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
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
68 #include <net/tcp_memcontrol.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 #define MEM_CGROUP_RECLAIM_RETRIES 5
79 static struct mem_cgroup *root_mem_cgroup __read_mostly;
80 struct cgroup_subsys_state *mem_cgroup_root_css __read_mostly;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 int do_swap_account __read_mostly;
86 #define do_swap_account 0
89 static const char * const mem_cgroup_stat_names[] = {
99 static const char * const mem_cgroup_events_names[] = {
106 static const char * const mem_cgroup_lru_names[] = {
114 #define THRESHOLDS_EVENTS_TARGET 128
115 #define SOFTLIMIT_EVENTS_TARGET 1024
116 #define NUMAINFO_EVENTS_TARGET 1024
119 * Cgroups above their limits are maintained in a RB-Tree, independent of
120 * their hierarchy representation
123 struct mem_cgroup_tree_per_zone {
124 struct rb_root rb_root;
128 struct mem_cgroup_tree_per_node {
129 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
132 struct mem_cgroup_tree {
133 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
136 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
139 struct mem_cgroup_eventfd_list {
140 struct list_head list;
141 struct eventfd_ctx *eventfd;
145 * cgroup_event represents events which userspace want to receive.
147 struct mem_cgroup_event {
149 * memcg which the event belongs to.
151 struct mem_cgroup *memcg;
153 * eventfd to signal userspace about the event.
155 struct eventfd_ctx *eventfd;
157 * Each of these stored in a list by the cgroup.
159 struct list_head list;
161 * register_event() callback will be used to add new userspace
162 * waiter for changes related to this event. Use eventfd_signal()
163 * on eventfd to send notification to userspace.
165 int (*register_event)(struct mem_cgroup *memcg,
166 struct eventfd_ctx *eventfd, const char *args);
168 * unregister_event() callback will be called when userspace closes
169 * the eventfd or on cgroup removing. This callback must be set,
170 * if you want provide notification functionality.
172 void (*unregister_event)(struct mem_cgroup *memcg,
173 struct eventfd_ctx *eventfd);
175 * All fields below needed to unregister event when
176 * userspace closes eventfd.
179 wait_queue_head_t *wqh;
181 struct work_struct remove;
184 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
185 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
187 /* Stuffs for move charges at task migration. */
189 * Types of charges to be moved.
191 #define MOVE_ANON 0x1U
192 #define MOVE_FILE 0x2U
193 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
195 /* "mc" and its members are protected by cgroup_mutex */
196 static struct move_charge_struct {
197 spinlock_t lock; /* for from, to */
198 struct mem_cgroup *from;
199 struct mem_cgroup *to;
201 unsigned long precharge;
202 unsigned long moved_charge;
203 unsigned long moved_swap;
204 struct task_struct *moving_task; /* a task moving charges */
205 wait_queue_head_t waitq; /* a waitq for other context */
207 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
208 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
212 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
213 * limit reclaim to prevent infinite loops, if they ever occur.
215 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
216 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
219 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
220 MEM_CGROUP_CHARGE_TYPE_ANON,
221 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
222 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
226 /* for encoding cft->private value on file */
234 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
235 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
236 #define MEMFILE_ATTR(val) ((val) & 0xffff)
237 /* Used for OOM nofiier */
238 #define OOM_CONTROL (0)
241 * The memcg_create_mutex will be held whenever a new cgroup is created.
242 * As a consequence, any change that needs to protect against new child cgroups
243 * appearing has to hold it as well.
245 static DEFINE_MUTEX(memcg_create_mutex);
247 /* Some nice accessors for the vmpressure. */
248 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
251 memcg = root_mem_cgroup;
252 return &memcg->vmpressure;
255 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
257 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
260 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
262 return (memcg == root_mem_cgroup);
266 * We restrict the id in the range of [1, 65535], so it can fit into
269 #define MEM_CGROUP_ID_MAX USHRT_MAX
271 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
273 return memcg->css.id;
277 * A helper function to get mem_cgroup from ID. must be called under
278 * rcu_read_lock(). The caller is responsible for calling
279 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
280 * refcnt from swap can be called against removed memcg.)
282 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
284 struct cgroup_subsys_state *css;
286 css = css_from_id(id, &memory_cgrp_subsys);
287 return mem_cgroup_from_css(css);
290 /* Writing them here to avoid exposing memcg's inner layout */
291 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
293 void sock_update_memcg(struct sock *sk)
295 if (mem_cgroup_sockets_enabled) {
296 struct mem_cgroup *memcg;
297 struct cg_proto *cg_proto;
299 BUG_ON(!sk->sk_prot->proto_cgroup);
301 /* Socket cloning can throw us here with sk_cgrp already
302 * filled. It won't however, necessarily happen from
303 * process context. So the test for root memcg given
304 * the current task's memcg won't help us in this case.
