible until the last reference is put.
3363  * A css which hasn't finished ->css_online() or already finished
3364  * ->css_offline() may show up during traversal.  It's each subsystem's
3365  * responsibility to synchronize against on/offlining.
3366  */
3367 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3368                                            struct cgroup_subsys_state *parent)
3369 {
3370         struct cgroup_subsys_state *next;
3371 
3372         cgroup_assert_mutex_or_rcu_locked();
3373 
3374         /*
3375          * @pos could already have been unlinked from the sibling list.
3376          * Once a cgroup is removed, its ->sibling.next is no longer
3377          * updated when its next sibling changes.  CSS_RELEASED is set when
3378          * @pos is taken off list, at which time its next pointer is valid,
3379          * and, as releases are serialized, the one pointed to by the next
3380          * pointer is guaranteed to not have started release yet.  This
3381          * implies that if we observe !CSS_RELEASED on @pos in this RCU
3382          * critical section, the one pointed to by its next pointer is
3383          * guaranteed to not have finished its RCU grace period even if we
3384          * have dropped rcu_read_lock() inbetween iterations.
3385          *
3386          * If @pos has CSS_RELEASED set, its next pointer can't be
3387          * dereferenced; however, as each css is given a monotonically
3388          * increasing unique serial number and always appended to the
3389          * sibling list, the next one can be found by walking the parent's
3390          * children until the first css with higher serial number than
3391          * @pos's.  While this path can be slower, it happens iff iteration
3392          * races against release and the race window is very small.
3393          */
3394         if (!pos) {
3395                 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3396         } else if (likely(!(pos->flags & CSS_RELEASED))) {
3397                 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3398         } else {
3399                 list_for_each_entry_rcu(next, &parent->children, sibling)
3400                         if (next->serial_nr > pos->serial_nr)
3401                                 break;
3402         }
3403 
3404         /*
3405          * @next, if not pointing to the head, can be dereferenced and is
3406          * the next sibling.
3407          */
3408         if (&next->sibling != &parent->children)
3409                 return next;
3410         return NULL;
3411 }
3412 
3413 /**
3414  * css_next_descendant_pre - find the next descendant for pre-order walk
3415  * @pos: the current position (%NULL to initiate traversal)
3416  * @root: css whose descendants to walk
3417  *
3418  * To be used by css_for_each_descendant_pre().  Find the next descendant
3419  * to visit for pre-order traversal of @root's descendants.  @root is
3420  * included in the iteration and the first node to be visited.
3421  *
3422  * While this function requires cgroup_mutex or RCU read locking, it
3423  * doesn't require the whole traversal to be contained in a single critical
3424  * section.  This function will return the correct next descendant as long
3425  * as both @pos and @root are accessible and @pos is a descendant of @root.
3426  *
3427  * If a subsystem synchronizes ->css_online() and the start of iteration, a
3428  * css which finished ->css_online() is guaranteed to be visible in the
3429  * future iterations and will stay visible until the last reference is put.
3430  * A css which hasn't finished ->css_online() or already finished
3431  * ->css_offline() may show up during traversal.  It's each subsystem's
3432  * responsibility to synchronize against on/offlining.
3433  */
3434 struct cgroup_subsys_state *
3435 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3436                         struct cgroup_subsys_state *root)
3437 {
3438         struct cgroup_subsys_state *next;
3439 
3440         cgroup_assert_mutex_or_rcu_locked();
3441 
3442         /* if first iteration, visit @root */
3443         if (!pos)
3444                 return root;
3445 
3446         /* visit the first child if exists */
3447         next = css_next_child(NULL, pos);
3448         if (next)
3449                 return next;
3450 
3451         /* no child, visit my or the closest ancestor's next sibling */
3452         while (pos != root) {
3453                 next = css_next_child(pos, pos->parent);
3454                 if (next)
3455                         return next;
3456                 pos = pos->parent;
3457         }
3458 
3459         return NULL;
3460 }
3461 
3462 /**
3463  * css_rightmost_descendant - return the rightmost descendant of a css
3464  * @pos: css of interest
3465  *
3466  * Return the rightmost descendant of @pos.  If there's no descendant, @pos
3467  * is returned.  This can be used during pre-order traversal to skip
3468  * subtree of @pos.
3469  *
3470  * While this function requires cgroup_mutex or RCU read locking, it
3471  * doesn't require the whole traversal to be contained in a single critical
3472  * section.  This function will return the correct rightmost descendant as
3473  * long as @pos is accessible.
3474  */
3475 struct cgroup_subsys_state *
3476 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3477 {
3478         struct cgroup_subsys_state *last, *tmp;
3479 
3480         cgroup_assert_mutex_or_rcu_locked();
3481 
3482         do {
3483                 last = pos;
3484                 /* ->prev isn't RCU safe, walk ->next till the end */
3485                 pos = NULL;
3486                 css_for_each_child(tmp, last)
3487                         pos = tmp;
3488         } while (pos);
3489 
3490         return last;
3491 }
3492 
3493 static struct cgroup_subsys_state *
3494 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3495 {
3496         struct cgroup_subsys_state *last;
3497 
3498         do {
3499                 last = pos;
3500                 pos = css_next_child(NULL, pos);
3501         } while (pos);
3502 
3503         return last;
3504 }
3505 
3506 /**
3507  * css_next_descendant_post - find the next descendant for post-order walk
3508  * @pos: the current position (%NULL to initiate traversal)
3509  * @root: css whose descendants to walk
3510  *
3511  * To be used by css_for_each_descendant_post().  Find the next descendant
3512  * to visit for post-order traversal of @root's descendants.  @root is
3513  * included in the iteration and the last node to be visited.
3514  *
3515  * While this function requires cgroup_mutex or RCU read locking, it
3516  * doesn't require the whole traversal to be contained in a single critical
3517  * section.  This function will return the correct next descendant as long
3518  * as both @pos and @cgroup are accessible and @pos is a descendant of
3519  * @cgroup.
3520  *
3521  * If a subsystem synchronizes ->css_online() and the start of iteration, a
3522  * css which finished ->css_online() is guaranteed to be visible in the
3523  * future iterations and will stay visible until the last reference is put.
3524  * A css which hasn't finished ->css_online() or already finished
3525  * ->css_offline() may show up during traversal.  It's each subsystem's
3526  * responsibility to synchronize against on/offlining.
3527  */
3528 struct cgroup_subsys_state *
3529 css_next_descendant_post(struct cgroup_subsys_state *pos,
3530                          struct cgroup_subsys_state *root)
3531 {
3532         struct cgroup_subsys_state *next;
3533 
3534         cgroup_assert_mutex_or_rcu_locked();
3535 
3536         /* if first iteration, visit leftmost descendant which may be @root */
3537         if (!pos)
3538                 return css_leftmost_descendant(root);
3539 
3540         /* if we visited @root, we're done */
3541         if (pos == root)
3542                 return NULL;
3543 
3544         /* if there's an unvisited sibling, visit its leftmost descendant */
3545         next = css_next_child(pos, pos->parent);
3546         if (next)
3547                 return css_leftmost_descendant(next);
3548 
3549         /* no sibling left, visit parent */
3550         return pos->parent;
3551 }
3552 
3553 /**
3554  * css_has_online_children - does a css have online children
3555  * @css: the target css
3556  *
3557  * Returns %true if @css has any online children; otherwise, %false.  This
3558  * function can be called from any context but the caller is responsible
3559  * for synchronizing against on/offlining as necessary.
3560  */
3561 bool css_has_online_children(struct cgroup_subsys_state *css)
3562 {
3563         struct cgroup_subsys_state *child;
3564         bool ret = false;
3565 
3566         rcu_read_lock();
3567         css_for_each_child(child, css) {
3568                 if (child->flags & CSS_ONLINE) {
3569                         ret = true;
3570                         break;
3571                 }
3572         }
3573         rcu_read_unlock();
3574         return ret;
3575 }
3576 
3577 /**
3578  * css_advance_task_iter - advance a task itererator to the next css_set
3579  * @it: the iterator to advance
3580  *
3581  * Advance @it to the next css_set to walk.
3582  */
3583 static void css_advance_task_iter(struct css_task_iter *it)
3584 {
3585         struct list_head *l = it->cset_pos;
3586         struct cgrp_cset_link *link;
3587         struct css_set *cset;
3588 
3589         /* Advance to the next non-empty css_set */
3590         do {
3591                 l = l->next;
3592                 if (l == it->cset_head) {
3593                         it->cset_pos = NULL;
3594                         return;
3595                 }
3596 
3597                 if (it->ss) {
3598                         cset = container_of(l, struct css_set,
3599                                             e_cset_node[it->ss->id]);
3600                 } else {
3601                         link = list_entry(l, struct cgrp_cset_link, cset_link);
3602                         cset = link->cset;
3603                 }
3604         } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3605 
3606         it->cset_pos = l;
3607 
3608         if (!list_empty(&cset->tasks))
3609                 it->task_pos = cset->tasks.next;
3610         else
3611                 it->task_pos = cset->mg_tasks.next;
3612 
3613         it->tasks_head = &cset->tasks;
3614         it->mg_tasks_head = &cset->mg_tasks;
3615 }
3616 
3617 /**
3618  * css_task_iter_start - initiate task iteration
3619  * @css: the css to walk tasks of
3620  * @it: the task iterator to use
3621  *
3622  * Initiate iteration through the tasks of @css.  The caller can call
3623  * css_task_iter_next() to walk through the tasks until the function
3624  * returns NULL.  On completion of iteration, css_task_iter_end() must be
3625  * called.
