1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
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.
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.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.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>
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>
60 #include <net/tcp_memcontrol.h>
62 #include <asm/uaccess.h>
64 #include <trace/events/vmscan.h>
66 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
67 EXPORT_SYMBOL(mem_cgroup_subsys);
69 #define MEM_CGROUP_RECLAIM_RETRIES 5
70 static struct mem_cgroup *root_mem_cgroup __read_mostly;
72 #ifdef CONFIG_MEMCG_SWAP
73 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
74 int do_swap_account __read_mostly;
76 /* for remember boot option*/
77 #ifdef CONFIG_MEMCG_SWAP_ENABLED
78 static int really_do_swap_account __initdata = 1;
80 static int really_do_swap_account __initdata = 0;
84 #define do_swap_account 0
88 static const char * const mem_cgroup_stat_names[] = {
97 enum mem_cgroup_events_index {
98 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
99 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
100 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
101 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
102 MEM_CGROUP_EVENTS_NSTATS,
105 static const char * const mem_cgroup_events_names[] = {
112 static const char * const mem_cgroup_lru_names[] = {
121 * Per memcg event counter is incremented at every pagein/pageout. With THP,
122 * it will be incremated by the number of pages. This counter is used for
123 * for trigger some periodic events. This is straightforward and better
124 * than using jiffies etc. to handle periodic memcg event.
126 enum mem_cgroup_events_target {
127 MEM_CGROUP_TARGET_THRESH,
128 MEM_CGROUP_TARGET_SOFTLIMIT,
129 MEM_CGROUP_TARGET_NUMAINFO,
132 #define THRESHOLDS_EVENTS_TARGET 128
133 #define SOFTLIMIT_EVENTS_TARGET 1024
134 #define NUMAINFO_EVENTS_TARGET 1024
136 struct mem_cgroup_stat_cpu {
137 long count[MEM_CGROUP_STAT_NSTATS];
138 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
139 unsigned long nr_page_events;
140 unsigned long targets[MEM_CGROUP_NTARGETS];
143 struct mem_cgroup_reclaim_iter {
145 * last scanned hierarchy member. Valid only if last_dead_count
146 * matches memcg->dead_count of the hierarchy root group.
148 struct mem_cgroup *last_visited;
149 unsigned long last_dead_count;
151 /* scan generation, increased every round-trip */
152 unsigned int generation;
156 * per-zone information in memory controller.
158 struct mem_cgroup_per_zone {
159 struct lruvec lruvec;
160 unsigned long lru_size[NR_LRU_LISTS];
162 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
164 struct rb_node tree_node; /* RB tree node */
165 unsigned long long usage_in_excess;/* Set to the value by which */
166 /* the soft limit is exceeded*/
168 struct mem_cgroup *memcg; /* Back pointer, we cannot */
169 /* use container_of */
172 struct mem_cgroup_per_node {
173 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
177 * Cgroups above their limits are maintained in a RB-Tree, independent of
178 * their hierarchy representation
181 struct mem_cgroup_tree_per_zone {
182 struct rb_root rb_root;
186 struct mem_cgroup_tree_per_node {
187 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
190 struct mem_cgroup_tree {
191 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
194 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
196 struct mem_cgroup_threshold {
197 struct eventfd_ctx *eventfd;
202 struct mem_cgroup_threshold_ary {
203 /* An array index points to threshold just below or equal to usage. */
204 int current_threshold;
205 /* Size of entries[] */
207 /* Array of thresholds */
208 struct mem_cgroup_threshold entries[0];
211 struct mem_cgroup_thresholds {
212 /* Primary thresholds array */
213 struct mem_cgroup_threshold_ary *primary;
215 * Spare threshold array.
216 * This is needed to make mem_cgroup_unregister_event() "never fail".
217 * It must be able to store at least primary->size - 1 entries.
219 struct mem_cgroup_threshold_ary *spare;
223 struct mem_cgroup_eventfd_list {
224 struct list_head list;
225 struct eventfd_ctx *eventfd;
228 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
229 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
232 * The memory controller data structure. The memory controller controls both
233 * page cache and RSS per cgroup. We would eventually like to provide
234 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
235 * to help the administrator determine what knobs to tune.
237 * TODO: Add a water mark for the memory controller. Reclaim will begin when
238 * we hit the water mark. May be even add a low water mark, such that
239 * no reclaim occurs from a cgroup at it's low water mark, this is
240 * a feature that will be implemented much later in the future.
243 struct cgroup_subsys_state css;
245 * the counter to account for memory usage
247 struct res_counter res;
249 /* vmpressure notifications */
250 struct vmpressure vmpressure;
253 * the counter to account for mem+swap usage.
255 struct res_counter memsw;
258 * the counter to account for kernel memory usage.
260 struct res_counter kmem;
262 * Should the accounting and control be hierarchical, per subtree?
265 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
269 atomic_t oom_wakeups;
272 /* OOM-Killer disable */
273 int oom_kill_disable;
275 /* set when res.limit == memsw.limit */
276 bool memsw_is_minimum;
278 /* protect arrays of thresholds */
279 struct mutex thresholds_lock;
281 /* thresholds for memory usage. RCU-protected */
282 struct mem_cgroup_thresholds thresholds;
284 /* thresholds for mem+swap usage. RCU-protected */
285 struct mem_cgroup_thresholds memsw_thresholds;
287 /* For oom notifier event fd */
288 struct list_head oom_notify;
291 * Should we move charges of a task when a task is moved into this
292 * mem_cgroup ? And what type of charges should we move ?
294 unsigned long move_charge_at_immigrate;
296 * set > 0 if pages under this cgroup are moving to other cgroup.
298 atomic_t moving_account;
299 /* taken only while moving_account > 0 */
300 spinlock_t move_lock;
304 struct mem_cgroup_stat_cpu __percpu *stat;
306 * used when a cpu is offlined or other synchronizations
307 * See mem_cgroup_read_stat().
309 struct mem_cgroup_stat_cpu nocpu_base;
310 spinlock_t pcp_counter_lock;
313 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
314 struct tcp_memcontrol tcp_mem;
316 #if defined(CONFIG_MEMCG_KMEM)
317 /* analogous to slab_common's slab_caches list. per-memcg */
318 struct list_head memcg_slab_caches;
319 /* Not a spinlock, we can take a lot of time walking the list */
320 struct mutex slab_caches_mutex;
321 /* Index in the kmem_cache->memcg_params->memcg_caches array */
325 int last_scanned_node;
327 nodemask_t scan_nodes;
328 atomic_t numainfo_events;
329 atomic_t numainfo_updating;
332 struct mem_cgroup_per_node *nodeinfo[0];
333 /* WARNING: nodeinfo must be the last member here */
336 static size_t memcg_size(void)
338 return sizeof(struct mem_cgroup) +
339 nr_node_ids * sizeof(struct mem_cgroup_per_node);
342 /* internal only representation about the status of kmem accounting. */
344 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
345 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
346 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
349 /* We account when limit is on, but only after call sites are patched */
350 #define KMEM_ACCOUNTED_MASK \
351 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
353 #ifdef CONFIG_MEMCG_KMEM
354 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
356 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
359 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
361 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
364 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
366 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
369 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
371 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
374 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
377 * Our caller must use css_get() first, because memcg_uncharge_kmem()
378 * will call css_put() if it sees the memcg is dead.
381 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
382 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
385 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
387 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
388 &memcg->kmem_account_flags);
392 /* Stuffs for move charges at task migration. */
394 * Types of charges to be moved. "move_charge_at_immitgrate" and
395 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
398 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
399 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
403 /* "mc" and its members are protected by cgroup_mutex */
404 static struct move_charge_struct {
405 spinlock_t lock; /* for from, to */
406 struct mem_cgroup *from;
407 struct mem_cgroup *to;
408 unsigned long immigrate_flags;
409 unsigned long precharge;
410 unsigned long moved_charge;
411 unsigned long moved_swap;
412 struct task_struct *moving_task; /* a task moving charges */
413 wait_queue_head_t waitq; /* a waitq for other context */
415 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
416 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
419 static bool move_anon(void)
421 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
424 static bool move_file(void)
426 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
430 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
431 * limit reclaim to prevent infinite loops, if they ever occur.
433 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
434 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
437 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
438 MEM_CGROUP_CHARGE_TYPE_ANON,
439 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
440 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
444 /* for encoding cft->private value on file */
452 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
453 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
454 #define MEMFILE_ATTR(val) ((val) & 0xffff)
455 /* Used for OOM nofiier */
456 #define OOM_CONTROL (0)
459 * Reclaim flags for mem_cgroup_hierarchical_reclaim
461 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
462 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
463 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
464 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
467 * The memcg_create_mutex will be held whenever a new cgroup is created.
468 * As a consequence, any change that needs to protect against new child cgroups
469 * appearing has to hold it as well.
