~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/memcontrol.c

Version: ~ [ linux-5.11-rc3 ] ~ [ linux-5.10.7 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.89 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.167 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.215 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.251 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.251 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /* memcontrol.c - Memory Controller
  2  *
  3  * Copyright IBM Corporation, 2007
  4  * Author Balbir Singh <balbir@linux.vnet.ibm.com>
  5  *
  6  * Copyright 2007 OpenVZ SWsoft Inc
  7  * Author: Pavel Emelianov <xemul@openvz.org>
  8  *
  9  * Memory thresholds
 10  * Copyright (C) 2009 Nokia Corporation
 11  * Author: Kirill A. Shutemov
 12  *
 13  * Kernel Memory Controller
 14  * Copyright (C) 2012 Parallels Inc. and Google Inc.
 15  * Authors: Glauber Costa and Suleiman Souhlal
 16  *
 17  * This program is free software; you can redistribute it and/or modify
 18  * it under the terms of the GNU General Public License as published by
 19  * the Free Software Foundation; either version 2 of the License, or
 20  * (at your option) any later version.
 21  *
 22  * This program is distributed in the hope that it will be useful,
 23  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 24  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 25  * GNU General Public License for more details.
 26  */
 27 
 28 #include <linux/res_counter.h>
 29 #include <linux/memcontrol.h>
 30 #include <linux/cgroup.h>
 31 #include <linux/mm.h>
 32 #include <linux/hugetlb.h>
 33 #include <linux/pagemap.h>
 34 #include <linux/smp.h>
 35 #include <linux/page-flags.h>
 36 #include <linux/backing-dev.h>
 37 #include <linux/bit_spinlock.h>
 38 #include <linux/rcupdate.h>
 39 #include <linux/limits.h>
 40 #include <linux/export.h>
 41 #include <linux/mutex.h>
 42 #include <linux/rbtree.h>
 43 #include <linux/slab.h>
 44 #include <linux/swap.h>
 45 #include <linux/swapops.h>
 46 #include <linux/spinlock.h>
 47 #include <linux/eventfd.h>
 48 #include <linux/sort.h>
 49 #include <linux/fs.h>
 50 #include <linux/seq_file.h>
 51 #include <linux/vmalloc.h>
 52 #include <linux/vmpressure.h>
 53 #include <linux/mm_inline.h>
 54 #include <linux/page_cgroup.h>
 55 #include <linux/cpu.h>
 56 #include <linux/oom.h>
 57 #include <linux/lockdep.h>
 58 #include "internal.h"
 59 #include <net/sock.h>
 60 #include <net/ip.h>
 61 #include <net/tcp_memcontrol.h>
 62 #include "slab.h"
 63 
 64 #include <asm/uaccess.h>
 65 
 66 #include <trace/events/vmscan.h>
 67 
 68 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
 69 EXPORT_SYMBOL(mem_cgroup_subsys);
 70 
 71 #define MEM_CGROUP_RECLAIM_RETRIES      5
 72 static struct mem_cgroup *root_mem_cgroup __read_mostly;
 73 
 74 #ifdef CONFIG_MEMCG_SWAP
 75 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
 76 int do_swap_account __read_mostly;
 77 
 78 /* for remember boot option*/
 79 #ifdef CONFIG_MEMCG_SWAP_ENABLED
 80 static int really_do_swap_account __initdata = 1;
 81 #else
 82 static int really_do_swap_account __initdata = 0;
 83 #endif
 84 
 85 #else
 86 #define do_swap_account         0
 87 #endif
 88 
 89 
 90 static const char * const mem_cgroup_stat_names[] = {
 91         "cache",
 92         "rss",
 93         "rss_huge",
 94         "mapped_file",
 95         "writeback",
 96         "swap",
 97 };
 98 
 99 enum mem_cgroup_events_index {
100         MEM_CGROUP_EVENTS_PGPGIN,       /* # of pages paged in */
101         MEM_CGROUP_EVENTS_PGPGOUT,      /* # of pages paged out */
102         MEM_CGROUP_EVENTS_PGFAULT,      /* # of page-faults */
103         MEM_CGROUP_EVENTS_PGMAJFAULT,   /* # of major page-faults */
104         MEM_CGROUP_EVENTS_NSTATS,
105 };
106 
107 static const char * const mem_cgroup_events_names[] = {
108         "pgpgin",
109         "pgpgout",
110         "pgfault",
111         "pgmajfault",
112 };
113 
114 static const char * const mem_cgroup_lru_names[] = {
115         "inactive_anon",
116         "active_anon",
117         "inactive_file",
118         "active_file",
119         "unevictable",
120 };
121 
122 /*
123  * Per memcg event counter is incremented at every pagein/pageout. With THP,
124  * it will be incremated by the number of pages. This counter is used for
125  * for trigger some periodic events. This is straightforward and better
126  * than using jiffies etc. to handle periodic memcg event.
127  */
128 enum mem_cgroup_events_target {
129         MEM_CGROUP_TARGET_THRESH,
130         MEM_CGROUP_TARGET_SOFTLIMIT,
131         MEM_CGROUP_TARGET_NUMAINFO,
132         MEM_CGROUP_NTARGETS,
133 };
134 #define THRESHOLDS_EVENTS_TARGET 128
135 #define SOFTLIMIT_EVENTS_TARGET 1024
136 #define NUMAINFO_EVENTS_TARGET  1024
137 
138 struct mem_cgroup_stat_cpu {
139         long count[MEM_CGROUP_STAT_NSTATS];
140         unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
141         unsigned long nr_page_events;
142         unsigned long targets[MEM_CGROUP_NTARGETS];
143 };
144 
145 struct mem_cgroup_reclaim_iter {
146         /*
147          * last scanned hierarchy member. Valid only if last_dead_count
148          * matches memcg->dead_count of the hierarchy root group.
149          */
150         struct mem_cgroup *last_visited;
151         unsigned long last_dead_count;
152 
153         /* scan generation, increased every round-trip */
154         unsigned int generation;
155 };
156 
157 /*
158  * per-zone information in memory controller.
159  */
160 struct mem_cgroup_per_zone {
161         struct lruvec           lruvec;
162         unsigned long           lru_size[NR_LRU_LISTS];
163 
164         struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
165 
166         struct rb_node          tree_node;      /* RB tree node */
167         unsigned long long      usage_in_excess;/* Set to the value by which */
168                                                 /* the soft limit is exceeded*/
169         bool                    on_tree;
170         struct mem_cgroup       *memcg;         /* Back pointer, we cannot */
171                                                 /* use container_of        */
172 };
173 
174 struct mem_cgroup_per_node {
175         struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
176 };
177 
178 /*
179  * Cgroups above their limits are maintained in a RB-Tree, independent of
180  * their hierarchy representation
181  */
182 
183 struct mem_cgroup_tree_per_zone {
184         struct rb_root rb_root;
185         spinlock_t lock;
186 };
187 
188 struct mem_cgroup_tree_per_node {
189         struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
190 };
191 
192 struct mem_cgroup_tree {
193         struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
194 };
195 
196 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
197 
198 struct mem_cgroup_threshold {
199         struct eventfd_ctx *eventfd;
200         u64 threshold;
201 };
202 
203 /* For threshold */
204 struct mem_cgroup_threshold_ary {
205         /* An array index points to threshold just below or equal to usage. */
206         int current_threshold;
207         /* Size of entries[] */
208         unsigned int size;
209         /* Array of thresholds */
210         struct mem_cgroup_threshold entries[0];
211 };
212 
213 struct mem_cgroup_thresholds {
214         /* Primary thresholds array */
215         struct mem_cgroup_threshold_ary *primary;
216         /*
217          * Spare threshold array.
218          * This is needed to make mem_cgroup_unregister_event() "never fail".
219          * It must be able to store at least primary->size - 1 entries.
220          */
221         struct mem_cgroup_threshold_ary *spare;
222 };
223 
224 /* for OOM */
225 struct mem_cgroup_eventfd_list {
226         struct list_head list;
227         struct eventfd_ctx *eventfd;
228 };
229 
230 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
231 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
232 
233 /*
234  * The memory controller data structure. The memory controller controls both
235  * page cache and RSS per cgroup. We would eventually like to provide
236  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
237  * to help the administrator determine what knobs to tune.
238  *
239  * TODO: Add a water mark for the memory controller. Reclaim will begin when
240  * we hit the water mark. May be even add a low water mark, such that
241  * no reclaim occurs from a cgroup at it's low water mark, this is
242  * a feature that will be implemented much later in the future.
243  */
244 struct mem_cgroup {
245         struct cgroup_subsys_state css;
246         /*
247          * the counter to account for memory usage
248          */
249         struct res_counter res;
250 
251         /* vmpressure notifications */
252         struct vmpressure vmpressure;
253 
254         /*
255          * the counter to account for mem+swap usage.
256          */
257         struct res_counter memsw;
258 
259         /*
260          * the counter to account for kernel memory usage.
261          */
262         struct res_counter kmem;
263         /*
264          * Should the accounting and control be hierarchical, per subtree?
265          */
266         bool use_hierarchy;
267         unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
268 
269         bool            oom_lock;
270         atomic_t        under_oom;
271         atomic_t        oom_wakeups;
272 
273         int     swappiness;
274         /* OOM-Killer disable */
275         int             oom_kill_disable;
276 
277         /* set when res.limit == memsw.limit */
278         bool            memsw_is_minimum;
279 
280         /* protect arrays of thresholds */
281         struct mutex thresholds_lock;
282 
283         /* thresholds for memory usage. RCU-protected */
284         struct mem_cgroup_thresholds thresholds;
285 
286         /* thresholds for mem+swap usage. RCU-protected */
287         struct mem_cgroup_thresholds memsw_thresholds;
288 
289         /* For oom notifier event fd */
290         struct list_head oom_notify;
291 
292         /*
293          * Should we move charges of a task when a task is moved into this
294          * mem_cgroup ? And what type of charges should we move ?
295          */
296         unsigned long move_charge_at_immigrate;
297         /*
298          * set > 0 if pages under this cgroup are moving to other cgroup.
299          */
300         atomic_t        moving_account;
301         /* taken only while moving_account > 0 */
302         spinlock_t      move_lock;
303         /*
304          * percpu counter.
305          */
306         struct mem_cgroup_stat_cpu __percpu *stat;
307         /*
308          * used when a cpu is offlined or other synchronizations
309          * See mem_cgroup_read_stat().
310          */
311         struct mem_cgroup_stat_cpu nocpu_base;
312         spinlock_t pcp_counter_lock;
313 
314         atomic_t        dead_count;
315 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
316         struct cg_proto tcp_mem;
317 #endif
318 #if defined(CONFIG_MEMCG_KMEM)
319         /* analogous to slab_common's slab_caches list. per-memcg */
320         struct list_head memcg_slab_caches;
321         /* Not a spinlock, we can take a lot of time walking the list */
322         struct mutex slab_caches_mutex;
323         /* Index in the kmem_cache->memcg_params->memcg_caches array */
324         int kmemcg_id;
325 #endif
326 
327         int last_scanned_node;
328 #if MAX_NUMNODES > 1
329         nodemask_t      scan_nodes;
330         atomic_t        numainfo_events;
331         atomic_t        numainfo_updating;
332 #endif
333 
334         struct mem_cgroup_per_node *nodeinfo[0];
335         /* WARNING: nodeinfo must be the last member here */
336 };
337 
338 static size_t memcg_size(void)
339 {
340         return sizeof(struct mem_cgroup) +
341                 nr_node_ids * sizeof(struct mem_cgroup_per_node *);
342 }
343 
344 /* internal only representation about the status of kmem accounting. */
345 enum {
346         KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
347         KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
348         KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
349 };
350 
351 /* We account when limit is on, but only after call sites are patched */
352 #define KMEM_ACCOUNTED_MASK \
353                 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
354 
355 #ifdef CONFIG_MEMCG_KMEM
356 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
357 {
358         set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
359 }
360 
361 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
362 {
363         return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
364 }
365 
366 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
367 {
368         set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
369 }
370 
371 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
372 {
373         clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
374 }
375 
376 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
377 {
378         /*
379          * Our caller must use css_get() first, because memcg_uncharge_kmem()
380          * will call css_put() if it sees the memcg is dead.
381          */
382         smp_wmb();
383         if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
384                 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
385 }
386 
387 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
388 {
389         return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
390                                   &memcg->kmem_account_flags);
391 }
392 #endif
393 
394 /* Stuffs for move charges at task migration. */
395 /*
396  * Types of charges to be moved. "move_charge_at_immitgrate" and
397  * "immigrate_flags" are treated as a left-shifted bitmap of these types.
398  */
399 enum move_type {
400         MOVE_CHARGE_TYPE_ANON,  /* private anonymous page and swap of it */
401         MOVE_CHARGE_TYPE_FILE,  /* file page(including tmpfs) and swap of it */
402         NR_MOVE_TYPE,
403 };
404 
405 /* "mc" and its members are protected by cgroup_mutex */
406 static struct move_charge_struct {
407         spinlock_t        lock; /* for from, to */
408         struct mem_cgroup *from;
409         struct mem_cgroup *to;
410         unsigned long immigrate_flags;
411         unsigned long precharge;
412         unsigned long moved_charge;
413         unsigned long moved_swap;
414         struct task_struct *moving_task;        /* a task moving charges */
415         wait_queue_head_t waitq;                /* a waitq for other context */
416 } mc = {
417         .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
418         .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
419 };
420 
421 static bool move_anon(void)
422 {
423         return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
424 }
425 
426 static bool move_file(void)
427 {
428         return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
429 }
430 
431 /*
432  * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
433  * limit reclaim to prevent infinite loops, if they ever occur.
434  */
435 #define MEM_CGROUP_MAX_RECLAIM_LOOPS            100
436 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
437 
438 enum charge_type {
439         MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
440         MEM_CGROUP_CHARGE_TYPE_ANON,
441         MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
442         MEM_CGROUP_CHARGE_TYPE_DROP,    /* a page was unused swap cache */
443         NR_CHARGE_TYPE,
444 };
445 
446 /* for encoding cft->private value on file */
447 enum res_type {
448         _MEM,
449         _MEMSWAP,
450         _OOM_TYPE,
451         _KMEM,
452 };
453 
454 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
455 #define MEMFILE_TYPE(val)       ((val) >> 16 & 0xffff)
456 #define MEMFILE_ATTR(val)       ((val) & 0xffff)
457 /* Used for OOM nofiier */
458 #define OOM_CONTROL             (0)
459 
460 /*
461  * Reclaim flags for mem_cgroup_hierarchical_reclaim
462  */
463 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT   0x0
464 #define MEM_CGROUP_RECLAIM_NOSWAP       (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
465 #define MEM_CGROUP_RECLAIM_SHRINK_BIT   0x1
466 #define MEM_CGROUP_RECLAIM_SHRINK       (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
467 
468 /*
469  * The memcg_create_mutex will be held whenever a new cgroup is created.
470  * As a consequence, any change that needs to protect against new child cgroups
471  * appearing has to hold it as well.
472  */
473 static DEFINE_MUTEX(memcg_create_mutex);
474 
475 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
476 {
477         return s ? container_of(s, struct mem_cgroup, css) : NULL;
478 }
479 
480 /* Some nice accessors for the vmpressure. */
481 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
482 {
483         if (!memcg)
484                 memcg = root_mem_cgroup;
485         return &memcg->vmpressure;
486 }
487 
488 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
489 {
490         return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
491 }
492 
493 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
494 {
495         return &mem_cgroup_from_css(css)->vmpressure;
496 }
497 
498 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
499 {
500         return (memcg == root_mem_cgroup);
501 }
502 
503 /*
504  * We restrict the id in the range of [1, 65535], so it can fit into
505  * an unsigned short.
506  */
507 #define MEM_CGROUP_ID_MAX       USHRT_MAX
508 
509 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
510 {
511         /*
512          * The ID of the root cgroup is 0, but memcg treat 0 as an
513          * invalid ID, so we return (cgroup_id + 1).
514          */
515         return memcg->css.cgroup->id + 1;
516 }
517 
518 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
519 {
520         struct cgroup_subsys_state *css;
521 
522         css = css_from_id(id - 1, &mem_cgroup_subsys);
523         return mem_cgroup_from_css(css);
524 }
525 
526 /* Writing them here to avoid exposing memcg's inner layout */
527 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
528 
529 void sock_update_memcg(struct sock *sk)
530 {
531         if (mem_cgroup_sockets_enabled) {
532                 struct mem_cgroup *memcg;
533                 struct cg_proto *cg_proto;
534 
535                 BUG_ON(!sk->sk_prot->proto_cgroup);
536 
537                 /* Socket cloning can throw us here with sk_cgrp already
538                  * filled. It won't however, necessarily happen from
539                  * process context. So the test for root memcg given
540                  * the current task's memcg won't help us in this case.
541                  *
542                  * Respecting the original socket's memcg is a better
543                  * decision in this case.
544                  */
545                 if (sk->sk_cgrp) {
546                         BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
547                         css_get(&sk->sk_cgrp->memcg->css);
548                         return;
549                 }
550 
551                 rcu_read_lock();
552                 memcg = mem_cgroup_from_task(current);
553                 cg_proto = sk->sk_prot->proto_cgroup(memcg);
554                 if (!mem_cgroup_is_root(memcg) &&
555                     memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
556                         sk->sk_cgrp = cg_proto;
557                 }
558                 rcu_read_unlock();
559         }
560 }
561 EXPORT_SYMBOL(sock_update_memcg);
562 
563 void sock_release_memcg(struct sock *sk)
564 {
565         if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
566                 struct mem_cgroup *memcg;
567                 WARN_ON(!sk->sk_cgrp->memcg);
568                 memcg = sk->sk_cgrp->memcg;
569                 css_put(&sk->sk_cgrp->memcg->css);
570         }
571 }
572 
573 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
574 {
575         if (!memcg || mem_cgroup_is_root(memcg))
576                 return NULL;
577 
578         return &memcg->tcp_mem;
579 }
580 EXPORT_SYMBOL(tcp_proto_cgroup);
581 
582 static void disarm_sock_keys(struct mem_cgroup *memcg)
583 {
584         if (!memcg_proto_activated(&memcg->tcp_mem))
585                 return;
586         static_key_slow_dec(&memcg_socket_limit_enabled);
587 }
588 #else
589 static void disarm_sock_keys(struct mem_cgroup *memcg)
590 {
591 }
592 #endif
593 
594 #ifdef CONFIG_MEMCG_KMEM
595 /*
596  * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
597  * The main reason for not using cgroup id for this:
598  *  this works better in sparse environments, where we have a lot of memcgs,
599  *  but only a few kmem-limited. Or also, if we have, for instance, 200
600  *  memcgs, and none but the 200th is kmem-limited, we'd have to have a
601  *  200 entry array for that.
602  *
603  * The current size of the caches array is stored in
604  * memcg_limited_groups_array_size.  It will double each time we have to
605  * increase it.
606  */
607 static DEFINE_IDA(kmem_limited_groups);
608 int memcg_limited_groups_array_size;
609 
610 /*
611  * MIN_SIZE is different than 1, because we would like to avoid going through
612  * the alloc/free process all the time. In a small machine, 4 kmem-limited
613  * cgroups is a reasonable guess. In the future, it could be a parameter or
614  * tunable, but that is strictly not necessary.
615  *
616  * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
617  * this constant directly from cgroup, but it is understandable that this is
618  * better kept as an internal representation in cgroup.c. In any case, the
619  * cgrp_id space is not getting any smaller, and we don't have to necessarily
620  * increase ours as well if it increases.
621  */
622 #define MEMCG_CACHES_MIN_SIZE 4
623 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
624 
625 /*
626  * A lot of the calls to the cache allocation functions are expected to be
627  * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
628  * conditional to this static branch, we'll have to allow modules that does
629  * kmem_cache_alloc and the such to see this symbol as well
630  */
631 struct static_key memcg_kmem_enabled_key;
632 EXPORT_SYMBOL(memcg_kmem_enabled_key);
633 
634 static void disarm_kmem_keys(struct mem_cgroup *memcg)
635 {
636         if (memcg_kmem_is_active(memcg)) {
637                 static_key_slow_dec(&memcg_kmem_enabled_key);
638                 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
639         }
640         /*
641          * This check can't live in kmem destruction function,
642          * since the charges will outlive the cgroup
643          */
644         WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
645 }
646 #else
647 static void disarm_kmem_keys(struct mem_cgroup *memcg)
648 {
649 }
650 #endif /* CONFIG_MEMCG_KMEM */
651 
652 static void disarm_static_keys(struct mem_cgroup *memcg)
653 {
654         disarm_sock_keys(memcg);
655         disarm_kmem_keys(memcg);
656 }
657 
658 static void drain_all_stock_async(struct mem_cgroup *memcg);
659 
660 static struct mem_cgroup_per_zone *
661 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
662 {
663         VM_BUG_ON((unsigned)nid >= nr_node_ids);
664         return &memcg->nodeinfo[nid]->zoneinfo[zid];
665 }
666 
667 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
668 {
669         return &memcg->css;
670 }
671 
672 static struct mem_cgroup_per_zone *
673 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
674 {
675         int nid = page_to_nid(page);
676         int zid = page_zonenum(page);
677 
678         return mem_cgroup_zoneinfo(memcg, nid, zid);
679 }
680 
681 static struct mem_cgroup_tree_per_zone *
682 soft_limit_tree_node_zone(int nid, int zid)
683 {
684         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
685 }
686 
687 static struct mem_cgroup_tree_per_zone *
688 soft_limit_tree_from_page(struct page *page)
689 {
690         int nid = page_to_nid(page);
691         int zid = page_zonenum(page);
692 
693         return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
694 }
695 
696 static void
697 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
698                                 struct mem_cgroup_per_zone *mz,
699                                 struct mem_cgroup_tree_per_zone *mctz,
700                                 unsigned long long new_usage_in_excess)
701 {
702         struct rb_node **p = &mctz->rb_root.rb_node;
703         struct rb_node *parent = NULL;
704         struct mem_cgroup_per_zone *mz_node;
705 
706         if (mz->on_tree)
707                 return;
708 
709         mz->usage_in_excess = new_usage_in_excess;
710         if (!mz->usage_in_excess)
711                 return;
712         while (*p) {
713                 parent = *p;
714                 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
715                                         tree_node);
716                 if (mz->usage_in_excess < mz_node->usage_in_excess)
717                         p = &(*p)->rb_left;
718                 /*
719                  * We can't avoid mem cgroups that are over their soft
720                  * limit by the same amount
721                  */
722                 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
723                         p = &(*p)->rb_right;
724         }
725         rb_link_node(&mz->tree_node, parent, p);
726         rb_insert_color(&mz->tree_node, &mctz->rb_root);
727         mz->on_tree = true;
728 }
729 
730 static void
731 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
732                                 struct mem_cgroup_per_zone *mz,
733                                 struct mem_cgroup_tree_per_zone *mctz)
734 {
735         if (!mz->on_tree)
736                 return;
737         rb_erase(&mz->tree_node, &mctz->rb_root);
738         mz->on_tree = false;
739 }
740 
741 static void
742 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
743                                 struct mem_cgroup_per_zone *mz,
744                                 struct mem_cgroup_tree_per_zone *mctz)
745 {
746         spin_lock(&mctz->lock);
747         __mem_cgroup_remove_exceeded(memcg, mz, mctz);
748         spin_unlock(&mctz->lock);
749 }
750 
751 
752 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
753 {
754         unsigned long long excess;
755         struct mem_cgroup_per_zone *mz;
756         struct mem_cgroup_tree_per_zone *mctz;
757         int nid = page_to_nid(page);
758         int zid = page_zonenum(page);
759         mctz = soft_limit_tree_from_page(page);
760 
761         /*
762          * Necessary to update all ancestors when hierarchy is used.
763          * because their event counter is not touched.
764          */
765         for (; memcg; memcg = parent_mem_cgroup(memcg)) {
766                 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
767                 excess = res_counter_soft_limit_excess(&memcg->res);
768                 /*
769                  * We have to update the tree if mz is on RB-tree or
770                  * mem is over its softlimit.
771                  */
772                 if (excess || mz->on_tree) {
773                         spin_lock(&mctz->lock);
774                         /* if on-tree, remove it */
775                         if (mz->on_tree)
776                                 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
777                         /*
778                          * Insert again. mz->usage_in_excess will be updated.
779                          * If excess is 0, no tree ops.
780                          */
781                         __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
782                         spin_unlock(&mctz->lock);
783                 }
784         }
785 }
786 
787 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
788 {
789         int node, zone;
790         struct mem_cgroup_per_zone *mz;
791         struct mem_cgroup_tree_per_zone *mctz;
792 
793         for_each_node(node) {
794                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
795                         mz = mem_cgroup_zoneinfo(memcg, node, zone);
796                         mctz = soft_limit_tree_node_zone(node, zone);
797                         mem_cgroup_remove_exceeded(memcg, mz, mctz);
798                 }
799         }
800 }
801 
802 static struct mem_cgroup_per_zone *
803 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
804 {
805         struct rb_node *rightmost = NULL;
806         struct mem_cgroup_per_zone *mz;
807 
808 retry:
809         mz = NULL;
810         rightmost = rb_last(&mctz->rb_root);
811         if (!rightmost)
812                 goto done;              /* Nothing to reclaim from */
813 
814         mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
815         /*
816          * Remove the node now but someone else can add it back,
817          * we will to add it back at the end of reclaim to its correct
818          * position in the tree.