3626  *
3627  * Note that this function acquires a lock which is released when the
3628  * iteration finishes.  The caller can't sleep while iteration is in
3629  * progress.
3630  */
3631 void css_task_iter_start(struct cgroup_subsys_state *css,
3632                          struct css_task_iter *it)
3633         __acquires(css_set_rwsem)
3634 {
3635         /* no one should try to iterate before mounting cgroups */
3636         WARN_ON_ONCE(!use_task_css_set_links);
3637 
3638         down_read(&css_set_rwsem);
3639 
3640         it->ss = css->ss;
3641 
3642         if (it->ss)
3643                 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3644         else
3645                 it->cset_pos = &css->cgroup->cset_links;
3646 
3647         it->cset_head = it->cset_pos;
3648 
3649         css_advance_task_iter(it);
3650 }
3651 
3652 /**
3653  * css_task_iter_next - return the next task for the iterator
3654  * @it: the task iterator being iterated
3655  *
3656  * The "next" function for task iteration.  @it should have been
3657  * initialized via css_task_iter_start().  Returns NULL when the iteration
3658  * reaches the end.
3659  */
3660 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3661 {
3662         struct task_struct *res;
3663         struct list_head *l = it->task_pos;
3664 
3665         /* If the iterator cg is NULL, we have no tasks */
3666         if (!it->cset_pos)
3667                 return NULL;
3668         res = list_entry(l, struct task_struct, cg_list);
3669 
3670         /*
3671          * Advance iterator to find next entry.  cset->tasks is consumed
3672          * first and then ->mg_tasks.  After ->mg_tasks, we move onto the
3673          * next cset.
3674          */
3675         l = l->next;
3676 
3677         if (l == it->tasks_head)
3678                 l = it->mg_tasks_head->next;
3679 
3680         if (l == it->mg_tasks_head)
3681                 css_advance_task_iter(it);
3682         else
3683                 it->task_pos = l;
3684 
3685         return res;
3686 }
3687 
3688 /**
3689  * css_task_iter_end - finish task iteration
3690  * @it: the task iterator to finish
3691  *
3692  * Finish task iteration started by css_task_iter_start().
3693  */
3694 void css_task_iter_end(struct css_task_iter *it)
3695         __releases(css_set_rwsem)
3696 {
3697         up_read(&css_set_rwsem);
3698 }
3699 
3700 /**
3701  * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3702  * @to: cgroup to which the tasks will be moved
3703  * @from: cgroup in which the tasks currently reside
3704  *
3705  * Locking rules between cgroup_post_fork() and the migration path
3706  * guarantee that, if a task is forking while being migrated, the new child
3707  * is guaranteed to be either visible in the source cgroup after the
3708  * parent's migration is complete or put into the target cgroup.  No task
3709  * can slip out of migration through forking.
3710  */
3711 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3712 {
3713         LIST_HEAD(preloaded_csets);
3714         struct cgrp_cset_link *link;
3715         struct css_task_iter it;
3716         struct task_struct *task;
3717         int ret;
3718 
3719         mutex_lock(&cgroup_mutex);
3720 
3721         /* all tasks in @from are being moved, all csets are source */
3722         down_read(&css_set_rwsem);
3723         list_for_each_entry(link, &from->cset_links, cset_link)
3724                 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3725         up_read(&css_set_rwsem);
3726 
3727         ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3728         if (ret)
3729                 goto out_err;
3730 
3731         /*
3732          * Migrate tasks one-by-one until @form is empty.  This fails iff
3733          * ->can_attach() fails.
3734          */
3735         do {
3736                 css_task_iter_start(&from->self, &it);
3737                 task = css_task_iter_next(&it);
3738                 if (task)
3739                         get_task_struct(task);
3740                 css_task_iter_end(&it);
3741 
3742                 if (task) {
3743                         ret = cgroup_migrate(to, task, false);
3744                         put_task_struct(task);
3745                 }
3746         } while (task && !ret);
3747 out_err:
3748         cgroup_migrate_finish(&preloaded_csets);
3749         mutex_unlock(&cgroup_mutex);
3750         return ret;
3751 }
3752 
3753 /*
3754  * Stuff for reading the 'tasks'/'procs' files.
3755  *
3756  * Reading this file can return large amounts of data if a cgroup has
3757  * *lots* of attached tasks. So it may need several calls to read(),
3758  * but we cannot guarantee that the information we produce is correct
3759  * unless we produce it entirely atomically.
3760  *
3761  */
3762 
3763 /* which pidlist file are we talking about? */
3764 enum cgroup_filetype {
3765         CGROUP_FILE_PROCS,
3766         CGROUP_FILE_TASKS,
3767 };
3768 
3769 /*
3770  * A pidlist is a list of pids that virtually represents the contents of one
3771  * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3772  * a pair (one each for procs, tasks) for each pid namespace that's relevant
3773  * to the cgroup.
3774  */
3775 struct cgroup_pidlist {
3776         /*
3777          * used to find which pidlist is wanted. doesn't change as long as
3778          * this particular list stays in the list.
3779         */
3780         struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3781         /* array of xids */
3782         pid_t *list;
3783         /* how many elements the above list has */
3784         int length;
3785         /* each of these stored in a list by its cgroup */
3786         struct list_head links;
3787         /* pointer to the cgroup we belong to, for list removal purposes */
3788         struct cgroup *owner;
3789         /* for delayed destruction */
3790         struct delayed_work destroy_dwork;
3791 };
3792 
3793 /*
3794  * The following two functions "fix" the issue where there are more pids
3795  * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3796  * TODO: replace with a kernel-wide solution to this problem
3797  */
3798 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3799 static void *pidlist_allocate(int count)
3800 {
3801         if (PIDLIST_TOO_LARGE(count))
3802                 return vmalloc(count * sizeof(pid_t));
3803         else
3804                 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3805 }
3806 
3807 static void pidlist_free(void *p)
3808 {
3809         if (is_vmalloc_addr(p))
3810                 vfree(p);
3811         else
3812                 kfree(p);
3813 }
3814 
3815 /*
3816  * Used to destroy all pidlists lingering waiting for destroy timer.  None
3817  * should be left afterwards.
3818  */
3819 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3820 {
3821         struct cgroup_pidlist *l, *tmp_l;
3822 
3823         mutex_lock(&cgrp->pidlist_mutex);
3824         list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3825                 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3826         mutex_unlock(&cgrp->pidlist_mutex);
3827 
3828         flush_workqueue(cgroup_pidlist_destroy_wq);
3829         BUG_ON(!list_empty(&cgrp->pidlists));
3830 }
3831 
3832 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3833 {
3834         struct delayed_work *dwork = to_delayed_work(work);
3835         struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3836                                                 destroy_dwork);
3837         struct cgroup_pidlist *tofree = NULL;
3838 
3839         mutex_lock(&l->owner->pidlist_mutex);
3840 
3841         /*
3842          * Destroy iff we didn't get queued again.  The state won't change
3843          * as destroy_dwork can only be queued while locked.
3844          */
3845         if (!delayed_work_pending(dwork)) {
3846                 list_del(&l->links);
3847                 pidlist_free(l->list);
3848                 put_pid_ns(l->key.ns);
3849                 tofree = l;
3850         }
3851 
3852         mutex_unlock(&l->owner->pidlist_mutex);
3853         kfree(tofree);
3854 }
3855 
3856 /*
3857  * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3858  * Returns the number of unique elements.
3859  */
3860 static int pidlist_uniq(pid_t *list, int length)
3861 {
3862         int src, dest = 1;
3863 
3864         /*
3865          * we presume the 0th element is unique, so i starts at 1. trivial
3866          * edge cases first; no work needs to be done for either
3867          */
3868         if (length == 0 || length == 1)
3869                 return length;
3870         /* src and dest walk down the list; dest counts unique elements */
3871         for (src = 1; src < length; src++) {
3872                 /* find next unique element */
3873                 while (list[src] == list[src-1]) {
3874                         src++;
3875                         if (src == length)
3876                                 goto after;
3877                 }
3878                 /* dest always points to where the next unique element goes */
3879                 list[dest] = list[src];
3880                 dest++;
3881         }
3882 after:
3883         return dest;
3884 }
3885 
3886 /*
3887  * The two pid files - task and cgroup.procs - guaranteed that the result
3888  * is sorted, which forced this whole pidlist fiasco.  As pid order is
3889  * different per namespace, each namespace needs differently sorted list,
3890  * making it impossible to use, for example, single rbtree of member tasks
3891  * sorted by task pointer.  As pidlists can be fairly large, allocating one
3892  * per open file is dangerous, so cgroup had to implement shared pool of
3893  * pidlists keyed by cgroup and namespace.
3894  *
3895  * All this extra complexity was caused by the original implementation
3896  * committing to an entirely unnecessary property.  In the long term, we
3897  * want to do away with it.  Explicitly scramble sort order if on the
3898  * default hierarchy so that no such expectation exists in the new
3899  * interface.
3900  *
3901  * Scrambling is done by swapping every two consecutive bits, which is
3902  * non-identity one-to-one mapping which disturbs sort order sufficiently.