306 * Respecting the original socket's memcg is a better
307 * decision in this case.
310 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
311 css_get(&sk->sk_cgrp->memcg->css);
316 memcg = mem_cgroup_from_task(current);
317 cg_proto = sk->sk_prot->proto_cgroup(memcg);
318 if (!mem_cgroup_is_root(memcg) &&
319 memcg_proto_active(cg_proto) &&
320 css_tryget_online(&memcg->css)) {
321 sk->sk_cgrp = cg_proto;
326 EXPORT_SYMBOL(sock_update_memcg);
328 void sock_release_memcg(struct sock *sk)
330 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
331 struct mem_cgroup *memcg;
332 WARN_ON(!sk->sk_cgrp->memcg);
333 memcg = sk->sk_cgrp->memcg;
334 css_put(&sk->sk_cgrp->memcg->css);
338 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
340 if (!memcg || mem_cgroup_is_root(memcg))
343 return &memcg->tcp_mem;
345 EXPORT_SYMBOL(tcp_proto_cgroup);
349 #ifdef CONFIG_MEMCG_KMEM
351 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida);
362 int memcg_nr_cache_ids;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 struct static_key memcg_kmem_enabled_key;
399 EXPORT_SYMBOL(memcg_kmem_enabled_key);
401 #endif /* CONFIG_MEMCG_KMEM */
403 static struct mem_cgroup_per_zone *
404 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
406 int nid = zone_to_nid(zone);
407 int zid = zone_idx(zone);
409 return &memcg->nodeinfo[nid]->zoneinfo[zid];
413 * mem_cgroup_css_from_page - css of the memcg associated with a page
414 * @page: page of interest
416 * If memcg is bound to the default hierarchy, css of the memcg associated
417 * with @page is returned. The returned css remains associated with @page
418 * until it is released.
420 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
423 * XXX: The above description of behavior on the default hierarchy isn't
424 * strictly true yet as replace_page_cache_page() can modify the
425 * association before @page is released even on the default hierarchy;
426 * however, the current and planned usages don't mix the the two functions
427 * and replace_page_cache_page() will soon be updated to make the invariant
430 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
432 struct mem_cgroup *memcg;
436 memcg = page->mem_cgroup;
438 if (!memcg || !cgroup_on_dfl(memcg->css.cgroup))
439 memcg = root_mem_cgroup;
445 static struct mem_cgroup_per_zone *
446 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
448 int nid = page_to_nid(page);
449 int zid = page_zonenum(page);
451 return &memcg->nodeinfo[nid]->zoneinfo[zid];
454 static struct mem_cgroup_tree_per_zone *
455 soft_limit_tree_node_zone(int nid, int zid)
457 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
460 static struct mem_cgroup_tree_per_zone *
461 soft_limit_tree_from_page(struct page *page)
463 int nid = page_to_nid(page);
464 int zid = page_zonenum(page);
466 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
469 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
470 struct mem_cgroup_tree_per_zone *mctz,
471 unsigned long new_usage_in_excess)
473 struct rb_node **p = &mctz->rb_root.rb_node;
474 struct rb_node *parent = NULL;
475 struct mem_cgroup_per_zone *mz_node;
480 mz->usage_in_excess = new_usage_in_excess;
481 if (!mz->usage_in_excess)
485 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
487 if (mz->usage_in_excess < mz_node->usage_in_excess)
490 * We can't avoid mem cgroups that are over their soft
491 * limit by the same amount
493 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
496 rb_link_node(&mz->tree_node, parent, p);
497 rb_insert_color(&mz->tree_node, &mctz->rb_root);
501 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
502 struct mem_cgroup_tree_per_zone *mctz)
506 rb_erase(&mz->tree_node, &mctz->rb_root);
510 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
511 struct mem_cgroup_tree_per_zone *mctz)
515 spin_lock_irqsave(&mctz->lock, flags);
516 __mem_cgroup_remove_exceeded(mz, mctz);
517 spin_unlock_irqrestore(&mctz->lock, flags);
520 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
522 unsigned long nr_pages = page_counter_read(&memcg->memory);
523 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
524 unsigned long excess = 0;
526 if (nr_pages > soft_limit)
527 excess = nr_pages - soft_limit;
532 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
534 unsigned long excess;
535 struct mem_cgroup_per_zone *mz;
536 struct mem_cgroup_tree_per_zone *mctz;
538 mctz = soft_limit_tree_from_page(page);
540 * Necessary to update all ancestors when hierarchy is used.