819          */
820         __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
821         if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
822                 !css_tryget(&mz->memcg->css))
823                 goto retry;
824 done:
825         return mz;
826 }
827 
828 static struct mem_cgroup_per_zone *
829 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
830 {
831         struct mem_cgroup_per_zone *mz;
832 
833         spin_lock(&mctz->lock);
834         mz = __mem_cgroup_largest_soft_limit_node(mctz);
835         spin_unlock(&mctz->lock);
836         return mz;
837 }
838 
839 /*
840  * Implementation Note: reading percpu statistics for memcg.
841  *
842  * Both of vmstat[] and percpu_counter has threshold and do periodic
843  * synchronization to implement "quick" read. There are trade-off between
844  * reading cost and precision of value. Then, we may have a chance to implement
845  * a periodic synchronizion of counter in memcg's counter.
846  *
847  * But this _read() function is used for user interface now. The user accounts
848  * memory usage by memory cgroup and he _always_ requires exact value because
849  * he accounts memory. Even if we provide quick-and-fuzzy read, we always
850  * have to visit all online cpus and make sum. So, for now, unnecessary
851  * synchronization is not implemented. (just implemented for cpu hotplug)
852  *
853  * If there are kernel internal actions which can make use of some not-exact
854  * value, and reading all cpu value can be performance bottleneck in some
855  * common workload, threashold and synchonization as vmstat[] should be
856  * implemented.
857  */
858 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
859                                  enum mem_cgroup_stat_index idx)
860 {
861         long val = 0;
862         int cpu;
863 
864         get_online_cpus();
865         for_each_online_cpu(cpu)
866                 val += per_cpu(memcg->stat->count[idx], cpu);
867 #ifdef CONFIG_HOTPLUG_CPU
868         spin_lock(&memcg->pcp_counter_lock);
869         val += memcg->nocpu_base.count[idx];
870         spin_unlock(&memcg->pcp_counter_lock);
871 #endif
872         put_online_cpus();
873         return val;
874 }
875 
876 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
877                                          bool charge)
878 {
879         int val = (charge) ? 1 : -1;
880         this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
881 }
882 
883 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
884                                             enum mem_cgroup_events_index idx)
885 {
886         unsigned long val = 0;
887         int cpu;
888 
889         get_online_cpus();
890         for_each_online_cpu(cpu)
891                 val += per_cpu(memcg->stat->events[idx], cpu);
892 #ifdef CONFIG_HOTPLUG_CPU
893         spin_lock(&memcg->pcp_counter_lock);
894         val += memcg->nocpu_base.events[idx];
895         spin_unlock(&memcg->pcp_counter_lock);
896 #endif
897         put_online_cpus();
898         return val;
899 }
900 
901 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
902                                          struct page *page,
903                                          bool anon, int nr_pages)
904 {
905         preempt_disable();
906 
907         /*
908          * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
909          * counted as CACHE even if it's on ANON LRU.
910          */
911         if (anon)
912                 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
913                                 nr_pages);
914         else
915                 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
916                                 nr_pages);
917 
918         if (PageTransHuge(page))
919                 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
920                                 nr_pages);
921 
922         /* pagein of a big page is an event. So, ignore page size */
923         if (nr_pages > 0)
924                 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
925         else {
926                 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
927                 nr_pages = -nr_pages; /* for event */
928         }
929 
930         __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
931 
932         preempt_enable();
933 }
934 
935 unsigned long
936 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
937 {
938         struct mem_cgroup_per_zone *mz;
939 
940         mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
941         return mz->lru_size[lru];
942 }
943 
944 static unsigned long
945 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
946                         unsigned int lru_mask)
947 {
948         struct mem_cgroup_per_zone *mz;
949         enum lru_list lru;
950         unsigned long ret = 0;
951 
952         mz = mem_cgroup_zoneinfo(memcg, nid, zid);
953 
954         for_each_lru(lru) {
955                 if (BIT(lru) & lru_mask)
956                         ret += mz->lru_size[lru];
957         }
958         return ret;
959 }
960 
961 static unsigned long
962 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
963                         int nid, unsigned int lru_mask)
964 {
965         u64 total = 0;
966         int zid;
967 
968         for (zid = 0; zid < MAX_NR_ZONES; zid++)
969                 total += mem_cgroup_zone_nr_lru_pages(memcg,
970                                                 nid, zid, lru_mask);
971 
972         return total;
973 }
974 
975 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
976                         unsigned int lru_mask)
977 {
978         int nid;
979         u64 total = 0;
980 
981         for_each_node_state(nid, N_MEMORY)
982                 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
983         return total;
984 }
985 
986 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
987                                        enum mem_cgroup_events_target target)
988 {
989         unsigned long val, next;
990 
991         val = __this_cpu_read(memcg->stat->nr_page_events);
992         next = __this_cpu_read(memcg->stat->targets[target]);
993         /* from time_after() in jiffies.h */
994         if ((long)next - (long)val < 0) {
995                 switch (target) {
996                 case MEM_CGROUP_TARGET_THRESH:
997                         next = val + THRESHOLDS_EVENTS_TARGET;
998                         break;
999                 case MEM_CGROUP_TARGET_SOFTLIMIT:
1000                         next = val + SOFTLIMIT_EVENTS_TARGET;
1001                         break;
1002                 case MEM_CGROUP_TARGET_NUMAINFO:
1003                         next = val + NUMAINFO_EVENTS_TARGET;
1004                         break;
1005                 default:
1006                         break;
1007                 }
1008                 __this_cpu_write(memcg->stat->targets[target], next);
1009                 return true;
1010         }
1011         return false;
1012 }
1013 
1014 /*
1015  * Check events in order.
1016  *
1017  */
1018 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1019 {
1020         preempt_disable();
1021         /* threshold event is triggered in finer grain than soft limit */
1022         if (unlikely(mem_cgroup_event_ratelimit(memcg,
1023                                                 MEM_CGROUP_TARGET_THRESH))) {
1024                 bool do_softlimit;
1025                 bool do_numainfo __maybe_unused;
1026 
1027                 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1028                                                 MEM_CGROUP_TARGET_SOFTLIMIT);
1029 #if MAX_NUMNODES > 1
1030                 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1031                                                 MEM_CGROUP_TARGET_NUMAINFO);
1032 #endif
1033                 preempt_enable();
1034 
1035                 mem_cgroup_threshold(memcg);
1036                 if (unlikely(do_softlimit))
1037                         mem_cgroup_update_tree(memcg, page);
1038 #if MAX_NUMNODES > 1
1039                 if (unlikely(do_numainfo))
1040                         atomic_inc(&memcg->numainfo_events);
1041 #endif
1042         } else
1043                 preempt_enable();
1044 }
1045 
1046 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1047 {
1048         /*
1049          * mm_update_next_owner() may clear mm->owner to NULL
1050          * if it races with swapoff, page migration, etc.
1051          * So this can be called with p == NULL.
1052          */
1053         if (unlikely(!p))
1054                 return NULL;
1055 
1056         return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1057 }
1058 
1059 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1060 {
1061         struct mem_cgroup *memcg = NULL;
1062 
1063         if (!mm)
1064                 return NULL;
1065         /*
1066          * Because we have no locks, mm->owner's may be being moved to other
1067          * cgroup. We use css_tryget() here even if this looks
1068          * pessimistic (rather than adding locks here).
1069          */
1070         rcu_read_lock();
1071         do {
1072                 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1073                 if (unlikely(!memcg))
1074                         break;
1075         } while (!css_tryget(&memcg->css));
1076         rcu_read_unlock();
1077         return memcg;
1078 }
1079 
1080 /*
1081  * Returns a next (in a pre-order walk) alive memcg (with elevated css
1082  * ref. count) or NULL if the whole root's subtree has been visited.
1083  *
1084  * helper function to be used by mem_cgroup_iter
1085  */
1086 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1087                 struct mem_cgroup *last_visited)
1088 {
1089         struct cgroup_subsys_state *prev_css, *next_css;
1090 
1091         prev_css = last_visited ? &last_visited->css : NULL;
1092 skip_node:
1093         next_css = css_next_descendant_pre(prev_css, &root->css);
1094 
1095         /*
1096          * Even if we found a group we have to make sure it is
1097          * alive. css && !memcg means that the groups should be
1098          * skipped and we should continue the tree walk.
1099          * last_visited css is safe to use because it is
1100          * protected by css_get and the tree walk is rcu safe.
1101          *
1102          * We do not take a reference on the root of the tree walk
1103          * because we might race with the root removal when it would
1104          * be the only node in the iterated hierarchy and mem_cgroup_iter
1105          * would end up in an endless loop because it expects that at
1106          * least one valid node will be returned. Root cannot disappear
1107          * because caller of the iterator should hold it already so
1108          * skipping css reference should be safe.
1109          */
1110         if (next_css) {
1111                 if ((next_css == &root->css) ||
1112                     ((next_css->flags & CSS_ONLINE) && css_tryget(next_css)))
1113                         return mem_cgroup_from_css(next_css);
1114 
1115                 prev_css = next_css;
1116                 goto skip_node;
1117         }
1118 
1119         return NULL;
1120 }
1121 
1122 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1123 {
1124         /*
1125          * When a group in the hierarchy below root is destroyed, the
1126          * hierarchy iterator can no longer be trusted since it might
1127          * have pointed to the destroyed group.  Invalidate it.
1128          */
1129         atomic_inc(&root->dead_count);
1130 }
1131 
1132 static struct mem_cgroup *
1133 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1134                      struct mem_cgroup *root,
1135                      int *sequence)
1136 {
1137         struct mem_cgroup *position = NULL;
1138         /*
1139          * A cgroup destruction happens in two stages: offlining and
1140          * release.  They are separated by a RCU grace period.
1141          *
1142          * If the iterator is valid, we may still race with an
1143          * offlining.  The RCU lock ensures the object won't be
1144          * released, tryget will fail if we lost the race.
1145          */
1146         *sequence = atomic_read(&root->dead_count);
1147         if (iter->last_dead_count == *sequence) {
1148                 smp_rmb();
1149                 position = iter->last_visited;
1150 
1151                 /*
1152                  * We cannot take a reference to root because we might race
1153                  * with root removal and returning NULL would end up in
1154                  * an endless loop on the iterator user level when root
1155                  * would be returned all the time.
1156                  */
1157                 if (position && position != root &&
1158                                 !css_tryget(&position->css))
1159                         position = NULL;
1160         }
1161         return position;
1162 }
1163 
1164 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1165                                    struct mem_cgroup *last_visited,
1166                                    struct mem_cgroup *new_position,
1167                                    struct mem_cgroup *root,
1168                                    int sequence)
1169 {
1170         /* root reference counting symmetric to mem_cgroup_iter_load */
1171         if (last_visited && last_visited != root)
1172                 css_put(&last_visited->css);
1173         /*
1174          * We store the sequence count from the time @last_visited was
1175          * loaded successfully instead of rereading it here so that we
1176          * don't lose destruction events in between.  We could have
1177          * raced with the destruction of @new_position after all.
1178          */
1179         iter->last_visited = new_position;
1180         smp_wmb();
1181         iter->last_dead_count = sequence;
1182 }
1183 
1184 /**
1185  * mem_cgroup_iter - iterate over memory cgroup hierarchy
1186  * @root: hierarchy root
1187  * @prev: previously returned memcg, NULL on first invocation
1188  * @reclaim: cookie for shared reclaim walks, NULL for full walks
1189  *
1190  * Returns references to children of the hierarchy below @root, or
1191  * @root itself, or %NULL after a full round-trip.
1192  *
1193  * Caller must pass the return value in @prev on subsequent
1194  * invocations for reference counting, or use mem_cgroup_iter_break()
1195  * to cancel a hierarchy walk before the round-trip is complete.
1196  *
1197  * Reclaimers can specify a zone and a priority level in @reclaim to
1198  * divide up the memcgs in the hierarchy among all concurrent
1199  * reclaimers operating on the same zone and priority.
1200  */
1201 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1202                                    struct mem_cgroup *prev,
1203                                    struct mem_cgroup_reclaim_cookie *reclaim)
1204 {
1205         struct mem_cgroup *memcg = NULL;
1206         struct mem_cgroup *last_visited = NULL;
1207 
1208         if (mem_cgroup_disabled())
1209                 return NULL;
1210 
1211         if (!root)
1212                 root = root_mem_cgroup;
1213 
1214         if (prev && !reclaim)
1215                 last_visited = prev;
1216 
1217         if (!root->use_hierarchy && root != root_mem_cgroup) {
1218                 if (prev)
1219                         goto out_css_put;
1220                 return root;
1221         }
1222 
1223         rcu_read_lock();
1224         while (!memcg) {
1225                 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1226                 int uninitialized_var(seq);
1227 
1228                 if (reclaim) {
1229                         int nid = zone_to_nid(reclaim->zone);
1230                         int zid = zone_idx(reclaim->zone);
1231                         struct mem_cgroup_per_zone *mz;
1232 
1233                         mz = mem_cgroup_zoneinfo(root, nid, zid);
1234                         iter = &mz->reclaim_iter[reclaim->priority];
1235                         if (prev && reclaim->generation != iter->generation) {
1236                                 iter->last_visited = NULL;
1237                                 goto out_unlock;
1238                         }
1239 
1240                         last_visited = mem_cgroup_iter_load(iter, root, &seq);
1241                 }
1242 
1243                 memcg = __mem_cgroup_iter_next(root, last_visited);
1244 
1245                 if (reclaim) {
1246                         mem_cgroup_iter_update(iter, last_visited, memcg, root,
1247                                         seq);
1248 
1249                         if (!memcg)
1250                                 iter->generation++;
1251                         else if (!prev && memcg)
1252                                 reclaim->generation = iter->generation;
1253                 }
1254 
1255                 if (prev && !memcg)
1256                         goto out_unlock;
1257         }
1258 out_unlock:
1259         rcu_read_unlock();
1260 out_css_put:
1261         if (prev && prev != root)
1262                 css_put(&prev->css);
1263 
1264         return memcg;
1265 }
1266 
1267 /**
1268  * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1269  * @root: hierarchy root
1270  * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1271  */
1272 void mem_cgroup_iter_break(struct mem_cgroup *root,
1273                            struct mem_cgroup *prev)
1274 {
1275         if (!root)
1276                 root = root_mem_cgroup;
1277         if (prev && prev != root)
1278                 css_put(&prev->css);
1279 }
1280 
1281 /*
1282  * Iteration constructs for visiting all cgroups (under a tree).  If
1283  * loops are exited prematurely (break), mem_cgroup_iter_break() must
1284  * be used for reference counting.
1285  */
1286 #define for_each_mem_cgroup_tree(iter, root)            \
1287         for (iter = mem_cgroup_iter(root, NULL, NULL);  \
1288              iter != NULL;                              \
1289              iter = mem_cgroup_iter(root, iter, NULL))
1290 
1291 #define for_each_mem_cgroup(iter)                       \
1292         for (iter = mem_cgroup_iter(NULL, NULL, NULL);  \
1293              iter != NULL;                              \
1294              iter = mem_cgroup_iter(NULL, iter, NULL))
1295 
1296 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1297 {
1298         struct mem_cgroup *memcg;
1299 
1300         rcu_read_lock();
1301         memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1302         if (unlikely(!memcg))
1303                 goto out;
1304 
1305         switch (idx) {
1306         case PGFAULT:
1307                 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1308                 break;
1309         case PGMAJFAULT:
1310                 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1311                 break;
1312         default:
1313                 BUG();
1314         }
1315 out:
1316         rcu_read_unlock();
1317 }
1318 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1319 
1320 /**
1321  * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1322  * @zone: zone of the wanted lruvec
1323  * @memcg: memcg of the wanted lruvec
1324  *
1325  * Returns the lru list vector holding pages for the given @zone and
1326  * @mem.  This can be the global zone lruvec, if the memory controller
1327  * is disabled.
1328  */
1329 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1330                                       struct mem_cgroup *memcg)
1331 {
1332         struct mem_cgroup_per_zone *mz;
1333         struct lruvec *lruvec;
1334 
1335         if (mem_cgroup_disabled()) {
1336                 lruvec = &zone->lruvec;
1337                 goto out;
1338         }
1339 
1340         mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1341         lruvec = &mz->lruvec;
1342 out:
1343         /*
1344          * Since a node can be onlined after the mem_cgroup was created,
1345          * we have to be prepared to initialize lruvec->zone here;
1346          * and if offlined then reonlined, we need to reinitialize it.
1347          */
1348         if (unlikely(lruvec->zone != zone))
1349                 lruvec->zone = zone;
1350         return lruvec;
1351 }
1352 
1353 /*
1354  * Following LRU functions are allowed to be used without PCG_LOCK.
1355  * Operations are called by routine of global LRU independently from memcg.
1356  * What we have to take care of here is validness of pc->mem_cgroup.
1357  *
1358  * Changes to pc->mem_cgroup happens when
1359  * 1. charge
1360  * 2. moving account
1361  * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1362  * It is added to LRU before charge.
1363  * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1364  * When moving account, the page is not on LRU. It's isolated.
1365  */
1366 
1367 /**
1368  * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1369  * @page: the page
1370  * @zone: zone of the page
1371  */
1372 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1373 {
1374         struct mem_cgroup_per_zone *mz;
1375         struct mem_cgroup *memcg;
1376         struct page_cgroup *pc;
1377         struct lruvec *lruvec;
1378 
1379         if (mem_cgroup_disabled()) {
1380                 lruvec = &zone->lruvec;
1381                 goto out;
1382         }
1383 
1384         pc = lookup_page_cgroup(page);
1385         memcg = pc->mem_cgroup;
1386 
1387         /*
1388          * Surreptitiously switch any uncharged offlist page to root:
1389          * an uncharged page off lru does nothing to secure
1390          * its former mem_cgroup from sudden removal.
1391          *
1392          * Our caller holds lru_lock, and PageCgroupUsed is updated
1393          * under page_cgroup lock: between them, they make all uses
1394          * of pc->mem_cgroup safe.
1395          */
1396         if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1397                 pc->mem_cgroup = memcg = root_mem_cgroup;
1398 
1399         mz = page_cgroup_zoneinfo(memcg, page);
1400         lruvec = &mz->lruvec;
1401 out:
1402         /*
1403          * Since a node can be onlined after the mem_cgroup was created,
1404          * we have to be prepared to initialize lruvec->zone here;
1405          * and if offlined then reonlined, we need to reinitialize it.
1406          */
1407         if (unlikely(lruvec->zone != zone))
1408                 lruvec->zone = zone;
1409         return lruvec;
1410 }
1411 
1412 /**
1413  * mem_cgroup_update_lru_size - account for adding or removing an lru page
1414  * @lruvec: mem_cgroup per zone lru vector
1415  * @lru: index of lru list the page is sitting on
1416  * @nr_pages: positive when adding or negative when removing
1417  *
1418  * This function must be called when a page is added to or removed from an
1419  * lru list.
1420  */
1421 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1422                                 int nr_pages)
1423 {
1424         struct mem_cgroup_per_zone *mz;
1425         unsigned long *lru_size;
1426 
1427         if (mem_cgroup_disabled())
1428                 return;
1429 
1430         mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1431         lru_size = mz->lru_size + lru;
1432         *lru_size += nr_pages;
1433         VM_BUG_ON((long)(*lru_size) < 0);
1434 }
1435 
1436 /*
1437  * Checks whether given mem is same or in the root_mem_cgroup's
1438  * hierarchy subtree
1439  */
1440 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1441                                   struct mem_cgroup *memcg)
1442 {
1443         if (root_memcg == memcg)
1444                 return true;
1445         if (!root_memcg->use_hierarchy || !memcg)
1446                 return false;
1447         return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1448 }
1449 
1450 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1451                                        struct mem_cgroup *memcg)
1452 {
1453         bool ret;
1454 
1455         rcu_read_lock();
1456         ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1457         rcu_read_unlock();
1458         return ret;
1459 }
1460 
1461 bool task_in_mem_cgroup(struct task_struct *task,
1462                         const struct mem_cgroup *memcg)
1463 {
1464         struct mem_cgroup *curr = NULL;
1465         struct task_struct *p;
1466         bool ret;
1467 
1468         p = find_lock_task_mm(task);
1469         if (p) {
1470                 curr = try_get_mem_cgroup_from_mm(p->mm);
1471                 task_unlock(p);
1472         } else {
1473                 /*
1474                  * All threads may have already detached their mm's, but the oom
1475                  * killer still needs to detect if they have already been oom
1476                  * killed to prevent needlessly killing additional tasks.
1477                  */
1478                 rcu_read_lock();
1479                 curr = mem_cgroup_from_task(task);
1480                 if (curr)
1481                         css_get(&curr->css);
1482                 rcu_read_unlock();
1483         }
1484         if (!curr)
1485                 return false;
1486         /*
1487          * We should check use_hierarchy of "memcg" not "curr". Because checking
1488          * use_hierarchy of "curr" here make this function true if hierarchy is
1489          * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1490          * hierarchy(even if use_hierarchy is disabled in "memcg").
1491          */
1492         ret = mem_cgroup_same_or_subtree(memcg, curr);
1493         css_put(&curr->css);
1494         return ret;
1495 }
1496 
1497 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1498 {
1499         unsigned long inactive_ratio;
1500         unsigned long inactive;
1501         unsigned long active;
1502         unsigned long gb;
1503 
1504         inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1505         active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1506 
1507         gb = (inactive + active) >> (30 - PAGE_SHIFT);
1508         if (gb)
1509                 inactive_ratio = int_sqrt(10 * gb);
1510         else
1511                 inactive_ratio = 1;
1512 
1513         return inactive * inactive_ratio < active;
1514 }
1515 
1516 #define mem_cgroup_from_res_counter(counter, member)    \
1517         container_of(counter, struct mem_cgroup, member)
1518 
1519 /**
1520  * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1521  * @memcg: the memory cgroup
1522  *
1523  * Returns the maximum amount of memory @mem can be charged with, in
1524  * pages.
1525  */
1526 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1527 {
1528         unsigned long long margin;
1529 
1530         margin = res_counter_margin(&memcg->res);
1531         if (do_swap_account)
1532                 margin = min(margin, res_counter_margin(&memcg->memsw));
1533         return margin >> PAGE_SHIFT;
1534 }
1535 
1536 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1537 {
1538         /* root ? */
1539         if (!css_parent(&memcg->css))
1540                 return vm_swappiness;
1541 
1542         return memcg->swappiness;
1543 }
1544 
1545 /*
1546  * memcg->moving_account is used for checking possibility that some thread is
1547  * calling move_account(). When a thread on CPU-A starts moving pages under
1548  * a memcg, other threads should check memcg->moving_account under
1549  * rcu_read_lock(), like this:
1550  *
1551  *         CPU-A                                    CPU-B
1552  *                                              rcu_read_lock()
1553  *         memcg->moving_account+1              if (memcg->mocing_account)
1554  *                                                   take heavy locks.
1555  *         synchronize_rcu()                    update something.
1556  *                                              rcu_read_unlock()
1557  *         start move here.
1558  */
1559 
1560 /* for quick checking without looking up memcg */
1561 atomic_t memcg_moving __read_mostly;
1562 
1563 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1564 {
1565         atomic_inc(&memcg_moving);
1566         atomic_inc(&memcg->moving_account);
1567         synchronize_rcu();
1568 }
1569 
1570 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1571 {
1572         /*
1573          * Now, mem_cgroup_clear_mc() may call this function with NULL.
1574          * We check NULL in callee rather than caller.
1575          */
1576         if (memcg) {
1577                 atomic_dec(&memcg_moving);
1578                 atomic_dec(&memcg->moving_account);
1579         }
1580 }
1581 
1582 /*
1583  * 2 routines for checking "mem" is under move_account() or not.
1584  *
1585  * mem_cgroup_stolen() -  checking whether a cgroup is mc.from or not. This
1586  *                        is used for avoiding races in accounting.  If true,
1587  *                        pc->mem_cgroup may be overwritten.
1588  *
1589  * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1590  *                        under hierarchy of moving cgroups. This is for
1591  *                        waiting at hith-memory prressure caused by "move".
1592  */
1593 
1594 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1595 {
1596         VM_BUG_ON(!rcu_read_lock_held());
1597         return atomic_read(&memcg->moving_account) > 0;
1598 }
1599 
1600 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1601 {
1602         struct mem_cgroup *from;
1603         struct mem_cgroup *to;
1604         bool ret = false;
1605         /*
1606          * Unlike task_move routines, we access mc.to, mc.from not under
1607          * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1608          */
1609         spin_lock(&mc.lock);
1610         from = mc.from;
1611         to = mc.to;
1612         if (!from)
1613                 goto unlock;
1614 
1615         ret = mem_cgroup_same_or_subtree(memcg, from)
1616                 || mem_cgroup_same_or_subtree(memcg, to);
1617 unlock:
1618         spin_unlock(&mc.lock);
1619         return ret;
1620 }
1621 
1622 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1623 {
1624         if (mc.moving_task && current != mc.moving_task) {
1625                 if (mem_cgroup_under_move(memcg)) {
1626                         DEFINE_WAIT(wait);
1627                         prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1628                         /* moving charge context might have finished. */
1629                         if (mc.moving_task)
1630                                 schedule();
1631                         finish_wait(&mc.waitq, &wait);
1632                         return true;
1633                 }
1634         }
1635         return false;
1636 }
1637 
1638 /*
1639  * Take this lock when
1640  * - a code tries to modify page's memcg while it's USED.
1641  * - a code tries to modify page state accounting in a memcg.
1642  * see mem_cgroup_stolen(), too.