3903  */
3904 static pid_t pid_fry(pid_t pid)
3905 {
3906         unsigned a = pid & 0x55555555;
3907         unsigned b = pid & 0xAAAAAAAA;
3908 
3909         return (a << 1) | (b >> 1);
3910 }
3911 
3912 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3913 {
3914         if (cgroup_on_dfl(cgrp))
3915                 return pid_fry(pid);
3916         else
3917                 return pid;
3918 }
3919 
3920 static int cmppid(const void *a, const void *b)
3921 {
3922         return *(pid_t *)a - *(pid_t *)b;
3923 }
3924 
3925 static int fried_cmppid(const void *a, const void *b)
3926 {
3927         return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3928 }
3929 
3930 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3931                                                   enum cgroup_filetype type)
3932 {
3933         struct cgroup_pidlist *l;
3934         /* don't need task_nsproxy() if we're looking at ourself */
3935         struct pid_namespace *ns = task_active_pid_ns(current);
3936 
3937         lockdep_assert_held(&cgrp->pidlist_mutex);
3938 
3939         list_for_each_entry(l, &cgrp->pidlists, links)
3940                 if (l->key.type == type && l->key.ns == ns)
3941                         return l;
3942         return NULL;
3943 }
3944 
3945 /*
3946  * find the appropriate pidlist for our purpose (given procs vs tasks)
3947  * returns with the lock on that pidlist already held, and takes care
3948  * of the use count, or returns NULL with no locks held if we're out of
3949  * memory.
3950  */
3951 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
3952                                                 enum cgroup_filetype type)
3953 {
3954         struct cgroup_pidlist *l;
3955 
3956         lockdep_assert_held(&cgrp->pidlist_mutex);
3957 
3958         l = cgroup_pidlist_find(cgrp, type);
3959         if (l)
3960                 return l;
3961 
3962         /* entry not found; create a new one */
3963         l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3964         if (!l)
3965                 return l;
3966 
3967         INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3968         l->key.type = type;
3969         /* don't need task_nsproxy() if we're looking at ourself */
3970         l->key.ns = get_pid_ns(task_active_pid_ns(current));
3971         l->owner = cgrp;
3972         list_add(&l->links, &cgrp->pidlists);
3973         return l;
3974 }
3975 
3976 /*
3977  * Load a cgroup's pidarray with either procs' tgids or tasks' pids
3978  */
3979 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3980                               struct cgroup_pidlist **lp)
3981 {
3982         pid_t *array;
3983         int length;
3984         int pid, n = 0; /* used for populating the array */
3985         struct css_task_iter it;
3986         struct task_struct *tsk;
3987         struct cgroup_pidlist *l;
3988 
3989         lockdep_assert_held(&cgrp->pidlist_mutex);
3990 
3991         /*
3992          * If cgroup gets more users after we read count, we won't have
3993          * enough space - tough.  This race is indistinguishable to the
3994          * caller from the case that the additional cgroup users didn't
3995          * show up until sometime later on.
3996          */
3997         length = cgroup_task_count(cgrp);
3998         array = pidlist_allocate(length);
3999         if (!array)
4000                 return -ENOMEM;
4001         /* now, populate the array */
4002         css_task_iter_start(&cgrp->self, &it);
4003         while ((tsk = css_task_iter_next(&it))) {
4004                 if (unlikely(n == length))
4005                         break;
4006                 /* get tgid or pid for procs or tasks file respectively */
4007                 if (type == CGROUP_FILE_PROCS)
4008                         pid = task_tgid_vnr(tsk);
4009                 else
4010                         pid = task_pid_vnr(tsk);
4011                 if (pid > 0) /* make sure to only use valid results */
4012                         array[n++] = pid;
4013         }
4014         css_task_iter_end(&it);
4015         length = n;
4016         /* now sort & (if procs) strip out duplicates */
4017         if (cgroup_on_dfl(cgrp))
4018                 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4019         else
4020                 sort(array, length, sizeof(pid_t), cmppid, NULL);
4021         if (type == CGROUP_FILE_PROCS)
4022                 length = pidlist_uniq(array, length);
4023 
4024         l = cgroup_pidlist_find_create(cgrp, type);
4025         if (!l) {
4026                 pidlist_free(array);
4027                 return -ENOMEM;
4028         }
4029 
4030         /* store array, freeing old if necessary */
4031         pidlist_free(l->list);
4032         l->list = array;
4033         l->length = length;
4034         *lp = l;
4035         return 0;
4036 }
4037 
4038 /**
4039  * cgroupstats_build - build and fill cgroupstats
4040  * @stats: cgroupstats to fill information into
4041  * @dentry: A dentry entry belonging to the cgroup for which stats have
4042  * been requested.
4043  *
4044  * Build and fill cgroupstats so that taskstats can export it to user
4045  * space.
4046  */
4047 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4048 {
4049         struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4050         struct cgroup *cgrp;
4051         struct css_task_iter it;
4052         struct task_struct *tsk;
4053 
4054         /* it should be kernfs_node belonging to cgroupfs and is a directory */
4055         if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4056             kernfs_type(kn) != KERNFS_DIR)
4057                 return -EINVAL;
4058 
4059         mutex_lock(&cgroup_mutex);
4060 
4061         /*
4062          * We aren't being called from kernfs and there's no guarantee on
4063          * @kn->priv's validity.  For this and css_tryget_online_from_dir(),
4064          * @kn->priv is RCU safe.  Let's do the RCU dancing.
4065          */
4066         rcu_read_lock();
4067         cgrp = rcu_dereference(kn->priv);
4068         if (!cgrp || cgroup_is_dead(cgrp)) {
4069                 rcu_read_unlock();
4070                 mutex_unlock(&cgroup_mutex);
4071                 return -ENOENT;
4072         }
4073         rcu_read_unlock();
4074 
4075         css_task_iter_start(&cgrp->self, &it);
4076         while ((tsk = css_task_iter_next(&it))) {
4077                 switch (tsk->state) {
4078                 case TASK_RUNNING:
4079                         stats->nr_running++;
4080                         break;
4081                 case TASK_INTERRUPTIBLE:
4082                         stats->nr_sleeping++;
4083                         break;
4084                 case TASK_UNINTERRUPTIBLE:
4085                         stats->nr_uninterruptible++;
4086                         break;
4087                 case TASK_STOPPED:
4088                         stats->nr_stopped++;
4089                         break;
4090                 default:
4091                         if (delayacct_is_task_waiting_on_io(tsk))
4092                                 stats->nr_io_wait++;
4093                         break;
4094                 }
4095         }
4096         css_task_iter_end(&it);
4097 
4098         mutex_unlock(&cgroup_mutex);
4099         return 0;
4100 }
4101 
4102 
4103 /*
4104  * seq_file methods for the tasks/procs files. The seq_file position is the
4105  * next pid to display; the seq_file iterator is a pointer to the pid
4106  * in the cgroup->l->list array.
4107  */
4108 
4109 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4110 {
4111         /*
4112          * Initially we receive a position value that corresponds to
4113          * one more than the last pid shown (or 0 on the first call or
4114          * after a seek to the start). Use a binary-search to find the
4115          * next pid to display, if any
4116          */
4117         struct kernfs_open_file *of = s->private;
4118         struct cgroup *cgrp = seq_css(s)->cgroup;
4119         struct cgroup_pidlist *l;
4120         enum cgroup_filetype type = seq_cft(s)->private;
4121         int index = 0, pid = *pos;
4122         int *iter, ret;
4123 
4124         mutex_lock(&cgrp->pidlist_mutex);
4125 
4126         /*
4127          * !NULL @of->priv indicates that this isn't the first start()
4128          * after open.  If the matching pidlist is around, we can use that.
4129          * Look for it.  Note that @of->priv can't be used directly.  It
4130          * could already have been destroyed.
4131          */
4132         if (of->priv)
4133                 of->priv = cgroup_pidlist_find(cgrp, type);
4134 
4135         /*
4136          * Either this is the first start() after open or the matching
4137          * pidlist has been destroyed inbetween.  Create a new one.