471 static DEFINE_MUTEX(memcg_create_mutex);
473 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
475 return s ? container_of(s, struct mem_cgroup, css) : NULL;
478 /* Some nice accessors for the vmpressure. */
479 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
482 memcg = root_mem_cgroup;
483 return &memcg->vmpressure;
486 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
488 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
491 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
493 return &mem_cgroup_from_css(css)->vmpressure;
496 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
498 return (memcg == root_mem_cgroup);
501 /* Writing them here to avoid exposing memcg's inner layout */
502 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
504 void sock_update_memcg(struct sock *sk)
506 if (mem_cgroup_sockets_enabled) {
507 struct mem_cgroup *memcg;
508 struct cg_proto *cg_proto;
510 BUG_ON(!sk->sk_prot->proto_cgroup);
512 /* Socket cloning can throw us here with sk_cgrp already
513 * filled. It won't however, necessarily happen from
514 * process context. So the test for root memcg given
515 * the current task's memcg won't help us in this case.
517 * Respecting the original socket's memcg is a better
518 * decision in this case.
521 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
522 css_get(&sk->sk_cgrp->memcg->css);
527 memcg = mem_cgroup_from_task(current);
528 cg_proto = sk->sk_prot->proto_cgroup(memcg);
529 if (!mem_cgroup_is_root(memcg) &&
530 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
531 sk->sk_cgrp = cg_proto;
536 EXPORT_SYMBOL(sock_update_memcg);
538 void sock_release_memcg(struct sock *sk)
540 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
541 struct mem_cgroup *memcg;
542 WARN_ON(!sk->sk_cgrp->memcg);
543 memcg = sk->sk_cgrp->memcg;
544 css_put(&sk->sk_cgrp->memcg->css);
548 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
550 if (!memcg || mem_cgroup_is_root(memcg))
553 return &memcg->tcp_mem.cg_proto;
555 EXPORT_SYMBOL(tcp_proto_cgroup);
557 static void disarm_sock_keys(struct mem_cgroup *memcg)
559 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
561 static_key_slow_dec(&memcg_socket_limit_enabled);
564 static void disarm_sock_keys(struct mem_cgroup *memcg)
569 #ifdef CONFIG_MEMCG_KMEM
571 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
572 * There are two main reasons for not using the css_id for this:
573 * 1) this works better in sparse environments, where we have a lot of memcgs,
574 * but only a few kmem-limited. Or also, if we have, for instance, 200
575 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
576 * 200 entry array for that.
578 * 2) In order not to violate the cgroup API, we would like to do all memory
579 * allocation in ->create(). At that point, we haven't yet allocated the
580 * css_id. Having a separate index prevents us from messing with the cgroup
583 * The current size of the caches array is stored in
584 * memcg_limited_groups_array_size. It will double each time we have to
587 static DEFINE_IDA(kmem_limited_groups);
588 int memcg_limited_groups_array_size;
591 * MIN_SIZE is different than 1, because we would like to avoid going through
592 * the alloc/free process all the time. In a small machine, 4 kmem-limited
593 * cgroups is a reasonable guess. In the future, it could be a parameter or
594 * tunable, but that is strictly not necessary.
596 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
597 * this constant directly from cgroup, but it is understandable that this is
598 * better kept as an internal representation in cgroup.c. In any case, the
599 * css_id space is not getting any smaller, and we don't have to necessarily
600 * increase ours as well if it increases.
602 #define MEMCG_CACHES_MIN_SIZE 4
603 #define MEMCG_CACHES_MAX_SIZE 65535
606 * A lot of the calls to the cache allocation functions are expected to be
607 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
608 * conditional to this static branch, we'll have to allow modules that does
609 * kmem_cache_alloc and the such to see this symbol as well
611 struct static_key memcg_kmem_enabled_key;
612 EXPORT_SYMBOL(memcg_kmem_enabled_key);
614 static void disarm_kmem_keys(struct mem_cgroup *memcg)
616 if (memcg_kmem_is_active(memcg)) {
617 static_key_slow_dec(&memcg_kmem_enabled_key);
618 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
621 * This check can't live in kmem destruction function,
622 * since the charges will outlive the cgroup
624 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
627 static void disarm_kmem_keys(struct mem_cgroup *memcg)
630 #endif /* CONFIG_MEMCG_KMEM */
632 static void disarm_static_keys(struct mem_cgroup *memcg)
634 disarm_sock_keys(memcg);
635 disarm_kmem_keys(memcg);
638 static void drain_all_stock_async(struct mem_cgroup *memcg);
640 static struct mem_cgroup_per_zone *
641 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
643 VM_BUG_ON((unsigned)nid >= nr_node_ids);
644 return &memcg->nodeinfo[nid]->zoneinfo[zid];
647 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
652 static struct mem_cgroup_per_zone *
653 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
655 int nid = page_to_nid(page);
656 int zid = page_zonenum(page);
658 return mem_cgroup_zoneinfo(memcg, nid, zid);
661 static struct mem_cgroup_tree_per_zone *
662 soft_limit_tree_node_zone(int nid, int zid)
664 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
667 static struct mem_cgroup_tree_per_zone *
668 soft_limit_tree_from_page(struct page *page)
670 int nid = page_to_nid(page);
671 int zid = page_zonenum(page);
673 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
677 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
678 struct mem_cgroup_per_zone *mz,
679 struct mem_cgroup_tree_per_zone *mctz,
680 unsigned long long new_usage_in_excess)
682 struct rb_node **p = &mctz->rb_root.rb_node;
683 struct rb_node *parent = NULL;
684 struct mem_cgroup_per_zone *mz_node;
689 mz->usage_in_excess = new_usage_in_excess;
690 if (!mz->usage_in_excess)
694 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
696 if (mz->usage_in_excess < mz_node->usage_in_excess)
699 * We can't avoid mem cgroups that are over their soft
700 * limit by the same amount
702 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
705 rb_link_node(&mz->tree_node, parent, p);
706 rb_insert_color(&mz->tree_node, &mctz->rb_root);
711 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
712 struct mem_cgroup_per_zone *mz,
713 struct mem_cgroup_tree_per_zone *mctz)
717 rb_erase(&mz->tree_node, &mctz->rb_root);
722 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
723 struct mem_cgroup_per_zone *mz,
724 struct mem_cgroup_tree_per_zone *mctz)
726 spin_lock(&mctz->lock);
727 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
728 spin_unlock(&mctz->lock);
732 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
734 unsigned long long excess;
735 struct mem_cgroup_per_zone *mz;
736 struct mem_cgroup_tree_per_zone *mctz;
737 int nid = page_to_nid(page);
738 int zid = page_zonenum(page);
739 mctz = soft_limit_tree_from_page(page);
742 * Necessary to update all ancestors when hierarchy is used.
743 * because their event counter is not touched.
745 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
746 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
747 excess = res_counter_soft_limit_excess(&memcg->res);
749 * We have to update the tree if mz is on RB-tree or
750 * mem is over its softlimit.
752 if (excess || mz->on_tree) {
753 spin_lock(&mctz->lock);
754 /* if on-tree, remove it */
756 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
758 * Insert again. mz->usage_in_excess will be updated.
759 * If excess is 0, no tree ops.
761 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
762 spin_unlock(&mctz->lock);
767 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
770 struct mem_cgroup_per_zone *mz;
771 struct mem_cgroup_tree_per_zone *mctz;
773 for_each_node(node) {
774 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
775 mz = mem_cgroup_zoneinfo(memcg, node, zone);
776 mctz = soft_limit_tree_node_zone(node, zone);
777 mem_cgroup_remove_exceeded(memcg, mz, mctz);
782 static struct mem_cgroup_per_zone *
783 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
785 struct rb_node *rightmost = NULL;
786 struct mem_cgroup_per_zone *mz;
790 rightmost = rb_last(&mctz->rb_root);
792 goto done; /* Nothing to reclaim from */
794 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
796 * Remove the node now but someone else can add it back,
797 * we will to add it back at the end of reclaim to its correct
798 * position in the tree.
800 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
801 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
802 !css_tryget(&mz->memcg->css))
808 static struct mem_cgroup_per_zone *
809 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
811 struct mem_cgroup_per_zone *mz;
813 spin_lock(&mctz->lock);
814 mz = __mem_cgroup_largest_soft_limit_node(mctz);
815 spin_unlock(&mctz->lock);
820 * Implementation Note: reading percpu statistics for memcg.
822 * Both of vmstat[] and percpu_counter has threshold and do periodic
823 * synchronization to implement "quick" read. There are trade-off between
824 * reading cost and precision of value. Then, we may have a chance to implement
825 * a periodic synchronizion of counter in memcg's counter.