1643  */
1644 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1645                                   unsigned long *flags)
1646 {
1647         spin_lock_irqsave(&memcg->move_lock, *flags);
1648 }
1649 
1650 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1651                                 unsigned long *flags)
1652 {
1653         spin_unlock_irqrestore(&memcg->move_lock, *flags);
1654 }
1655 
1656 #define K(x) ((x) << (PAGE_SHIFT-10))
1657 /**
1658  * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1659  * @memcg: The memory cgroup that went over limit
1660  * @p: Task that is going to be killed
1661  *
1662  * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1663  * enabled
1664  */
1665 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1666 {
1667         struct cgroup *task_cgrp;
1668         struct cgroup *mem_cgrp;
1669         /*
1670          * Need a buffer in BSS, can't rely on allocations. The code relies
1671          * on the assumption that OOM is serialized for memory controller.
1672          * If this assumption is broken, revisit this code.
1673          */
1674         static char memcg_name[PATH_MAX];
1675         int ret;
1676         struct mem_cgroup *iter;
1677         unsigned int i;
1678 
1679         if (!p)
1680                 return;
1681 
1682         rcu_read_lock();
1683 
1684         mem_cgrp = memcg->css.cgroup;
1685         task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1686 
1687         ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1688         if (ret < 0) {
1689                 /*
1690                  * Unfortunately, we are unable to convert to a useful name
1691                  * But we'll still print out the usage information
1692                  */
1693                 rcu_read_unlock();
1694                 goto done;
1695         }
1696         rcu_read_unlock();
1697 
1698         pr_info("Task in %s killed", memcg_name);
1699 
1700         rcu_read_lock();
1701         ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1702         if (ret < 0) {
1703                 rcu_read_unlock();
1704                 goto done;
1705         }
1706         rcu_read_unlock();
1707 
1708         /*
1709          * Continues from above, so we don't need an KERN_ level
1710          */
1711         pr_cont(" as a result of limit of %s\n", memcg_name);
1712 done:
1713 
1714         pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1715                 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1716                 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1717                 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1718         pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1719                 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1720                 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1721                 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1722         pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1723                 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1724                 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1725                 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1726 
1727         for_each_mem_cgroup_tree(iter, memcg) {
1728                 pr_info("Memory cgroup stats");
1729 
1730                 rcu_read_lock();
1731                 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1732                 if (!ret)
1733                         pr_cont(" for %s", memcg_name);
1734                 rcu_read_unlock();
1735                 pr_cont(":");
1736 
1737                 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1738                         if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1739                                 continue;
1740                         pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1741                                 K(mem_cgroup_read_stat(iter, i)));
1742                 }
1743 
1744                 for (i = 0; i < NR_LRU_LISTS; i++)
1745                         pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1746                                 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1747 
1748                 pr_cont("\n");
1749         }
1750 }
1751 
1752 /*
1753  * This function returns the number of memcg under hierarchy tree. Returns
1754  * 1(self count) if no children.
1755  */
1756 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1757 {
1758         int num = 0;
1759         struct mem_cgroup *iter;
1760 
1761         for_each_mem_cgroup_tree(iter, memcg)
1762                 num++;
1763         return num;
1764 }
1765 
1766 /*
1767  * Return the memory (and swap, if configured) limit for a memcg.
1768  */
1769 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1770 {
1771         u64 limit;
1772 
1773         limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1774 
1775         /*
1776          * Do not consider swap space if we cannot swap due to swappiness
1777          */
1778         if (mem_cgroup_swappiness(memcg)) {
1779                 u64 memsw;
1780 
1781                 limit += total_swap_pages << PAGE_SHIFT;
1782                 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1783 
1784                 /*
1785                  * If memsw is finite and limits the amount of swap space
1786                  * available to this memcg, return that limit.
1787                  */
1788                 limit = min(limit, memsw);
1789         }
1790 
1791         return limit;
1792 }
1793 
1794 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1795                                      int order)
1796 {
1797         struct mem_cgroup *iter;
1798         unsigned long chosen_points = 0;
1799         unsigned long totalpages;
1800         unsigned int points = 0;
1801         struct task_struct *chosen = NULL;
1802 
1803         /*
1804          * If current has a pending SIGKILL or is exiting, then automatically
1805          * select it.  The goal is to allow it to allocate so that it may
1806          * quickly exit and free its memory.
1807          */
1808         if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1809                 set_thread_flag(TIF_MEMDIE);
1810                 return;
1811         }
1812 
1813         check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1814         totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1815         for_each_mem_cgroup_tree(iter, memcg) {
1816                 struct css_task_iter it;
1817                 struct task_struct *task;
1818 
1819                 css_task_iter_start(&iter->css, &it);
1820                 while ((task = css_task_iter_next(&it))) {
1821                         switch (oom_scan_process_thread(task, totalpages, NULL,
1822                                                         false)) {
1823                         case OOM_SCAN_SELECT:
1824                                 if (chosen)
1825                                         put_task_struct(chosen);
1826                                 chosen = task;
1827                                 chosen_points = ULONG_MAX;
1828                                 get_task_struct(chosen);
1829                                 /* fall through */
1830                         case OOM_SCAN_CONTINUE:
1831                                 continue;
1832                         case OOM_SCAN_ABORT:
1833                                 css_task_iter_end(&it);
1834                                 mem_cgroup_iter_break(memcg, iter);
1835                                 if (chosen)
1836                                         put_task_struct(chosen);
1837                                 return;
1838                         case OOM_SCAN_OK:
1839                                 break;
1840                         };
1841                         points = oom_badness(task, memcg, NULL, totalpages);
1842                         if (points > chosen_points) {
1843                                 if (chosen)
1844                                         put_task_struct(chosen);
1845                                 chosen = task;
1846                                 chosen_points = points;
1847                                 get_task_struct(chosen);
1848                         }
1849                 }
1850                 css_task_iter_end(&it);
1851         }
1852 
1853         if (!chosen)
1854                 return;
1855         points = chosen_points * 1000 / totalpages;
1856         oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1857                          NULL, "Memory cgroup out of memory");
1858 }
1859 
1860 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1861                                         gfp_t gfp_mask,
1862                                         unsigned long flags)
1863 {
1864         unsigned long total = 0;
1865         bool noswap = false;
1866         int loop;
1867 
1868         if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1869                 noswap = true;
1870         if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1871                 noswap = true;
1872 
1873         for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1874                 if (loop)
1875                         drain_all_stock_async(memcg);
1876                 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1877                 /*
1878                  * Allow limit shrinkers, which are triggered directly
1879                  * by userspace, to catch signals and stop reclaim
1880                  * after minimal progress, regardless of the margin.
1881                  */
1882                 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1883                         break;
1884                 if (mem_cgroup_margin(memcg))
1885                         break;
1886                 /*
1887                  * If nothing was reclaimed after two attempts, there
1888                  * may be no reclaimable pages in this hierarchy.
1889                  */
1890                 if (loop && !total)
1891                         break;
1892         }
1893         return total;
1894 }
1895 
1896 /**
1897  * test_mem_cgroup_node_reclaimable
1898  * @memcg: the target memcg
1899  * @nid: the node ID to be checked.
1900  * @noswap : specify true here if the user wants flle only information.
1901  *
1902  * This function returns whether the specified memcg contains any
1903  * reclaimable pages on a node. Returns true if there are any reclaimable
1904  * pages in the node.
1905  */
1906 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1907                 int nid, bool noswap)
1908 {
1909         if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1910                 return true;
1911         if (noswap || !total_swap_pages)
1912                 return false;
1913         if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1914                 return true;
1915         return false;
1916 
1917 }
1918 #if MAX_NUMNODES > 1
1919 
1920 /*
1921  * Always updating the nodemask is not very good - even if we have an empty
1922  * list or the wrong list here, we can start from some node and traverse all
1923  * nodes based on the zonelist. So update the list loosely once per 10 secs.
1924  *
1925  */
1926 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1927 {
1928         int nid;
1929         /*
1930          * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1931          * pagein/pageout changes since the last update.
1932          */
1933         if (!atomic_read(&memcg->numainfo_events))
1934                 return;
1935         if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1936                 return;
1937 
1938         /* make a nodemask where this memcg uses memory from */
1939         memcg->scan_nodes = node_states[N_MEMORY];
1940 
1941         for_each_node_mask(nid, node_states[N_MEMORY]) {
1942 
1943                 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1944                         node_clear(nid, memcg->scan_nodes);
1945         }
1946 
1947         atomic_set(&memcg->numainfo_events, 0);
1948         atomic_set(&memcg->numainfo_updating, 0);
1949 }
1950 
1951 /*
1952  * Selecting a node where we start reclaim from. Because what we need is just
1953  * reducing usage counter, start from anywhere is O,K. Considering
1954  * memory reclaim from current node, there are pros. and cons.
1955  *
1956  * Freeing memory from current node means freeing memory from a node which
1957  * we'll use or we've used. So, it may make LRU bad. And if several threads
1958  * hit limits, it will see a contention on a node. But freeing from remote
1959  * node means more costs for memory reclaim because of memory latency.
1960  *
1961  * Now, we use round-robin. Better algorithm is welcomed.
1962  */
1963 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1964 {
1965         int node;
1966 
1967         mem_cgroup_may_update_nodemask(memcg);
1968         node = memcg->last_scanned_node;
1969 
1970         node = next_node(node, memcg->scan_nodes);
1971         if (node == MAX_NUMNODES)
1972                 node = first_node(memcg->scan_nodes);
1973         /*
1974          * We call this when we hit limit, not when pages are added to LRU.
1975          * No LRU may hold pages because all pages are UNEVICTABLE or
1976          * memcg is too small and all pages are not on LRU. In that case,
1977          * we use curret node.
1978          */
1979         if (unlikely(node == MAX_NUMNODES))
1980                 node = numa_node_id();
1981 
1982         memcg->last_scanned_node = node;
1983         return node;
1984 }
1985 
1986 /*
1987  * Check all nodes whether it contains reclaimable pages or not.
1988  * For quick scan, we make use of scan_nodes. This will allow us to skip
1989  * unused nodes. But scan_nodes is lazily updated and may not cotain
1990  * enough new information. We need to do double check.
1991  */
1992 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1993 {
1994         int nid;
1995 
1996         /*
1997          * quick check...making use of scan_node.
1998          * We can skip unused nodes.
1999          */
2000         if (!nodes_empty(memcg->scan_nodes)) {
2001                 for (nid = first_node(memcg->scan_nodes);
2002                      nid < MAX_NUMNODES;
2003                      nid = next_node(nid, memcg->scan_nodes)) {
2004 
2005                         if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2006                                 return true;
2007                 }
2008         }
2009         /*
2010          * Check rest of nodes.
2011          */
2012         for_each_node_state(nid, N_MEMORY) {
2013                 if (node_isset(nid, memcg->scan_nodes))
2014                         continue;
2015                 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2016                         return true;
2017         }
2018         return false;
2019 }
2020 
2021 #else
2022 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2023 {
2024         return 0;
2025 }
2026 
2027 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2028 {
2029         return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2030 }
2031 #endif
2032 
2033 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2034                                    struct zone *zone,
2035                                    gfp_t gfp_mask,
2036                                    unsigned long *total_scanned)
2037 {
2038         struct mem_cgroup *victim = NULL;
2039         int total = 0;
2040         int loop = 0;
2041         unsigned long excess;
2042         unsigned long nr_scanned;
2043         struct mem_cgroup_reclaim_cookie reclaim = {
2044                 .zone = zone,
2045                 .priority = 0,
2046         };
2047 
2048         excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2049 
2050         while (1) {
2051                 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2052                 if (!victim) {
2053                         loop++;
2054                         if (loop >= 2) {
2055                                 /*
2056                                  * If we have not been able to reclaim
2057                                  * anything, it might because there are
2058                                  * no reclaimable pages under this hierarchy
2059                                  */
2060                                 if (!total)
2061                                         break;
2062                                 /*
2063                                  * We want to do more targeted reclaim.
2064                                  * excess >> 2 is not to excessive so as to
2065                                  * reclaim too much, nor too less that we keep
2066                                  * coming back to reclaim from this cgroup
2067                                  */
2068                                 if (total >= (excess >> 2) ||
2069                                         (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2070                                         break;
2071                         }
2072                         continue;
2073                 }
2074                 if (!mem_cgroup_reclaimable(victim, false))
2075                         continue;
2076                 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2077                                                      zone, &nr_scanned);
2078                 *total_scanned += nr_scanned;
2079                 if (!res_counter_soft_limit_excess(&root_memcg->res))
2080                         break;
2081         }
2082         mem_cgroup_iter_break(root_memcg, victim);
2083         return total;
2084 }
2085 
2086 #ifdef CONFIG_LOCKDEP
2087 static struct lockdep_map memcg_oom_lock_dep_map = {
2088         .name = "memcg_oom_lock",
2089 };
2090 #endif
2091 
2092 static DEFINE_SPINLOCK(memcg_oom_lock);
2093 
2094 /*
2095  * Check OOM-Killer is already running under our hierarchy.
2096  * If someone is running, return false.
2097  */
2098 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2099 {
2100         struct mem_cgroup *iter, *failed = NULL;
2101 
2102         spin_lock(&memcg_oom_lock);
2103 
2104         for_each_mem_cgroup_tree(iter, memcg) {
2105                 if (iter->oom_lock) {
2106                         /*
2107                          * this subtree of our hierarchy is already locked
2108                          * so we cannot give a lock.
2109                          */
2110                         failed = iter;
2111                         mem_cgroup_iter_break(memcg, iter);
2112                         break;
2113                 } else
2114                         iter->oom_lock = true;
2115         }
2116 
2117         if (failed) {
2118                 /*
2119                  * OK, we failed to lock the whole subtree so we have
2120                  * to clean up what we set up to the failing subtree
2121                  */
2122                 for_each_mem_cgroup_tree(iter, memcg) {
2123                         if (iter == failed) {
2124                                 mem_cgroup_iter_break(memcg, iter);
2125                                 break;
2126                         }
2127                         iter->oom_lock = false;
2128                 }
2129         } else
2130                 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2131 
2132         spin_unlock(&memcg_oom_lock);
2133 
2134         return !failed;
2135 }
2136 
2137 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2138 {
2139         struct mem_cgroup *iter;
2140 
2141         spin_lock(&memcg_oom_lock);
2142         mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2143         for_each_mem_cgroup_tree(iter, memcg)
2144                 iter->oom_lock = false;
2145         spin_unlock(&memcg_oom_lock);
2146 }
2147 
2148 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2149 {
2150         struct mem_cgroup *iter;
2151 
2152         for_each_mem_cgroup_tree(iter, memcg)
2153                 atomic_inc(&iter->under_oom);
2154 }
2155 
2156 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2157 {
2158         struct mem_cgroup *iter;
2159 
2160         /*
2161          * When a new child is created while the hierarchy is under oom,
2162          * mem_cgroup_oom_lock() may not be called. We have to use
2163          * atomic_add_unless() here.
2164          */
2165         for_each_mem_cgroup_tree(iter, memcg)
2166                 atomic_add_unless(&iter->under_oom, -1, 0);
2167 }
2168 
2169 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2170 
2171 struct oom_wait_info {
2172         struct mem_cgroup *memcg;
2173         wait_queue_t    wait;
2174 };
2175 
2176 static int memcg_oom_wake_function(wait_queue_t *wait,
2177         unsigned mode, int sync, void *arg)
2178 {
2179         struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2180         struct mem_cgroup *oom_wait_memcg;
2181         struct oom_wait_info *oom_wait_info;
2182 
2183         oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2184         oom_wait_memcg = oom_wait_info->memcg;
2185 
2186         /*
2187          * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2188          * Then we can use css_is_ancestor without taking care of RCU.
2189          */
2190         if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2191                 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2192                 return 0;
2193         return autoremove_wake_function(wait, mode, sync, arg);
2194 }
2195 
2196 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2197 {
2198         atomic_inc(&memcg->oom_wakeups);
2199         /* for filtering, pass "memcg" as argument. */
2200         __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2201 }
2202 
2203 static void memcg_oom_recover(struct mem_cgroup *memcg)
2204 {
2205         if (memcg && atomic_read(&memcg->under_oom))
2206                 memcg_wakeup_oom(memcg);
2207 }
2208 
2209 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2210 {
2211         if (!current->memcg_oom.may_oom)
2212                 return;
2213         /*
2214          * We are in the middle of the charge context here, so we
2215          * don't want to block when potentially sitting on a callstack
2216          * that holds all kinds of filesystem and mm locks.
2217          *
2218          * Also, the caller may handle a failed allocation gracefully
2219          * (like optional page cache readahead) and so an OOM killer
2220          * invocation might not even be necessary.
2221          *
2222          * That's why we don't do anything here except remember the
2223          * OOM context and then deal with it at the end of the page
2224          * fault when the stack is unwound, the locks are released,
2225          * and when we know whether the fault was overall successful.
2226          */
2227         css_get(&memcg->css);
2228         current->memcg_oom.memcg = memcg;
2229         current->memcg_oom.gfp_mask = mask;
2230         current->memcg_oom.order = order;
2231 }
2232 
2233 /**
2234  * mem_cgroup_oom_synchronize - complete memcg OOM handling
2235  * @handle: actually kill/wait or just clean up the OOM state
2236  *
2237  * This has to be called at the end of a page fault if the memcg OOM
2238  * handler was enabled.
2239  *
2240  * Memcg supports userspace OOM handling where failed allocations must
2241  * sleep on a waitqueue until the userspace task resolves the
2242  * situation.  Sleeping directly in the charge context with all kinds
2243  * of locks held is not a good idea, instead we remember an OOM state
2244  * in the task and mem_cgroup_oom_synchronize() has to be called at
2245  * the end of the page fault to complete the OOM handling.
2246  *
2247  * Returns %true if an ongoing memcg OOM situation was detected and
2248  * completed, %false otherwise.
2249  */
2250 bool mem_cgroup_oom_synchronize(bool handle)
2251 {
2252         struct mem_cgroup *memcg = current->memcg_oom.memcg;
2253         struct oom_wait_info owait;
2254         bool locked;
2255 
2256         /* OOM is global, do not handle */
2257         if (!memcg)
2258                 return false;
2259 
2260         if (!handle)
2261                 goto cleanup;
2262 
2263         owait.memcg = memcg;
2264         owait.wait.flags = 0;
2265         owait.wait.func = memcg_oom_wake_function;
2266         owait.wait.private = current;
2267         INIT_LIST_HEAD(&owait.wait.task_list);
2268 
2269         prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2270         mem_cgroup_mark_under_oom(memcg);
2271 
2272         locked = mem_cgroup_oom_trylock(memcg);
2273 
2274         if (locked)
2275                 mem_cgroup_oom_notify(memcg);
2276 
2277         if (locked && !memcg->oom_kill_disable) {
2278                 mem_cgroup_unmark_under_oom(memcg);
2279                 finish_wait(&memcg_oom_waitq, &owait.wait);
2280                 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2281                                          current->memcg_oom.order);
2282         } else {
2283                 schedule();
2284                 mem_cgroup_unmark_under_oom(memcg);
2285                 finish_wait(&memcg_oom_waitq, &owait.wait);
2286         }
2287 
2288         if (locked) {
2289                 mem_cgroup_oom_unlock(memcg);
2290                 /*
2291                  * There is no guarantee that an OOM-lock contender
2292                  * sees the wakeups triggered by the OOM kill
2293                  * uncharges.  Wake any sleepers explicitely.
2294                  */
2295                 memcg_oom_recover(memcg);
2296         }
2297 cleanup:
2298         current->memcg_oom.memcg = NULL;
2299         css_put(&memcg->css);
2300         return true;
2301 }
2302 
2303 /*
2304  * Currently used to update mapped file statistics, but the routine can be
2305  * generalized to update other statistics as well.
2306  *
2307  * Notes: Race condition
2308  *
2309  * We usually use page_cgroup_lock() for accessing page_cgroup member but
2310  * it tends to be costly. But considering some conditions, we doesn't need
2311  * to do so _always_.
2312  *
2313  * Considering "charge", lock_page_cgroup() is not required because all
2314  * file-stat operations happen after a page is attached to radix-tree. There
2315  * are no race with "charge".
2316  *
2317  * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2318  * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2319  * if there are race with "uncharge". Statistics itself is properly handled
2320  * by flags.
2321  *
2322  * Considering "move", this is an only case we see a race. To make the race
2323  * small, we check mm->moving_account and detect there are possibility of race
2324  * If there is, we take a lock.
2325  */
2326 
2327 void __mem_cgroup_begin_update_page_stat(struct page *page,
2328                                 bool *locked, unsigned long *flags)
2329 {
2330         struct mem_cgroup *memcg;
2331         struct page_cgroup *pc;
2332 
2333         pc = lookup_page_cgroup(page);
2334 again:
2335         memcg = pc->mem_cgroup;
2336         if (unlikely(!memcg || !PageCgroupUsed(pc)))
2337                 return;
2338         /*
2339          * If this memory cgroup is not under account moving, we don't
2340          * need to take move_lock_mem_cgroup(). Because we already hold
2341          * rcu_read_lock(), any calls to move_account will be delayed until
2342          * rcu_read_unlock() if mem_cgroup_stolen() == true.
2343          */
2344         if (!mem_cgroup_stolen(memcg))
2345                 return;
2346 
2347         move_lock_mem_cgroup(memcg, flags);
2348         if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2349                 move_unlock_mem_cgroup(memcg, flags);
2350                 goto again;
2351         }
2352         *locked = true;
2353 }
2354 
2355 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2356 {
2357         struct page_cgroup *pc = lookup_page_cgroup(page);
2358 
2359         /*
2360          * It's guaranteed that pc->mem_cgroup never changes while
2361          * lock is held because a routine modifies pc->mem_cgroup
2362          * should take move_lock_mem_cgroup().
2363          */
2364         move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2365 }
2366 
2367 void mem_cgroup_update_page_stat(struct page *page,
2368                                  enum mem_cgroup_stat_index idx, int val)
2369 {
2370         struct mem_cgroup *memcg;
2371         struct page_cgroup *pc = lookup_page_cgroup(page);
2372         unsigned long uninitialized_var(flags);
2373 
2374         if (mem_cgroup_disabled())
2375                 return;
2376 
2377         VM_BUG_ON(!rcu_read_lock_held());
2378         memcg = pc->mem_cgroup;
2379         if (unlikely(!memcg || !PageCgroupUsed(pc)))
2380                 return;
2381 
2382         this_cpu_add(memcg->stat->count[idx], val);
2383 }
2384 
2385 /*
2386  * size of first charge trial. "32" comes from vmscan.c's magic value.
2387  * TODO: maybe necessary to use big numbers in big irons.
2388  */
2389 #define CHARGE_BATCH    32U
2390 struct memcg_stock_pcp {
2391         struct mem_cgroup *cached; /* this never be root cgroup */
2392         unsigned int nr_pages;
2393         struct work_struct work;
2394         unsigned long flags;
2395 #define FLUSHING_CACHED_CHARGE  0
2396 };
2397 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2398 static DEFINE_MUTEX(percpu_charge_mutex);
2399 
2400 /**
2401  * consume_stock: Try to consume stocked charge on this cpu.
2402  * @memcg: memcg to consume from.
2403  * @nr_pages: how many pages to charge.
2404  *
2405  * The charges will only happen if @memcg matches the current cpu's memcg
2406  * stock, and at least @nr_pages are available in that stock.  Failure to
2407  * service an allocation will refill the stock.
2408  *
2409  * returns true if successful, false otherwise.
2410  */
2411 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2412 {
2413         struct memcg_stock_pcp *stock;
2414         bool ret = true;
2415 
2416         if (nr_pages > CHARGE_BATCH)
2417                 return false;
2418 
2419         stock = &get_cpu_var(memcg_stock);
2420         if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2421                 stock->nr_pages -= nr_pages;
2422         else /* need to call res_counter_charge */
2423                 ret = false;
2424         put_cpu_var(memcg_stock);
2425         return ret;
2426 }
2427 
2428 /*
2429  * Returns stocks cached in percpu to res_counter and reset cached information.
2430  */
2431 static void drain_stock(struct memcg_stock_pcp *stock)
2432 {
2433         struct mem_cgroup *old = stock->cached;
2434 
2435         if (stock->nr_pages) {
2436                 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2437 
2438                 res_counter_uncharge(&old->res, bytes);
2439                 if (do_swap_account)
2440                         res_counter_uncharge(&old->memsw, bytes);
2441                 stock->nr_pages = 0;
2442         }
2443         stock->cached = NULL;
2444 }
2445 
2446 /*
2447  * This must be called under preempt disabled or must be called by
2448  * a thread which is pinned to local cpu.
2449  */
2450 static void drain_local_stock(struct work_struct *dummy)
2451 {
2452         struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2453         drain_stock(stock);
2454         clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2455 }
2456 
2457 static void __init memcg_stock_init(void)
2458 {
2459         int cpu;
2460 
2461         for_each_possible_cpu(cpu) {
2462                 struct memcg_stock_pcp *stock =
2463                                         &per_cpu(memcg_stock, cpu);
2464                 INIT_WORK(&stock->work, drain_local_stock);
2465         }
2466 }
2467 
2468 /*
2469  * Cache charges(val) which is from res_counter, to local per_cpu area.
2470  * This will be consumed by consume_stock() function, later.
2471  */
2472 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2473 {
2474         struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2475 
2476         if (stock->cached != memcg) { /* reset if necessary */
2477                 drain_stock(stock);
2478                 stock->cached = memcg;
2479         }
2480         stock->nr_pages += nr_pages;
2481         put_cpu_var(memcg_stock);
2482 }
2483 
2484 /*
2485  * Drains all per-CPU charge caches for given root_memcg resp. subtree
2486  * of the hierarchy under it. sync flag says whether we should block
2487  * until the work is done.