4138          */
4139         if (!of->priv) {
4140                 ret = pidlist_array_load(cgrp, type,
4141                                          (struct cgroup_pidlist **)&of->priv);
4142                 if (ret)
4143                         return ERR_PTR(ret);
4144         }
4145         l = of->priv;
4146 
4147         if (pid) {
4148                 int end = l->length;
4149 
4150                 while (index < end) {
4151                         int mid = (index + end) / 2;
4152                         if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4153                                 index = mid;
4154                                 break;
4155                         } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4156                                 index = mid + 1;
4157                         else
4158                                 end = mid;
4159                 }
4160         }
4161         /* If we're off the end of the array, we're done */
4162         if (index >= l->length)
4163                 return NULL;
4164         /* Update the abstract position to be the actual pid that we found */
4165         iter = l->list + index;
4166         *pos = cgroup_pid_fry(cgrp, *iter);
4167         return iter;
4168 }
4169 
4170 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4171 {
4172         struct kernfs_open_file *of = s->private;
4173         struct cgroup_pidlist *l = of->priv;
4174 
4175         if (l)
4176                 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4177                                  CGROUP_PIDLIST_DESTROY_DELAY);
4178         mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4179 }
4180 
4181 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4182 {
4183         struct kernfs_open_file *of = s->private;
4184         struct cgroup_pidlist *l = of->priv;
4185         pid_t *p = v;
4186         pid_t *end = l->list + l->length;
4187         /*
4188          * Advance to the next pid in the array. If this goes off the
4189          * end, we're done
4190          */
4191         p++;
4192         if (p >= end) {
4193                 return NULL;
4194         } else {
4195                 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4196                 return p;
4197         }
4198 }
4199 
4200 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4201 {
4202         return seq_printf(s, "%d\n", *(int *)v);
4203 }
4204 
4205 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4206                                          struct cftype *cft)
4207 {
4208         return notify_on_release(css->cgroup);
4209 }
4210 
4211 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4212                                           struct cftype *cft, u64 val)
4213 {
4214         if (val)
4215                 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4216         else
4217                 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4218         return 0;
4219 }
4220 
4221 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4222                                       struct cftype *cft)
4223 {
4224         return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4225 }
4226 
4227 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4228                                        struct cftype *cft, u64 val)
4229 {
4230         if (val)
4231                 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4232         else
4233                 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4234         return 0;
4235 }
4236 
4237 /* cgroup core interface files for the default hierarchy */
4238 static struct cftype cgroup_dfl_base_files[] = {
4239         {
4240                 .name = "cgroup.procs",
4241                 .seq_start = cgroup_pidlist_start,
4242                 .seq_next = cgroup_pidlist_next,
4243                 .seq_stop = cgroup_pidlist_stop,
4244                 .seq_show = cgroup_pidlist_show,
4245                 .private = CGROUP_FILE_PROCS,
4246                 .write = cgroup_procs_write,
4247                 .mode = S_IRUGO | S_IWUSR,
4248         },
4249         {
4250                 .name = "cgroup.controllers",
4251                 .flags = CFTYPE_ONLY_ON_ROOT,
4252                 .seq_show = cgroup_root_controllers_show,
4253         },
4254         {
4255                 .name = "cgroup.controllers",
4256                 .flags = CFTYPE_NOT_ON_ROOT,
4257                 .seq_show = cgroup_controllers_show,
4258         },
4259         {
4260                 .name = "cgroup.subtree_control",
4261                 .seq_show = cgroup_subtree_control_show,
4262                 .write = cgroup_subtree_control_write,
4263         },
4264         {
4265                 .name = "cgroup.populated",
4266                 .flags = CFTYPE_NOT_ON_ROOT,
4267                 .seq_show = cgroup_populated_show,
4268         },
4269         { }     /* terminate */
4270 };
4271 
4272 /* cgroup core interface files for the legacy hierarchies */
4273 static struct cftype cgroup_legacy_base_files[] = {
4274         {
4275                 .name = "cgroup.procs",
4276                 .seq_start = cgroup_pidlist_start,
4277                 .seq_next = cgroup_pidlist_next,
4278                 .seq_stop = cgroup_pidlist_stop,
4279                 .seq_show = cgroup_pidlist_show,
4280                 .private = CGROUP_FILE_PROCS,
4281                 .write = cgroup_procs_write,
4282                 .mode = S_IRUGO | S_IWUSR,
4283         },
4284         {
4285                 .name = "cgroup.clone_children",
4286                 .read_u64 = cgroup_clone_children_read,
4287                 .write_u64 = cgroup_clone_children_write,
4288         },
4289         {
4290                 .name = "cgroup.sane_behavior",
4291                 .flags = CFTYPE_ONLY_ON_ROOT,
4292                 .seq_show = cgroup_sane_behavior_show,
4293         },
4294         {
4295                 .name = "tasks",
4296                 .seq_start = cgroup_pidlist_start,
4297                 .seq_next = cgroup_pidlist_next,
4298                 .seq_stop = cgroup_pidlist_stop,
4299                 .seq_show = cgroup_pidlist_show,
4300                 .private = CGROUP_FILE_TASKS,
4301                 .write = cgroup_tasks_write,
4302                 .mode = S_IRUGO | S_IWUSR,
4303         },
4304         {
4305                 .name = "notify_on_release",
4306                 .read_u64 = cgroup_read_notify_on_release,
4307                 .write_u64 = cgroup_write_notify_on_release,
4308         },
4309         {
4310                 .name = "release_agent",
4311                 .flags = CFTYPE_ONLY_ON_ROOT,
4312                 .seq_show = cgroup_release_agent_show,
4313                 .write = cgroup_release_agent_write,
4314                 .max_write_len = PATH_MAX - 1,
4315         },
4316         { }     /* terminate */
4317 };
4318 
4319 /**
4320  * cgroup_populate_dir - create subsys files in a cgroup directory
4321  * @cgrp: target cgroup
4322  * @subsys_mask: mask of the subsystem ids whose files should be added
4323  *
4324  * On failure, no file is added.
4325  */
4326 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned int subsys_mask)
4327 {
4328         struct cgroup_subsys *ss;
4329         int i, ret = 0;
4330 
4331         /* process cftsets of each subsystem */
4332         for_each_subsys(ss, i) {
4333                 struct cftype *cfts;
4334 
4335                 if (!(subsys_mask & (1 << i)))
4336                         continue;
4337 
4338                 list_for_each_entry(cfts, &ss->cfts, node) {
4339                         ret = cgroup_addrm_files(cgrp, cfts, true);
4340                         if (ret < 0)
4341                                 goto err;
4342                 }
4343         }
4344         return 0;
4345 err:
4346         cgroup_clear_dir(cgrp, subsys_mask);
4347         return ret;
4348 }
4349 
4350 /*
4351  * css destruction is four-stage process.
4352  *
4353  * 1. Destruction starts.  Killing of the percpu_ref is initiated.
4354  *    Implemented in kill_css().
4355  *
4356  * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4357  *    and thus css_tryget_online() is guaranteed to fail, the css can be
4358  *    offlined by invoking offline_css().  After offlining, the base ref is
4359  *    put.  Implemented in css_killed_work_fn().
4360  *
4361  * 3. When the percpu_ref reaches zero, the only possible remaining
4362  *    accessors are inside RCU read sections.  css_release() schedules the
4363  *    RCU callback.
4364  *
4365  * 4. After the grace period, the css can be freed.  Implemented in
4366  *    css_free_work_fn().
4367  *
4368  * It is actually hairier because both step 2 and 4 require process context
4369  * and thus involve punting to css->destroy_work adding two additional
4370  * steps to the already complex sequence.
4371  */
4372 static void css_free_work_fn(struct work_struct *work)
4373 {
4374         struct cgroup_subsys_state *css =
4375                 container_of(work, struct cgroup_subsys_state, destroy_work);
4376         struct cgroup_subsys *ss = css->ss;
4377         struct cgroup *cgrp = css->cgroup;
4378 
4379         percpu_ref_exit(&css->refcnt);
4380 
4381         if (ss) {
4382                 /* css free path */
4383                 int id = css->id;
4384 
4385                 if (css->parent)
4386                         css_put(css->parent);
4387 
4388                 ss->css_free(css);
4389                 cgroup_idr_remove(&ss->css_idr, id);
4390                 cgroup_put(cgrp);
4391         } else {
4392                 /* cgroup free path */
4393                 atomic_dec(&cgrp->root->nr_cgrps);
4394                 cgroup_pidlist_destroy_all(cgrp);
4395                 cancel_work_sync(&cgrp->release_agent_work);
4396 
4397                 if (cgroup_parent(cgrp)) {
4398                         /*
4399                          * We get a ref to the parent, and put the ref when
4400                          * this cgroup is being freed, so it's guaranteed
4401                          * that the parent won't be destroyed before its
4402                          * children.
4403                          */
4404                         cgroup_put(cgroup_parent(cgrp));
4405                         kernfs_put(cgrp->kn);
4406                         kfree(cgrp);
4407                 } else {
4408                         /*
4409                          * This is root cgroup's refcnt reaching zero,
4410                          * which indicates that the root should be
4411                          * released.
4412                          */
4413                         cgroup_destroy_root(cgrp->root);
4414                 }
4415         }
4416 }
4417 
4418 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4419 {
4420         struct cgroup_subsys_state *css =
4421                 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4422 
4423         INIT_WORK(&css->destroy_work, css_free_work_fn);
4424         queue_work(cgroup_destroy_wq, &css->destroy_work);
4425 }
4426 
4427 static void css_release_work_fn(struct work_struct *work)
4428 {
4429         struct cgroup_subsys_state *css =
4430                 container_of(work, struct cgroup_subsys_state, destroy_work);
4431         struct cgroup_subsys *ss = css->ss;
4432         struct cgroup *cgrp = css->cgroup;
4433 
4434         mutex_lock(&cgroup_mutex);
4435 
4436         css->flags |= CSS_RELEASED;
4437         list_del_rcu(&css->sibling);
4438 
4439         if (ss) {
4440                 /* css release path */
4441                 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4442                 if (ss->css_released)
4443                         ss->css_released(css);
4444         } else {
4445                 /* cgroup release path */
4446                 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4447                 cgrp->id = -1;
4448 
4449                 /*
4450                  * There are two control paths which try to determine
4451                  * cgroup from dentry without going through kernfs -
4452                  * cgroupstats_build() and css_tryget_online_from_dir().
4453                  * Those are supported by RCU protecting clearing of
4454                  * cgrp->kn->priv backpointer.