827 * But this _read() function is used for user interface now. The user accounts
828 * memory usage by memory cgroup and he _always_ requires exact value because
829 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
830 * have to visit all online cpus and make sum. So, for now, unnecessary
831 * synchronization is not implemented. (just implemented for cpu hotplug)
833 * If there are kernel internal actions which can make use of some not-exact
834 * value, and reading all cpu value can be performance bottleneck in some
835 * common workload, threashold and synchonization as vmstat[] should be
838 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
839 enum mem_cgroup_stat_index idx)
845 for_each_online_cpu(cpu)
846 val += per_cpu(memcg->stat->count[idx], cpu);
847 #ifdef CONFIG_HOTPLUG_CPU
848 spin_lock(&memcg->pcp_counter_lock);
849 val += memcg->nocpu_base.count[idx];
850 spin_unlock(&memcg->pcp_counter_lock);
856 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
859 int val = (charge) ? 1 : -1;
860 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
863 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
864 enum mem_cgroup_events_index idx)
866 unsigned long val = 0;
870 for_each_online_cpu(cpu)
871 val += per_cpu(memcg->stat->events[idx], cpu);
872 #ifdef CONFIG_HOTPLUG_CPU
873 spin_lock(&memcg->pcp_counter_lock);
874 val += memcg->nocpu_base.events[idx];
875 spin_unlock(&memcg->pcp_counter_lock);
881 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
883 bool anon, int nr_pages)
888 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
889 * counted as CACHE even if it's on ANON LRU.
892 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
895 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
898 if (PageTransHuge(page))
899 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
902 /* pagein of a big page is an event. So, ignore page size */
904 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
906 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
907 nr_pages = -nr_pages; /* for event */
910 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
916 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
918 struct mem_cgroup_per_zone *mz;
920 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
921 return mz->lru_size[lru];
925 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
926 unsigned int lru_mask)
928 struct mem_cgroup_per_zone *mz;
930 unsigned long ret = 0;
932 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
935 if (BIT(lru) & lru_mask)
936 ret += mz->lru_size[lru];
942 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
943 int nid, unsigned int lru_mask)
948 for (zid = 0; zid < MAX_NR_ZONES; zid++)
949 total += mem_cgroup_zone_nr_lru_pages(memcg,
955 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
956 unsigned int lru_mask)
961 for_each_node_state(nid, N_MEMORY)
962 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
966 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
967 enum mem_cgroup_events_target target)
969 unsigned long val, next;
971 val = __this_cpu_read(memcg->stat->nr_page_events);
972 next = __this_cpu_read(memcg->stat->targets[target]);
973 /* from time_after() in jiffies.h */
974 if ((long)next - (long)val < 0) {
976 case MEM_CGROUP_TARGET_THRESH:
977 next = val + THRESHOLDS_EVENTS_TARGET;
979 case MEM_CGROUP_TARGET_SOFTLIMIT:
980 next = val + SOFTLIMIT_EVENTS_TARGET;
982 case MEM_CGROUP_TARGET_NUMAINFO:
983 next = val + NUMAINFO_EVENTS_TARGET;
988 __this_cpu_write(memcg->stat->targets[target], next);
995 * Check events in order.
998 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1001 /* threshold event is triggered in finer grain than soft limit */
1002 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1003 MEM_CGROUP_TARGET_THRESH))) {
1005 bool do_numainfo __maybe_unused;
1007 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1008 MEM_CGROUP_TARGET_SOFTLIMIT);
1009 #if MAX_NUMNODES > 1
1010 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1011 MEM_CGROUP_TARGET_NUMAINFO);
1015 mem_cgroup_threshold(memcg);
1016 if (unlikely(do_softlimit))
1017 mem_cgroup_update_tree(memcg, page);
1018 #if MAX_NUMNODES > 1
1019 if (unlikely(do_numainfo))
1020 atomic_inc(&memcg->numainfo_events);
1026 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1029 * mm_update_next_owner() may clear mm->owner to NULL
1030 * if it races with swapoff, page migration, etc.
1031 * So this can be called with p == NULL.
1036 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1039 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1041 struct mem_cgroup *memcg = NULL;
1046 * Because we have no locks, mm->owner's may be being moved to other
1047 * cgroup. We use css_tryget() here even if this looks
1048 * pessimistic (rather than adding locks here).
1052 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1053 if (unlikely(!memcg))
1055 } while (!css_tryget(&memcg->css));
1061 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1062 * ref. count) or NULL if the whole root's subtree has been visited.
1064 * helper function to be used by mem_cgroup_iter
1066 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1067 struct mem_cgroup *last_visited)
1069 struct cgroup_subsys_state *prev_css, *next_css;
1071 prev_css = last_visited ? &last_visited->css : NULL;
1073 next_css = css_next_descendant_pre(prev_css, &root->css);
1076 * Even if we found a group we have to make sure it is
1077 * alive. css && !memcg means that the groups should be
1078 * skipped and we should continue the tree walk.
1079 * last_visited css is safe to use because it is
1080 * protected by css_get and the tree walk is rcu safe.
1083 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1085 if (css_tryget(&mem->css))
1088 prev_css = next_css;
1096 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1099 * When a group in the hierarchy below root is destroyed, the
1100 * hierarchy iterator can no longer be trusted since it might
1101 * have pointed to the destroyed group. Invalidate it.
1103 atomic_inc(&root->dead_count);
1106 static struct mem_cgroup *
1107 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1108 struct mem_cgroup *root,
1111 struct mem_cgroup *position = NULL;
1113 * A cgroup destruction happens in two stages: offlining and
1114 * release. They are separated by a RCU grace period.
1116 * If the iterator is valid, we may still race with an
1117 * offlining. The RCU lock ensures the object won't be
1118 * released, tryget will fail if we lost the race.
1120 *sequence = atomic_read(&root->dead_count);
1121 if (iter->last_dead_count == *sequence) {
1123 position = iter->last_visited;
1124 if (position && !css_tryget(&position->css))
1130 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1131 struct mem_cgroup *last_visited,
1132 struct mem_cgroup *new_position,
1136 css_put(&last_visited->css);
1138 * We store the sequence count from the time @last_visited was
1139 * loaded successfully instead of rereading it here so that we
1140 * don't lose destruction events in between. We could have
1141 * raced with the destruction of @new_position after all.
1143 iter->last_visited = new_position;
1145 iter->last_dead_count = sequence;
1149 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1150 * @root: hierarchy root
1151 * @prev: previously returned memcg, NULL on first invocation
1152 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1154 * Returns references to children of the hierarchy below @root, or
1155 * @root itself, or %NULL after a full round-trip.
1157 * Caller must pass the return value in @prev on subsequent
1158 * invocations for reference counting, or use mem_cgroup_iter_break()
1159 * to cancel a hierarchy walk before the round-trip is complete.
1161 * Reclaimers can specify a zone and a priority level in @reclaim to
1162 * divide up the memcgs in the hierarchy among all concurrent
1163 * reclaimers operating on the same zone and priority.
1165 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1166 struct mem_cgroup *prev,
1167 struct mem_cgroup_reclaim_cookie *reclaim)
1169 struct mem_cgroup *memcg = NULL;
1170 struct mem_cgroup *last_visited = NULL;
1172 if (mem_cgroup_disabled())
1176 root = root_mem_cgroup;
1178 if (prev && !reclaim)
1179 last_visited = prev;
1181 if (!root->use_hierarchy && root != root_mem_cgroup) {
1189 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1190 int uninitialized_var(seq);
1193 int nid = zone_to_nid(reclaim->zone);
1194 int zid = zone_idx(reclaim->zone);
1195 struct mem_cgroup_per_zone *mz;
1197 mz = mem_cgroup_zoneinfo(root, nid, zid);
1198 iter = &mz->reclaim_iter[reclaim->priority];
1199 if (prev && reclaim->generation != iter->generation) {
1200 iter->last_visited = NULL;
1204 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1207 memcg = __mem_cgroup_iter_next(root, last_visited);
1210 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1214 else if (!prev && memcg)
1215 reclaim->generation = iter->generation;
1224 if (prev && prev != root)
1225 css_put(&prev->css);
1231 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1232 * @root: hierarchy root
1233 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1235 void mem_cgroup_iter_break(struct mem_cgroup *root,
1236 struct mem_cgroup *prev)
1239 root = root_mem_cgroup;
1240 if (prev && prev != root)
1241 css_put(&prev->css);
1245 * Iteration constructs for visiting all cgroups (under a tree). If
1246 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1247 * be used for reference counting.
1249 #define for_each_mem_cgroup_tree(iter, root) \
1250 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1252 iter = mem_cgroup_iter(root, iter, NULL))
1254 #define for_each_mem_cgroup(iter) \
1255 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1257 iter = mem_cgroup_iter(NULL, iter, NULL))
1259 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1261 struct mem_cgroup *memcg;
1264 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1265 if (unlikely(!memcg))
1270 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1273 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1281 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1284 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1285 * @zone: zone of the wanted lruvec
1286 * @memcg: memcg of the wanted lruvec
1288 * Returns the lru list vector holding pages for the given @zone and
1289 * @mem. This can be the global zone lruvec, if the memory controller
1292 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1293 struct mem_cgroup *memcg)
1295 struct mem_cgroup_per_zone *mz;
1296 struct lruvec *lruvec;
1298 if (mem_cgroup_disabled()) {
1299 lruvec = &zone->lruvec;
1303 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1304 lruvec = &mz->lruvec;
1307 * Since a node can be onlined after the mem_cgroup was created,
1308 * we have to be prepared to initialize lruvec->zone here;
1309 * and if offlined then reonlined, we need to reinitialize it.