2488  */
2489 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2490 {
2491         int cpu, curcpu;
2492 
2493         /* Notify other cpus that system-wide "drain" is running */
2494         get_online_cpus();
2495         curcpu = get_cpu();
2496         for_each_online_cpu(cpu) {
2497                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2498                 struct mem_cgroup *memcg;
2499 
2500                 memcg = stock->cached;
2501                 if (!memcg || !stock->nr_pages)
2502                         continue;
2503                 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2504                         continue;
2505                 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2506                         if (cpu == curcpu)
2507                                 drain_local_stock(&stock->work);
2508                         else
2509                                 schedule_work_on(cpu, &stock->work);
2510                 }
2511         }
2512         put_cpu();
2513 
2514         if (!sync)
2515                 goto out;
2516 
2517         for_each_online_cpu(cpu) {
2518                 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2519                 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2520                         flush_work(&stock->work);
2521         }
2522 out:
2523         put_online_cpus();
2524 }
2525 
2526 /*
2527  * Tries to drain stocked charges in other cpus. This function is asynchronous
2528  * and just put a work per cpu for draining localy on each cpu. Caller can
2529  * expects some charges will be back to res_counter later but cannot wait for
2530  * it.
2531  */
2532 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2533 {
2534         /*
2535          * If someone calls draining, avoid adding more kworker runs.
2536          */
2537         if (!mutex_trylock(&percpu_charge_mutex))
2538                 return;
2539         drain_all_stock(root_memcg, false);
2540         mutex_unlock(&percpu_charge_mutex);
2541 }
2542 
2543 /* This is a synchronous drain interface. */
2544 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2545 {
2546         /* called when force_empty is called */
2547         mutex_lock(&percpu_charge_mutex);
2548         drain_all_stock(root_memcg, true);
2549         mutex_unlock(&percpu_charge_mutex);
2550 }
2551 
2552 /*
2553  * This function drains percpu counter value from DEAD cpu and
2554  * move it to local cpu. Note that this function can be preempted.
2555  */
2556 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2557 {
2558         int i;
2559 
2560         spin_lock(&memcg->pcp_counter_lock);
2561         for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2562                 long x = per_cpu(memcg->stat->count[i], cpu);
2563 
2564                 per_cpu(memcg->stat->count[i], cpu) = 0;
2565                 memcg->nocpu_base.count[i] += x;
2566         }
2567         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2568                 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2569 
2570                 per_cpu(memcg->stat->events[i], cpu) = 0;
2571                 memcg->nocpu_base.events[i] += x;
2572         }
2573         spin_unlock(&memcg->pcp_counter_lock);
2574 }
2575 
2576 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2577                                         unsigned long action,
2578                                         void *hcpu)
2579 {
2580         int cpu = (unsigned long)hcpu;
2581         struct memcg_stock_pcp *stock;
2582         struct mem_cgroup *iter;
2583 
2584         if (action == CPU_ONLINE)
2585                 return NOTIFY_OK;
2586 
2587         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2588                 return NOTIFY_OK;
2589 
2590         for_each_mem_cgroup(iter)
2591                 mem_cgroup_drain_pcp_counter(iter, cpu);
2592 
2593         stock = &per_cpu(memcg_stock, cpu);
2594         drain_stock(stock);
2595         return NOTIFY_OK;
2596 }
2597 
2598 
2599 /* See __mem_cgroup_try_charge() for details */
2600 enum {
2601         CHARGE_OK,              /* success */
2602         CHARGE_RETRY,           /* need to retry but retry is not bad */
2603         CHARGE_NOMEM,           /* we can't do more. return -ENOMEM */
2604         CHARGE_WOULDBLOCK,      /* GFP_WAIT wasn't set and no enough res. */
2605 };
2606 
2607 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2608                                 unsigned int nr_pages, unsigned int min_pages,
2609                                 bool invoke_oom)
2610 {
2611         unsigned long csize = nr_pages * PAGE_SIZE;
2612         struct mem_cgroup *mem_over_limit;
2613         struct res_counter *fail_res;
2614         unsigned long flags = 0;
2615         int ret;
2616 
2617         ret = res_counter_charge(&memcg->res, csize, &fail_res);
2618 
2619         if (likely(!ret)) {
2620                 if (!do_swap_account)
2621                         return CHARGE_OK;
2622                 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2623                 if (likely(!ret))
2624                         return CHARGE_OK;
2625 
2626                 res_counter_uncharge(&memcg->res, csize);
2627                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2628                 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2629         } else
2630                 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2631         /*
2632          * Never reclaim on behalf of optional batching, retry with a
2633          * single page instead.
2634          */
2635         if (nr_pages > min_pages)
2636                 return CHARGE_RETRY;
2637 
2638         if (!(gfp_mask & __GFP_WAIT))
2639                 return CHARGE_WOULDBLOCK;
2640 
2641         if (gfp_mask & __GFP_NORETRY)
2642                 return CHARGE_NOMEM;
2643 
2644         ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2645         if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2646                 return CHARGE_RETRY;
2647         /*
2648          * Even though the limit is exceeded at this point, reclaim
2649          * may have been able to free some pages.  Retry the charge
2650          * before killing the task.
2651          *
2652          * Only for regular pages, though: huge pages are rather
2653          * unlikely to succeed so close to the limit, and we fall back
2654          * to regular pages anyway in case of failure.
2655          */
2656         if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2657                 return CHARGE_RETRY;
2658 
2659         /*
2660          * At task move, charge accounts can be doubly counted. So, it's
2661          * better to wait until the end of task_move if something is going on.
2662          */
2663         if (mem_cgroup_wait_acct_move(mem_over_limit))
2664                 return CHARGE_RETRY;
2665 
2666         if (invoke_oom)
2667                 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2668 
2669         return CHARGE_NOMEM;
2670 }
2671 
2672 /*
2673  * __mem_cgroup_try_charge() does
2674  * 1. detect memcg to be charged against from passed *mm and *ptr,
2675  * 2. update res_counter
2676  * 3. call memory reclaim if necessary.
2677  *
2678  * In some special case, if the task is fatal, fatal_signal_pending() or
2679  * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2680  * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2681  * as possible without any hazards. 2: all pages should have a valid
2682  * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2683  * pointer, that is treated as a charge to root_mem_cgroup.
2684  *
2685  * So __mem_cgroup_try_charge() will return
2686  *  0       ...  on success, filling *ptr with a valid memcg pointer.
2687  *  -ENOMEM ...  charge failure because of resource limits.
2688  *  -EINTR  ...  if thread is fatal. *ptr is filled with root_mem_cgroup.
2689  *
2690  * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2691  * the oom-killer can be invoked.
2692  */
2693 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2694                                    gfp_t gfp_mask,
2695                                    unsigned int nr_pages,
2696                                    struct mem_cgroup **ptr,
2697                                    bool oom)
2698 {
2699         unsigned int batch = max(CHARGE_BATCH, nr_pages);
2700         int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2701         struct mem_cgroup *memcg = NULL;
2702         int ret;
2703 
2704         /*
2705          * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2706          * in system level. So, allow to go ahead dying process in addition to
2707          * MEMDIE process.
2708          */
2709         if (unlikely(test_thread_flag(TIF_MEMDIE)
2710                      || fatal_signal_pending(current)))
2711                 goto bypass;
2712 
2713         if (unlikely(task_in_memcg_oom(current)))
2714                 goto nomem;
2715 
2716         if (gfp_mask & __GFP_NOFAIL)
2717                 oom = false;
2718 
2719         /*
2720          * We always charge the cgroup the mm_struct belongs to.
2721          * The mm_struct's mem_cgroup changes on task migration if the
2722          * thread group leader migrates. It's possible that mm is not
2723          * set, if so charge the root memcg (happens for pagecache usage).
2724          */
2725         if (!*ptr && !mm)
2726                 *ptr = root_mem_cgroup;
2727 again:
2728         if (*ptr) { /* css should be a valid one */
2729                 memcg = *ptr;
2730                 if (mem_cgroup_is_root(memcg))
2731                         goto done;
2732                 if (consume_stock(memcg, nr_pages))
2733                         goto done;
2734                 css_get(&memcg->css);
2735         } else {
2736                 struct task_struct *p;
2737 
2738                 rcu_read_lock();
2739                 p = rcu_dereference(mm->owner);
2740                 /*
2741                  * Because we don't have task_lock(), "p" can exit.
2742                  * In that case, "memcg" can point to root or p can be NULL with
2743                  * race with swapoff. Then, we have small risk of mis-accouning.
2744                  * But such kind of mis-account by race always happens because
2745                  * we don't have cgroup_mutex(). It's overkill and we allo that
2746                  * small race, here.
2747                  * (*) swapoff at el will charge against mm-struct not against
2748                  * task-struct. So, mm->owner can be NULL.
2749                  */
2750                 memcg = mem_cgroup_from_task(p);
2751                 if (!memcg)
2752                         memcg = root_mem_cgroup;
2753                 if (mem_cgroup_is_root(memcg)) {
2754                         rcu_read_unlock();
2755                         goto done;
2756                 }
2757                 if (consume_stock(memcg, nr_pages)) {
2758                         /*
2759                          * It seems dagerous to access memcg without css_get().
2760                          * But considering how consume_stok works, it's not
2761                          * necessary. If consume_stock success, some charges
2762                          * from this memcg are cached on this cpu. So, we
2763                          * don't need to call css_get()/css_tryget() before
2764                          * calling consume_stock().
2765                          */
2766                         rcu_read_unlock();
2767                         goto done;
2768                 }
2769                 /* after here, we may be blocked. we need to get refcnt */
2770                 if (!css_tryget(&memcg->css)) {
2771                         rcu_read_unlock();
2772                         goto again;
2773                 }
2774                 rcu_read_unlock();
2775         }
2776 
2777         do {
2778                 bool invoke_oom = oom && !nr_oom_retries;
2779 
2780                 /* If killed, bypass charge */
2781                 if (fatal_signal_pending(current)) {
2782                         css_put(&memcg->css);
2783                         goto bypass;
2784                 }
2785 
2786                 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2787                                            nr_pages, invoke_oom);
2788                 switch (ret) {
2789                 case CHARGE_OK:
2790                         break;
2791                 case CHARGE_RETRY: /* not in OOM situation but retry */
2792                         batch = nr_pages;
2793                         css_put(&memcg->css);
2794                         memcg = NULL;
2795                         goto again;
2796                 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2797                         css_put(&memcg->css);
2798                         goto nomem;
2799                 case CHARGE_NOMEM: /* OOM routine works */
2800                         if (!oom || invoke_oom) {
2801                                 css_put(&memcg->css);
2802                                 goto nomem;
2803                         }
2804                         nr_oom_retries--;
2805                         break;
2806                 }
2807         } while (ret != CHARGE_OK);
2808 
2809         if (batch > nr_pages)
2810                 refill_stock(memcg, batch - nr_pages);
2811         css_put(&memcg->css);
2812 done:
2813         *ptr = memcg;
2814         return 0;
2815 nomem:
2816         if (!(gfp_mask & __GFP_NOFAIL)) {
2817                 *ptr = NULL;
2818                 return -ENOMEM;
2819         }
2820 bypass:
2821         *ptr = root_mem_cgroup;
2822         return -EINTR;
2823 }
2824 
2825 /*
2826  * Somemtimes we have to undo a charge we got by try_charge().
2827  * This function is for that and do uncharge, put css's refcnt.
2828  * gotten by try_charge().
2829  */
2830 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2831                                        unsigned int nr_pages)
2832 {
2833         if (!mem_cgroup_is_root(memcg)) {
2834                 unsigned long bytes = nr_pages * PAGE_SIZE;
2835 
2836                 res_counter_uncharge(&memcg->res, bytes);
2837                 if (do_swap_account)
2838                         res_counter_uncharge(&memcg->memsw, bytes);
2839         }
2840 }
2841 
2842 /*
2843  * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2844  * This is useful when moving usage to parent cgroup.
2845  */
2846 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2847                                         unsigned int nr_pages)
2848 {
2849         unsigned long bytes = nr_pages * PAGE_SIZE;
2850 
2851         if (mem_cgroup_is_root(memcg))
2852                 return;
2853 
2854         res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2855         if (do_swap_account)
2856                 res_counter_uncharge_until(&memcg->memsw,
2857                                                 memcg->memsw.parent, bytes);
2858 }
2859 
2860 /*
2861  * A helper function to get mem_cgroup from ID. must be called under
2862  * rcu_read_lock().  The caller is responsible for calling css_tryget if
2863  * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2864  * called against removed memcg.)
2865  */
2866 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2867 {
2868         /* ID 0 is unused ID */
2869         if (!id)
2870                 return NULL;
2871         return mem_cgroup_from_id(id);
2872 }
2873 
2874 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2875 {
2876         struct mem_cgroup *memcg = NULL;
2877         struct page_cgroup *pc;
2878         unsigned short id;
2879         swp_entry_t ent;
2880 
2881         VM_BUG_ON(!PageLocked(page));
2882 
2883         pc = lookup_page_cgroup(page);
2884         lock_page_cgroup(pc);
2885         if (PageCgroupUsed(pc)) {
2886                 memcg = pc->mem_cgroup;
2887                 if (memcg && !css_tryget(&memcg->css))
2888                         memcg = NULL;
2889         } else if (PageSwapCache(page)) {
2890                 ent.val = page_private(page);
2891                 id = lookup_swap_cgroup_id(ent);
2892                 rcu_read_lock();
2893                 memcg = mem_cgroup_lookup(id);
2894                 if (memcg && !css_tryget(&memcg->css))
2895                         memcg = NULL;
2896                 rcu_read_unlock();
2897         }
2898         unlock_page_cgroup(pc);
2899         return memcg;
2900 }
2901 
2902 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2903                                        struct page *page,
2904                                        unsigned int nr_pages,
2905                                        enum charge_type ctype,
2906                                        bool lrucare)
2907 {
2908         struct page_cgroup *pc = lookup_page_cgroup(page);
2909         struct zone *uninitialized_var(zone);
2910         struct lruvec *lruvec;
2911         bool was_on_lru = false;
2912         bool anon;
2913 
2914         lock_page_cgroup(pc);
2915         VM_BUG_ON(PageCgroupUsed(pc));
2916         /*
2917          * we don't need page_cgroup_lock about tail pages, becase they are not
2918          * accessed by any other context at this point.
2919          */
2920 
2921         /*
2922          * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2923          * may already be on some other mem_cgroup's LRU.  Take care of it.
2924          */
2925         if (lrucare) {
2926                 zone = page_zone(page);
2927                 spin_lock_irq(&zone->lru_lock);
2928                 if (PageLRU(page)) {
2929                         lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2930                         ClearPageLRU(page);
2931                         del_page_from_lru_list(page, lruvec, page_lru(page));
2932                         was_on_lru = true;
2933                 }
2934         }
2935 
2936         pc->mem_cgroup = memcg;
2937         /*
2938          * We access a page_cgroup asynchronously without lock_page_cgroup().
2939          * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2940          * is accessed after testing USED bit. To make pc->mem_cgroup visible
2941          * before USED bit, we need memory barrier here.
2942          * See mem_cgroup_add_lru_list(), etc.
2943          */
2944         smp_wmb();
2945         SetPageCgroupUsed(pc);
2946 
2947         if (lrucare) {
2948                 if (was_on_lru) {
2949                         lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2950                         VM_BUG_ON(PageLRU(page));
2951                         SetPageLRU(page);
2952                         add_page_to_lru_list(page, lruvec, page_lru(page));
2953                 }
2954                 spin_unlock_irq(&zone->lru_lock);
2955         }
2956 
2957         if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2958                 anon = true;
2959         else
2960                 anon = false;
2961 
2962         mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2963         unlock_page_cgroup(pc);
2964 
2965         /*
2966          * "charge_statistics" updated event counter. Then, check it.
2967          * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2968          * if they exceeds softlimit.
2969          */
2970         memcg_check_events(memcg, page);
2971 }
2972 
2973 static DEFINE_MUTEX(set_limit_mutex);
2974 
2975 #ifdef CONFIG_MEMCG_KMEM
2976 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2977 {
2978         return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2979                 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2980 }
2981 
2982 /*
2983  * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2984  * in the memcg_cache_params struct.
2985  */
2986 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2987 {
2988         struct kmem_cache *cachep;
2989 
2990         VM_BUG_ON(p->is_root_cache);
2991         cachep = p->root_cache;
2992         return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2993 }
2994 
2995 #ifdef CONFIG_SLABINFO
2996 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2997                                     struct cftype *cft, struct seq_file *m)
2998 {
2999         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3000         struct memcg_cache_params *params;
3001 
3002         if (!memcg_can_account_kmem(memcg))
3003                 return -EIO;
3004 
3005         print_slabinfo_header(m);
3006 
3007         mutex_lock(&memcg->slab_caches_mutex);
3008         list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3009                 cache_show(memcg_params_to_cache(params), m);
3010         mutex_unlock(&memcg->slab_caches_mutex);
3011 
3012         return 0;
3013 }
3014 #endif
3015 
3016 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3017 {
3018         struct res_counter *fail_res;
3019         struct mem_cgroup *_memcg;
3020         int ret = 0;
3021 
3022         ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3023         if (ret)
3024                 return ret;
3025 
3026         _memcg = memcg;
3027         ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3028                                       &_memcg, oom_gfp_allowed(gfp));
3029 
3030         if (ret == -EINTR)  {
3031                 /*
3032                  * __mem_cgroup_try_charge() chosed to bypass to root due to
3033                  * OOM kill or fatal signal.  Since our only options are to
3034                  * either fail the allocation or charge it to this cgroup, do
3035                  * it as a temporary condition. But we can't fail. From a
3036                  * kmem/slab perspective, the cache has already been selected,
3037                  * by mem_cgroup_kmem_get_cache(), so it is too late to change
3038                  * our minds.
3039                  *
3040                  * This condition will only trigger if the task entered
3041                  * memcg_charge_kmem in a sane state, but was OOM-killed during
3042                  * __mem_cgroup_try_charge() above. Tasks that were already
3043                  * dying when the allocation triggers should have been already
3044                  * directed to the root cgroup in memcontrol.h
3045                  */
3046                 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3047                 if (do_swap_account)
3048                         res_counter_charge_nofail(&memcg->memsw, size,
3049                                                   &fail_res);
3050                 ret = 0;
3051         } else if (ret)
3052                 res_counter_uncharge(&memcg->kmem, size);
3053 
3054         return ret;
3055 }
3056 
3057 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3058 {
3059         res_counter_uncharge(&memcg->res, size);
3060         if (do_swap_account)
3061                 res_counter_uncharge(&memcg->memsw, size);
3062 
3063         /* Not down to 0 */
3064         if (res_counter_uncharge(&memcg->kmem, size))
3065                 return;
3066 
3067         /*
3068          * Releases a reference taken in kmem_cgroup_css_offline in case
3069          * this last uncharge is racing with the offlining code or it is
3070          * outliving the memcg existence.
3071          *
3072          * The memory barrier imposed by test&clear is paired with the
3073          * explicit one in memcg_kmem_mark_dead().
3074          */
3075         if (memcg_kmem_test_and_clear_dead(memcg))
3076                 css_put(&memcg->css);
3077 }
3078 
3079 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3080 {
3081         if (!memcg)
3082                 return;
3083 
3084         mutex_lock(&memcg->slab_caches_mutex);
3085         list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3086         mutex_unlock(&memcg->slab_caches_mutex);
3087 }
3088 
3089 /*
3090  * helper for acessing a memcg's index. It will be used as an index in the
3091  * child cache array in kmem_cache, and also to derive its name. This function
3092  * will return -1 when this is not a kmem-limited memcg.
3093  */
3094 int memcg_cache_id(struct mem_cgroup *memcg)
3095 {
3096         return memcg ? memcg->kmemcg_id : -1;
3097 }
3098 
3099 /*
3100  * This ends up being protected by the set_limit mutex, during normal
3101  * operation, because that is its main call site.
3102  *
3103  * But when we create a new cache, we can call this as well if its parent
3104  * is kmem-limited. That will have to hold set_limit_mutex as well.
3105  */
3106 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3107 {
3108         int num, ret;
3109 
3110         num = ida_simple_get(&kmem_limited_groups,
3111                                 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3112         if (num < 0)
3113                 return num;
3114         /*
3115          * After this point, kmem_accounted (that we test atomically in
3116          * the beginning of this conditional), is no longer 0. This
3117          * guarantees only one process will set the following boolean
3118          * to true. We don't need test_and_set because we're protected
3119          * by the set_limit_mutex anyway.
3120          */
3121         memcg_kmem_set_activated(memcg);
3122 
3123         ret = memcg_update_all_caches(num+1);
3124         if (ret) {
3125                 ida_simple_remove(&kmem_limited_groups, num);
3126                 memcg_kmem_clear_activated(memcg);
3127                 return ret;
3128         }
3129 
3130         memcg->kmemcg_id = num;
3131         INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3132         mutex_init(&memcg->slab_caches_mutex);
3133         return 0;
3134 }
3135 
3136 static size_t memcg_caches_array_size(int num_groups)
3137 {
3138         ssize_t size;
3139         if (num_groups <= 0)
3140                 return 0;
3141 
3142         size = 2 * num_groups;
3143         if (size < MEMCG_CACHES_MIN_SIZE)
3144                 size = MEMCG_CACHES_MIN_SIZE;
3145         else if (size > MEMCG_CACHES_MAX_SIZE)
3146                 size = MEMCG_CACHES_MAX_SIZE;
3147 
3148         return size;
3149 }
3150 
3151 /*
3152  * We should update the current array size iff all caches updates succeed. This
3153  * can only be done from the slab side. The slab mutex needs to be held when
3154  * calling this.
3155  */
3156 void memcg_update_array_size(int num)
3157 {
3158         if (num > memcg_limited_groups_array_size)
3159                 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3160 }
3161 
3162 static void kmem_cache_destroy_work_func(struct work_struct *w);
3163 
3164 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3165 {
3166         struct memcg_cache_params *cur_params = s->memcg_params;
3167 
3168         VM_BUG_ON(!is_root_cache(s));
3169 
3170         if (num_groups > memcg_limited_groups_array_size) {
3171                 int i;
3172                 ssize_t size = memcg_caches_array_size(num_groups);
3173 
3174                 size *= sizeof(void *);
3175                 size += offsetof(struct memcg_cache_params, memcg_caches);
3176 
3177                 s->memcg_params = kzalloc(size, GFP_KERNEL);
3178                 if (!s->memcg_params) {
3179                         s->memcg_params = cur_params;
3180                         return -ENOMEM;
3181                 }
3182 
3183                 s->memcg_params->is_root_cache = true;
3184 
3185                 /*
3186                  * There is the chance it will be bigger than
3187                  * memcg_limited_groups_array_size, if we failed an allocation
3188                  * in a cache, in which case all caches updated before it, will
3189                  * have a bigger array.
3190                  *
3191                  * But if that is the case, the data after
3192                  * memcg_limited_groups_array_size is certainly unused
3193                  */
3194                 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3195                         if (!cur_params->memcg_caches[i])
3196                                 continue;
3197                         s->memcg_params->memcg_caches[i] =
3198                                                 cur_params->memcg_caches[i];
3199                 }
3200 
3201                 /*
3202                  * Ideally, we would wait until all caches succeed, and only
3203                  * then free the old one. But this is not worth the extra
3204                  * pointer per-cache we'd have to have for this.
3205                  *
3206                  * It is not a big deal if some caches are left with a size
3207                  * bigger than the others. And all updates will reset this
3208                  * anyway.
3209                  */
3210                 kfree(cur_params);
3211         }
3212         return 0;
3213 }
3214 
3215 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3216                          struct kmem_cache *root_cache)
3217 {
3218         size_t size;
3219 
3220         if (!memcg_kmem_enabled())
3221                 return 0;
3222 
3223         if (!memcg) {
3224                 size = offsetof(struct memcg_cache_params, memcg_caches);
3225                 size += memcg_limited_groups_array_size * sizeof(void *);
3226         } else
3227                 size = sizeof(struct memcg_cache_params);
3228 
3229         s->memcg_params = kzalloc(size, GFP_KERNEL);
3230         if (!s->memcg_params)
3231                 return -ENOMEM;
3232 
3233         if (memcg) {
3234                 s->memcg_params->memcg = memcg;
3235                 s->memcg_params->root_cache = root_cache;
3236                 INIT_WORK(&s->memcg_params->destroy,
3237                                 kmem_cache_destroy_work_func);
3238         } else
3239                 s->memcg_params->is_root_cache = true;
3240 
3241         return 0;
3242 }
3243 
3244 void memcg_release_cache(struct kmem_cache *s)
3245 {
3246         struct kmem_cache *root;
3247         struct mem_cgroup *memcg;
3248         int id;
3249 
3250         /*
3251          * This happens, for instance, when a root cache goes away before we
3252          * add any memcg.
3253          */
3254         if (!s->memcg_params)
3255                 return;
3256 
3257         if (s->memcg_params->is_root_cache)
3258                 goto out;
3259 
3260         memcg = s->memcg_params->memcg;
3261         id  = memcg_cache_id(memcg);
3262 
3263         root = s->memcg_params->root_cache;
3264         root->memcg_params->memcg_caches[id] = NULL;
3265 
3266         mutex_lock(&memcg->slab_caches_mutex);
3267         list_del(&s->memcg_params->list);
3268         mutex_unlock(&memcg->slab_caches_mutex);
3269 
3270         css_put(&memcg->css);
3271 out:
3272         kfree(s->memcg_params);
3273 }
3274 
3275 /*
3276  * During the creation a new cache, we need to disable our accounting mechanism
3277  * altogether. This is true even if we are not creating, but rather just
3278  * enqueing new caches to be created.