4455                  */
4456                 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4457         }
4458 
4459         mutex_unlock(&cgroup_mutex);
4460 
4461         call_rcu(&css->rcu_head, css_free_rcu_fn);
4462 }
4463 
4464 static void css_release(struct percpu_ref *ref)
4465 {
4466         struct cgroup_subsys_state *css =
4467                 container_of(ref, struct cgroup_subsys_state, refcnt);
4468 
4469         INIT_WORK(&css->destroy_work, css_release_work_fn);
4470         queue_work(cgroup_destroy_wq, &css->destroy_work);
4471 }
4472 
4473 static void init_and_link_css(struct cgroup_subsys_state *css,
4474                               struct cgroup_subsys *ss, struct cgroup *cgrp)
4475 {
4476         lockdep_assert_held(&cgroup_mutex);
4477 
4478         cgroup_get(cgrp);
4479 
4480         memset(css, 0, sizeof(*css));
4481         css->cgroup = cgrp;
4482         css->ss = ss;
4483         INIT_LIST_HEAD(&css->sibling);
4484         INIT_LIST_HEAD(&css->children);
4485         css->serial_nr = css_serial_nr_next++;
4486 
4487         if (cgroup_parent(cgrp)) {
4488                 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4489                 css_get(css->parent);
4490         }
4491 
4492         BUG_ON(cgroup_css(cgrp, ss));
4493 }
4494 
4495 /* invoke ->css_online() on a new CSS and mark it online if successful */
4496 static int online_css(struct cgroup_subsys_state *css)
4497 {
4498         struct cgroup_subsys *ss = css->ss;
4499         int ret = 0;
4500 
4501         lockdep_assert_held(&cgroup_mutex);
4502 
4503         if (ss->css_online)
4504                 ret = ss->css_online(css);
4505         if (!ret) {
4506                 css->flags |= CSS_ONLINE;
4507                 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4508         }
4509         return ret;
4510 }
4511 
4512 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4513 static void offline_css(struct cgroup_subsys_state *css)
4514 {
4515         struct cgroup_subsys *ss = css->ss;
4516 
4517         lockdep_assert_held(&cgroup_mutex);
4518 
4519         if (!(css->flags & CSS_ONLINE))
4520                 return;
4521 
4522         if (ss->css_offline)
4523                 ss->css_offline(css);
4524 
4525         css->flags &= ~CSS_ONLINE;
4526         RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4527 
4528         wake_up_all(&css->cgroup->offline_waitq);
4529 }
4530 
4531 /**
4532  * create_css - create a cgroup_subsys_state
4533  * @cgrp: the cgroup new css will be associated with
4534  * @ss: the subsys of new css
4535  * @visible: whether to create control knobs for the new css or not
4536  *
4537  * Create a new css associated with @cgrp - @ss pair.  On success, the new
4538  * css is online and installed in @cgrp with all interface files created if
4539  * @visible.  Returns 0 on success, -errno on failure.
4540  */
4541 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4542                       bool visible)
4543 {
4544         struct cgroup *parent = cgroup_parent(cgrp);
4545         struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4546         struct cgroup_subsys_state *css;
4547         int err;
4548 
4549         lockdep_assert_held(&cgroup_mutex);
4550 
4551         css = ss->css_alloc(parent_css);
4552         if (IS_ERR(css))
4553                 return PTR_ERR(css);
4554 
4555         init_and_link_css(css, ss, cgrp);
4556 
4557         err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4558         if (err)
4559                 goto err_free_css;
4560 
4561         err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4562         if (err < 0)
4563                 goto err_free_percpu_ref;
4564         css->id = err;
4565 
4566         if (visible) {
4567                 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4568                 if (err)
4569                         goto err_free_id;
4570         }
4571 
4572         /* @css is ready to be brought online now, make it visible */
4573         list_add_tail_rcu(&css->sibling, &parent_css->children);
4574         cgroup_idr_replace(&ss->css_idr, css, css->id);
4575 
4576         err = online_css(css);
4577         if (err)
4578                 goto err_list_del;
4579 
4580         if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4581             cgroup_parent(parent)) {
4582                 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4583                         current->comm, current->pid, ss->name);
4584                 if (!strcmp(ss->name, "memory"))
4585                         pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4586                 ss->warned_broken_hierarchy = true;
4587         }
4588 
4589         return 0;
4590 
4591 err_list_del:
4592         list_del_rcu(&css->sibling);
4593         cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4594 err_free_id:
4595         cgroup_idr_remove(&ss->css_idr, css->id);
4596 err_free_percpu_ref:
4597         percpu_ref_exit(&css->refcnt);
4598 err_free_css:
4599         call_rcu(&css->rcu_head, css_free_rcu_fn);
4600         return err;
4601 }
4602 
4603 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4604                         umode_t mode)
4605 {
4606         struct cgroup *parent, *cgrp;
4607         struct cgroup_root *root;
4608         struct cgroup_subsys *ss;
4609         struct kernfs_node *kn;
4610         struct cftype *base_files;
4611         int ssid, ret;
4612 
4613         /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4614          */
4615         if (strchr(name, '\n'))
4616                 return -EINVAL;
4617 
4618         parent = cgroup_kn_lock_live(parent_kn);
4619         if (!parent)
4620                 return -ENODEV;
4621         root = parent->root;
4622 
4623         /* allocate the cgroup and its ID, 0 is reserved for the root */
4624         cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4625         if (!cgrp) {
4626                 ret = -ENOMEM;
4627                 goto out_unlock;
4628         }
4629 
4630         ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4631         if (ret)
4632                 goto out_free_cgrp;
4633 
4634         /*
4635          * Temporarily set the pointer to NULL, so idr_find() won't return
4636          * a half-baked cgroup.
4637          */
4638         cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4639         if (cgrp->id < 0) {
4640                 ret = -ENOMEM;
4641                 goto out_cancel_ref;
4642         }
4643 
4644         init_cgroup_housekeeping(cgrp);
4645 
4646         cgrp->self.parent = &parent->self;
4647         cgrp->root = root;
4648 
4649         if (notify_on_release(parent))
4650                 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4651 
4652         if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4653                 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4654 
4655         /* create the directory */
4656         kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4657         if (IS_ERR(kn)) {
4658                 ret = PTR_ERR(kn);
4659                 goto out_free_id;
4660         }
4661         cgrp->kn = kn;
4662 
4663         /*
4664          * This extra ref will be put in cgroup_free_fn() and guarantees
4665          * that @cgrp->kn is always accessible.
4666          */
4667         kernfs_get(kn);
4668 
4669         cgrp->self.serial_nr = css_serial_nr_next++;
4670 
4671         /* allocation complete, commit to creation */
4672         list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4673         atomic_inc(&root->nr_cgrps);
4674         cgroup_get(parent);
4675 
4676         /*
4677          * @cgrp is now fully operational.  If something fails after this
4678          * point, it'll be released via the normal destruction path.
4679          */
4680         cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4681 
4682         ret = cgroup_kn_set_ugid(kn);
4683         if (ret)
4684                 goto out_destroy;
4685 
4686         if (cgroup_on_dfl(cgrp))
4687                 base_files = cgroup_dfl_base_files;
4688         else
4689                 base_files = cgroup_legacy_base_files;
4690 
4691         ret = cgroup_addrm_files(cgrp, base_files, true);
4692         if (ret)
4693                 goto out_destroy;
4694 
4695         /* let's create and online css's */
4696         for_each_subsys(ss, ssid) {
4697                 if (parent->child_subsys_mask & (1 << ssid)) {
4698                         ret = create_css(cgrp, ss,
4699                                          parent->subtree_control & (1 << ssid));
4700                         if (ret)
4701                                 goto out_destroy;
4702                 }
4703         }
4704 
4705         /*
4706          * On the default hierarchy, a child doesn't automatically inherit
4707          * subtree_control from the parent.  Each is configured manually.
4708          */
4709         if (!cgroup_on_dfl(cgrp)) {
4710                 cgrp->subtree_control = parent->subtree_control;
4711                 cgroup_refresh_child_subsys_mask(cgrp);
4712         }
4713 
4714         kernfs_activate(kn);
4715 
4716         ret = 0;
4717         goto out_unlock;
4718 
4719 out_free_id:
4720         cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4721 out_cancel_ref:
4722         percpu_ref_exit(&cgrp->self.refcnt);
4723 out_free_cgrp:
4724         kfree(cgrp);
4725 out_unlock:
4726         cgroup_kn_unlock(parent_kn);
4727         return ret;
4728 
4729 out_destroy:
4730         cgroup_destroy_locked(cgrp);
4731         goto out_unlock;
4732 }
4733 
4734 /*
4735  * This is called when the refcnt of a css is confirmed to be killed.
4736  * css_tryget_online() is now guaranteed to fail.  Tell the subsystem to
4737  * initate destruction and put the css ref from kill_css().