1311 if (unlikely(lruvec->zone != zone))
1312 lruvec->zone = zone;
1317 * Following LRU functions are allowed to be used without PCG_LOCK.
1318 * Operations are called by routine of global LRU independently from memcg.
1319 * What we have to take care of here is validness of pc->mem_cgroup.
1321 * Changes to pc->mem_cgroup happens when
1324 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1325 * It is added to LRU before charge.
1326 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1327 * When moving account, the page is not on LRU. It's isolated.
1331 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1333 * @zone: zone of the page
1335 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1337 struct mem_cgroup_per_zone *mz;
1338 struct mem_cgroup *memcg;
1339 struct page_cgroup *pc;
1340 struct lruvec *lruvec;
1342 if (mem_cgroup_disabled()) {
1343 lruvec = &zone->lruvec;
1347 pc = lookup_page_cgroup(page);
1348 memcg = pc->mem_cgroup;
1351 * Surreptitiously switch any uncharged offlist page to root:
1352 * an uncharged page off lru does nothing to secure
1353 * its former mem_cgroup from sudden removal.
1355 * Our caller holds lru_lock, and PageCgroupUsed is updated
1356 * under page_cgroup lock: between them, they make all uses
1357 * of pc->mem_cgroup safe.
1359 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1360 pc->mem_cgroup = memcg = root_mem_cgroup;
1362 mz = page_cgroup_zoneinfo(memcg, page);
1363 lruvec = &mz->lruvec;
1366 * Since a node can be onlined after the mem_cgroup was created,
1367 * we have to be prepared to initialize lruvec->zone here;
1368 * and if offlined then reonlined, we need to reinitialize it.
1370 if (unlikely(lruvec->zone != zone))
1371 lruvec->zone = zone;
1376 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1377 * @lruvec: mem_cgroup per zone lru vector
1378 * @lru: index of lru list the page is sitting on
1379 * @nr_pages: positive when adding or negative when removing
1381 * This function must be called when a page is added to or removed from an
1384 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1387 struct mem_cgroup_per_zone *mz;
1388 unsigned long *lru_size;
1390 if (mem_cgroup_disabled())
1393 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1394 lru_size = mz->lru_size + lru;
1395 *lru_size += nr_pages;
1396 VM_BUG_ON((long)(*lru_size) < 0);
1400 * Checks whether given mem is same or in the root_mem_cgroup's
1403 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1404 struct mem_cgroup *memcg)
1406 if (root_memcg == memcg)
1408 if (!root_memcg->use_hierarchy || !memcg)
1410 return css_is_ancestor(&memcg->css, &root_memcg->css);
1413 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1414 struct mem_cgroup *memcg)
1419 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1424 bool task_in_mem_cgroup(struct task_struct *task,
1425 const struct mem_cgroup *memcg)
1427 struct mem_cgroup *curr = NULL;
1428 struct task_struct *p;
1431 p = find_lock_task_mm(task);
1433 curr = try_get_mem_cgroup_from_mm(p->mm);
1437 * All threads may have already detached their mm's, but the oom
1438 * killer still needs to detect if they have already been oom
1439 * killed to prevent needlessly killing additional tasks.
1442 curr = mem_cgroup_from_task(task);
1444 css_get(&curr->css);
1450 * We should check use_hierarchy of "memcg" not "curr". Because checking
1451 * use_hierarchy of "curr" here make this function true if hierarchy is
1452 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1453 * hierarchy(even if use_hierarchy is disabled in "memcg").
1455 ret = mem_cgroup_same_or_subtree(memcg, curr);
1456 css_put(&curr->css);
1460 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1462 unsigned long inactive_ratio;
1463 unsigned long inactive;
1464 unsigned long active;
1467 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1468 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1470 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1472 inactive_ratio = int_sqrt(10 * gb);
1476 return inactive * inactive_ratio < active;
1479 #define mem_cgroup_from_res_counter(counter, member) \
1480 container_of(counter, struct mem_cgroup, member)
1483 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1484 * @memcg: the memory cgroup
1486 * Returns the maximum amount of memory @mem can be charged with, in
1489 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1491 unsigned long long margin;
1493 margin = res_counter_margin(&memcg->res);
1494 if (do_swap_account)
1495 margin = min(margin, res_counter_margin(&memcg->memsw));
1496 return margin >> PAGE_SHIFT;
1499 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1502 if (!css_parent(&memcg->css))
1503 return vm_swappiness;
1505 return memcg->swappiness;
1509 * memcg->moving_account is used for checking possibility that some thread is
1510 * calling move_account(). When a thread on CPU-A starts moving pages under
1511 * a memcg, other threads should check memcg->moving_account under
1512 * rcu_read_lock(), like this:
1516 * memcg->moving_account+1 if (memcg->mocing_account)
1518 * synchronize_rcu() update something.
1523 /* for quick checking without looking up memcg */
1524 atomic_t memcg_moving __read_mostly;
1526 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1528 atomic_inc(&memcg_moving);
1529 atomic_inc(&memcg->moving_account);
1533 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1536 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1537 * We check NULL in callee rather than caller.
1540 atomic_dec(&memcg_moving);
1541 atomic_dec(&memcg->moving_account);
1546 * 2 routines for checking "mem" is under move_account() or not.
1548 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1549 * is used for avoiding races in accounting. If true,
1550 * pc->mem_cgroup may be overwritten.
1552 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1553 * under hierarchy of moving cgroups. This is for
1554 * waiting at hith-memory prressure caused by "move".
1557 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1559 VM_BUG_ON(!rcu_read_lock_held());
1560 return atomic_read(&memcg->moving_account) > 0;
1563 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1565 struct mem_cgroup *from;
1566 struct mem_cgroup *to;
1569 * Unlike task_move routines, we access mc.to, mc.from not under
1570 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1572 spin_lock(&mc.lock);
1578 ret = mem_cgroup_same_or_subtree(memcg, from)
1579 || mem_cgroup_same_or_subtree(memcg, to);
1581 spin_unlock(&mc.lock);
1585 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1587 if (mc.moving_task && current != mc.moving_task) {
1588 if (mem_cgroup_under_move(memcg)) {
1590 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1591 /* moving charge context might have finished. */
1594 finish_wait(&mc.waitq, &wait);
1602 * Take this lock when
1603 * - a code tries to modify page's memcg while it's USED.
1604 * - a code tries to modify page state accounting in a memcg.
1605 * see mem_cgroup_stolen(), too.
1607 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1608 unsigned long *flags)
1610 spin_lock_irqsave(&memcg->move_lock, *flags);
1613 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1614 unsigned long *flags)
1616 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1619 #define K(x) ((x) << (PAGE_SHIFT-10))
1621 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1622 * @memcg: The memory cgroup that went over limit
1623 * @p: Task that is going to be killed
1625 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1628 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1630 struct cgroup *task_cgrp;
1631 struct cgroup *mem_cgrp;
1633 * Need a buffer in BSS, can't rely on allocations. The code relies
1634 * on the assumption that OOM is serialized for memory controller.
1635 * If this assumption is broken, revisit this code.
1637 static char memcg_name[PATH_MAX];
1639 struct mem_cgroup *iter;
1647 mem_cgrp = memcg->css.cgroup;
1648 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1650 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1653 * Unfortunately, we are unable to convert to a useful name
1654 * But we'll still print out the usage information
1661 pr_info("Task in %s killed", memcg_name);
1664 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1672 * Continues from above, so we don't need an KERN_ level
1674 pr_cont(" as a result of limit of %s\n", memcg_name);
1677 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1678 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1679 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1680 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1681 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1682 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1683 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1684 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1685 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1686 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1687 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1688 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1690 for_each_mem_cgroup_tree(iter, memcg) {
1691 pr_info("Memory cgroup stats");
1694 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1696 pr_cont(" for %s", memcg_name);
1700 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1701 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1703 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1704 K(mem_cgroup_read_stat(iter, i)));
1707 for (i = 0; i < NR_LRU_LISTS; i++)
1708 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1709 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1716 * This function returns the number of memcg under hierarchy tree. Returns
1717 * 1(self count) if no children.
1719 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1722 struct mem_cgroup *iter;
1724 for_each_mem_cgroup_tree(iter, memcg)
1730 * Return the memory (and swap, if configured) limit for a memcg.
1732 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1736 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1739 * Do not consider swap space if we cannot swap due to swappiness
1741 if (mem_cgroup_swappiness(memcg)) {
1744 limit += total_swap_pages << PAGE_SHIFT;
1745 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1748 * If memsw is finite and limits the amount of swap space
1749 * available to this memcg, return that limit.
1751 limit = min(limit, memsw);
1757 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1760 struct mem_cgroup *iter;
1761 unsigned long chosen_points = 0;
1762 unsigned long totalpages;
1763 unsigned int points = 0;
1764 struct task_struct *chosen = NULL;
1767 * If current has a pending SIGKILL or is exiting, then automatically
1768 * select it. The goal is to allow it to allocate so that it may
1769 * quickly exit and free its memory.