3279  *
3280  * This is because that process will trigger allocations; some visible, like
3281  * explicit kmallocs to auxiliary data structures, name strings and internal
3282  * cache structures; some well concealed, like INIT_WORK() that can allocate
3283  * objects during debug.
3284  *
3285  * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3286  * to it. This may not be a bounded recursion: since the first cache creation
3287  * failed to complete (waiting on the allocation), we'll just try to create the
3288  * cache again, failing at the same point.
3289  *
3290  * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3291  * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3292  * inside the following two functions.
3293  */
3294 static inline void memcg_stop_kmem_account(void)
3295 {
3296         VM_BUG_ON(!current->mm);
3297         current->memcg_kmem_skip_account++;
3298 }
3299 
3300 static inline void memcg_resume_kmem_account(void)
3301 {
3302         VM_BUG_ON(!current->mm);
3303         current->memcg_kmem_skip_account--;
3304 }
3305 
3306 static void kmem_cache_destroy_work_func(struct work_struct *w)
3307 {
3308         struct kmem_cache *cachep;
3309         struct memcg_cache_params *p;
3310 
3311         p = container_of(w, struct memcg_cache_params, destroy);
3312 
3313         cachep = memcg_params_to_cache(p);
3314 
3315         /*
3316          * If we get down to 0 after shrink, we could delete right away.
3317          * However, memcg_release_pages() already puts us back in the workqueue
3318          * in that case. If we proceed deleting, we'll get a dangling
3319          * reference, and removing the object from the workqueue in that case
3320          * is unnecessary complication. We are not a fast path.
3321          *
3322          * Note that this case is fundamentally different from racing with
3323          * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3324          * kmem_cache_shrink, not only we would be reinserting a dead cache
3325          * into the queue, but doing so from inside the worker racing to
3326          * destroy it.
3327          *
3328          * So if we aren't down to zero, we'll just schedule a worker and try
3329          * again
3330          */
3331         if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3332                 kmem_cache_shrink(cachep);
3333                 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3334                         return;
3335         } else
3336                 kmem_cache_destroy(cachep);
3337 }
3338 
3339 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3340 {
3341         if (!cachep->memcg_params->dead)
3342                 return;
3343 
3344         /*
3345          * There are many ways in which we can get here.
3346          *
3347          * We can get to a memory-pressure situation while the delayed work is
3348          * still pending to run. The vmscan shrinkers can then release all
3349          * cache memory and get us to destruction. If this is the case, we'll
3350          * be executed twice, which is a bug (the second time will execute over
3351          * bogus data). In this case, cancelling the work should be fine.
3352          *
3353          * But we can also get here from the worker itself, if
3354          * kmem_cache_shrink is enough to shake all the remaining objects and
3355          * get the page count to 0. In this case, we'll deadlock if we try to
3356          * cancel the work (the worker runs with an internal lock held, which
3357          * is the same lock we would hold for cancel_work_sync().)
3358          *
3359          * Since we can't possibly know who got us here, just refrain from
3360          * running if there is already work pending
3361          */
3362         if (work_pending(&cachep->memcg_params->destroy))
3363                 return;
3364         /*
3365          * We have to defer the actual destroying to a workqueue, because
3366          * we might currently be in a context that cannot sleep.
3367          */
3368         schedule_work(&cachep->memcg_params->destroy);
3369 }
3370 
3371 /*
3372  * This lock protects updaters, not readers. We want readers to be as fast as
3373  * they can, and they will either see NULL or a valid cache value. Our model
3374  * allow them to see NULL, in which case the root memcg will be selected.
3375  *
3376  * We need this lock because multiple allocations to the same cache from a non
3377  * will span more than one worker. Only one of them can create the cache.
3378  */
3379 static DEFINE_MUTEX(memcg_cache_mutex);
3380 
3381 /*
3382  * Called with memcg_cache_mutex held
3383  */
3384 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3385                                          struct kmem_cache *s)
3386 {
3387         struct kmem_cache *new;
3388         static char *tmp_name = NULL;
3389 
3390         lockdep_assert_held(&memcg_cache_mutex);
3391 
3392         /*
3393          * kmem_cache_create_memcg duplicates the given name and
3394          * cgroup_name for this name requires RCU context.
3395          * This static temporary buffer is used to prevent from
3396          * pointless shortliving allocation.
3397          */
3398         if (!tmp_name) {
3399                 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3400                 if (!tmp_name)
3401                         return NULL;
3402         }
3403 
3404         rcu_read_lock();
3405         snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3406                          memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3407         rcu_read_unlock();
3408 
3409         new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3410                                       (s->flags & ~SLAB_PANIC), s->ctor, s);
3411 
3412         if (new)
3413                 new->allocflags |= __GFP_KMEMCG;
3414 
3415         return new;
3416 }
3417 
3418 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3419                                                   struct kmem_cache *cachep)
3420 {
3421         struct kmem_cache *new_cachep;
3422         int idx;
3423 
3424         BUG_ON(!memcg_can_account_kmem(memcg));
3425 
3426         idx = memcg_cache_id(memcg);
3427 
3428         mutex_lock(&memcg_cache_mutex);
3429         new_cachep = cache_from_memcg_idx(cachep, idx);
3430         if (new_cachep) {
3431                 css_put(&memcg->css);
3432                 goto out;
3433         }
3434 
3435         new_cachep = kmem_cache_dup(memcg, cachep);
3436         if (new_cachep == NULL) {
3437                 new_cachep = cachep;
3438                 css_put(&memcg->css);
3439                 goto out;
3440         }
3441 
3442         atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3443 
3444         cachep->memcg_params->memcg_caches[idx] = new_cachep;
3445         /*
3446          * the readers won't lock, make sure everybody sees the updated value,
3447          * so they won't put stuff in the queue again for no reason
3448          */
3449         wmb();
3450 out:
3451         mutex_unlock(&memcg_cache_mutex);
3452         return new_cachep;
3453 }
3454 
3455 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3456 {
3457         struct kmem_cache *c;
3458         int i;
3459 
3460         if (!s->memcg_params)
3461                 return;
3462         if (!s->memcg_params->is_root_cache)
3463                 return;
3464 
3465         /*
3466          * If the cache is being destroyed, we trust that there is no one else
3467          * requesting objects from it. Even if there are, the sanity checks in
3468          * kmem_cache_destroy should caught this ill-case.
3469          *
3470          * Still, we don't want anyone else freeing memcg_caches under our
3471          * noses, which can happen if a new memcg comes to life. As usual,
3472          * we'll take the set_limit_mutex to protect ourselves against this.
3473          */
3474         mutex_lock(&set_limit_mutex);
3475         for_each_memcg_cache_index(i) {
3476                 c = cache_from_memcg_idx(s, i);
3477                 if (!c)
3478                         continue;
3479 
3480                 /*
3481                  * We will now manually delete the caches, so to avoid races
3482                  * we need to cancel all pending destruction workers and
3483                  * proceed with destruction ourselves.
3484                  *
3485                  * kmem_cache_destroy() will call kmem_cache_shrink internally,
3486                  * and that could spawn the workers again: it is likely that
3487                  * the cache still have active pages until this very moment.
3488                  * This would lead us back to mem_cgroup_destroy_cache.
3489                  *
3490                  * But that will not execute at all if the "dead" flag is not
3491                  * set, so flip it down to guarantee we are in control.
3492                  */
3493                 c->memcg_params->dead = false;
3494                 cancel_work_sync(&c->memcg_params->destroy);
3495                 kmem_cache_destroy(c);
3496         }
3497         mutex_unlock(&set_limit_mutex);
3498 }
3499 
3500 struct create_work {
3501         struct mem_cgroup *memcg;
3502         struct kmem_cache *cachep;
3503         struct work_struct work;
3504 };
3505 
3506 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3507 {
3508         struct kmem_cache *cachep;
3509         struct memcg_cache_params *params;
3510 
3511         if (!memcg_kmem_is_active(memcg))
3512                 return;
3513 
3514         mutex_lock(&memcg->slab_caches_mutex);
3515         list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3516                 cachep = memcg_params_to_cache(params);
3517                 cachep->memcg_params->dead = true;
3518                 schedule_work(&cachep->memcg_params->destroy);
3519         }
3520         mutex_unlock(&memcg->slab_caches_mutex);
3521 }
3522 
3523 static void memcg_create_cache_work_func(struct work_struct *w)
3524 {
3525         struct create_work *cw;
3526 
3527         cw = container_of(w, struct create_work, work);
3528         memcg_create_kmem_cache(cw->memcg, cw->cachep);
3529         kfree(cw);
3530 }
3531 
3532 /*
3533  * Enqueue the creation of a per-memcg kmem_cache.
3534  */
3535 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3536                                          struct kmem_cache *cachep)
3537 {
3538         struct create_work *cw;
3539 
3540         cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3541         if (cw == NULL) {
3542                 css_put(&memcg->css);
3543                 return;
3544         }
3545 
3546         cw->memcg = memcg;
3547         cw->cachep = cachep;
3548 
3549         INIT_WORK(&cw->work, memcg_create_cache_work_func);
3550         schedule_work(&cw->work);
3551 }
3552 
3553 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3554                                        struct kmem_cache *cachep)
3555 {
3556         /*
3557          * We need to stop accounting when we kmalloc, because if the
3558          * corresponding kmalloc cache is not yet created, the first allocation
3559          * in __memcg_create_cache_enqueue will recurse.
3560          *
3561          * However, it is better to enclose the whole function. Depending on
3562          * the debugging options enabled, INIT_WORK(), for instance, can
3563          * trigger an allocation. This too, will make us recurse. Because at
3564          * this point we can't allow ourselves back into memcg_kmem_get_cache,
3565          * the safest choice is to do it like this, wrapping the whole function.
3566          */
3567         memcg_stop_kmem_account();
3568         __memcg_create_cache_enqueue(memcg, cachep);
3569         memcg_resume_kmem_account();
3570 }
3571 /*
3572  * Return the kmem_cache we're supposed to use for a slab allocation.
3573  * We try to use the current memcg's version of the cache.
3574  *
3575  * If the cache does not exist yet, if we are the first user of it,
3576  * we either create it immediately, if possible, or create it asynchronously
3577  * in a workqueue.
3578  * In the latter case, we will let the current allocation go through with
3579  * the original cache.
3580  *
3581  * Can't be called in interrupt context or from kernel threads.
3582  * This function needs to be called with rcu_read_lock() held.
3583  */
3584 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3585                                           gfp_t gfp)
3586 {
3587         struct mem_cgroup *memcg;
3588         int idx;
3589 
3590         VM_BUG_ON(!cachep->memcg_params);
3591         VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3592 
3593         if (!current->mm || current->memcg_kmem_skip_account)
3594                 return cachep;
3595 
3596         rcu_read_lock();
3597         memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3598 
3599         if (!memcg_can_account_kmem(memcg))
3600                 goto out;
3601 
3602         idx = memcg_cache_id(memcg);
3603 
3604         /*
3605          * barrier to mare sure we're always seeing the up to date value.  The
3606          * code updating memcg_caches will issue a write barrier to match this.
3607          */
3608         read_barrier_depends();
3609         if (likely(cache_from_memcg_idx(cachep, idx))) {
3610                 cachep = cache_from_memcg_idx(cachep, idx);
3611                 goto out;
3612         }
3613 
3614         /* The corresponding put will be done in the workqueue. */
3615         if (!css_tryget(&memcg->css))
3616                 goto out;
3617         rcu_read_unlock();
3618 
3619         /*
3620          * If we are in a safe context (can wait, and not in interrupt
3621          * context), we could be be predictable and return right away.
3622          * This would guarantee that the allocation being performed
3623          * already belongs in the new cache.
3624          *
3625          * However, there are some clashes that can arrive from locking.
3626          * For instance, because we acquire the slab_mutex while doing
3627          * kmem_cache_dup, this means no further allocation could happen
3628          * with the slab_mutex held.
3629          *
3630          * Also, because cache creation issue get_online_cpus(), this
3631          * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3632          * that ends up reversed during cpu hotplug. (cpuset allocates
3633          * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3634          * better to defer everything.
3635          */
3636         memcg_create_cache_enqueue(memcg, cachep);
3637         return cachep;
3638 out:
3639         rcu_read_unlock();
3640         return cachep;
3641 }
3642 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3643 
3644 /*
3645  * We need to verify if the allocation against current->mm->owner's memcg is
3646  * possible for the given order. But the page is not allocated yet, so we'll
3647  * need a further commit step to do the final arrangements.
3648  *
3649  * It is possible for the task to switch cgroups in this mean time, so at
3650  * commit time, we can't rely on task conversion any longer.  We'll then use
3651  * the handle argument to return to the caller which cgroup we should commit
3652  * against. We could also return the memcg directly and avoid the pointer
3653  * passing, but a boolean return value gives better semantics considering
3654  * the compiled-out case as well.
3655  *
3656  * Returning true means the allocation is possible.
3657  */
3658 bool
3659 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3660 {
3661         struct mem_cgroup *memcg;
3662         int ret;
3663 
3664         *_memcg = NULL;
3665 
3666         /*
3667          * Disabling accounting is only relevant for some specific memcg
3668          * internal allocations. Therefore we would initially not have such
3669          * check here, since direct calls to the page allocator that are marked
3670          * with GFP_KMEMCG only happen outside memcg core. We are mostly
3671          * concerned with cache allocations, and by having this test at
3672          * memcg_kmem_get_cache, we are already able to relay the allocation to
3673          * the root cache and bypass the memcg cache altogether.
3674          *
3675          * There is one exception, though: the SLUB allocator does not create
3676          * large order caches, but rather service large kmallocs directly from
3677          * the page allocator. Therefore, the following sequence when backed by
3678          * the SLUB allocator:
3679          *
3680          *      memcg_stop_kmem_account();
3681          *      kmalloc(<large_number>)
3682          *      memcg_resume_kmem_account();
3683          *
3684          * would effectively ignore the fact that we should skip accounting,
3685          * since it will drive us directly to this function without passing
3686          * through the cache selector memcg_kmem_get_cache. Such large
3687          * allocations are extremely rare but can happen, for instance, for the
3688          * cache arrays. We bring this test here.
3689          */
3690         if (!current->mm || current->memcg_kmem_skip_account)
3691                 return true;
3692 
3693         memcg = try_get_mem_cgroup_from_mm(current->mm);
3694 
3695         /*
3696          * very rare case described in mem_cgroup_from_task. Unfortunately there
3697          * isn't much we can do without complicating this too much, and it would
3698          * be gfp-dependent anyway. Just let it go
3699          */
3700         if (unlikely(!memcg))
3701                 return true;
3702 
3703         if (!memcg_can_account_kmem(memcg)) {
3704                 css_put(&memcg->css);
3705                 return true;
3706         }
3707 
3708         ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3709         if (!ret)
3710                 *_memcg = memcg;
3711 
3712         css_put(&memcg->css);
3713         return (ret == 0);
3714 }
3715 
3716 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3717                               int order)
3718 {
3719         struct page_cgroup *pc;
3720 
3721         VM_BUG_ON(mem_cgroup_is_root(memcg));
3722 
3723         /* The page allocation failed. Revert */
3724         if (!page) {
3725                 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3726                 return;
3727         }
3728 
3729         pc = lookup_page_cgroup(page);
3730         lock_page_cgroup(pc);
3731         pc->mem_cgroup = memcg;
3732         SetPageCgroupUsed(pc);
3733         unlock_page_cgroup(pc);
3734 }
3735 
3736 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3737 {
3738         struct mem_cgroup *memcg = NULL;
3739         struct page_cgroup *pc;
3740 
3741 
3742         pc = lookup_page_cgroup(page);
3743         /*
3744          * Fast unlocked return. Theoretically might have changed, have to
3745          * check again after locking.
3746          */
3747         if (!PageCgroupUsed(pc))
3748                 return;
3749 
3750         lock_page_cgroup(pc);
3751         if (PageCgroupUsed(pc)) {
3752                 memcg = pc->mem_cgroup;
3753                 ClearPageCgroupUsed(pc);
3754         }
3755         unlock_page_cgroup(pc);
3756 
3757         /*
3758          * We trust that only if there is a memcg associated with the page, it
3759          * is a valid allocation
3760          */
3761         if (!memcg)
3762                 return;
3763 
3764         VM_BUG_ON(mem_cgroup_is_root(memcg));
3765         memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3766 }
3767 #else
3768 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3769 {
3770 }
3771 #endif /* CONFIG_MEMCG_KMEM */
3772 
3773 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3774 
3775 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3776 /*
3777  * Because tail pages are not marked as "used", set it. We're under
3778  * zone->lru_lock, 'splitting on pmd' and compound_lock.
3779  * charge/uncharge will be never happen and move_account() is done under
3780  * compound_lock(), so we don't have to take care of races.
3781  */
3782 void mem_cgroup_split_huge_fixup(struct page *head)
3783 {
3784         struct page_cgroup *head_pc = lookup_page_cgroup(head);
3785         struct page_cgroup *pc;
3786         struct mem_cgroup *memcg;
3787         int i;
3788 
3789         if (mem_cgroup_disabled())
3790                 return;
3791 
3792         memcg = head_pc->mem_cgroup;
3793         for (i = 1; i < HPAGE_PMD_NR; i++) {
3794                 pc = head_pc + i;
3795                 pc->mem_cgroup = memcg;
3796                 smp_wmb();/* see __commit_charge() */
3797                 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3798         }
3799         __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3800                        HPAGE_PMD_NR);
3801 }
3802 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3803 
3804 static inline
3805 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3806                                         struct mem_cgroup *to,
3807                                         unsigned int nr_pages,
3808                                         enum mem_cgroup_stat_index idx)
3809 {
3810         /* Update stat data for mem_cgroup */
3811         preempt_disable();
3812         __this_cpu_sub(from->stat->count[idx], nr_pages);
3813         __this_cpu_add(to->stat->count[idx], nr_pages);
3814         preempt_enable();
3815 }
3816 
3817 /**
3818  * mem_cgroup_move_account - move account of the page
3819  * @page: the page
3820  * @nr_pages: number of regular pages (>1 for huge pages)
3821  * @pc: page_cgroup of the page.
3822  * @from: mem_cgroup which the page is moved from.
3823  * @to: mem_cgroup which the page is moved to. @from != @to.
3824  *
3825  * The caller must confirm following.
3826  * - page is not on LRU (isolate_page() is useful.)
3827  * - compound_lock is held when nr_pages > 1
3828  *
3829  * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3830  * from old cgroup.
3831  */
3832 static int mem_cgroup_move_account(struct page *page,
3833                                    unsigned int nr_pages,
3834                                    struct page_cgroup *pc,
3835                                    struct mem_cgroup *from,
3836                                    struct mem_cgroup *to)
3837 {
3838         unsigned long flags;
3839         int ret;
3840         bool anon = PageAnon(page);
3841 
3842         VM_BUG_ON(from == to);
3843         VM_BUG_ON(PageLRU(page));
3844         /*
3845          * The page is isolated from LRU. So, collapse function
3846          * will not handle this page. But page splitting can happen.
3847          * Do this check under compound_page_lock(). The caller should
3848          * hold it.
3849          */
3850         ret = -EBUSY;
3851         if (nr_pages > 1 && !PageTransHuge(page))
3852                 goto out;
3853 
3854         lock_page_cgroup(pc);
3855 
3856         ret = -EINVAL;
3857         if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3858                 goto unlock;
3859 
3860         move_lock_mem_cgroup(from, &flags);
3861 
3862         if (!anon && page_mapped(page))
3863                 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3864                         MEM_CGROUP_STAT_FILE_MAPPED);
3865 
3866         if (PageWriteback(page))
3867                 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3868                         MEM_CGROUP_STAT_WRITEBACK);
3869 
3870         mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3871 
3872         /* caller should have done css_get */
3873         pc->mem_cgroup = to;
3874         mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3875         move_unlock_mem_cgroup(from, &flags);
3876         ret = 0;
3877 unlock:
3878         unlock_page_cgroup(pc);
3879         /*
3880          * check events
3881          */
3882         memcg_check_events(to, page);
3883         memcg_check_events(from, page);
3884 out:
3885         return ret;
3886 }
3887 
3888 /**
3889  * mem_cgroup_move_parent - moves page to the parent group
3890  * @page: the page to move
3891  * @pc: page_cgroup of the page
3892  * @child: page's cgroup
3893  *
3894  * move charges to its parent or the root cgroup if the group has no
3895  * parent (aka use_hierarchy==0).
3896  * Although this might fail (get_page_unless_zero, isolate_lru_page or
3897  * mem_cgroup_move_account fails) the failure is always temporary and
3898  * it signals a race with a page removal/uncharge or migration. In the
3899  * first case the page is on the way out and it will vanish from the LRU
3900  * on the next attempt and the call should be retried later.
3901  * Isolation from the LRU fails only if page has been isolated from
3902  * the LRU since we looked at it and that usually means either global
3903  * reclaim or migration going on. The page will either get back to the
3904  * LRU or vanish.
3905  * Finaly mem_cgroup_move_account fails only if the page got uncharged
3906  * (!PageCgroupUsed) or moved to a different group. The page will
3907  * disappear in the next attempt.
3908  */
3909 static int mem_cgroup_move_parent(struct page *page,
3910                                   struct page_cgroup *pc,
3911                                   struct mem_cgroup *child)
3912 {
3913         struct mem_cgroup *parent;
3914         unsigned int nr_pages;
3915         unsigned long uninitialized_var(flags);
3916         int ret;
3917 
3918         VM_BUG_ON(mem_cgroup_is_root(child));
3919 
3920         ret = -EBUSY;
3921         if (!get_page_unless_zero(page))
3922                 goto out;
3923         if (isolate_lru_page(page))
3924                 goto put;
3925 
3926         nr_pages = hpage_nr_pages(page);
3927 
3928         parent = parent_mem_cgroup(child);
3929         /*
3930          * If no parent, move charges to root cgroup.
3931          */
3932         if (!parent)
3933                 parent = root_mem_cgroup;
3934 
3935         if (nr_pages > 1) {
3936                 VM_BUG_ON(!PageTransHuge(page));
3937                 flags = compound_lock_irqsave(page);
3938         }
3939 
3940         ret = mem_cgroup_move_account(page, nr_pages,
3941                                 pc, child, parent);
3942         if (!ret)
3943                 __mem_cgroup_cancel_local_charge(child, nr_pages);
3944 
3945         if (nr_pages > 1)
3946                 compound_unlock_irqrestore(page, flags);
3947         putback_lru_page(page);
3948 put:
3949         put_page(page);
3950 out:
3951         return ret;
3952 }
3953 
3954 /*
3955  * Charge the memory controller for page usage.
3956  * Return
3957  * 0 if the charge was successful
3958  * < 0 if the cgroup is over its limit
3959  */
3960 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3961                                 gfp_t gfp_mask, enum charge_type ctype)
3962 {
3963         struct mem_cgroup *memcg = NULL;
3964         unsigned int nr_pages = 1;
3965         bool oom = true;
3966         int ret;
3967 
3968         if (PageTransHuge(page)) {
3969                 nr_pages <<= compound_order(page);
3970                 VM_BUG_ON(!PageTransHuge(page));
3971                 /*
3972                  * Never OOM-kill a process for a huge page.  The
3973                  * fault handler will fall back to regular pages.
3974                  */
3975                 oom = false;
3976         }
3977 
3978         ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3979         if (ret == -ENOMEM)
3980                 return ret;
3981         __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3982         return 0;
3983 }
3984 
3985 int mem_cgroup_newpage_charge(struct page *page,
3986                               struct mm_struct *mm, gfp_t gfp_mask)
3987 {
3988         if (mem_cgroup_disabled())
3989                 return 0;
3990         VM_BUG_ON(page_mapped(page));
3991         VM_BUG_ON(page->mapping && !PageAnon(page));
3992         VM_BUG_ON(!mm);
3993         return mem_cgroup_charge_common(page, mm, gfp_mask,
3994                                         MEM_CGROUP_CHARGE_TYPE_ANON);
3995 }
3996 
3997 /*
3998  * While swap-in, try_charge -> commit or cancel, the page is locked.
3999  * And when try_charge() successfully returns, one refcnt to memcg without
4000  * struct page_cgroup is acquired. This refcnt will be consumed by
4001  * "commit()" or removed by "cancel()"
4002  */
4003 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
4004                                           struct page *page,
4005                                           gfp_t mask,
4006                                           struct mem_cgroup **memcgp)
4007 {
4008         struct mem_cgroup *memcg;
4009         struct page_cgroup *pc;
4010         int ret;
4011 
4012         pc = lookup_page_cgroup(page);
4013         /*
4014          * Every swap fault against a single page tries to charge the
4015          * page, bail as early as possible.  shmem_unuse() encounters
4016          * already charged pages, too.  The USED bit is protected by
4017          * the page lock, which serializes swap cache removal, which
4018          * in turn serializes uncharging.
4019          */
4020         if (PageCgroupUsed(pc))
4021                 return 0;
4022         if (!do_swap_account)
4023                 goto charge_cur_mm;
4024         memcg = try_get_mem_cgroup_from_page(page);
4025         if (!memcg)
4026                 goto charge_cur_mm;
4027         *memcgp = memcg;
4028         ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4029         css_put(&memcg->css);
4030         if (ret == -EINTR)
4031                 ret = 0;
4032         return ret;
4033 charge_cur_mm:
4034         ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4035         if (ret == -EINTR)
4036                 ret = 0;
4037         return ret;
4038 }
4039 
4040 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4041                                  gfp_t gfp_mask, struct mem_cgroup **memcgp)
4042 {
4043         *memcgp = NULL;
4044         if (mem_cgroup_disabled())
4045                 return 0;
4046         /*
4047          * A racing thread's fault, or swapoff, may have already
4048          * updated the pte, and even removed page from swap cache: in
4049          * those cases unuse_pte()'s pte_same() test will fail; but
4050          * there's also a KSM case which does need to charge the page.
