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

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