541 * because their event counter is not touched.
543 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
544 mz = mem_cgroup_page_zoneinfo(memcg, page);
545 excess = soft_limit_excess(memcg);
547 * We have to update the tree if mz is on RB-tree or
548 * mem is over its softlimit.
550 if (excess || mz->on_tree) {
553 spin_lock_irqsave(&mctz->lock, flags);
554 /* if on-tree, remove it */
556 __mem_cgroup_remove_exceeded(mz, mctz);
558 * Insert again. mz->usage_in_excess will be updated.
559 * If excess is 0, no tree ops.
561 __mem_cgroup_insert_exceeded(mz, mctz, excess);
562 spin_unlock_irqrestore(&mctz->lock, flags);
567 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
569 struct mem_cgroup_tree_per_zone *mctz;
570 struct mem_cgroup_per_zone *mz;
574 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
575 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
576 mctz = soft_limit_tree_node_zone(nid, zid);
577 mem_cgroup_remove_exceeded(mz, mctz);
582 static struct mem_cgroup_per_zone *
583 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
585 struct rb_node *rightmost = NULL;
586 struct mem_cgroup_per_zone *mz;
590 rightmost = rb_last(&mctz->rb_root);
592 goto done; /* Nothing to reclaim from */
594 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
596 * Remove the node now but someone else can add it back,
597 * we will to add it back at the end of reclaim to its correct
598 * position in the tree.
600 __mem_cgroup_remove_exceeded(mz, mctz);
601 if (!soft_limit_excess(mz->memcg) ||
602 !css_tryget_online(&mz->memcg->css))
608 static struct mem_cgroup_per_zone *
609 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
611 struct mem_cgroup_per_zone *mz;
613 spin_lock_irq(&mctz->lock);
614 mz = __mem_cgroup_largest_soft_limit_node(mctz);
615 spin_unlock_irq(&mctz->lock);
620 * Implementation Note: reading percpu statistics for memcg.
622 * Both of vmstat[] and percpu_counter has threshold and do periodic
623 * synchronization to implement "quick" read. There are trade-off between
624 * reading cost and precision of value. Then, we may have a chance to implement
625 * a periodic synchronizion of counter in memcg's counter.
627 * But this _read() function is used for user interface now. The user accounts
628 * memory usage by memory cgroup and he _always_ requires exact value because
629 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
630 * have to visit all online cpus and make sum. So, for now, unnecessary
631 * synchronization is not implemented. (just implemented for cpu hotplug)
633 * If there are kernel internal actions which can make use of some not-exact
634 * value, and reading all cpu value can be performance bottleneck in some
635 * common workload, threashold and synchonization as vmstat[] should be
638 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
639 enum mem_cgroup_stat_index idx)
644 for_each_possible_cpu(cpu)
645 val += per_cpu(memcg->stat->count[idx], cpu);
649 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
650 enum mem_cgroup_events_index idx)
652 unsigned long val = 0;
655 for_each_possible_cpu(cpu)
656 val += per_cpu(memcg->stat->events[idx], cpu);
660 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
665 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
666 * counted as CACHE even if it's on ANON LRU.
669 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
672 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
675 if (PageTransHuge(page))
676 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
679 /* pagein of a big page is an event. So, ignore page size */
681 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
683 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
684 nr_pages = -nr_pages; /* for event */
687 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
690 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
692 unsigned int lru_mask)
694 unsigned long nr = 0;
697 VM_BUG_ON((unsigned)nid >= nr_node_ids);
699 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
700 struct mem_cgroup_per_zone *mz;
704 if (!(BIT(lru) & lru_mask))
706 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
707 nr += mz->lru_size[lru];
713 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
714 unsigned int lru_mask)
716 unsigned long nr = 0;
719 for_each_node_state(nid, N_MEMORY)
720 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
724 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
725 enum mem_cgroup_events_target target)
727 unsigned long val, next;
729 val = __this_cpu_read(memcg->stat->nr_page_events);
730 next = __this_cpu_read(memcg->stat->targets[target]);
731 /* from time_after() in jiffies.h */
732 if ((long)next - (long)val < 0) {
734 case MEM_CGROUP_TARGET_THRESH:
735 next = val + THRESHOLDS_EVENTS_TARGET;
737 case MEM_CGROUP_TARGET_SOFTLIMIT:
738 next = val + SOFTLIMIT_EVENTS_TARGET;
740 case MEM_CGROUP_TARGET_NUMAINFO:
741 next = val + NUMAINFO_EVENTS_TARGET;
746 __this_cpu_write(memcg->stat->targets[target], next);
753 * Check events in order.