4051          */
4052         if (!PageSwapCache(page)) {
4053                 int ret;
4054 
4055                 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4056                 if (ret == -EINTR)
4057                         ret = 0;
4058                 return ret;
4059         }
4060         return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4061 }
4062 
4063 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4064 {
4065         if (mem_cgroup_disabled())
4066                 return;
4067         if (!memcg)
4068                 return;
4069         __mem_cgroup_cancel_charge(memcg, 1);
4070 }
4071 
4072 static void
4073 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4074                                         enum charge_type ctype)
4075 {
4076         if (mem_cgroup_disabled())
4077                 return;
4078         if (!memcg)
4079                 return;
4080 
4081         __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4082         /*
4083          * Now swap is on-memory. This means this page may be
4084          * counted both as mem and swap....double count.
4085          * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4086          * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4087          * may call delete_from_swap_cache() before reach here.
4088          */
4089         if (do_swap_account && PageSwapCache(page)) {
4090                 swp_entry_t ent = {.val = page_private(page)};
4091                 mem_cgroup_uncharge_swap(ent);
4092         }
4093 }
4094 
4095 void mem_cgroup_commit_charge_swapin(struct page *page,
4096                                      struct mem_cgroup *memcg)
4097 {
4098         __mem_cgroup_commit_charge_swapin(page, memcg,
4099                                           MEM_CGROUP_CHARGE_TYPE_ANON);
4100 }
4101 
4102 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4103                                 gfp_t gfp_mask)
4104 {
4105         struct mem_cgroup *memcg = NULL;
4106         enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4107         int ret;
4108 
4109         if (mem_cgroup_disabled())
4110                 return 0;
4111         if (PageCompound(page))
4112                 return 0;
4113 
4114         if (!PageSwapCache(page))
4115                 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4116         else { /* page is swapcache/shmem */
4117                 ret = __mem_cgroup_try_charge_swapin(mm, page,
4118                                                      gfp_mask, &memcg);
4119                 if (!ret)
4120                         __mem_cgroup_commit_charge_swapin(page, memcg, type);
4121         }
4122         return ret;
4123 }
4124 
4125 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4126                                    unsigned int nr_pages,
4127                                    const enum charge_type ctype)
4128 {
4129         struct memcg_batch_info *batch = NULL;
4130         bool uncharge_memsw = true;
4131 
4132         /* If swapout, usage of swap doesn't decrease */
4133         if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4134                 uncharge_memsw = false;
4135 
4136         batch = &current->memcg_batch;
4137         /*
4138          * In usual, we do css_get() when we remember memcg pointer.
4139          * But in this case, we keep res->usage until end of a series of
4140          * uncharges. Then, it's ok to ignore memcg's refcnt.
4141          */
4142         if (!batch->memcg)
4143                 batch->memcg = memcg;
4144         /*
4145          * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4146          * In those cases, all pages freed continuously can be expected to be in
4147          * the same cgroup and we have chance to coalesce uncharges.
4148          * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4149          * because we want to do uncharge as soon as possible.
4150          */
4151 
4152         if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4153                 goto direct_uncharge;
4154 
4155         if (nr_pages > 1)
4156                 goto direct_uncharge;
4157 
4158         /*
4159          * In typical case, batch->memcg == mem. This means we can
4160          * merge a series of uncharges to an uncharge of res_counter.
4161          * If not, we uncharge res_counter ony by one.
4162          */
4163         if (batch->memcg != memcg)
4164                 goto direct_uncharge;
4165         /* remember freed charge and uncharge it later */
4166         batch->nr_pages++;
4167         if (uncharge_memsw)
4168                 batch->memsw_nr_pages++;
4169         return;
4170 direct_uncharge:
4171         res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4172         if (uncharge_memsw)
4173                 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4174         if (unlikely(batch->memcg != memcg))
4175                 memcg_oom_recover(memcg);
4176 }
4177 
4178 /*
4179  * uncharge if !page_mapped(page)
4180  */
4181 static struct mem_cgroup *
4182 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4183                              bool end_migration)
4184 {
4185         struct mem_cgroup *memcg = NULL;
4186         unsigned int nr_pages = 1;
4187         struct page_cgroup *pc;
4188         bool anon;
4189 
4190         if (mem_cgroup_disabled())
4191                 return NULL;
4192 
4193         if (PageTransHuge(page)) {
4194                 nr_pages <<= compound_order(page);
4195                 VM_BUG_ON(!PageTransHuge(page));
4196         }
4197         /*
4198          * Check if our page_cgroup is valid
4199          */
4200         pc = lookup_page_cgroup(page);
4201         if (unlikely(!PageCgroupUsed(pc)))
4202                 return NULL;
4203 
4204         lock_page_cgroup(pc);
4205 
4206         memcg = pc->mem_cgroup;
4207 
4208         if (!PageCgroupUsed(pc))
4209                 goto unlock_out;
4210 
4211         anon = PageAnon(page);
4212 
4213         switch (ctype) {
4214         case MEM_CGROUP_CHARGE_TYPE_ANON:
4215                 /*
4216                  * Generally PageAnon tells if it's the anon statistics to be
4217                  * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4218                  * used before page reached the stage of being marked PageAnon.
4219                  */
4220                 anon = true;
4221                 /* fallthrough */
4222         case MEM_CGROUP_CHARGE_TYPE_DROP:
4223                 /* See mem_cgroup_prepare_migration() */
4224                 if (page_mapped(page))
4225                         goto unlock_out;
4226                 /*
4227                  * Pages under migration may not be uncharged.  But
4228                  * end_migration() /must/ be the one uncharging the
4229                  * unused post-migration page and so it has to call
4230                  * here with the migration bit still set.  See the
4231                  * res_counter handling below.
4232                  */
4233                 if (!end_migration && PageCgroupMigration(pc))
4234                         goto unlock_out;
4235                 break;
4236         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4237                 if (!PageAnon(page)) {  /* Shared memory */
4238                         if (page->mapping && !page_is_file_cache(page))
4239                                 goto unlock_out;
4240                 } else if (page_mapped(page)) /* Anon */
4241                                 goto unlock_out;
4242                 break;
4243         default:
4244                 break;
4245         }
4246 
4247         mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4248 
4249         ClearPageCgroupUsed(pc);
4250         /*
4251          * pc->mem_cgroup is not cleared here. It will be accessed when it's
4252          * freed from LRU. This is safe because uncharged page is expected not
4253          * to be reused (freed soon). Exception is SwapCache, it's handled by
4254          * special functions.
4255          */
4256 
4257         unlock_page_cgroup(pc);
4258         /*
4259          * even after unlock, we have memcg->res.usage here and this memcg
4260          * will never be freed, so it's safe to call css_get().
4261          */
4262         memcg_check_events(memcg, page);
4263         if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4264                 mem_cgroup_swap_statistics(memcg, true);
4265                 css_get(&memcg->css);
4266         }
4267         /*
4268          * Migration does not charge the res_counter for the
4269          * replacement page, so leave it alone when phasing out the
4270          * page that is unused after the migration.
4271          */
4272         if (!end_migration && !mem_cgroup_is_root(memcg))
4273                 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4274 
4275         return memcg;
4276 
4277 unlock_out:
4278         unlock_page_cgroup(pc);
4279         return NULL;
4280 }
4281 
4282 void mem_cgroup_uncharge_page(struct page *page)
4283 {
4284         /* early check. */
4285         if (page_mapped(page))
4286                 return;
4287         VM_BUG_ON(page->mapping && !PageAnon(page));
4288         /*
4289          * If the page is in swap cache, uncharge should be deferred
4290          * to the swap path, which also properly accounts swap usage
4291          * and handles memcg lifetime.
4292          *
4293          * Note that this check is not stable and reclaim may add the
4294          * page to swap cache at any time after this.  However, if the
4295          * page is not in swap cache by the time page->mapcount hits
4296          * 0, there won't be any page table references to the swap
4297          * slot, and reclaim will free it and not actually write the
4298          * page to disk.
4299          */
4300         if (PageSwapCache(page))
4301                 return;
4302         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4303 }
4304 
4305 void mem_cgroup_uncharge_cache_page(struct page *page)
4306 {
4307         VM_BUG_ON(page_mapped(page));
4308         VM_BUG_ON(page->mapping);
4309         __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4310 }
4311 
4312 /*
4313  * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4314  * In that cases, pages are freed continuously and we can expect pages
4315  * are in the same memcg. All these calls itself limits the number of
4316  * pages freed at once, then uncharge_start/end() is called properly.
4317  * This may be called prural(2) times in a context,
4318  */
4319 
4320 void mem_cgroup_uncharge_start(void)
4321 {
4322         current->memcg_batch.do_batch++;
4323         /* We can do nest. */
4324         if (current->memcg_batch.do_batch == 1) {
4325                 current->memcg_batch.memcg = NULL;
4326                 current->memcg_batch.nr_pages = 0;
4327                 current->memcg_batch.memsw_nr_pages = 0;
4328         }
4329 }
4330 
4331 void mem_cgroup_uncharge_end(void)
4332 {
4333         struct memcg_batch_info *batch = &current->memcg_batch;
4334 
4335         if (!batch->do_batch)
4336                 return;
4337 
4338         batch->do_batch--;
4339         if (batch->do_batch) /* If stacked, do nothing. */
4340                 return;
4341 
4342         if (!batch->memcg)
4343                 return;
4344         /*
4345          * This "batch->memcg" is valid without any css_get/put etc...
4346          * bacause we hide charges behind us.
4347          */
4348         if (batch->nr_pages)
4349                 res_counter_uncharge(&batch->memcg->res,
4350                                      batch->nr_pages * PAGE_SIZE);
4351         if (batch->memsw_nr_pages)
4352                 res_counter_uncharge(&batch->memcg->memsw,
4353                                      batch->memsw_nr_pages * PAGE_SIZE);
4354         memcg_oom_recover(batch->memcg);
4355         /* forget this pointer (for sanity check) */
4356         batch->memcg = NULL;
4357 }
4358 
4359 #ifdef CONFIG_SWAP
4360 /*
4361  * called after __delete_from_swap_cache() and drop "page" account.
4362  * memcg information is recorded to swap_cgroup of "ent"
4363  */
4364 void
4365 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4366 {
4367         struct mem_cgroup *memcg;
4368         int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4369 
4370         if (!swapout) /* this was a swap cache but the swap is unused ! */
4371                 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4372 
4373         memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4374 
4375         /*
4376          * record memcg information,  if swapout && memcg != NULL,
4377          * css_get() was called in uncharge().
4378          */
4379         if (do_swap_account && swapout && memcg)
4380                 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4381 }
4382 #endif
4383 
4384 #ifdef CONFIG_MEMCG_SWAP
4385 /*
4386  * called from swap_entry_free(). remove record in swap_cgroup and
4387  * uncharge "memsw" account.
4388  */
4389 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4390 {
4391         struct mem_cgroup *memcg;
4392         unsigned short id;
4393 
4394         if (!do_swap_account)
4395                 return;
4396 
4397         id = swap_cgroup_record(ent, 0);
4398         rcu_read_lock();
4399         memcg = mem_cgroup_lookup(id);
4400         if (memcg) {
4401                 /*
4402                  * We uncharge this because swap is freed.
4403                  * This memcg can be obsolete one. We avoid calling css_tryget
4404                  */
4405                 if (!mem_cgroup_is_root(memcg))
4406                         res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4407                 mem_cgroup_swap_statistics(memcg, false);
4408                 css_put(&memcg->css);
4409         }
4410         rcu_read_unlock();
4411 }
4412 
4413 /**
4414  * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4415  * @entry: swap entry to be moved
4416  * @from:  mem_cgroup which the entry is moved from
4417  * @to:  mem_cgroup which the entry is moved to
4418  *
4419  * It succeeds only when the swap_cgroup's record for this entry is the same
4420  * as the mem_cgroup's id of @from.
4421  *
4422  * Returns 0 on success, -EINVAL on failure.
4423  *
4424  * The caller must have charged to @to, IOW, called res_counter_charge() about
4425  * both res and memsw, and called css_get().
4426  */
4427 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4428                                 struct mem_cgroup *from, struct mem_cgroup *to)
4429 {
4430         unsigned short old_id, new_id;
4431 
4432         old_id = mem_cgroup_id(from);
4433         new_id = mem_cgroup_id(to);
4434 
4435         if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4436                 mem_cgroup_swap_statistics(from, false);
4437                 mem_cgroup_swap_statistics(to, true);
4438                 /*
4439                  * This function is only called from task migration context now.
4440                  * It postpones res_counter and refcount handling till the end
4441                  * of task migration(mem_cgroup_clear_mc()) for performance
4442                  * improvement. But we cannot postpone css_get(to)  because if
4443                  * the process that has been moved to @to does swap-in, the
4444                  * refcount of @to might be decreased to 0.
4445                  *
4446                  * We are in attach() phase, so the cgroup is guaranteed to be
4447                  * alive, so we can just call css_get().
4448                  */
4449                 css_get(&to->css);
4450                 return 0;
4451         }
4452         return -EINVAL;
4453 }
4454 #else
4455 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4456                                 struct mem_cgroup *from, struct mem_cgroup *to)
4457 {
4458         return -EINVAL;
4459 }
4460 #endif
4461 
4462 /*
4463  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4464  * page belongs to.
4465  */
4466 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4467                                   struct mem_cgroup **memcgp)
4468 {
4469         struct mem_cgroup *memcg = NULL;
4470         unsigned int nr_pages = 1;
4471         struct page_cgroup *pc;
4472         enum charge_type ctype;
4473 
4474         *memcgp = NULL;
4475 
4476         if (mem_cgroup_disabled())
4477                 return;
4478 
4479         if (PageTransHuge(page))
4480                 nr_pages <<= compound_order(page);
4481 
4482         pc = lookup_page_cgroup(page);
4483         lock_page_cgroup(pc);
4484         if (PageCgroupUsed(pc)) {
4485                 memcg = pc->mem_cgroup;
4486                 css_get(&memcg->css);
4487                 /*
4488                  * At migrating an anonymous page, its mapcount goes down
4489                  * to 0 and uncharge() will be called. But, even if it's fully
4490                  * unmapped, migration may fail and this page has to be
4491                  * charged again. We set MIGRATION flag here and delay uncharge
4492                  * until end_migration() is called
4493                  *
4494                  * Corner Case Thinking
4495                  * A)
4496                  * When the old page was mapped as Anon and it's unmap-and-freed
4497                  * while migration was ongoing.
4498                  * If unmap finds the old page, uncharge() of it will be delayed
4499                  * until end_migration(). If unmap finds a new page, it's
4500                  * uncharged when it make mapcount to be 1->0. If unmap code
4501                  * finds swap_migration_entry, the new page will not be mapped
4502                  * and end_migration() will find it(mapcount==0).
4503                  *
4504                  * B)
4505                  * When the old page was mapped but migraion fails, the kernel
4506                  * remaps it. A charge for it is kept by MIGRATION flag even
4507                  * if mapcount goes down to 0. We can do remap successfully
4508                  * without charging it again.
4509                  *
4510                  * C)
4511                  * The "old" page is under lock_page() until the end of
4512                  * migration, so, the old page itself will not be swapped-out.
4513                  * If the new page is swapped out before end_migraton, our
4514                  * hook to usual swap-out path will catch the event.
4515                  */
4516                 if (PageAnon(page))
4517                         SetPageCgroupMigration(pc);
4518         }
4519         unlock_page_cgroup(pc);
4520         /*
4521          * If the page is not charged at this point,
4522          * we return here.
4523          */
4524         if (!memcg)
4525                 return;
4526 
4527         *memcgp = memcg;
4528         /*
4529          * We charge new page before it's used/mapped. So, even if unlock_page()
4530          * is called before end_migration, we can catch all events on this new
4531          * page. In the case new page is migrated but not remapped, new page's
4532          * mapcount will be finally 0 and we call uncharge in end_migration().
4533          */
4534         if (PageAnon(page))
4535                 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4536         else
4537                 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4538         /*
4539          * The page is committed to the memcg, but it's not actually
4540          * charged to the res_counter since we plan on replacing the
4541          * old one and only one page is going to be left afterwards.
4542          */
4543         __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4544 }
4545 
4546 /* remove redundant charge if migration failed*/
4547 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4548         struct page *oldpage, struct page *newpage, bool migration_ok)
4549 {
4550         struct page *used, *unused;
4551         struct page_cgroup *pc;
4552         bool anon;
4553 
4554         if (!memcg)
4555                 return;
4556 
4557         if (!migration_ok) {
4558                 used = oldpage;
4559                 unused = newpage;
4560         } else {
4561                 used = newpage;
4562                 unused = oldpage;
4563         }
4564         anon = PageAnon(used);
4565         __mem_cgroup_uncharge_common(unused,
4566                                      anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4567                                      : MEM_CGROUP_CHARGE_TYPE_CACHE,
4568                                      true);
4569         css_put(&memcg->css);
4570         /*
4571          * We disallowed uncharge of pages under migration because mapcount
4572          * of the page goes down to zero, temporarly.
4573          * Clear the flag and check the page should be charged.
4574          */
4575         pc = lookup_page_cgroup(oldpage);
4576         lock_page_cgroup(pc);
4577         ClearPageCgroupMigration(pc);
4578         unlock_page_cgroup(pc);
4579 
4580         /*
4581          * If a page is a file cache, radix-tree replacement is very atomic
4582          * and we can skip this check. When it was an Anon page, its mapcount
4583          * goes down to 0. But because we added MIGRATION flage, it's not
4584          * uncharged yet. There are several case but page->mapcount check
4585          * and USED bit check in mem_cgroup_uncharge_page() will do enough
4586          * check. (see prepare_charge() also)
4587          */
4588         if (anon)
4589                 mem_cgroup_uncharge_page(used);
4590 }
4591 
4592 /*
4593  * At replace page cache, newpage is not under any memcg but it's on
4594  * LRU. So, this function doesn't touch res_counter but handles LRU
4595  * in correct way. Both pages are locked so we cannot race with uncharge.
4596  */
4597 void mem_cgroup_replace_page_cache(struct page *oldpage,
4598                                   struct page *newpage)
4599 {
4600         struct mem_cgroup *memcg = NULL;
4601         struct page_cgroup *pc;
4602         enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4603 
4604         if (mem_cgroup_disabled())
4605                 return;
4606 
4607         pc = lookup_page_cgroup(oldpage);
4608         /* fix accounting on old pages */
4609         lock_page_cgroup(pc);
4610         if (PageCgroupUsed(pc)) {
4611                 memcg = pc->mem_cgroup;
4612                 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4613                 ClearPageCgroupUsed(pc);
4614         }
4615         unlock_page_cgroup(pc);
4616 
4617         /*
4618          * When called from shmem_replace_page(), in some cases the
4619          * oldpage has already been charged, and in some cases not.
4620          */
4621         if (!memcg)
4622                 return;
4623         /*
4624          * Even if newpage->mapping was NULL before starting replacement,
4625          * the newpage may be on LRU(or pagevec for LRU) already. We lock
4626          * LRU while we overwrite pc->mem_cgroup.
4627          */
4628         __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4629 }
4630 
4631 #ifdef CONFIG_DEBUG_VM
4632 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4633 {
4634         struct page_cgroup *pc;
4635 
4636         pc = lookup_page_cgroup(page);
4637         /*
4638          * Can be NULL while feeding pages into the page allocator for
4639          * the first time, i.e. during boot or memory hotplug;
4640          * or when mem_cgroup_disabled().
4641          */
4642         if (likely(pc) && PageCgroupUsed(pc))
4643                 return pc;
4644         return NULL;
4645 }
4646 
4647 bool mem_cgroup_bad_page_check(struct page *page)
4648 {
4649         if (mem_cgroup_disabled())
4650                 return false;
4651 
4652         return lookup_page_cgroup_used(page) != NULL;
4653 }
4654 
4655 void mem_cgroup_print_bad_page(struct page *page)
4656 {
4657         struct page_cgroup *pc;
4658 
4659         pc = lookup_page_cgroup_used(page);
4660         if (pc) {
4661                 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4662                          pc, pc->flags, pc->mem_cgroup);
4663         }
4664 }
4665 #endif
4666 
4667 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4668                                 unsigned long long val)
4669 {
4670         int retry_count;
4671         u64 memswlimit, memlimit;
4672         int ret = 0;
4673         int children = mem_cgroup_count_children(memcg);
4674         u64 curusage, oldusage;
4675         int enlarge;
4676 
4677         /*
4678          * For keeping hierarchical_reclaim simple, how long we should retry
4679          * is depends on callers. We set our retry-count to be function
4680          * of # of children which we should visit in this loop.
4681          */
4682         retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4683 
4684         oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4685 
4686         enlarge = 0;
4687         while (retry_count) {
4688                 if (signal_pending(current)) {
4689                         ret = -EINTR;
4690                         break;
4691                 }
4692                 /*
4693                  * Rather than hide all in some function, I do this in
4694                  * open coded manner. You see what this really does.
4695                  * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4696                  */
4697                 mutex_lock(&set_limit_mutex);
4698                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4699                 if (memswlimit < val) {
4700                         ret = -EINVAL;
4701                         mutex_unlock(&set_limit_mutex);
4702                         break;
4703                 }
4704 
4705                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4706                 if (memlimit < val)
4707                         enlarge = 1;
4708 
4709                 ret = res_counter_set_limit(&memcg->res, val);
4710                 if (!ret) {
4711                         if (memswlimit == val)
4712                                 memcg->memsw_is_minimum = true;
4713                         else
4714                                 memcg->memsw_is_minimum = false;
4715                 }
4716                 mutex_unlock(&set_limit_mutex);
4717 
4718                 if (!ret)
4719                         break;
4720 
4721                 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4722                                    MEM_CGROUP_RECLAIM_SHRINK);
4723                 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4724                 /* Usage is reduced ? */
4725                 if (curusage >= oldusage)
4726                         retry_count--;
4727                 else
4728                         oldusage = curusage;
4729         }
4730         if (!ret && enlarge)
4731                 memcg_oom_recover(memcg);
4732 
4733         return ret;
4734 }
4735 
4736 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4737                                         unsigned long long val)
4738 {
4739         int retry_count;
4740         u64 memlimit, memswlimit, oldusage, curusage;
4741         int children = mem_cgroup_count_children(memcg);
4742         int ret = -EBUSY;
4743         int enlarge = 0;
4744 
4745         /* see mem_cgroup_resize_res_limit */
4746         retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4747         oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4748         while (retry_count) {
4749                 if (signal_pending(current)) {
4750                         ret = -EINTR;
4751                         break;
4752                 }
4753                 /*
4754                  * Rather than hide all in some function, I do this in
4755                  * open coded manner. You see what this really does.
4756                  * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4757                  */
4758                 mutex_lock(&set_limit_mutex);
4759                 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4760                 if (memlimit > val) {
4761                         ret = -EINVAL;
4762                         mutex_unlock(&set_limit_mutex);
4763                         break;
4764                 }
4765                 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4766                 if (memswlimit < val)
4767                         enlarge = 1;
4768                 ret = res_counter_set_limit(&memcg->memsw, val);
4769                 if (!ret) {
4770                         if (memlimit == val)
4771                                 memcg->memsw_is_minimum = true;
4772                         else
4773                                 memcg->memsw_is_minimum = false;
4774                 }
4775                 mutex_unlock(&set_limit_mutex);
4776 
4777                 if (!ret)
4778                         break;
4779 
4780                 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4781                                    MEM_CGROUP_RECLAIM_NOSWAP |
4782                                    MEM_CGROUP_RECLAIM_SHRINK);
4783                 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4784                 /* Usage is reduced ? */
4785                 if (curusage >= oldusage)
4786                         retry_count--;
4787                 else
4788                         oldusage = curusage;
4789         }
4790         if (!ret && enlarge)
4791                 memcg_oom_recover(memcg);
4792         return ret;
4793 }
4794 
4795 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4796                                             gfp_t gfp_mask,
4797                                             unsigned long *total_scanned)
4798 {
4799         unsigned long nr_reclaimed = 0;
4800         struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4801         unsigned long reclaimed;
4802         int loop = 0;
4803         struct mem_cgroup_tree_per_zone *mctz;
4804         unsigned long long excess;
4805         unsigned long nr_scanned;
4806 
4807         if (order > 0)
4808                 return 0;
4809 
4810         mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4811         /*
4812          * This loop can run a while, specially if mem_cgroup's continuously
4813          * keep exceeding their soft limit and putting the system under
4814          * pressure
4815          */
4816         do {
4817                 if (next_mz)
4818                         mz = next_mz;
4819                 else
4820                         mz = mem_cgroup_largest_soft_limit_node(mctz);
4821                 if (!mz)
4822                         break;
4823 
4824                 nr_scanned = 0;
4825                 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4826                                                     gfp_mask, &nr_scanned);
4827                 nr_reclaimed += reclaimed;
4828                 *total_scanned += nr_scanned;
4829                 spin_lock(&mctz->lock);
4830 
4831                 /*
4832                  * If we failed to reclaim anything from this memory cgroup
4833                  * it is time to move on to the next cgroup
4834                  */
4835                 next_mz = NULL;
4836                 if (!reclaimed) {
4837                         do {
4838                                 /*
4839                                  * Loop until we find yet another one.