4738  */
4739 static void css_killed_work_fn(struct work_struct *work)
4740 {
4741         struct cgroup_subsys_state *css =
4742                 container_of(work, struct cgroup_subsys_state, destroy_work);
4743 
4744         mutex_lock(&cgroup_mutex);
4745         offline_css(css);
4746         mutex_unlock(&cgroup_mutex);
4747 
4748         css_put(css);
4749 }
4750 
4751 /* css kill confirmation processing requires process context, bounce */
4752 static void css_killed_ref_fn(struct percpu_ref *ref)
4753 {
4754         struct cgroup_subsys_state *css =
4755                 container_of(ref, struct cgroup_subsys_state, refcnt);
4756 
4757         INIT_WORK(&css->destroy_work, css_killed_work_fn);
4758         queue_work(cgroup_destroy_wq, &css->destroy_work);
4759 }
4760 
4761 /**
4762  * kill_css - destroy a css
4763  * @css: css to destroy
4764  *
4765  * This function initiates destruction of @css by removing cgroup interface
4766  * files and putting its base reference.  ->css_offline() will be invoked
4767  * asynchronously once css_tryget_online() is guaranteed to fail and when
4768  * the reference count reaches zero, @css will be released.
4769  */
4770 static void kill_css(struct cgroup_subsys_state *css)
4771 {
4772         lockdep_assert_held(&cgroup_mutex);
4773 
4774         /*
4775          * This must happen before css is disassociated with its cgroup.
4776          * See seq_css() for details.
4777          */
4778         cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4779 
4780         /*
4781          * Killing would put the base ref, but we need to keep it alive
4782          * until after ->css_offline().
4783          */
4784         css_get(css);
4785 
4786         /*
4787          * cgroup core guarantees that, by the time ->css_offline() is
4788          * invoked, no new css reference will be given out via
4789          * css_tryget_online().  We can't simply call percpu_ref_kill() and
4790          * proceed to offlining css's because percpu_ref_kill() doesn't
4791          * guarantee that the ref is seen as killed on all CPUs on return.
4792          *
4793          * Use percpu_ref_kill_and_confirm() to get notifications as each
4794          * css is confirmed to be seen as killed on all CPUs.
4795          */
4796         percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4797 }
4798 
4799 /**
4800  * cgroup_destroy_locked - the first stage of cgroup destruction
4801  * @cgrp: cgroup to be destroyed
4802  *
4803  * css's make use of percpu refcnts whose killing latency shouldn't be
4804  * exposed to userland and are RCU protected.  Also, cgroup core needs to
4805  * guarantee that css_tryget_online() won't succeed by the time
4806  * ->css_offline() is invoked.  To satisfy all the requirements,
4807  * destruction is implemented in the following two steps.
4808  *
4809  * s1. Verify @cgrp can be destroyed and mark it dying.  Remove all
4810  *     userland visible parts and start killing the percpu refcnts of
4811  *     css's.  Set up so that the next stage will be kicked off once all
4812  *     the percpu refcnts are confirmed to be killed.
4813  *
4814  * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4815  *     rest of destruction.  Once all cgroup references are gone, the
4816  *     cgroup is RCU-freed.
4817  *
4818  * This function implements s1.  After this step, @cgrp is gone as far as
4819  * the userland is concerned and a new cgroup with the same name may be
4820  * created.  As cgroup doesn't care about the names internally, this
4821  * doesn't cause any problem.
4822  */
4823 static int cgroup_destroy_locked(struct cgroup *cgrp)
4824         __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4825 {
4826         struct cgroup_subsys_state *css;
4827         bool empty;
4828         int ssid;
4829 
4830         lockdep_assert_held(&cgroup_mutex);
4831 
4832         /*
4833          * css_set_rwsem synchronizes access to ->cset_links and prevents
4834          * @cgrp from being removed while put_css_set() is in progress.
4835          */
4836         down_read(&css_set_rwsem);
4837         empty = list_empty(&cgrp->cset_links);
4838         up_read(&css_set_rwsem);
4839         if (!empty)
4840                 return -EBUSY;
4841 
4842         /*
4843          * Make sure there's no live children.  We can't test emptiness of
4844          * ->self.children as dead children linger on it while being
4845          * drained; otherwise, "rmdir parent/child parent" may fail.
4846          */
4847         if (css_has_online_children(&cgrp->self))
4848                 return -EBUSY;
4849 
4850         /*
4851          * Mark @cgrp dead.  This prevents further task migration and child
4852          * creation by disabling cgroup_lock_live_group().
4853          */
4854         cgrp->self.flags &= ~CSS_ONLINE;
4855 
4856         /* initiate massacre of all css's */
4857         for_each_css(css, ssid, cgrp)
4858                 kill_css(css);
4859 
4860         /*
4861          * Remove @cgrp directory along with the base files.  @cgrp has an
4862          * extra ref on its kn.
4863          */
4864         kernfs_remove(cgrp->kn);
4865 
4866         check_for_release(cgroup_parent(cgrp));
4867 
4868         /* put the base reference */
4869         percpu_ref_kill(&cgrp->self.refcnt);
4870 
4871         return 0;
4872 };
4873 
4874 static int cgroup_rmdir(struct kernfs_node *kn)
4875 {
4876         struct cgroup *cgrp;
4877         int ret = 0;
4878 
4879         cgrp = cgroup_kn_lock_live(kn);
4880         if (!cgrp)
4881                 return 0;
4882 
4883         ret = cgroup_destroy_locked(cgrp);
4884 
4885         cgroup_kn_unlock(kn);
4886         return ret;
4887 }
4888 
4889 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4890         .remount_fs             = cgroup_remount,
4891         .show_options           = cgroup_show_options,
4892         .mkdir                  = cgroup_mkdir,
4893         .rmdir                  = cgroup_rmdir,
4894         .rename                 = cgroup_rename,
4895 };
4896 
4897 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4898 {
4899         struct cgroup_subsys_state *css;
4900 
4901         printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4902 
4903         mutex_lock(&cgroup_mutex);
4904 
4905         idr_init(&ss->css_idr);
4906         INIT_LIST_HEAD(&ss->cfts);
4907 
4908         /* Create the root cgroup state for this subsystem */
4909         ss->root = &cgrp_dfl_root;
4910         css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4911         /* We don't handle early failures gracefully */
4912         BUG_ON(IS_ERR(css));
4913         init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4914 
4915         /*
4916          * Root csses are never destroyed and we can't initialize
4917          * percpu_ref during early init.  Disable refcnting.
4918          */
4919         css->flags |= CSS_NO_REF;
4920 
4921         if (early) {
4922                 /* allocation can't be done safely during early init */
4923                 css->id = 1;
4924         } else {
4925                 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4926                 BUG_ON(css->id < 0);
4927         }
4928 
4929         /* Update the init_css_set to contain a subsys
4930          * pointer to this state - since the subsystem is
4931          * newly registered, all tasks and hence the
4932          * init_css_set is in the subsystem's root cgroup. */
4933         init_css_set.subsys[ss->id] = css;
4934 
4935         need_forkexit_callback |= ss->fork || ss->exit;
4936 
4937         /* At system boot, before all subsystems have been
4938          * registered, no tasks have been forked, so we don't
4939          * need to invoke fork callbacks here. */
4940         BUG_ON(!list_empty(&init_task.tasks));
4941 
4942         BUG_ON(online_css(css));
4943 
4944         mutex_unlock(&cgroup_mutex);
4945 }
4946 
4947 /**
4948  * cgroup_init_early - cgroup initialization at system boot
4949  *
4950  * Initialize cgroups at system boot, and initialize any
4951  * subsystems that request early init.
4952  */
4953 int __init cgroup_init_early(void)
4954 {
4955         static struct cgroup_sb_opts __initdata opts;
4956         struct cgroup_subsys *ss;
4957         int i;
4958 
4959         init_cgroup_root(&cgrp_dfl_root, &opts);
4960         cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
4961 
4962         RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4963 
4964         for_each_subsys(ss, i) {
4965                 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
4966                      "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4967                      i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4968                      ss->id, ss->name);
4969                 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4970                      "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4971 
4972                 ss->id = i;
4973                 ss->name = cgroup_subsys_name[i];
4974 
4975                 if (ss->early_init)
4976                         cgroup_init_subsys(ss, true);
4977         }
4978         return 0;
4979 }
4980 
4981 /**
4982  * cgroup_init - cgroup initialization
4983  *
4984  * Register cgroup filesystem and /proc file, and initialize
4985  * any subsystems that didn't request early init.
4986  */
4987 int __init cgroup_init(void)
4988 {
4989         struct cgroup_subsys *ss;
4990         unsigned long key;
4991         int ssid, err;
4992 
4993         BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
4994         BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
4995 
4996         mutex_lock(&cgroup_mutex);
4997 
4998         /* Add init_css_set to the hash table */
4999         key = css_set_hash(init_css_set.subsys);
5000         hash_add(css_set_table, &init_css_set.hlist, key);
5001 
5002         BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5003 
5004         mutex_unlock(&cgroup_mutex);
5005 
5006         for_each_subsys(ss, ssid) {
5007                 if (ss->early_init) {
5008                         struct cgroup_subsys_state *css =
5009                                 init_css_set.subsys[ss->id];
5010 
5011                         css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5012                                                    GFP_KERNEL);
5013                         BUG_ON(css->id < 0);
5014                 } else {
5015                         cgroup_init_subsys(ss, false);
5016                 }
5017 
5018                 list_add_tail(&init_css_set.e_cset_node[ssid],
5019                               &cgrp_dfl_root.cgrp.e_csets[ssid]);
5020 
5021                 /*
5022                  * Setting dfl_root subsys_mask needs to consider the
5023                  * disabled flag and cftype registration needs kmalloc,
5024                  * both of which aren't available during early_init.