1771 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1772 set_thread_flag(TIF_MEMDIE);
1776 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1777 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1778 for_each_mem_cgroup_tree(iter, memcg) {
1779 struct css_task_iter it;
1780 struct task_struct *task;
1782 css_task_iter_start(&iter->css, &it);
1783 while ((task = css_task_iter_next(&it))) {
1784 switch (oom_scan_process_thread(task, totalpages, NULL,
1786 case OOM_SCAN_SELECT:
1788 put_task_struct(chosen);
1790 chosen_points = ULONG_MAX;
1791 get_task_struct(chosen);
1793 case OOM_SCAN_CONTINUE:
1795 case OOM_SCAN_ABORT:
1796 css_task_iter_end(&it);
1797 mem_cgroup_iter_break(memcg, iter);
1799 put_task_struct(chosen);
1804 points = oom_badness(task, memcg, NULL, totalpages);
1805 if (points > chosen_points) {
1807 put_task_struct(chosen);
1809 chosen_points = points;
1810 get_task_struct(chosen);
1813 css_task_iter_end(&it);
1818 points = chosen_points * 1000 / totalpages;
1819 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1820 NULL, "Memory cgroup out of memory");
1823 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1825 unsigned long flags)
1827 unsigned long total = 0;
1828 bool noswap = false;
1831 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1833 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1836 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1838 drain_all_stock_async(memcg);
1839 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1841 * Allow limit shrinkers, which are triggered directly
1842 * by userspace, to catch signals and stop reclaim
1843 * after minimal progress, regardless of the margin.
1845 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1847 if (mem_cgroup_margin(memcg))
1850 * If nothing was reclaimed after two attempts, there
1851 * may be no reclaimable pages in this hierarchy.
1860 * test_mem_cgroup_node_reclaimable
1861 * @memcg: the target memcg
1862 * @nid: the node ID to be checked.
1863 * @noswap : specify true here if the user wants flle only information.
1865 * This function returns whether the specified memcg contains any
1866 * reclaimable pages on a node. Returns true if there are any reclaimable
1867 * pages in the node.
1869 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1870 int nid, bool noswap)
1872 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1874 if (noswap || !total_swap_pages)
1876 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1881 #if MAX_NUMNODES > 1
1884 * Always updating the nodemask is not very good - even if we have an empty
1885 * list or the wrong list here, we can start from some node and traverse all
1886 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1889 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1893 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1894 * pagein/pageout changes since the last update.
1896 if (!atomic_read(&memcg->numainfo_events))
1898 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1901 /* make a nodemask where this memcg uses memory from */
1902 memcg->scan_nodes = node_states[N_MEMORY];
1904 for_each_node_mask(nid, node_states[N_MEMORY]) {
1906 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1907 node_clear(nid, memcg->scan_nodes);
1910 atomic_set(&memcg->numainfo_events, 0);
1911 atomic_set(&memcg->numainfo_updating, 0);
1915 * Selecting a node where we start reclaim from. Because what we need is just
1916 * reducing usage counter, start from anywhere is O,K. Considering
1917 * memory reclaim from current node, there are pros. and cons.
1919 * Freeing memory from current node means freeing memory from a node which
1920 * we'll use or we've used. So, it may make LRU bad. And if several threads
1921 * hit limits, it will see a contention on a node. But freeing from remote
1922 * node means more costs for memory reclaim because of memory latency.
1924 * Now, we use round-robin. Better algorithm is welcomed.
1926 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1930 mem_cgroup_may_update_nodemask(memcg);
1931 node = memcg->last_scanned_node;
1933 node = next_node(node, memcg->scan_nodes);
1934 if (node == MAX_NUMNODES)
1935 node = first_node(memcg->scan_nodes);
1937 * We call this when we hit limit, not when pages are added to LRU.
1938 * No LRU may hold pages because all pages are UNEVICTABLE or
1939 * memcg is too small and all pages are not on LRU. In that case,
1940 * we use curret node.
1942 if (unlikely(node == MAX_NUMNODES))
1943 node = numa_node_id();
1945 memcg->last_scanned_node = node;
1950 * Check all nodes whether it contains reclaimable pages or not.
1951 * For quick scan, we make use of scan_nodes. This will allow us to skip
1952 * unused nodes. But scan_nodes is lazily updated and may not cotain
1953 * enough new information. We need to do double check.
1955 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1960 * quick check...making use of scan_node.
1961 * We can skip unused nodes.
1963 if (!nodes_empty(memcg->scan_nodes)) {
1964 for (nid = first_node(memcg->scan_nodes);
1966 nid = next_node(nid, memcg->scan_nodes)) {
1968 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1973 * Check rest of nodes.
1975 for_each_node_state(nid, N_MEMORY) {
1976 if (node_isset(nid, memcg->scan_nodes))
1978 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1985 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1990 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1992 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1996 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1999 unsigned long *total_scanned)
2001 struct mem_cgroup *victim = NULL;
2004 unsigned long excess;
2005 unsigned long nr_scanned;
2006 struct mem_cgroup_reclaim_cookie reclaim = {
2011 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2014 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2019 * If we have not been able to reclaim
2020 * anything, it might because there are
2021 * no reclaimable pages under this hierarchy
2026 * We want to do more targeted reclaim.
2027 * excess >> 2 is not to excessive so as to
2028 * reclaim too much, nor too less that we keep
2029 * coming back to reclaim from this cgroup
2031 if (total >= (excess >> 2) ||
2032 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2037 if (!mem_cgroup_reclaimable(victim, false))
2039 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2041 *total_scanned += nr_scanned;
2042 if (!res_counter_soft_limit_excess(&root_memcg->res))
2045 mem_cgroup_iter_break(root_memcg, victim);
2049 static DEFINE_SPINLOCK(memcg_oom_lock);
2052 * Check OOM-Killer is already running under our hierarchy.
2053 * If someone is running, return false.
2055 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2057 struct mem_cgroup *iter, *failed = NULL;
2059 spin_lock(&memcg_oom_lock);
2061 for_each_mem_cgroup_tree(iter, memcg) {
2062 if (iter->oom_lock) {
2064 * this subtree of our hierarchy is already locked
2065 * so we cannot give a lock.
2068 mem_cgroup_iter_break(memcg, iter);
2071 iter->oom_lock = true;
2076 * OK, we failed to lock the whole subtree so we have
2077 * to clean up what we set up to the failing subtree
2079 for_each_mem_cgroup_tree(iter, memcg) {
2080 if (iter == failed) {
2081 mem_cgroup_iter_break(memcg, iter);
2084 iter->oom_lock = false;
2088 spin_unlock(&memcg_oom_lock);
2093 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2095 struct mem_cgroup *iter;
2097 spin_lock(&memcg_oom_lock);
2098 for_each_mem_cgroup_tree(iter, memcg)
2099 iter->oom_lock = false;
2100 spin_unlock(&memcg_oom_lock);
2103 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2105 struct mem_cgroup *iter;
2107 for_each_mem_cgroup_tree(iter, memcg)
2108 atomic_inc(&iter->under_oom);
2111 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2113 struct mem_cgroup *iter;
2116 * When a new child is created while the hierarchy is under oom,
2117 * mem_cgroup_oom_lock() may not be called. We have to use
2118 * atomic_add_unless() here.
2120 for_each_mem_cgroup_tree(iter, memcg)
2121 atomic_add_unless(&iter->under_oom, -1, 0);
2124 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2126 struct oom_wait_info {
2127 struct mem_cgroup *memcg;
2131 static int memcg_oom_wake_function(wait_queue_t *wait,
2132 unsigned mode, int sync, void *arg)
2134 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2135 struct mem_cgroup *oom_wait_memcg;
2136 struct oom_wait_info *oom_wait_info;
2138 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2139 oom_wait_memcg = oom_wait_info->memcg;
2142 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2143 * Then we can use css_is_ancestor without taking care of RCU.
2145 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2146 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2148 return autoremove_wake_function(wait, mode, sync, arg);
2151 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2153 atomic_inc(&memcg->oom_wakeups);
2154 /* for filtering, pass "memcg" as argument. */
2155 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2158 static void memcg_oom_recover(struct mem_cgroup *memcg)
2160 if (memcg && atomic_read(&memcg->under_oom))
2161 memcg_wakeup_oom(memcg);
2164 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2166 if (!current->memcg_oom.may_oom)
2169 * We are in the middle of the charge context here, so we
2170 * don't want to block when potentially sitting on a callstack
2171 * that holds all kinds of filesystem and mm locks.
2173 * Also, the caller may handle a failed allocation gracefully
2174 * (like optional page cache readahead) and so an OOM killer
2175 * invocation might not even be necessary.
2177 * That's why we don't do anything here except remember the
2178 * OOM context and then deal with it at the end of the page
2179 * fault when the stack is unwound, the locks are released,
2180 * and when we know whether the fault was overall successful.