756 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
758 /* threshold event is triggered in finer grain than soft limit */
759 if (unlikely(mem_cgroup_event_ratelimit(memcg,
760 MEM_CGROUP_TARGET_THRESH))) {
762 bool do_numainfo __maybe_unused;
764 do_softlimit = mem_cgroup_event_ratelimit(memcg,
765 MEM_CGROUP_TARGET_SOFTLIMIT);
767 do_numainfo = mem_cgroup_event_ratelimit(memcg,
768 MEM_CGROUP_TARGET_NUMAINFO);
770 mem_cgroup_threshold(memcg);
771 if (unlikely(do_softlimit))
772 mem_cgroup_update_tree(memcg, page);
774 if (unlikely(do_numainfo))
775 atomic_inc(&memcg->numainfo_events);
780 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
783 * mm_update_next_owner() may clear mm->owner to NULL
784 * if it races with swapoff, page migration, etc.
785 * So this can be called with p == NULL.
790 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
792 EXPORT_SYMBOL(mem_cgroup_from_task);
794 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
796 struct mem_cgroup *memcg = NULL;
801 * Page cache insertions can happen withou an
802 * actual mm context, e.g. during disk probing
803 * on boot, loopback IO, acct() writes etc.
806 memcg = root_mem_cgroup;
808 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
809 if (unlikely(!memcg))
810 memcg = root_mem_cgroup;
812 } while (!css_tryget_online(&memcg->css));
818 * mem_cgroup_iter - iterate over memory cgroup hierarchy
819 * @root: hierarchy root
820 * @prev: previously returned memcg, NULL on first invocation
821 * @reclaim: cookie for shared reclaim walks, NULL for full walks
823 * Returns references to children of the hierarchy below @root, or
824 * @root itself, or %NULL after a full round-trip.
826 * Caller must pass the return value in @prev on subsequent
827 * invocations for reference counting, or use mem_cgroup_iter_break()
828 * to cancel a hierarchy walk before the round-trip is complete.
830 * Reclaimers can specify a zone and a priority level in @reclaim to
831 * divide up the memcgs in the hierarchy among all concurrent
832 * reclaimers operating on the same zone and priority.
834 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
835 struct mem_cgroup *prev,
836 struct mem_cgroup_reclaim_cookie *reclaim)
838 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
839 struct cgroup_subsys_state *css = NULL;
840 struct mem_cgroup *memcg = NULL;
841 struct mem_cgroup *pos = NULL;
843 if (mem_cgroup_disabled())
847 root = root_mem_cgroup;
849 if (prev && !reclaim)
852 if (!root->use_hierarchy && root != root_mem_cgroup) {
861 struct mem_cgroup_per_zone *mz;
863 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
864 iter = &mz->iter[reclaim->priority];
866 if (prev && reclaim->generation != iter->generation)
870 pos = READ_ONCE(iter->position);
872 * A racing update may change the position and
873 * put the last reference, hence css_tryget(),
874 * or retry to see the updated position.
876 } while (pos && !css_tryget(&pos->css));
883 css = css_next_descendant_pre(css, &root->css);
886 * Reclaimers share the hierarchy walk, and a
887 * new one might jump in right at the end of
888 * the hierarchy - make sure they see at least
889 * one group and restart from the beginning.
897 * Verify the css and acquire a reference. The root
898 * is provided by the caller, so we know it's alive
899 * and kicking, and don't take an extra reference.
901 memcg = mem_cgroup_from_css(css);
903 if (css == &root->css)
906 if (css_tryget(css)) {
908 * Make sure the memcg is initialized:
909 * mem_cgroup_css_online() orders the the
910 * initialization against setting the flag.
912 if (smp_load_acquire(&memcg->initialized))
922 if (cmpxchg(&iter->position, pos, memcg) == pos) {
924 css_get(&memcg->css);
930 * pairs with css_tryget when dereferencing iter->position
939 reclaim->generation = iter->generation;
945 if (prev && prev != root)
952 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
953 * @root: hierarchy root
954 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
956 void mem_cgroup_iter_break(struct mem_cgroup *root,
957 struct mem_cgroup *prev)
960 root = root_mem_cgroup;
961 if (prev && prev != root)
966 * Iteration constructs for visiting all cgroups (under a tree). If
967 * loops are exited prematurely (break), mem_cgroup_iter_break() must
968 * be used for reference counting.