5025                  */
5026                 if (ss->disabled)
5027                         continue;
5028 
5029                 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5030 
5031                 if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
5032                         ss->dfl_cftypes = ss->legacy_cftypes;
5033 
5034                 if (!ss->dfl_cftypes)
5035                         cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5036 
5037                 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5038                         WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5039                 } else {
5040                         WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5041                         WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5042                 }
5043         }
5044 
5045         cgroup_kobj = kobject_create_and_add("cgroup", fs_kobj);
5046         if (!cgroup_kobj)
5047                 return -ENOMEM;
5048 
5049         err = register_filesystem(&cgroup_fs_type);
5050         if (err < 0) {
5051                 kobject_put(cgroup_kobj);
5052                 return err;
5053         }
5054 
5055         proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5056         return 0;
5057 }
5058 
5059 static int __init cgroup_wq_init(void)
5060 {
5061         /*
5062          * There isn't much point in executing destruction path in
5063          * parallel.  Good chunk is serialized with cgroup_mutex anyway.
5064          * Use 1 for @max_active.
5065          *
5066          * We would prefer to do this in cgroup_init() above, but that
5067          * is called before init_workqueues(): so leave this until after.
5068          */
5069         cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5070         BUG_ON(!cgroup_destroy_wq);
5071 
5072         /*
5073          * Used to destroy pidlists and separate to serve as flush domain.
5074          * Cap @max_active to 1 too.
5075          */
5076         cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5077                                                     0, 1);
5078         BUG_ON(!cgroup_pidlist_destroy_wq);
5079 
5080         return 0;
5081 }
5082 core_initcall(cgroup_wq_init);
5083 
5084 /*
5085  * proc_cgroup_show()
5086  *  - Print task's cgroup paths into seq_file, one line for each hierarchy
5087  *  - Used for /proc/<pid>/cgroup.
5088  */
5089 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5090                      struct pid *pid, struct task_struct *tsk)
5091 {
5092         char *buf, *path;
5093         int retval;
5094         struct cgroup_root *root;
5095 
5096         retval = -ENOMEM;
5097         buf = kmalloc(PATH_MAX, GFP_KERNEL);
5098         if (!buf)
5099                 goto out;
5100 
5101         mutex_lock(&cgroup_mutex);
5102         down_read(&css_set_rwsem);
5103 
5104         for_each_root(root) {
5105                 struct cgroup_subsys *ss;
5106                 struct cgroup *cgrp;
5107                 int ssid, count = 0;
5108 
5109                 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5110                         continue;
5111 
5112                 seq_printf(m, "%d:", root->hierarchy_id);
5113                 for_each_subsys(ss, ssid)
5114                         if (root->subsys_mask & (1 << ssid))
5115                                 seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
5116                 if (strlen(root->name))
5117                         seq_printf(m, "%sname=%s", count ? "," : "",
5118                                    root->name);
5119                 seq_putc(m, ':');
5120                 cgrp = task_cgroup_from_root(tsk, root);
5121                 path = cgroup_path(cgrp, buf, PATH_MAX);
5122                 if (!path) {
5123                         retval = -ENAMETOOLONG;
5124                         goto out_unlock;
5125                 }
5126                 seq_puts(m, path);
5127                 seq_putc(m, '\n');
5128         }
5129 
5130         retval = 0;
5131 out_unlock:
5132         up_read(&css_set_rwsem);
5133         mutex_unlock(&cgroup_mutex);
5134         kfree(buf);
5135 out:
5136         return retval;
5137 }
5138 
5139 /* Display information about each subsystem and each hierarchy */
5140 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5141 {
5142         struct cgroup_subsys *ss;
5143         int i;
5144 
5145         seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5146         /*
5147          * ideally we don't want subsystems moving around while we do this.
5148          * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5149          * subsys/hierarchy state.
5150          */
5151         mutex_lock(&cgroup_mutex);
5152 
5153         for_each_subsys(ss, i)
5154                 seq_printf(m, "%s\t%d\t%d\t%d\n",
5155                            ss->name, ss->root->hierarchy_id,
5156                            atomic_read(&ss->root->nr_cgrps), !ss->disabled);
5157 
5158         mutex_unlock(&cgroup_mutex);
5159         return 0;
5160 }
5161 
5162 static int cgroupstats_open(struct inode *inode, struct file *file)
5163 {
5164         return single_open(file, proc_cgroupstats_show, NULL);
5165 }
5166 
5167 static const struct file_operations proc_cgroupstats_operations = {
5168         .open = cgroupstats_open,
5169         .read = seq_read,
5170         .llseek = seq_lseek,
5171         .release = single_release,
5172 };
5173 
5174 /**
5175  * cgroup_fork - initialize cgroup related fields during copy_process()
5176  * @child: pointer to task_struct of forking parent process.
5177  *
5178  * A task is associated with the init_css_set until cgroup_post_fork()
5179  * attaches it to the parent's css_set.  Empty cg_list indicates that
5180  * @child isn't holding reference to its css_set.
5181  */
5182 void cgroup_fork(struct task_struct *child)
5183 {
5184         RCU_INIT_POINTER(child->cgroups, &init_css_set);
5185         INIT_LIST_HEAD(&child->cg_list);
5186 }
5187 
5188 /**
5189  * cgroup_post_fork - called on a new task after adding it to the task list
5190  * @child: the task in question
5191  *
5192  * Adds the task to the list running through its css_set if necessary and
5193  * call the subsystem fork() callbacks.  Has to be after the task is
5194  * visible on the task list in case we race with the first call to
5195  * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5196  * list.
5197  */
5198 void cgroup_post_fork(struct task_struct *child)
5199 {
5200         struct cgroup_subsys *ss;
5201         int i;
5202 
5203         /*
5204          * This may race against cgroup_enable_task_cg_lists().  As that
5205          * function sets use_task_css_set_links before grabbing
5206          * tasklist_lock and we just went through tasklist_lock to add
5207          * @child, it's guaranteed that either we see the set
5208          * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5209          * @child during its iteration.
5210          *
5211          * If we won the race, @child is associated with %current's
5212          * css_set.  Grabbing css_set_rwsem guarantees both that the
5213          * association is stable, and, on completion of the parent's
5214          * migration, @child is visible in the source of migration or
5215          * already in the destination cgroup.  This guarantee is necessary
5216          * when implementing operations which need to migrate all tasks of
5217          * a cgroup to another.
5218          *
5219          * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5220          * will remain in init_css_set.  This is safe because all tasks are
5221          * in the init_css_set before cg_links is enabled and there's no
5222          * operation which transfers all tasks out of init_css_set.
5223          */
5224         if (use_task_css_set_links) {
5225                 struct css_set *cset;
5226 
5227                 down_write(&css_set_rwsem);
5228                 cset = task_css_set(current);
5229                 if (list_empty(&child->cg_list)) {
5230                         rcu_assign_pointer(child->cgroups, cset);
5231                         list_add(&child->cg_list, &cset->tasks);
5232                         get_css_set(cset);
5233                 }
5234                 up_write(&css_set_rwsem);
5235         }
5236 
5237         /*
5238          * Call ss->fork().  This must happen after @child is linked on
5239          * css_set; otherwise, @child might change state between ->fork()
5240          * and addition to css_set.
5241          */
5242         if (need_forkexit_callback) {
5243                 for_each_subsys(ss, i)
5244                         if (ss->fork)
5245                                 ss->fork(child);
5246         }
5247 }
5248 
5249 /**
5250  * cgroup_exit - detach cgroup from exiting task
5251  * @tsk: pointer to task_struct of exiting process
5252  *
5253  * Description: Detach cgroup from @tsk and release it.
5254  *
5255  * Note that cgroups marked notify_on_release force every task in
5256  * them to take the global cgroup_mutex mutex when exiting.
5257  * This could impact scaling on very large systems.  Be reluctant to
5258  * use notify_on_release cgroups where very high task exit scaling
5259  * is required on large systems.
5260  *
5261  * We set the exiting tasks cgroup to the root cgroup (top_cgroup).  We
5262  * call cgroup_exit() while the task is still competent to handle
5263  * notify_on_release(), then leave the task attached to the root cgroup in
5264  * each hierarchy for the remainder of its exit.  No need to bother with
5265  * init_css_set refcnting.  init_css_set never goes away and we can't race
5266  * with migration path - PF_EXITING is visible to migration path.
5267  */
5268 void cgroup_exit(struct task_struct *tsk)
5269 {
5270         struct cgroup_subsys *ss;
5271         struct css_set *cset;
5272         bool put_cset = false;
5273         int i;
5274 
5275         /*
5276          * Unlink from @tsk from its css_set.  As migration path can't race
5277          * with us, we can check cg_list without grabbing css_set_rwsem.
5278          */
5279         if (!list_empty(&tsk->cg_list)) {
5280                 down_write(&css_set_rwsem);
5281                 list_del_init(&tsk->cg_list);
5282                 up_write(&css_set_rwsem);
5283                 put_cset = true;
5284         }
5285 
5286         /* Reassign the task to the init_css_set. */
5287         cset = task_css_set(tsk);
5288         RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5289 
5290         if (need_forkexit_callback) {
5291                 /* see cgroup_post_fork() for details */
5292                 for_each_subsys(ss, i) {
5293                         if (ss->exit) {
5294                                 struct cgroup_subsys_state *old_css = cset->subsys[i];
5295                                 struct cgroup_subsys_state *css = task_css(tsk, i);
5296 
5297                                 ss->exit(css, old_css, tsk);
5298                         }
5299                 }
5300         }
5301 
5302         if (put_cset)
5303                 put_css_set(cset);
5304 }
5305 
5306 static void check_for_release(struct cgroup *cgrp)
5307 {
5308         if (notify_on_release(cgrp) && !cgroup_has_tasks(cgrp) &&
5309             !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5310                 schedule_work(&cgrp->release_agent_work);
5311 }
5312 
5313 /*
5314  * Notify userspace when a cgroup is released, by running the
5315  * configured release agent with the name of the cgroup (path
5316  * relative to the root of cgroup file system) as the argument.