2182 css_get(&memcg->css);
2183 current->memcg_oom.memcg = memcg;
2184 current->memcg_oom.gfp_mask = mask;
2185 current->memcg_oom.order = order;
2189 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2190 * @handle: actually kill/wait or just clean up the OOM state
2192 * This has to be called at the end of a page fault if the memcg OOM
2193 * handler was enabled.
2195 * Memcg supports userspace OOM handling where failed allocations must
2196 * sleep on a waitqueue until the userspace task resolves the
2197 * situation. Sleeping directly in the charge context with all kinds
2198 * of locks held is not a good idea, instead we remember an OOM state
2199 * in the task and mem_cgroup_oom_synchronize() has to be called at
2200 * the end of the page fault to complete the OOM handling.
2202 * Returns %true if an ongoing memcg OOM situation was detected and
2203 * completed, %false otherwise.
2205 bool mem_cgroup_oom_synchronize(bool handle)
2207 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2208 struct oom_wait_info owait;
2211 /* OOM is global, do not handle */
2218 owait.memcg = memcg;
2219 owait.wait.flags = 0;
2220 owait.wait.func = memcg_oom_wake_function;
2221 owait.wait.private = current;
2222 INIT_LIST_HEAD(&owait.wait.task_list);
2224 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2225 mem_cgroup_mark_under_oom(memcg);
2227 locked = mem_cgroup_oom_trylock(memcg);
2230 mem_cgroup_oom_notify(memcg);
2232 if (locked && !memcg->oom_kill_disable) {
2233 mem_cgroup_unmark_under_oom(memcg);
2234 finish_wait(&memcg_oom_waitq, &owait.wait);
2235 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2236 current->memcg_oom.order);
2239 mem_cgroup_unmark_under_oom(memcg);
2240 finish_wait(&memcg_oom_waitq, &owait.wait);
2244 mem_cgroup_oom_unlock(memcg);
2246 * There is no guarantee that an OOM-lock contender
2247 * sees the wakeups triggered by the OOM kill
2248 * uncharges. Wake any sleepers explicitely.
2250 memcg_oom_recover(memcg);
2253 current->memcg_oom.memcg = NULL;
2254 css_put(&memcg->css);
2259 * Currently used to update mapped file statistics, but the routine can be
2260 * generalized to update other statistics as well.
2262 * Notes: Race condition
2264 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2265 * it tends to be costly. But considering some conditions, we doesn't need
2266 * to do so _always_.
2268 * Considering "charge", lock_page_cgroup() is not required because all
2269 * file-stat operations happen after a page is attached to radix-tree. There
2270 * are no race with "charge".
2272 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2273 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2274 * if there are race with "uncharge". Statistics itself is properly handled
2277 * Considering "move", this is an only case we see a race. To make the race
2278 * small, we check mm->moving_account and detect there are possibility of race
2279 * If there is, we take a lock.
2282 void __mem_cgroup_begin_update_page_stat(struct page *page,
2283 bool *locked, unsigned long *flags)
2285 struct mem_cgroup *memcg;
2286 struct page_cgroup *pc;
2288 pc = lookup_page_cgroup(page);
2290 memcg = pc->mem_cgroup;
2291 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2294 * If this memory cgroup is not under account moving, we don't
2295 * need to take move_lock_mem_cgroup(). Because we already hold
2296 * rcu_read_lock(), any calls to move_account will be delayed until
2297 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2299 if (!mem_cgroup_stolen(memcg))
2302 move_lock_mem_cgroup(memcg, flags);
2303 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2304 move_unlock_mem_cgroup(memcg, flags);
2310 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2312 struct page_cgroup *pc = lookup_page_cgroup(page);
2315 * It's guaranteed that pc->mem_cgroup never changes while
2316 * lock is held because a routine modifies pc->mem_cgroup
2317 * should take move_lock_mem_cgroup().
2319 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2322 void mem_cgroup_update_page_stat(struct page *page,
2323 enum mem_cgroup_stat_index idx, int val)
2325 struct mem_cgroup *memcg;
2326 struct page_cgroup *pc = lookup_page_cgroup(page);
2327 unsigned long uninitialized_var(flags);
2329 if (mem_cgroup_disabled())
2332 VM_BUG_ON(!rcu_read_lock_held());
2333 memcg = pc->mem_cgroup;
2334 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2337 this_cpu_add(memcg->stat->count[idx], val);
2341 * size of first charge trial. "32" comes from vmscan.c's magic value.
2342 * TODO: maybe necessary to use big numbers in big irons.
2344 #define CHARGE_BATCH 32U
2345 struct memcg_stock_pcp {
2346 struct mem_cgroup *cached; /* this never be root cgroup */
2347 unsigned int nr_pages;
2348 struct work_struct work;
2349 unsigned long flags;
2350 #define FLUSHING_CACHED_CHARGE 0
2352 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2353 static DEFINE_MUTEX(percpu_charge_mutex);
2356 * consume_stock: Try to consume stocked charge on this cpu.
2357 * @memcg: memcg to consume from.
2358 * @nr_pages: how many pages to charge.
2360 * The charges will only happen if @memcg matches the current cpu's memcg
2361 * stock, and at least @nr_pages are available in that stock. Failure to
2362 * service an allocation will refill the stock.
2364 * returns true if successful, false otherwise.
2366 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2368 struct memcg_stock_pcp *stock;
2371 if (nr_pages > CHARGE_BATCH)
2374 stock = &get_cpu_var(memcg_stock);
2375 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2376 stock->nr_pages -= nr_pages;
2377 else /* need to call res_counter_charge */
2379 put_cpu_var(memcg_stock);
2384 * Returns stocks cached in percpu to res_counter and reset cached information.
2386 static void drain_stock(struct memcg_stock_pcp *stock)
2388 struct mem_cgroup *old = stock->cached;
2390 if (stock->nr_pages) {
2391 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2393 res_counter_uncharge(&old->res, bytes);
2394 if (do_swap_account)
2395 res_counter_uncharge(&old->memsw, bytes);
2396 stock->nr_pages = 0;
2398 stock->cached = NULL;
2402 * This must be called under preempt disabled or must be called by
2403 * a thread which is pinned to local cpu.
2405 static void drain_local_stock(struct work_struct *dummy)
2407 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2409 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2412 static void __init memcg_stock_init(void)
2416 for_each_possible_cpu(cpu) {
2417 struct memcg_stock_pcp *stock =
2418 &per_cpu(memcg_stock, cpu);
2419 INIT_WORK(&stock->work, drain_local_stock);
2424 * Cache charges(val) which is from res_counter, to local per_cpu area.
2425 * This will be consumed by consume_stock() function, later.
2427 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2429 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2431 if (stock->cached != memcg) { /* reset if necessary */
2433 stock->cached = memcg;
2435 stock->nr_pages += nr_pages;
2436 put_cpu_var(memcg_stock);
2440 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2441 * of the hierarchy under it. sync flag says whether we should block
2442 * until the work is done.
2444 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2448 /* Notify other cpus that system-wide "drain" is running */
2451 for_each_online_cpu(cpu) {
2452 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2453 struct mem_cgroup *memcg;
2455 memcg = stock->cached;
2456 if (!memcg || !stock->nr_pages)
2458 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2460 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2462 drain_local_stock(&stock->work);
2464 schedule_work_on(cpu, &stock->work);
2472 for_each_online_cpu(cpu) {
2473 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2474 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2475 flush_work(&stock->work);
2482 * Tries to drain stocked charges in other cpus. This function is asynchronous
2483 * and just put a work per cpu for draining localy on each cpu. Caller can
2484 * expects some charges will be back to res_counter later but cannot wait for
2487 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2490 * If someone calls draining, avoid adding more kworker runs.
2492 if (!mutex_trylock(&percpu_charge_mutex))
2494 drain_all_stock(root_memcg, false);
2495 mutex_unlock(&percpu_charge_mutex);
2498 /* This is a synchronous drain interface. */
2499 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2501 /* called when force_empty is called */
2502 mutex_lock(&percpu_charge_mutex);
2503 drain_all_stock(root_memcg, true);
2504 mutex_unlock(&percpu_charge_mutex);
2508 * This function drains percpu counter value from DEAD cpu and
2509 * move it to local cpu. Note that this function can be preempted.