4840                                  *
4841                                  * By the time we get the soft_limit lock
4842                                  * again, someone might have aded the
4843                                  * group back on the RB tree. Iterate to
4844                                  * make sure we get a different mem.
4845                                  * mem_cgroup_largest_soft_limit_node returns
4846                                  * NULL if no other cgroup is present on
4847                                  * the tree
4848                                  */
4849                                 next_mz =
4850                                 __mem_cgroup_largest_soft_limit_node(mctz);
4851                                 if (next_mz == mz)
4852                                         css_put(&next_mz->memcg->css);
4853                                 else /* next_mz == NULL or other memcg */
4854                                         break;
4855                         } while (1);
4856                 }
4857                 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4858                 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4859                 /*
4860                  * One school of thought says that we should not add
4861                  * back the node to the tree if reclaim returns 0.
4862                  * But our reclaim could return 0, simply because due
4863                  * to priority we are exposing a smaller subset of
4864                  * memory to reclaim from. Consider this as a longer
4865                  * term TODO.
4866                  */
4867                 /* If excess == 0, no tree ops */
4868                 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4869                 spin_unlock(&mctz->lock);
4870                 css_put(&mz->memcg->css);
4871                 loop++;
4872                 /*
4873                  * Could not reclaim anything and there are no more
4874                  * mem cgroups to try or we seem to be looping without
4875                  * reclaiming anything.
4876                  */
4877                 if (!nr_reclaimed &&
4878                         (next_mz == NULL ||
4879                         loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4880                         break;
4881         } while (!nr_reclaimed);
4882         if (next_mz)
4883                 css_put(&next_mz->memcg->css);
4884         return nr_reclaimed;
4885 }
4886 
4887 /**
4888  * mem_cgroup_force_empty_list - clears LRU of a group
4889  * @memcg: group to clear
4890  * @node: NUMA node
4891  * @zid: zone id
4892  * @lru: lru to to clear
4893  *
4894  * Traverse a specified page_cgroup list and try to drop them all.  This doesn't
4895  * reclaim the pages page themselves - pages are moved to the parent (or root)
4896  * group.
4897  */
4898 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4899                                 int node, int zid, enum lru_list lru)
4900 {
4901         struct lruvec *lruvec;
4902         unsigned long flags;
4903         struct list_head *list;
4904         struct page *busy;
4905         struct zone *zone;
4906 
4907         zone = &NODE_DATA(node)->node_zones[zid];
4908         lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4909         list = &lruvec->lists[lru];
4910 
4911         busy = NULL;
4912         do {
4913                 struct page_cgroup *pc;
4914                 struct page *page;
4915 
4916                 spin_lock_irqsave(&zone->lru_lock, flags);
4917                 if (list_empty(list)) {
4918                         spin_unlock_irqrestore(&zone->lru_lock, flags);
4919                         break;
4920                 }
4921                 page = list_entry(list->prev, struct page, lru);
4922                 if (busy == page) {
4923                         list_move(&page->lru, list);
4924                         busy = NULL;
4925                         spin_unlock_irqrestore(&zone->lru_lock, flags);
4926                         continue;
4927                 }
4928                 spin_unlock_irqrestore(&zone->lru_lock, flags);
4929 
4930                 pc = lookup_page_cgroup(page);
4931 
4932                 if (mem_cgroup_move_parent(page, pc, memcg)) {
4933                         /* found lock contention or "pc" is obsolete. */
4934                         busy = page;
4935                         cond_resched();
4936                 } else
4937                         busy = NULL;
4938         } while (!list_empty(list));
4939 }
4940 
4941 /*
4942  * make mem_cgroup's charge to be 0 if there is no task by moving
4943  * all the charges and pages to the parent.
4944  * This enables deleting this mem_cgroup.
4945  *
4946  * Caller is responsible for holding css reference on the memcg.
4947  */
4948 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4949 {
4950         int node, zid;
4951         u64 usage;
4952 
4953         do {
4954                 /* This is for making all *used* pages to be on LRU. */
4955                 lru_add_drain_all();
4956                 drain_all_stock_sync(memcg);
4957                 mem_cgroup_start_move(memcg);
4958                 for_each_node_state(node, N_MEMORY) {
4959                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4960                                 enum lru_list lru;
4961                                 for_each_lru(lru) {
4962                                         mem_cgroup_force_empty_list(memcg,
4963                                                         node, zid, lru);
4964                                 }
4965                         }
4966                 }
4967                 mem_cgroup_end_move(memcg);
4968                 memcg_oom_recover(memcg);
4969                 cond_resched();
4970 
4971                 /*
4972                  * Kernel memory may not necessarily be trackable to a specific
4973                  * process. So they are not migrated, and therefore we can't
4974                  * expect their value to drop to 0 here.
4975                  * Having res filled up with kmem only is enough.
4976                  *
4977                  * This is a safety check because mem_cgroup_force_empty_list
4978                  * could have raced with mem_cgroup_replace_page_cache callers
4979                  * so the lru seemed empty but the page could have been added
4980                  * right after the check. RES_USAGE should be safe as we always
4981                  * charge before adding to the LRU.
4982                  */
4983                 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4984                         res_counter_read_u64(&memcg->kmem, RES_USAGE);
4985         } while (usage > 0);
4986 }
4987 
4988 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4989 {
4990         lockdep_assert_held(&memcg_create_mutex);
4991         /*
4992          * The lock does not prevent addition or deletion to the list
4993          * of children, but it prevents a new child from being
4994          * initialized based on this parent in css_online(), so it's
4995          * enough to decide whether hierarchically inherited
4996          * attributes can still be changed or not.
4997          */
4998         return memcg->use_hierarchy &&
4999                 !list_empty(&memcg->css.cgroup->children);
5000 }
5001 
5002 /*
5003  * Reclaims as many pages from the given memcg as possible and moves
5004  * the rest to the parent.
5005  *
5006  * Caller is responsible for holding css reference for memcg.
5007  */
5008 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
5009 {
5010         int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5011         struct cgroup *cgrp = memcg->css.cgroup;
5012 
5013         /* returns EBUSY if there is a task or if we come here twice. */
5014         if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5015                 return -EBUSY;
5016 
5017         /* we call try-to-free pages for make this cgroup empty */
5018         lru_add_drain_all();
5019         /* try to free all pages in this cgroup */
5020         while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5021                 int progress;
5022 
5023                 if (signal_pending(current))
5024                         return -EINTR;
5025 
5026                 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5027                                                 false);
5028                 if (!progress) {
5029                         nr_retries--;
5030                         /* maybe some writeback is necessary */
5031                         congestion_wait(BLK_RW_ASYNC, HZ/10);
5032                 }
5033 
5034         }
5035         lru_add_drain();
5036         mem_cgroup_reparent_charges(memcg);
5037 
5038         return 0;
5039 }
5040 
5041 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5042                                         unsigned int event)
5043 {
5044         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5045 
5046         if (mem_cgroup_is_root(memcg))
5047                 return -EINVAL;
5048         return mem_cgroup_force_empty(memcg);
5049 }
5050 
5051 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5052                                      struct cftype *cft)
5053 {
5054         return mem_cgroup_from_css(css)->use_hierarchy;
5055 }
5056 
5057 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5058                                       struct cftype *cft, u64 val)
5059 {
5060         int retval = 0;
5061         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5062         struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5063 
5064         mutex_lock(&memcg_create_mutex);
5065 
5066         if (memcg->use_hierarchy == val)
5067                 goto out;
5068 
5069         /*
5070          * If parent's use_hierarchy is set, we can't make any modifications
5071          * in the child subtrees. If it is unset, then the change can
5072          * occur, provided the current cgroup has no children.
5073          *
5074          * For the root cgroup, parent_mem is NULL, we allow value to be
5075          * set if there are no children.
5076          */
5077         if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5078                                 (val == 1 || val == 0)) {
5079                 if (list_empty(&memcg->css.cgroup->children))
5080                         memcg->use_hierarchy = val;
5081                 else
5082                         retval = -EBUSY;
5083         } else
5084                 retval = -EINVAL;
5085 
5086 out:
5087         mutex_unlock(&memcg_create_mutex);
5088 
5089         return retval;
5090 }
5091 
5092 
5093 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5094                                                enum mem_cgroup_stat_index idx)
5095 {
5096         struct mem_cgroup *iter;
5097         long val = 0;
5098 
5099         /* Per-cpu values can be negative, use a signed accumulator */
5100         for_each_mem_cgroup_tree(iter, memcg)
5101                 val += mem_cgroup_read_stat(iter, idx);
5102 
5103         if (val < 0) /* race ? */
5104                 val = 0;
5105         return val;
5106 }
5107 
5108 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5109 {
5110         u64 val;
5111 
5112         if (!mem_cgroup_is_root(memcg)) {
5113                 if (!swap)
5114                         return res_counter_read_u64(&memcg->res, RES_USAGE);
5115                 else
5116                         return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5117         }
5118 
5119         /*
5120          * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5121          * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5122          */
5123         val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5124         val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5125 
5126         if (swap)
5127                 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5128 
5129         return val << PAGE_SHIFT;
5130 }
5131 
5132 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
5133                                struct cftype *cft, struct file *file,
5134                                char __user *buf, size_t nbytes, loff_t *ppos)
5135 {
5136         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5137         char str[64];
5138         u64 val;
5139         int name, len;
5140         enum res_type type;
5141 
5142         type = MEMFILE_TYPE(cft->private);
5143         name = MEMFILE_ATTR(cft->private);
5144 
5145         switch (type) {
5146         case _MEM:
5147                 if (name == RES_USAGE)
5148                         val = mem_cgroup_usage(memcg, false);
5149                 else
5150                         val = res_counter_read_u64(&memcg->res, name);
5151                 break;
5152         case _MEMSWAP:
5153                 if (name == RES_USAGE)
5154                         val = mem_cgroup_usage(memcg, true);
5155                 else
5156                         val = res_counter_read_u64(&memcg->memsw, name);
5157                 break;
5158         case _KMEM:
5159                 val = res_counter_read_u64(&memcg->kmem, name);
5160                 break;
5161         default:
5162                 BUG();
5163         }
5164 
5165         len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5166         return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5167 }
5168 
5169 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5170 {
5171         int ret = -EINVAL;
5172 #ifdef CONFIG_MEMCG_KMEM
5173         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5174         /*
5175          * For simplicity, we won't allow this to be disabled.  It also can't
5176          * be changed if the cgroup has children already, or if tasks had
5177          * already joined.
5178          *
5179          * If tasks join before we set the limit, a person looking at
5180          * kmem.usage_in_bytes will have no way to determine when it took
5181          * place, which makes the value quite meaningless.
5182          *
5183          * After it first became limited, changes in the value of the limit are
5184          * of course permitted.
5185          */
5186         mutex_lock(&memcg_create_mutex);
5187         mutex_lock(&set_limit_mutex);
5188         if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5189                 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5190                         ret = -EBUSY;
5191                         goto out;
5192                 }
5193                 ret = res_counter_set_limit(&memcg->kmem, val);
5194                 VM_BUG_ON(ret);
5195 
5196                 ret = memcg_update_cache_sizes(memcg);
5197                 if (ret) {
5198                         res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5199                         goto out;
5200                 }
5201                 static_key_slow_inc(&memcg_kmem_enabled_key);
5202                 /*
5203                  * setting the active bit after the inc will guarantee no one
5204                  * starts accounting before all call sites are patched
5205                  */
5206                 memcg_kmem_set_active(memcg);
5207         } else
5208                 ret = res_counter_set_limit(&memcg->kmem, val);
5209 out:
5210         mutex_unlock(&set_limit_mutex);
5211         mutex_unlock(&memcg_create_mutex);
5212 #endif
5213         return ret;
5214 }
5215 
5216 #ifdef CONFIG_MEMCG_KMEM
5217 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5218 {
5219         int ret = 0;
5220         struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5221         if (!parent)
5222                 goto out;
5223 
5224         memcg->kmem_account_flags = parent->kmem_account_flags;
5225         /*
5226          * When that happen, we need to disable the static branch only on those
5227          * memcgs that enabled it. To achieve this, we would be forced to
5228          * complicate the code by keeping track of which memcgs were the ones
5229          * that actually enabled limits, and which ones got it from its
5230          * parents.
5231          *
5232          * It is a lot simpler just to do static_key_slow_inc() on every child
5233          * that is accounted.
5234          */
5235         if (!memcg_kmem_is_active(memcg))
5236                 goto out;
5237 
5238         /*
5239          * __mem_cgroup_free() will issue static_key_slow_dec() because this
5240          * memcg is active already. If the later initialization fails then the
5241          * cgroup core triggers the cleanup so we do not have to do it here.
5242          */
5243         static_key_slow_inc(&memcg_kmem_enabled_key);
5244 
5245         mutex_lock(&set_limit_mutex);
5246         memcg_stop_kmem_account();
5247         ret = memcg_update_cache_sizes(memcg);
5248         memcg_resume_kmem_account();
5249         mutex_unlock(&set_limit_mutex);
5250 out:
5251         return ret;
5252 }
5253 #endif /* CONFIG_MEMCG_KMEM */
5254 
5255 /*
5256  * The user of this function is...
5257  * RES_LIMIT.
5258  */
5259 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5260                             const char *buffer)
5261 {
5262         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5263         enum res_type type;
5264         int name;
5265         unsigned long long val;
5266         int ret;
5267 
5268         type = MEMFILE_TYPE(cft->private);
5269         name = MEMFILE_ATTR(cft->private);
5270 
5271         switch (name) {
5272         case RES_LIMIT:
5273                 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5274                         ret = -EINVAL;
5275                         break;
5276                 }
5277                 /* This function does all necessary parse...reuse it */
5278                 ret = res_counter_memparse_write_strategy(buffer, &val);
5279                 if (ret)
5280                         break;
5281                 if (type == _MEM)
5282                         ret = mem_cgroup_resize_limit(memcg, val);
5283                 else if (type == _MEMSWAP)
5284                         ret = mem_cgroup_resize_memsw_limit(memcg, val);
5285                 else if (type == _KMEM)
5286                         ret = memcg_update_kmem_limit(css, val);
5287                 else
5288                         return -EINVAL;
5289                 break;
5290         case RES_SOFT_LIMIT:
5291                 ret = res_counter_memparse_write_strategy(buffer, &val);
5292                 if (ret)
5293                         break;
5294                 /*
5295                  * For memsw, soft limits are hard to implement in terms
5296                  * of semantics, for now, we support soft limits for
5297                  * control without swap
5298                  */
5299                 if (type == _MEM)
5300                         ret = res_counter_set_soft_limit(&memcg->res, val);
5301                 else
5302                         ret = -EINVAL;
5303                 break;
5304         default:
5305                 ret = -EINVAL; /* should be BUG() ? */
5306                 break;
5307         }
5308         return ret;
5309 }
5310 
5311 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5312                 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5313 {
5314         unsigned long long min_limit, min_memsw_limit, tmp;
5315 
5316         min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5317         min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5318         if (!memcg->use_hierarchy)
5319                 goto out;
5320 
5321         while (css_parent(&memcg->css)) {
5322                 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5323                 if (!memcg->use_hierarchy)
5324                         break;
5325                 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5326                 min_limit = min(min_limit, tmp);
5327                 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5328                 min_memsw_limit = min(min_memsw_limit, tmp);
5329         }
5330 out:
5331         *mem_limit = min_limit;
5332         *memsw_limit = min_memsw_limit;
5333 }
5334 
5335 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5336 {
5337         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5338         int name;
5339         enum res_type type;
5340 
5341         type = MEMFILE_TYPE(event);
5342         name = MEMFILE_ATTR(event);
5343 
5344         switch (name) {
5345         case RES_MAX_USAGE:
5346                 if (type == _MEM)
5347                         res_counter_reset_max(&memcg->res);
5348                 else if (type == _MEMSWAP)
5349                         res_counter_reset_max(&memcg->memsw);
5350                 else if (type == _KMEM)
5351                         res_counter_reset_max(&memcg->kmem);
5352                 else
5353                         return -EINVAL;
5354                 break;
5355         case RES_FAILCNT:
5356                 if (type == _MEM)
5357                         res_counter_reset_failcnt(&memcg->res);
5358                 else if (type == _MEMSWAP)
5359                         res_counter_reset_failcnt(&memcg->memsw);
5360                 else if (type == _KMEM)
5361                         res_counter_reset_failcnt(&memcg->kmem);
5362                 else
5363                         return -EINVAL;
5364                 break;
5365         }
5366 
5367         return 0;
5368 }
5369 
5370 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5371                                         struct cftype *cft)
5372 {
5373         return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5374 }
5375 
5376 #ifdef CONFIG_MMU
5377 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5378                                         struct cftype *cft, u64 val)
5379 {
5380         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5381 
5382         if (val >= (1 << NR_MOVE_TYPE))
5383                 return -EINVAL;
5384 
5385         /*
5386          * No kind of locking is needed in here, because ->can_attach() will
5387          * check this value once in the beginning of the process, and then carry
5388          * on with stale data. This means that changes to this value will only
5389          * affect task migrations starting after the change.
5390          */
5391         memcg->move_charge_at_immigrate = val;
5392         return 0;
5393 }
5394 #else
5395 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5396                                         struct cftype *cft, u64 val)
5397 {
5398         return -ENOSYS;
5399 }
5400 #endif
5401 
5402 #ifdef CONFIG_NUMA
5403 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5404                                 struct cftype *cft, struct seq_file *m)
5405 {
5406         struct numa_stat {
5407                 const char *name;
5408                 unsigned int lru_mask;
5409         };
5410 
5411         static const struct numa_stat stats[] = {
5412                 { "total", LRU_ALL },
5413                 { "file", LRU_ALL_FILE },
5414                 { "anon", LRU_ALL_ANON },
5415                 { "unevictable", BIT(LRU_UNEVICTABLE) },
5416         };
5417         const struct numa_stat *stat;
5418         int nid;
5419         unsigned long nr;
5420         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5421 
5422         for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5423                 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5424                 seq_printf(m, "%s=%lu", stat->name, nr);
5425                 for_each_node_state(nid, N_MEMORY) {
5426                         nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5427                                                           stat->lru_mask);
5428                         seq_printf(m, " N%d=%lu", nid, nr);
5429                 }
5430                 seq_putc(m, '\n');
5431         }
5432 
5433         for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5434                 struct mem_cgroup *iter;
5435 
5436                 nr = 0;
5437                 for_each_mem_cgroup_tree(iter, memcg)
5438                         nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5439                 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5440                 for_each_node_state(nid, N_MEMORY) {
5441                         nr = 0;
5442                         for_each_mem_cgroup_tree(iter, memcg)
5443                                 nr += mem_cgroup_node_nr_lru_pages(
5444                                         iter, nid, stat->lru_mask);
5445                         seq_printf(m, " N%d=%lu", nid, nr);
5446                 }
5447                 seq_putc(m, '\n');
5448         }
5449 
5450         return 0;
5451 }
5452 #endif /* CONFIG_NUMA */
5453 
5454 static inline void mem_cgroup_lru_names_not_uptodate(void)
5455 {
5456         BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5457 }
5458 
5459 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5460                                  struct seq_file *m)
5461 {
5462         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5463         struct mem_cgroup *mi;
5464         unsigned int i;
5465 
5466         for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5467                 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5468                         continue;
5469                 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5470                            mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5471         }
5472 
5473         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5474                 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5475                            mem_cgroup_read_events(memcg, i));
5476 
5477         for (i = 0; i < NR_LRU_LISTS; i++)
5478                 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5479                            mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5480 
5481         /* Hierarchical information */
5482         {
5483                 unsigned long long limit, memsw_limit;
5484                 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5485                 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5486                 if (do_swap_account)
5487                         seq_printf(m, "hierarchical_memsw_limit %llu\n",
5488                                    memsw_limit);
5489         }
5490 
5491         for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5492                 long long val = 0;
5493 
5494                 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5495                         continue;
5496                 for_each_mem_cgroup_tree(mi, memcg)
5497                         val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5498                 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5499         }
5500 
5501         for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5502                 unsigned long long val = 0;
5503 
5504                 for_each_mem_cgroup_tree(mi, memcg)
5505                         val += mem_cgroup_read_events(mi, i);
5506                 seq_printf(m, "total_%s %llu\n",
5507                            mem_cgroup_events_names[i], val);
5508         }
5509 
5510         for (i = 0; i < NR_LRU_LISTS; i++) {
5511                 unsigned long long val = 0;
5512 
5513                 for_each_mem_cgroup_tree(mi, memcg)
5514                         val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5515                 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5516         }
5517 
5518 #ifdef CONFIG_DEBUG_VM
5519         {
5520                 int nid, zid;
5521                 struct mem_cgroup_per_zone *mz;
5522                 struct zone_reclaim_stat *rstat;
5523                 unsigned long recent_rotated[2] = {0, 0};
5524                 unsigned long recent_scanned[2] = {0, 0};
5525 
5526                 for_each_online_node(nid)
5527                         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5528                                 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5529                                 rstat = &mz->lruvec.reclaim_stat;
5530 
5531                                 recent_rotated[0] += rstat->recent_rotated[0];
5532                                 recent_rotated[1] += rstat->recent_rotated[1];
5533                                 recent_scanned[0] += rstat->recent_scanned[0];
5534                                 recent_scanned[1] += rstat->recent_scanned[1];
5535                         }
5536                 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5537                 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5538                 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5539                 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5540         }
5541 #endif
5542 
5543         return 0;
5544 }
5545 
5546 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5547                                       struct cftype *cft)
5548 {
5549         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5550 
5551         return mem_cgroup_swappiness(memcg);
5552 }
5553 
5554 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5555                                        struct cftype *cft, u64 val)
5556 {
5557         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5558         struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5559 
5560         if (val > 100 || !parent)
5561                 return -EINVAL;
5562 
5563         mutex_lock(&memcg_create_mutex);
5564 
5565         /* If under hierarchy, only empty-root can set this value */
5566         if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5567                 mutex_unlock(&memcg_create_mutex);
5568                 return -EINVAL;
5569         }
5570 
5571         memcg->swappiness = val;
5572 
5573         mutex_unlock(&memcg_create_mutex);
5574 
5575         return 0;
5576 }
5577 
5578 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5579 {
5580         struct mem_cgroup_threshold_ary *t;
5581         u64 usage;
5582         int i;
5583 
5584         rcu_read_lock();
5585         if (!swap)
5586                 t = rcu_dereference(memcg->thresholds.primary);
5587         else
5588                 t = rcu_dereference(memcg->memsw_thresholds.primary);
5589 
5590         if (!t)
5591                 goto unlock;
5592 
5593         usage = mem_cgroup_usage(memcg, swap);
5594 
5595         /*
5596          * current_threshold points to threshold just below or equal to usage.
5597          * If it's not true, a threshold was crossed after last
5598          * call of __mem_cgroup_threshold().
5599          */
5600         i = t->current_threshold;
5601 
5602         /*
5603          * Iterate backward over array of thresholds starting from
5604          * current_threshold and check if a threshold is crossed.
5605          * If none of thresholds below usage is crossed, we read
5606          * only one element of the array here.
5607          */
5608         for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5609                 eventfd_signal(t->entries[i].eventfd, 1);
5610 
5611         /* i = current_threshold + 1 */
5612         i++;
5613 
5614         /*
5615          * Iterate forward over array of thresholds starting from
5616          * current_threshold+1 and check if a threshold is crossed.
5617          * If none of thresholds above usage is crossed, we read
5618          * only one element of the array here.
5619          */
5620         for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5621                 eventfd_signal(t->entries[i].eventfd, 1);
5622 
5623         /* Update current_threshold */
5624         t->current_threshold = i - 1;
5625 unlock:
5626         rcu_read_unlock();
5627 }
5628 
5629 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5630 {
5631         while (memcg) {
5632                 __mem_cgroup_threshold(memcg, false);
5633                 if (do_swap_account)
5634                         __mem_cgroup_threshold(memcg, true);
5635 
5636                 memcg = parent_mem_cgroup(memcg);
5637         }
5638 }
5639 
5640 static int compare_thresholds(const void *a, const void *b)
5641 {
5642         const struct mem_cgroup_threshold *_a = a;
5643         const struct mem_cgroup_threshold *_b = b;
5644 
5645         if (_a->threshold > _b->threshold)
5646                 return 1;
5647 
5648         if (_a->threshold < _b->threshold)
5649                 return -1;
5650 
5651         return 0;
5652 }
5653 
5654 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5655 {
5656         struct mem_cgroup_eventfd_list *ev;
5657 
5658         list_for_each_entry(ev, &memcg->oom_notify, list)
5659                 eventfd_signal(ev->eventfd, 1);
5660         return 0;
5661 }
5662 
5663 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5664 {
5665         struct mem_cgroup *iter;
5666 
5667         for_each_mem_cgroup_tree(iter, memcg)
5668                 mem_cgroup_oom_notify_cb(iter);
5669 }
5670 
5671 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5672         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5673 {
5674         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5675         struct mem_cgroup_thresholds *thresholds;
5676         struct mem_cgroup_threshold_ary *new;
5677         enum res_type type = MEMFILE_TYPE(cft->private);
5678         u64 threshold, usage;
5679         int i, size, ret;
5680 
5681         ret = res_counter_memparse_write_strategy(args, &threshold);
5682         if (ret)
5683                 return ret;
5684 
5685         mutex_lock(&memcg->thresholds_lock);
5686 
5687         if (type == _MEM)
5688                 thresholds = &memcg->thresholds;
5689         else if (type == _MEMSWAP)
5690                 thresholds = &memcg->memsw_thresholds;
5691         else
5692                 BUG();
5693 
5694         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5695 
5696         /* Check if a threshold crossed before adding a new one */
5697         if (thresholds->primary)
5698                 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5699 
5700         size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5701 
5702         /* Allocate memory for new array of thresholds */
5703         new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5704                         GFP_KERNEL);
5705         if (!new) {
5706                 ret = -ENOMEM;
5707                 goto unlock;
5708         }
5709         new->size = size;
5710 
5711         /* Copy thresholds (if any) to new array */
5712         if (thresholds->primary) {
5713                 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5714                                 sizeof(struct mem_cgroup_threshold));
5715         }
5716 
5717         /* Add new threshold */
5718         new->entries[size - 1].eventfd = eventfd;
5719         new->entries[size - 1].threshold = threshold;
5720 
5721         /* Sort thresholds. Registering of new threshold isn't time-critical */
5722         sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5723                         compare_thresholds, NULL);
5724 
5725         /* Find current threshold */
5726         new->current_threshold = -1;
5727         for (i = 0; i < size; i++) {
5728                 if (new->entries[i].threshold <= usage) {
5729                         /*
5730                          * new->current_threshold will not be used until
5731                          * rcu_assign_pointer(), so it's safe to increment
5732                          * it here.