5317  *
5318  * Most likely, this user command will try to rmdir this cgroup.
5319  *
5320  * This races with the possibility that some other task will be
5321  * attached to this cgroup before it is removed, or that some other
5322  * user task will 'mkdir' a child cgroup of this cgroup.  That's ok.
5323  * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5324  * unused, and this cgroup will be reprieved from its death sentence,
5325  * to continue to serve a useful existence.  Next time it's released,
5326  * we will get notified again, if it still has 'notify_on_release' set.
5327  *
5328  * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5329  * means only wait until the task is successfully execve()'d.  The
5330  * separate release agent task is forked by call_usermodehelper(),
5331  * then control in this thread returns here, without waiting for the
5332  * release agent task.  We don't bother to wait because the caller of
5333  * this routine has no use for the exit status of the release agent
5334  * task, so no sense holding our caller up for that.
5335  */
5336 static void cgroup_release_agent(struct work_struct *work)
5337 {
5338         struct cgroup *cgrp =
5339                 container_of(work, struct cgroup, release_agent_work);
5340         char *pathbuf = NULL, *agentbuf = NULL, *path;
5341         char *argv[3], *envp[3];
5342 
5343         mutex_lock(&cgroup_mutex);
5344 
5345         pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5346         agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5347         if (!pathbuf || !agentbuf)
5348                 goto out;
5349 
5350         path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5351         if (!path)
5352                 goto out;
5353 
5354         argv[0] = agentbuf;
5355         argv[1] = path;
5356         argv[2] = NULL;
5357 
5358         /* minimal command environment */
5359         envp[0] = "HOME=/";
5360         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5361         envp[2] = NULL;
5362 
5363         mutex_unlock(&cgroup_mutex);
5364         call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5365         goto out_free;
5366 out:
5367         mutex_unlock(&cgroup_mutex);
5368 out_free:
5369         kfree(agentbuf);
5370         kfree(pathbuf);
5371 }
5372 
5373 static int __init cgroup_disable(char *str)
5374 {
5375         struct cgroup_subsys *ss;
5376         char *token;
5377         int i;
5378 
5379         while ((token = strsep(&str, ",")) != NULL) {
5380                 if (!*token)
5381                         continue;
5382 
5383                 for_each_subsys(ss, i) {
5384                         if (!strcmp(token, ss->name)) {
5385                                 ss->disabled = 1;
5386                                 printk(KERN_INFO "Disabling %s control group"
5387                                         " subsystem\n", ss->name);
5388                                 break;
5389                         }
5390                 }
5391         }
5392         return 1;
5393 }
5394 __setup("cgroup_disable=", cgroup_disable);
5395 
5396 static int __init cgroup_set_legacy_files_on_dfl(char *str)
5397 {
5398         printk("cgroup: using legacy files on the default hierarchy\n");
5399         cgroup_legacy_files_on_dfl = true;
5400         return 0;
5401 }
5402 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5403 
5404 /**
5405  * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5406  * @dentry: directory dentry of interest
5407  * @ss: subsystem of interest
5408  *
5409  * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5410  * to get the corresponding css and return it.  If such css doesn't exist
5411  * or can't be pinned, an ERR_PTR value is returned.
5412  */
5413 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5414                                                        struct cgroup_subsys *ss)
5415 {
5416         struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5417         struct cgroup_subsys_state *css = NULL;
5418         struct cgroup *cgrp;
5419 
5420         /* is @dentry a cgroup dir? */
5421         if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5422             kernfs_type(kn) != KERNFS_DIR)
5423                 return ERR_PTR(-EBADF);
5424 
5425         rcu_read_lock();
5426 
5427         /*
5428          * This path doesn't originate from kernfs and @kn could already
5429          * have been or be removed at any point.  @kn->priv is RCU
5430          * protected for this access.  See css_release_work_fn() for details.
5431          */
5432         cgrp = rcu_dereference(kn->priv);
5433         if (cgrp)
5434                 css = cgroup_css(cgrp, ss);
5435 
5436         if (!css || !css_tryget_online(css))
5437                 css = ERR_PTR(-ENOENT);
5438 
5439         rcu_read_unlock();
5440         return css;
5441 }
5442 
5443 /**
5444  * css_from_id - lookup css by id
5445  * @id: the cgroup id
5446  * @ss: cgroup subsys to be looked into
5447  *
5448  * Returns the css if there's valid one with @id, otherwise returns NULL.
5449  * Should be called under rcu_read_lock().
5450  */
5451 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5452 {
5453         WARN_ON_ONCE(!rcu_read_lock_held());
5454         return idr_find(&ss->css_idr, id);
5455 }
5456 
5457 #ifdef CONFIG_CGROUP_DEBUG
5458 static struct cgroup_subsys_state *
5459 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5460 {
5461         struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5462 
5463         if (!css)
5464                 return ERR_PTR(-ENOMEM);
5465 
5466         return css;
5467 }
5468 
5469 static void debug_css_free(struct cgroup_subsys_state *css)
5470 {
5471         kfree(css);
5472 }
5473 
5474 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5475                                 struct cftype *cft)
5476 {
5477         return cgroup_task_count(css->cgroup);
5478 }
5479 
5480 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5481                                 struct cftype *cft)
5482 {
5483         return (u64)(unsigned long)current->cgroups;
5484 }
5485 
5486 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5487                                          struct cftype *cft)
5488 {
5489         u64 count;
5490 
5491         rcu_read_lock();
5492         count = atomic_read(&task_css_set(current)->refcount);
5493         rcu_read_unlock();
5494         return count;
5495 }
5496 
5497 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5498 {
5499         struct cgrp_cset_link *link;
5500         struct css_set *cset;
5501         char *name_buf;
5502 
5503         name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5504         if (!name_buf)
5505                 return -ENOMEM;
5506 
5507         down_read(&css_set_rwsem);
5508         rcu_read_lock();
5509         cset = rcu_dereference(current->cgroups);
5510         list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5511                 struct cgroup *c = link->cgrp;
5512 
5513                 cgroup_name(c, name_buf, NAME_MAX + 1);
5514                 seq_printf(seq, "Root %d group %s\n",
5515                            c->root->hierarchy_id, name_buf);
5516         }
5517         rcu_read_unlock();
5518         up_read(&css_set_rwsem);
5519         kfree(name_buf);
5520         return 0;
5521 }
5522 
5523 #define MAX_TASKS_SHOWN_PER_CSS 25
5524 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5525 {
5526         struct cgroup_subsys_state *css = seq_css(seq);
5527         struct cgrp_cset_link *link;
5528 
5529         down_read(&css_set_rwsem);
5530         list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5531                 struct css_set *cset = link->cset;
5532                 struct task_struct *task;
5533                 int count = 0;
5534 
5535                 seq_printf(seq, "css_set %p\n", cset);
5536 
5537                 list_for_each_entry(task, &cset->tasks, cg_list) {
5538                         if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5539                                 goto overflow;
5540                         seq_printf(seq, "  task %d\n", task_pid_vnr(task));
5541                 }
5542 
5543                 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5544                         if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5545                                 goto overflow;
5546                         seq_printf(seq, "  task %d\n", task_pid_vnr(task));
5547                 }
5548                 continue;
5549         overflow:
5550                 seq_puts(seq, "  ...\n");
5551         }
5552         up_read(&css_set_rwsem);
5553         return 0;
5554 }
5555 
5556 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5557 {
5558         return (!cgroup_has_tasks(css->cgroup) &&
5559                 !css_has_online_children(&css->cgroup->self));
5560 }
5561 
5562 static struct cftype debug_files[] =  {
5563         {
5564                 .name = "taskcount",
5565                 .read_u64 = debug_taskcount_read,
5566         },
5567 
5568         {
5569                 .name = "current_css_set",
5570                 .read_u64 = current_css_set_read,
5571         },
5572 
5573         {
5574                 .name = "current_css_set_refcount",
5575                 .read_u64 = current_css_set_refcount_read,
5576         },
5577 
5578         {
5579                 .name = "current_css_set_cg_links",
5580                 .seq_show = current_css_set_cg_links_read,
5581         },
5582 
5583         {
5584                 .name = "cgroup_css_links",
5585                 .seq_show = cgroup_css_links_read,
5586         },
5587 
5588         {
5589                 .name = "releasable",
5590                 .read_u64 = releasable_read,
5591         },
5592 
5593         { }     /* terminate */
5594 };
5595 
5596 struct cgroup_subsys debug_cgrp_subsys = {
5597         .css_alloc = debug_css_alloc,
5598         .css_free = debug_css_free,
5599         .legacy_cftypes = debug_files,
5600 };
5601 #endif /* CONFIG_CGROUP_DEBUG */
5602 

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