2511 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2515 spin_lock(&memcg->pcp_counter_lock);
2516 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2517 long x = per_cpu(memcg->stat->count[i], cpu);
2519 per_cpu(memcg->stat->count[i], cpu) = 0;
2520 memcg->nocpu_base.count[i] += x;
2522 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2523 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2525 per_cpu(memcg->stat->events[i], cpu) = 0;
2526 memcg->nocpu_base.events[i] += x;
2528 spin_unlock(&memcg->pcp_counter_lock);
2531 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2532 unsigned long action,
2535 int cpu = (unsigned long)hcpu;
2536 struct memcg_stock_pcp *stock;
2537 struct mem_cgroup *iter;
2539 if (action == CPU_ONLINE)
2542 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2545 for_each_mem_cgroup(iter)
2546 mem_cgroup_drain_pcp_counter(iter, cpu);
2548 stock = &per_cpu(memcg_stock, cpu);
2554 /* See __mem_cgroup_try_charge() for details */
2556 CHARGE_OK, /* success */
2557 CHARGE_RETRY, /* need to retry but retry is not bad */
2558 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2559 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2562 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2563 unsigned int nr_pages, unsigned int min_pages,
2566 unsigned long csize = nr_pages * PAGE_SIZE;
2567 struct mem_cgroup *mem_over_limit;
2568 struct res_counter *fail_res;
2569 unsigned long flags = 0;
2572 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2575 if (!do_swap_account)
2577 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2581 res_counter_uncharge(&memcg->res, csize);
2582 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2583 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2585 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2587 * Never reclaim on behalf of optional batching, retry with a
2588 * single page instead.
2590 if (nr_pages > min_pages)
2591 return CHARGE_RETRY;
2593 if (!(gfp_mask & __GFP_WAIT))
2594 return CHARGE_WOULDBLOCK;
2596 if (gfp_mask & __GFP_NORETRY)
2597 return CHARGE_NOMEM;
2599 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2600 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2601 return CHARGE_RETRY;
2603 * Even though the limit is exceeded at this point, reclaim
2604 * may have been able to free some pages. Retry the charge
2605 * before killing the task.
2607 * Only for regular pages, though: huge pages are rather
2608 * unlikely to succeed so close to the limit, and we fall back
2609 * to regular pages anyway in case of failure.
2611 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2612 return CHARGE_RETRY;
2615 * At task move, charge accounts can be doubly counted. So, it's
2616 * better to wait until the end of task_move if something is going on.
2618 if (mem_cgroup_wait_acct_move(mem_over_limit))
2619 return CHARGE_RETRY;
2622 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2624 return CHARGE_NOMEM;
2628 * __mem_cgroup_try_charge() does
2629 * 1. detect memcg to be charged against from passed *mm and *ptr,
2630 * 2. update res_counter
2631 * 3. call memory reclaim if necessary.
2633 * In some special case, if the task is fatal, fatal_signal_pending() or
2634 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2635 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2636 * as possible without any hazards. 2: all pages should have a valid
2637 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2638 * pointer, that is treated as a charge to root_mem_cgroup.
2640 * So __mem_cgroup_try_charge() will return
2641 * 0 ... on success, filling *ptr with a valid memcg pointer.
2642 * -ENOMEM ... charge failure because of resource limits.
2643 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2645 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2646 * the oom-killer can be invoked.
2648 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2650 unsigned int nr_pages,
2651 struct mem_cgroup **ptr,
2654 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2655 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2656 struct mem_cgroup *memcg = NULL;
2660 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2661 * in system level. So, allow to go ahead dying process in addition to
2664 if (unlikely(test_thread_flag(TIF_MEMDIE)
2665 || fatal_signal_pending(current)))
2668 if (unlikely(task_in_memcg_oom(current)))
2672 * We always charge the cgroup the mm_struct belongs to.
2673 * The mm_struct's mem_cgroup changes on task migration if the
2674 * thread group leader migrates. It's possible that mm is not
2675 * set, if so charge the root memcg (happens for pagecache usage).
2678 *ptr = root_mem_cgroup;
2680 if (*ptr) { /* css should be a valid one */
2682 if (mem_cgroup_is_root(memcg))
2684 if (consume_stock(memcg, nr_pages))
2686 css_get(&memcg->css);
2688 struct task_struct *p;
2691 p = rcu_dereference(mm->owner);
2693 * Because we don't have task_lock(), "p" can exit.
2694 * In that case, "memcg" can point to root or p can be NULL with
2695 * race with swapoff. Then, we have small risk of mis-accouning.
2696 * But such kind of mis-account by race always happens because
2697 * we don't have cgroup_mutex(). It's overkill and we allo that
2699 * (*) swapoff at el will charge against mm-struct not against
2700 * task-struct. So, mm->owner can be NULL.
2702 memcg = mem_cgroup_from_task(p);
2704 memcg = root_mem_cgroup;
2705 if (mem_cgroup_is_root(memcg)) {
2709 if (consume_stock(memcg, nr_pages)) {
2711 * It seems dagerous to access memcg without css_get().
2712 * But considering how consume_stok works, it's not
2713 * necessary. If consume_stock success, some charges
2714 * from this memcg are cached on this cpu. So, we
2715 * don't need to call css_get()/css_tryget() before
2716 * calling consume_stock().
2721 /* after here, we may be blocked. we need to get refcnt */
2722 if (!css_tryget(&memcg->css)) {
2730 bool invoke_oom = oom && !nr_oom_retries;
2732 /* If killed, bypass charge */
2733 if (fatal_signal_pending(current)) {
2734 css_put(&memcg->css);
2738 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2739 nr_pages, invoke_oom);
2743 case CHARGE_RETRY: /* not in OOM situation but retry */
2745 css_put(&memcg->css);
2748 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2749 css_put(&memcg->css);
2751 case CHARGE_NOMEM: /* OOM routine works */
2752 if (!oom || invoke_oom) {
2753 css_put(&memcg->css);
2759 } while (ret != CHARGE_OK);
2761 if (batch > nr_pages)
2762 refill_stock(memcg, batch - nr_pages);
2763 css_put(&memcg->css);
2769 if (gfp_mask & __GFP_NOFAIL)
2773 *ptr = root_mem_cgroup;
2778 * Somemtimes we have to undo a charge we got by try_charge().
2779 * This function is for that and do uncharge, put css's refcnt.
2780 * gotten by try_charge().
2782 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2783 unsigned int nr_pages)
2785 if (!mem_cgroup_is_root(memcg)) {
2786 unsigned long bytes = nr_pages * PAGE_SIZE;
2788 res_counter_uncharge(&memcg->res, bytes);
2789 if (do_swap_account)
2790 res_counter_uncharge(&memcg->memsw, bytes);
2795 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2796 * This is useful when moving usage to parent cgroup.
2798 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2799 unsigned int nr_pages)
2801 unsigned long bytes = nr_pages * PAGE_SIZE;
2803 if (mem_cgroup_is_root(memcg))
2806 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2807 if (do_swap_account)
2808 res_counter_uncharge_until(&memcg->memsw,
2809 memcg->memsw.parent, bytes);
2813 * A helper function to get mem_cgroup from ID. must be called under
2814 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2815 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2816 * called against removed memcg.)
2818 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2820 struct cgroup_subsys_state *css;
2822 /* ID 0 is unused ID */
2825 css = css_lookup(&mem_cgroup_subsys, id);
2828 return mem_cgroup_from_css(css);
2831 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2833 struct mem_cgroup *memcg = NULL;
2834 struct page_cgroup *pc;
2838 VM_BUG_ON(!PageLocked(page));
2840 pc = lookup_page_cgroup(page);
2841 lock_page_cgroup(pc);
2842 if (PageCgroupUsed(pc)) {
2843 memcg = pc->mem_cgroup;
2844 if (memcg && !css_tryget(&memcg->css))
2846 } else if (PageSwapCache(page)) {
2847 ent.val = page_private(page);
2848 id = lookup_swap_cgroup_id(ent);
2850 memcg = mem_cgroup_lookup(id);
2851 if (memcg && !css_tryget(&memcg->css))
2855 unlock_page_cgroup(pc);
2859 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2861 unsigned int nr_pages,
2862 enum charge_type ctype,
2865 struct page_cgroup *pc = lookup_page_cgroup(page);
2866 struct zone *uninitialized_var(zone);
2867 struct lruvec *lruvec;
2868 bool was_on_lru = false;
2871 lock_page_cgroup(pc);
2872 VM_BUG_ON(PageCgroupUsed(pc));
2874 * we don't need page_cgroup_lock about tail pages, becase they are not
2875 * accessed by any other context at this point.
2879 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2880 * may already be on some other mem_cgroup's LRU. Take care of it.
2883 zone = page_zone(page);
2884 spin_lock_irq(&zone->lru_lock);
2885 if (PageLRU(page)) {
2886 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2888 del_page_from_lru_list(page, lruvec, page_lru(page));
2893 pc->mem_cgroup = memcg;
2895 * We access a page_cgroup asynchronously without lock_page_cgroup().
2896 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2897 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2898 * before USED bit, we need memory barrier here.
2899 * See mem_cgroup_add_lru_list(), etc.
2902 SetPageCgroupUsed(pc);
2906 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2907 VM_BUG_ON(PageLRU(page));
2909 add_page_to_lru_list(page, lruvec, page_lru(page));
2911 spin_unlock_irq(&zone->lru_lock);
2914 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2919 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2920 unlock_page_cgroup(pc);
2923 * "charge_statistics" updated event counter. Then, check it.
2924 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2925 * if they exceeds softlimit.
2927 memcg_check_events(memcg, page);
2930 static DEFINE_MUTEX(set_limit_mutex);
2932 #ifdef CONFIG_MEMCG_KMEM
2933 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2935 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2936 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2940 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2941 * in the memcg_cache_params struct.