5733                          */
5734                         ++new->current_threshold;
5735                 } else
5736                         break;
5737         }
5738 
5739         /* Free old spare buffer and save old primary buffer as spare */
5740         kfree(thresholds->spare);
5741         thresholds->spare = thresholds->primary;
5742 
5743         rcu_assign_pointer(thresholds->primary, new);
5744 
5745         /* To be sure that nobody uses thresholds */
5746         synchronize_rcu();
5747 
5748 unlock:
5749         mutex_unlock(&memcg->thresholds_lock);
5750 
5751         return ret;
5752 }
5753 
5754 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5755         struct cftype *cft, struct eventfd_ctx *eventfd)
5756 {
5757         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5758         struct mem_cgroup_thresholds *thresholds;
5759         struct mem_cgroup_threshold_ary *new;
5760         enum res_type type = MEMFILE_TYPE(cft->private);
5761         u64 usage;
5762         int i, j, size;
5763 
5764         mutex_lock(&memcg->thresholds_lock);
5765         if (type == _MEM)
5766                 thresholds = &memcg->thresholds;
5767         else if (type == _MEMSWAP)
5768                 thresholds = &memcg->memsw_thresholds;
5769         else
5770                 BUG();
5771 
5772         if (!thresholds->primary)
5773                 goto unlock;
5774 
5775         usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5776 
5777         /* Check if a threshold crossed before removing */
5778         __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5779 
5780         /* Calculate new number of threshold */
5781         size = 0;
5782         for (i = 0; i < thresholds->primary->size; i++) {
5783                 if (thresholds->primary->entries[i].eventfd != eventfd)
5784                         size++;
5785         }
5786 
5787         new = thresholds->spare;
5788 
5789         /* Set thresholds array to NULL if we don't have thresholds */
5790         if (!size) {
5791                 kfree(new);
5792                 new = NULL;
5793                 goto swap_buffers;
5794         }
5795 
5796         new->size = size;
5797 
5798         /* Copy thresholds and find current threshold */
5799         new->current_threshold = -1;
5800         for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5801                 if (thresholds->primary->entries[i].eventfd == eventfd)
5802                         continue;
5803 
5804                 new->entries[j] = thresholds->primary->entries[i];
5805                 if (new->entries[j].threshold <= usage) {
5806                         /*
5807                          * new->current_threshold will not be used
5808                          * until rcu_assign_pointer(), so it's safe to increment
5809                          * it here.
5810                          */
5811                         ++new->current_threshold;
5812                 }
5813                 j++;
5814         }
5815 
5816 swap_buffers:
5817         /* Swap primary and spare array */
5818         thresholds->spare = thresholds->primary;
5819         /* If all events are unregistered, free the spare array */
5820         if (!new) {
5821                 kfree(thresholds->spare);
5822                 thresholds->spare = NULL;
5823         }
5824 
5825         rcu_assign_pointer(thresholds->primary, new);
5826 
5827         /* To be sure that nobody uses thresholds */
5828         synchronize_rcu();
5829 unlock:
5830         mutex_unlock(&memcg->thresholds_lock);
5831 }
5832 
5833 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5834         struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5835 {
5836         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5837         struct mem_cgroup_eventfd_list *event;
5838         enum res_type type = MEMFILE_TYPE(cft->private);
5839 
5840         BUG_ON(type != _OOM_TYPE);
5841         event = kmalloc(sizeof(*event), GFP_KERNEL);
5842         if (!event)
5843                 return -ENOMEM;
5844 
5845         spin_lock(&memcg_oom_lock);
5846 
5847         event->eventfd = eventfd;
5848         list_add(&event->list, &memcg->oom_notify);
5849 
5850         /* already in OOM ? */
5851         if (atomic_read(&memcg->under_oom))
5852                 eventfd_signal(eventfd, 1);
5853         spin_unlock(&memcg_oom_lock);
5854 
5855         return 0;
5856 }
5857 
5858 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5859         struct cftype *cft, struct eventfd_ctx *eventfd)
5860 {
5861         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5862         struct mem_cgroup_eventfd_list *ev, *tmp;
5863         enum res_type type = MEMFILE_TYPE(cft->private);
5864 
5865         BUG_ON(type != _OOM_TYPE);
5866 
5867         spin_lock(&memcg_oom_lock);
5868 
5869         list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5870                 if (ev->eventfd == eventfd) {
5871                         list_del(&ev->list);
5872                         kfree(ev);
5873                 }
5874         }
5875 
5876         spin_unlock(&memcg_oom_lock);
5877 }
5878 
5879 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5880         struct cftype *cft,  struct cgroup_map_cb *cb)
5881 {
5882         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5883 
5884         cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5885 
5886         if (atomic_read(&memcg->under_oom))
5887                 cb->fill(cb, "under_oom", 1);
5888         else
5889                 cb->fill(cb, "under_oom", 0);
5890         return 0;
5891 }
5892 
5893 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5894         struct cftype *cft, u64 val)
5895 {
5896         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5897         struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5898 
5899         /* cannot set to root cgroup and only 0 and 1 are allowed */
5900         if (!parent || !((val == 0) || (val == 1)))
5901                 return -EINVAL;
5902 
5903         mutex_lock(&memcg_create_mutex);
5904         /* oom-kill-disable is a flag for subhierarchy. */
5905         if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5906                 mutex_unlock(&memcg_create_mutex);
5907                 return -EINVAL;
5908         }
5909         memcg->oom_kill_disable = val;
5910         if (!val)
5911                 memcg_oom_recover(memcg);
5912         mutex_unlock(&memcg_create_mutex);
5913         return 0;
5914 }
5915 
5916 #ifdef CONFIG_MEMCG_KMEM
5917 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5918 {
5919         int ret;
5920 
5921         memcg->kmemcg_id = -1;
5922         ret = memcg_propagate_kmem(memcg);
5923         if (ret)
5924                 return ret;
5925 
5926         return mem_cgroup_sockets_init(memcg, ss);
5927 }
5928 
5929 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5930 {
5931         mem_cgroup_sockets_destroy(memcg);
5932 }
5933 
5934 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5935 {
5936         if (!memcg_kmem_is_active(memcg))
5937                 return;
5938 
5939         /*
5940          * kmem charges can outlive the cgroup. In the case of slab
5941          * pages, for instance, a page contain objects from various
5942          * processes. As we prevent from taking a reference for every
5943          * such allocation we have to be careful when doing uncharge
5944          * (see memcg_uncharge_kmem) and here during offlining.
5945          *
5946          * The idea is that that only the _last_ uncharge which sees
5947          * the dead memcg will drop the last reference. An additional
5948          * reference is taken here before the group is marked dead
5949          * which is then paired with css_put during uncharge resp. here.
5950          *
5951          * Although this might sound strange as this path is called from
5952          * css_offline() when the referencemight have dropped down to 0
5953          * and shouldn't be incremented anymore (css_tryget would fail)
5954          * we do not have other options because of the kmem allocations
5955          * lifetime.
5956          */
5957         css_get(&memcg->css);
5958 
5959         memcg_kmem_mark_dead(memcg);
5960 
5961         if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5962                 return;
5963 
5964         if (memcg_kmem_test_and_clear_dead(memcg))
5965                 css_put(&memcg->css);
5966 }
5967 #else
5968 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5969 {
5970         return 0;
5971 }
5972 
5973 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5974 {
5975 }
5976 
5977 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5978 {
5979 }
5980 #endif
5981 
5982 static struct cftype mem_cgroup_files[] = {
5983         {
5984                 .name = "usage_in_bytes",
5985                 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5986                 .read = mem_cgroup_read,
5987                 .register_event = mem_cgroup_usage_register_event,
5988                 .unregister_event = mem_cgroup_usage_unregister_event,
5989         },
5990         {
5991                 .name = "max_usage_in_bytes",
5992                 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5993                 .trigger = mem_cgroup_reset,
5994                 .read = mem_cgroup_read,
5995         },
5996         {
5997                 .name = "limit_in_bytes",
5998                 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5999                 .write_string = mem_cgroup_write,
6000                 .read = mem_cgroup_read,
6001         },
6002         {
6003                 .name = "soft_limit_in_bytes",
6004                 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6005                 .write_string = mem_cgroup_write,
6006                 .read = mem_cgroup_read,
6007         },
6008         {
6009                 .name = "failcnt",
6010                 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6011                 .trigger = mem_cgroup_reset,
6012                 .read = mem_cgroup_read,
6013         },
6014         {
6015                 .name = "stat",
6016                 .read_seq_string = memcg_stat_show,
6017         },
6018         {
6019                 .name = "force_empty",
6020                 .trigger = mem_cgroup_force_empty_write,
6021         },
6022         {
6023                 .name = "use_hierarchy",
6024                 .flags = CFTYPE_INSANE,
6025                 .write_u64 = mem_cgroup_hierarchy_write,
6026                 .read_u64 = mem_cgroup_hierarchy_read,
6027         },
6028         {
6029                 .name = "swappiness",
6030                 .read_u64 = mem_cgroup_swappiness_read,
6031                 .write_u64 = mem_cgroup_swappiness_write,
6032         },
6033         {
6034                 .name = "move_charge_at_immigrate",
6035                 .read_u64 = mem_cgroup_move_charge_read,
6036                 .write_u64 = mem_cgroup_move_charge_write,
6037         },
6038         {
6039                 .name = "oom_control",
6040                 .read_map = mem_cgroup_oom_control_read,
6041                 .write_u64 = mem_cgroup_oom_control_write,
6042                 .register_event = mem_cgroup_oom_register_event,
6043                 .unregister_event = mem_cgroup_oom_unregister_event,
6044                 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6045         },
6046         {
6047                 .name = "pressure_level",
6048                 .register_event = vmpressure_register_event,
6049                 .unregister_event = vmpressure_unregister_event,
6050         },
6051 #ifdef CONFIG_NUMA
6052         {
6053                 .name = "numa_stat",
6054                 .read_seq_string = memcg_numa_stat_show,
6055         },
6056 #endif
6057 #ifdef CONFIG_MEMCG_KMEM
6058         {
6059                 .name = "kmem.limit_in_bytes",
6060                 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6061                 .write_string = mem_cgroup_write,
6062                 .read = mem_cgroup_read,
6063         },
6064         {
6065                 .name = "kmem.usage_in_bytes",
6066                 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6067                 .read = mem_cgroup_read,
6068         },
6069         {
6070                 .name = "kmem.failcnt",
6071                 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6072                 .trigger = mem_cgroup_reset,
6073                 .read = mem_cgroup_read,
6074         },
6075         {
6076                 .name = "kmem.max_usage_in_bytes",
6077                 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6078                 .trigger = mem_cgroup_reset,
6079                 .read = mem_cgroup_read,
6080         },
6081 #ifdef CONFIG_SLABINFO
6082         {
6083                 .name = "kmem.slabinfo",
6084                 .read_seq_string = mem_cgroup_slabinfo_read,
6085         },
6086 #endif
6087 #endif
6088         { },    /* terminate */
6089 };
6090 
6091 #ifdef CONFIG_MEMCG_SWAP
6092 static struct cftype memsw_cgroup_files[] = {
6093         {
6094                 .name = "memsw.usage_in_bytes",
6095                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6096                 .read = mem_cgroup_read,
6097                 .register_event = mem_cgroup_usage_register_event,
6098                 .unregister_event = mem_cgroup_usage_unregister_event,
6099         },
6100         {
6101                 .name = "memsw.max_usage_in_bytes",
6102                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6103                 .trigger = mem_cgroup_reset,
6104                 .read = mem_cgroup_read,
6105         },
6106         {
6107                 .name = "memsw.limit_in_bytes",
6108                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6109                 .write_string = mem_cgroup_write,
6110                 .read = mem_cgroup_read,
6111         },
6112         {
6113                 .name = "memsw.failcnt",
6114                 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6115                 .trigger = mem_cgroup_reset,
6116                 .read = mem_cgroup_read,
6117         },
6118         { },    /* terminate */
6119 };
6120 #endif
6121 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6122 {
6123         struct mem_cgroup_per_node *pn;
6124         struct mem_cgroup_per_zone *mz;
6125         int zone, tmp = node;
6126         /*
6127          * This routine is called against possible nodes.
6128          * But it's BUG to call kmalloc() against offline node.
6129          *
6130          * TODO: this routine can waste much memory for nodes which will
6131          *       never be onlined. It's better to use memory hotplug callback
6132          *       function.
6133          */
6134         if (!node_state(node, N_NORMAL_MEMORY))
6135                 tmp = -1;
6136         pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6137         if (!pn)
6138                 return 1;
6139 
6140         for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6141                 mz = &pn->zoneinfo[zone];
6142                 lruvec_init(&mz->lruvec);
6143                 mz->usage_in_excess = 0;
6144                 mz->on_tree = false;
6145                 mz->memcg = memcg;
6146         }
6147         memcg->nodeinfo[node] = pn;
6148         return 0;
6149 }
6150 
6151 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6152 {
6153         kfree(memcg->nodeinfo[node]);
6154 }
6155 
6156 static struct mem_cgroup *mem_cgroup_alloc(void)
6157 {
6158         struct mem_cgroup *memcg;
6159         size_t size = memcg_size();
6160 
6161         /* Can be very big if nr_node_ids is very big */
6162         if (size < PAGE_SIZE)
6163                 memcg = kzalloc(size, GFP_KERNEL);
6164         else
6165                 memcg = vzalloc(size);
6166 
6167         if (!memcg)
6168                 return NULL;
6169 
6170         memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6171         if (!memcg->stat)
6172                 goto out_free;
6173         spin_lock_init(&memcg->pcp_counter_lock);
6174         return memcg;
6175 
6176 out_free:
6177         if (size < PAGE_SIZE)
6178                 kfree(memcg);
6179         else
6180                 vfree(memcg);
6181         return NULL;
6182 }
6183 
6184 /*
6185  * At destroying mem_cgroup, references from swap_cgroup can remain.
6186  * (scanning all at force_empty is too costly...)
6187  *
6188  * Instead of clearing all references at force_empty, we remember
6189  * the number of reference from swap_cgroup and free mem_cgroup when
6190  * it goes down to 0.
6191  *
6192  * Removal of cgroup itself succeeds regardless of refs from swap.
6193  */
6194 
6195 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6196 {
6197         int node;
6198         size_t size = memcg_size();
6199 
6200         mem_cgroup_remove_from_trees(memcg);
6201 
6202         for_each_node(node)
6203                 free_mem_cgroup_per_zone_info(memcg, node);
6204 
6205         free_percpu(memcg->stat);
6206 
6207         /*
6208          * We need to make sure that (at least for now), the jump label
6209          * destruction code runs outside of the cgroup lock. This is because
6210          * get_online_cpus(), which is called from the static_branch update,
6211          * can't be called inside the cgroup_lock. cpusets are the ones
6212          * enforcing this dependency, so if they ever change, we might as well.
6213          *
6214          * schedule_work() will guarantee this happens. Be careful if you need
6215          * to move this code around, and make sure it is outside
6216          * the cgroup_lock.
6217          */
6218         disarm_static_keys(memcg);
6219         if (size < PAGE_SIZE)
6220                 kfree(memcg);
6221         else
6222                 vfree(memcg);
6223 }
6224 
6225 /*
6226  * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6227  */
6228 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6229 {
6230         if (!memcg->res.parent)
6231                 return NULL;
6232         return mem_cgroup_from_res_counter(memcg->res.parent, res);
6233 }
6234 EXPORT_SYMBOL(parent_mem_cgroup);
6235 
6236 static void __init mem_cgroup_soft_limit_tree_init(void)
6237 {
6238         struct mem_cgroup_tree_per_node *rtpn;
6239         struct mem_cgroup_tree_per_zone *rtpz;
6240         int tmp, node, zone;
6241 
6242         for_each_node(node) {
6243                 tmp = node;
6244                 if (!node_state(node, N_NORMAL_MEMORY))
6245                         tmp = -1;
6246                 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6247                 BUG_ON(!rtpn);
6248 
6249                 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6250 
6251                 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6252                         rtpz = &rtpn->rb_tree_per_zone[zone];
6253                         rtpz->rb_root = RB_ROOT;
6254                         spin_lock_init(&rtpz->lock);
6255                 }
6256         }
6257 }
6258 
6259 static struct cgroup_subsys_state * __ref
6260 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6261 {
6262         struct mem_cgroup *memcg;
6263         long error = -ENOMEM;
6264         int node;
6265 
6266         memcg = mem_cgroup_alloc();
6267         if (!memcg)
6268                 return ERR_PTR(error);
6269 
6270         for_each_node(node)
6271                 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6272                         goto free_out;
6273 
6274         /* root ? */
6275         if (parent_css == NULL) {
6276                 root_mem_cgroup = memcg;
6277                 res_counter_init(&memcg->res, NULL);
6278                 res_counter_init(&memcg->memsw, NULL);
6279                 res_counter_init(&memcg->kmem, NULL);
6280         }
6281 
6282         memcg->last_scanned_node = MAX_NUMNODES;
6283         INIT_LIST_HEAD(&memcg->oom_notify);
6284         memcg->move_charge_at_immigrate = 0;
6285         mutex_init(&memcg->thresholds_lock);
6286         spin_lock_init(&memcg->move_lock);
6287         vmpressure_init(&memcg->vmpressure);
6288 
6289         return &memcg->css;
6290 
6291 free_out:
6292         __mem_cgroup_free(memcg);
6293         return ERR_PTR(error);
6294 }
6295 
6296 static int
6297 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6298 {
6299         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6300         struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6301         int error = 0;
6302 
6303         if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6304                 return -ENOSPC;
6305 
6306         if (!parent)
6307                 return 0;
6308 
6309         mutex_lock(&memcg_create_mutex);
6310 
6311         memcg->use_hierarchy = parent->use_hierarchy;
6312         memcg->oom_kill_disable = parent->oom_kill_disable;
6313         memcg->swappiness = mem_cgroup_swappiness(parent);
6314 
6315         if (parent->use_hierarchy) {
6316                 res_counter_init(&memcg->res, &parent->res);
6317                 res_counter_init(&memcg->memsw, &parent->memsw);
6318                 res_counter_init(&memcg->kmem, &parent->kmem);
6319 
6320                 /*
6321                  * No need to take a reference to the parent because cgroup
6322                  * core guarantees its existence.
6323                  */
6324         } else {
6325                 res_counter_init(&memcg->res, NULL);
6326                 res_counter_init(&memcg->memsw, NULL);
6327                 res_counter_init(&memcg->kmem, NULL);
6328                 /*
6329                  * Deeper hierachy with use_hierarchy == false doesn't make
6330                  * much sense so let cgroup subsystem know about this
6331                  * unfortunate state in our controller.
6332                  */
6333                 if (parent != root_mem_cgroup)
6334                         mem_cgroup_subsys.broken_hierarchy = true;
6335         }
6336 
6337         error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6338         mutex_unlock(&memcg_create_mutex);
6339         return error;
6340 }
6341 
6342 /*
6343  * Announce all parents that a group from their hierarchy is gone.
6344  */
6345 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6346 {
6347         struct mem_cgroup *parent = memcg;
6348 
6349         while ((parent = parent_mem_cgroup(parent)))
6350                 mem_cgroup_iter_invalidate(parent);
6351 
6352         /*
6353          * if the root memcg is not hierarchical we have to check it
6354          * explicitely.
6355          */
6356         if (!root_mem_cgroup->use_hierarchy)
6357                 mem_cgroup_iter_invalidate(root_mem_cgroup);
6358 }
6359 
6360 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6361 {
6362         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6363         struct cgroup_subsys_state *iter;
6364 
6365         kmem_cgroup_css_offline(memcg);
6366 
6367         mem_cgroup_invalidate_reclaim_iterators(memcg);
6368 
6369         /*
6370          * This requires that offlining is serialized.  Right now that is
6371          * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6372          */
6373         rcu_read_lock();
6374         css_for_each_descendant_post(iter, css) {
6375                 rcu_read_unlock();
6376                 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6377                 rcu_read_lock();
6378         }
6379         rcu_read_unlock();
6380 
6381         mem_cgroup_destroy_all_caches(memcg);
6382         vmpressure_cleanup(&memcg->vmpressure);
6383 }
6384 
6385 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6386 {
6387         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6388         /*
6389          * XXX: css_offline() would be where we should reparent all
6390          * memory to prepare the cgroup for destruction.  However,
6391          * memcg does not do css_tryget() and res_counter charging
6392          * under the same RCU lock region, which means that charging
6393          * could race with offlining.  Offlining only happens to
6394          * cgroups with no tasks in them but charges can show up
6395          * without any tasks from the swapin path when the target
6396          * memcg is looked up from the swapout record and not from the
6397          * current task as it usually is.  A race like this can leak
6398          * charges and put pages with stale cgroup pointers into
6399          * circulation:
6400          *
6401          * #0                        #1
6402          *                           lookup_swap_cgroup_id()
6403          *                           rcu_read_lock()
6404          *                           mem_cgroup_lookup()
6405          *                           css_tryget()
6406          *                           rcu_read_unlock()
6407          * disable css_tryget()
6408          * call_rcu()
6409          *   offline_css()
6410          *     reparent_charges()
6411          *                           res_counter_charge()
6412          *                           css_put()
6413          *                             css_free()
6414          *                           pc->mem_cgroup = dead memcg
6415          *                           add page to lru
6416          *
6417          * The bulk of the charges are still moved in offline_css() to
6418          * avoid pinning a lot of pages in case a long-term reference
6419          * like a swapout record is deferring the css_free() to long
6420          * after offlining.  But this makes sure we catch any charges
6421          * made after offlining:
6422          */
6423         mem_cgroup_reparent_charges(memcg);
6424 
6425         memcg_destroy_kmem(memcg);
6426         __mem_cgroup_free(memcg);
6427 }
6428 
6429 #ifdef CONFIG_MMU
6430 /* Handlers for move charge at task migration. */
6431 #define PRECHARGE_COUNT_AT_ONCE 256
6432 static int mem_cgroup_do_precharge(unsigned long count)
6433 {
6434         int ret = 0;
6435         int batch_count = PRECHARGE_COUNT_AT_ONCE;
6436         struct mem_cgroup *memcg = mc.to;
6437 
6438         if (mem_cgroup_is_root(memcg)) {
6439                 mc.precharge += count;
6440                 /* we don't need css_get for root */
6441                 return ret;
6442         }
6443         /* try to charge at once */
6444         if (count > 1) {
6445                 struct res_counter *dummy;
6446                 /*
6447                  * "memcg" cannot be under rmdir() because we've already checked
6448                  * by cgroup_lock_live_cgroup() that it is not removed and we
6449                  * are still under the same cgroup_mutex. So we can postpone
6450                  * css_get().
6451                  */
6452                 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6453                         goto one_by_one;
6454                 if (do_swap_account && res_counter_charge(&memcg->memsw,
6455                                                 PAGE_SIZE * count, &dummy)) {
6456                         res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6457                         goto one_by_one;
6458                 }
6459                 mc.precharge += count;
6460                 return ret;
6461         }
6462 one_by_one:
6463         /* fall back to one by one charge */
6464         while (count--) {
6465                 if (signal_pending(current)) {
6466                         ret = -EINTR;
6467                         break;
6468                 }
6469                 if (!batch_count--) {
6470                         batch_count = PRECHARGE_COUNT_AT_ONCE;
6471                         cond_resched();
6472                 }
6473                 ret = __mem_cgroup_try_charge(NULL,
6474                                         GFP_KERNEL, 1, &memcg, false);
6475                 if (ret)
6476                         /* mem_cgroup_clear_mc() will do uncharge later */
6477                         return ret;
6478                 mc.precharge++;
6479         }
6480         return ret;
6481 }
6482 
6483 /**
6484  * get_mctgt_type - get target type of moving charge
6485  * @vma: the vma the pte to be checked belongs
6486  * @addr: the address corresponding to the pte to be checked
6487  * @ptent: the pte to be checked
6488  * @target: the pointer the target page or swap ent will be stored(can be NULL)
6489  *
6490  * Returns
6491  *   0(MC_TARGET_NONE): if the pte is not a target for move charge.
6492  *   1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6493  *     move charge. if @target is not NULL, the page is stored in target->page
6494  *     with extra refcnt got(Callers should handle it).
6495  *   2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6496  *     target for charge migration. if @target is not NULL, the entry is stored
6497  *     in target->ent.
6498  *
6499  * Called with pte lock held.
6500  */
6501 union mc_target {
6502         struct page     *page;
6503