2943 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2945 struct kmem_cache *cachep;
2947 VM_BUG_ON(p->is_root_cache);
2948 cachep = p->root_cache;
2949 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2952 #ifdef CONFIG_SLABINFO
2953 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2954 struct cftype *cft, struct seq_file *m)
2956 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2957 struct memcg_cache_params *params;
2959 if (!memcg_can_account_kmem(memcg))
2962 print_slabinfo_header(m);
2964 mutex_lock(&memcg->slab_caches_mutex);
2965 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2966 cache_show(memcg_params_to_cache(params), m);
2967 mutex_unlock(&memcg->slab_caches_mutex);
2973 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2975 struct res_counter *fail_res;
2976 struct mem_cgroup *_memcg;
2980 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2985 * Conditions under which we can wait for the oom_killer. Those are
2986 * the same conditions tested by the core page allocator
2988 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2991 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2994 if (ret == -EINTR) {
2996 * __mem_cgroup_try_charge() chosed to bypass to root due to
2997 * OOM kill or fatal signal. Since our only options are to
2998 * either fail the allocation or charge it to this cgroup, do
2999 * it as a temporary condition. But we can't fail. From a
3000 * kmem/slab perspective, the cache has already been selected,
3001 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3004 * This condition will only trigger if the task entered
3005 * memcg_charge_kmem in a sane state, but was OOM-killed during
3006 * __mem_cgroup_try_charge() above. Tasks that were already
3007 * dying when the allocation triggers should have been already
3008 * directed to the root cgroup in memcontrol.h
3010 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3011 if (do_swap_account)
3012 res_counter_charge_nofail(&memcg->memsw, size,
3016 res_counter_uncharge(&memcg->kmem, size);
3021 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3023 res_counter_uncharge(&memcg->res, size);
3024 if (do_swap_account)
3025 res_counter_uncharge(&memcg->memsw, size);
3028 if (res_counter_uncharge(&memcg->kmem, size))
3032 * Releases a reference taken in kmem_cgroup_css_offline in case
3033 * this last uncharge is racing with the offlining code or it is
3034 * outliving the memcg existence.
3036 * The memory barrier imposed by test&clear is paired with the
3037 * explicit one in memcg_kmem_mark_dead().
3039 if (memcg_kmem_test_and_clear_dead(memcg))
3040 css_put(&memcg->css);
3043 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3048 mutex_lock(&memcg->slab_caches_mutex);
3049 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3050 mutex_unlock(&memcg->slab_caches_mutex);
3054 * helper for acessing a memcg's index. It will be used as an index in the
3055 * child cache array in kmem_cache, and also to derive its name. This function
3056 * will return -1 when this is not a kmem-limited memcg.
3058 int memcg_cache_id(struct mem_cgroup *memcg)
3060 return memcg ? memcg->kmemcg_id : -1;
3064 * This ends up being protected by the set_limit mutex, during normal
3065 * operation, because that is its main call site.
3067 * But when we create a new cache, we can call this as well if its parent
3068 * is kmem-limited. That will have to hold set_limit_mutex as well.
3070 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3074 num = ida_simple_get(&kmem_limited_groups,
3075 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3079 * After this point, kmem_accounted (that we test atomically in
3080 * the beginning of this conditional), is no longer 0. This
3081 * guarantees only one process will set the following boolean
3082 * to true. We don't need test_and_set because we're protected
3083 * by the set_limit_mutex anyway.
3085 memcg_kmem_set_activated(memcg);
3087 ret = memcg_update_all_caches(num+1);
3089 ida_simple_remove(&kmem_limited_groups, num);
3090 memcg_kmem_clear_activated(memcg);
3094 memcg->kmemcg_id = num;
3095 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3096 mutex_init(&memcg->slab_caches_mutex);
3100 static size_t memcg_caches_array_size(int num_groups)
3103 if (num_groups <= 0)
3106 size = 2 * num_groups;
3107 if (size < MEMCG_CACHES_MIN_SIZE)
3108 size = MEMCG_CACHES_MIN_SIZE;
3109 else if (size > MEMCG_CACHES_MAX_SIZE)
3110 size = MEMCG_CACHES_MAX_SIZE;
3116 * We should update the current array size iff all caches updates succeed. This
3117 * can only be done from the slab side. The slab mutex needs to be held when
3120 void memcg_update_array_size(int num)
3122 if (num > memcg_limited_groups_array_size)
3123 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3126 static void kmem_cache_destroy_work_func(struct work_struct *w);
3128 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3130 struct memcg_cache_params *cur_params = s->memcg_params;
3132 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3134 if (num_groups > memcg_limited_groups_array_size) {
3136 ssize_t size = memcg_caches_array_size(num_groups);
3138 size *= sizeof(void *);
3139 size += offsetof(struct memcg_cache_params, memcg_caches);
3141 s->memcg_params = kzalloc(size, GFP_KERNEL);
3142 if (!s->memcg_params) {
3143 s->memcg_params = cur_params;
3147 s->memcg_params->is_root_cache = true;
3150 * There is the chance it will be bigger than
3151 * memcg_limited_groups_array_size, if we failed an allocation
3152 * in a cache, in which case all caches updated before it, will
3153 * have a bigger array.
3155 * But if that is the case, the data after
3156 * memcg_limited_groups_array_size is certainly unused
3158 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3159 if (!cur_params->memcg_caches[i])
3161 s->memcg_params->memcg_caches[i] =
3162 cur_params->memcg_caches[i];
3166 * Ideally, we would wait until all caches succeed, and only
3167 * then free the old one. But this is not worth the extra
3168 * pointer per-cache we'd have to have for this.
3170 * It is not a big deal if some caches are left with a size
3171 * bigger than the others. And all updates will reset this
3179 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3180 struct kmem_cache *root_cache)
3184 if (!memcg_kmem_enabled())
3188 size = offsetof(struct memcg_cache_params, memcg_caches);
3189 size += memcg_limited_groups_array_size * sizeof(void *);
3191 size = sizeof(struct memcg_cache_params);
3193 s->memcg_params = kzalloc(size, GFP_KERNEL);
3194 if (!s->memcg_params)
3198 s->memcg_params->memcg = memcg;
3199 s->memcg_params->root_cache = root_cache;
3200 INIT_WORK(&s->memcg_params->destroy,
3201 kmem_cache_destroy_work_func);
3203 s->memcg_params->is_root_cache = true;
3208 void memcg_release_cache(struct kmem_cache *s)
3210 struct kmem_cache *root;
3211 struct mem_cgroup *memcg;
3215 * This happens, for instance, when a root cache goes away before we
3218 if (!s->memcg_params)
3221 if (s->memcg_params->is_root_cache)
3224 memcg = s->memcg_params->memcg;
3225 id = memcg_cache_id(memcg);
3227 root = s->memcg_params->root_cache;
3228 root->memcg_params->memcg_caches[id] = NULL;
3230 mutex_lock(&memcg->slab_caches_mutex);
3231 list_del(&s->memcg_params->list);
3232 mutex_unlock(&memcg->slab_caches_mutex);
3234 css_put(&memcg->css);
3236 kfree(s->memcg_params);
3240 * During the creation a new cache, we need to disable our accounting mechanism
3241 * altogether. This is true even if we are not creating, but rather just
3242 * enqueing new caches to be created.
3244 * This is because that process will trigger allocations; some visible, like
3245 * explicit kmallocs to auxiliary data structures, name strings and internal
3246 * cache structures; some well concealed, like INIT_WORK() that can allocate
3247 * objects during debug.
3249 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3250 * to it. This may not be a bounded recursion: since the first cache creation
3251 * failed to complete (waiting on the allocation), we'll just try to create the
3252 * cache again, failing at the same point.
3254 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3255 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3256 * inside the following two functions.
3258 static inline void memcg_stop_kmem_account(void)
3260 VM_BUG_ON(!current->mm);
3261 current->memcg_kmem_skip_account++;
3264 static inline void memcg_resume_kmem_account(void)
3266 VM_BUG_ON(!current->mm);
3267 current->memcg_kmem_skip_account--;
3270 static void kmem_cache_destroy_work_func(struct work_struct *w)
3272 struct kmem_cache *cachep;
3273 struct memcg_cache_params *p;
3275 p = container_of(w, struct memcg_cache_params, destroy);
3277 cachep = memcg_params_to_cache(p);
3280 * If we get down to 0 after shrink, we could delete right away.
3281 * However, memcg_release_pages() already puts us back in the workqueue
3282 * in that case. If we proceed deleting, we'll get a dangling
3283 * reference, and removing the object from the workqueue in that case
3284 * is unnecessary complication. We are not a fast path.
3286 * Note that this case is fundamentally different from racing with
3287 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3288 * kmem_cache_shrink, not only we would be reinserting a dead cache
3289 * into the queue, but doing so from inside the worker racing to
3292 * So if we aren't down to zero, we'll just schedule a worker and try
3295 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3296 kmem_cache_shrink(cachep);
3297 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)