970 #define for_each_mem_cgroup_tree(iter, root) \
971 for (iter = mem_cgroup_iter(root, NULL, NULL); \
973 iter = mem_cgroup_iter(root, iter, NULL))
975 #define for_each_mem_cgroup(iter) \
976 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
978 iter = mem_cgroup_iter(NULL, iter, NULL))
981 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
982 * @zone: zone of the wanted lruvec
983 * @memcg: memcg of the wanted lruvec
985 * Returns the lru list vector holding pages for the given @zone and
986 * @mem. This can be the global zone lruvec, if the memory controller
989 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
990 struct mem_cgroup *memcg)
992 struct mem_cgroup_per_zone *mz;
993 struct lruvec *lruvec;
995 if (mem_cgroup_disabled()) {
996 lruvec = &zone->lruvec;
1000 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1001 lruvec = &mz->lruvec;
1004 * Since a node can be onlined after the mem_cgroup was created,
1005 * we have to be prepared to initialize lruvec->zone here;
1006 * and if offlined then reonlined, we need to reinitialize it.
1008 if (unlikely(lruvec->zone != zone))
1009 lruvec->zone = zone;
1014 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1016 * @zone: zone of the page
1018 * This function is only safe when following the LRU page isolation
1019 * and putback protocol: the LRU lock must be held, and the page must
1020 * either be PageLRU() or the caller must have isolated/allocated it.
1022 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1024 struct mem_cgroup_per_zone *mz;
1025 struct mem_cgroup *memcg;
1026 struct lruvec *lruvec;
1028 if (mem_cgroup_disabled()) {
1029 lruvec = &zone->lruvec;
1033 memcg = page->mem_cgroup;
1035 * Swapcache readahead pages are added to the LRU - and
1036 * possibly migrated - before they are charged.
1039 memcg = root_mem_cgroup;
1041 mz = mem_cgroup_page_zoneinfo(memcg, page);
1042 lruvec = &mz->lruvec;
1045 * Since a node can be onlined after the mem_cgroup was created,
1046 * we have to be prepared to initialize lruvec->zone here;
1047 * and if offlined then reonlined, we need to reinitialize it.
1049 if (unlikely(lruvec->zone != zone))
1050 lruvec->zone = zone;
1055 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1056 * @lruvec: mem_cgroup per zone lru vector
1057 * @lru: index of lru list the page is sitting on
1058 * @nr_pages: positive when adding or negative when removing
1060 * This function must be called when a page is added to or removed from an
1063 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1066 struct mem_cgroup_per_zone *mz;
1067 unsigned long *lru_size;
1069 if (mem_cgroup_disabled())
1072 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1073 lru_size = mz->lru_size + lru;
1074 *lru_size += nr_pages;
1075 VM_BUG_ON((long)(*lru_size) < 0);
1078 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1080 struct mem_cgroup *task_memcg;
1081 struct task_struct *p;
1084 p = find_lock_task_mm(task);
1086 task_memcg = get_mem_cgroup_from_mm(p->mm);
1090 * All threads may have already detached their mm's, but the oom
1091 * killer still needs to detect if they have already been oom
1092 * killed to prevent needlessly killing additional tasks.
1095 task_memcg = mem_cgroup_from_task(task);
1096 css_get(&task_memcg->css);
1099 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1100 css_put(&task_memcg->css);
1104 #define mem_cgroup_from_counter(counter, member) \
1105 container_of(counter, struct mem_cgroup, member)
1108 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1109 * @memcg: the memory cgroup
1111 * Returns the maximum amount of memory @mem can be charged with, in
1114 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1116 unsigned long margin = 0;
1117 unsigned long count;
1118 unsigned long limit;
1120 count = page_counter_read(&memcg->memory);
1121 limit = READ_ONCE(memcg->memory.limit);
1123 margin = limit - count;
1125 if (do_swap_account) {
1126 count = page_counter_read(&memcg->memsw);
1127 limit = READ_ONCE(memcg->memsw.limit);
1129 margin = min(margin, limit - count);
1136 * A routine for checking "mem" is under move_account() or not.
1138 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1139 * moving cgroups. This is for waiting at high-memory pressure
1142 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1144 struct mem_cgroup *from;
1145 struct mem_cgroup *to;
1148 * Unlike task_move routines, we access mc.to, mc.from not under
1149 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1151 spin_lock(&mc.lock);
1157 ret = mem_cgroup_is_descendant(from, memcg) ||
1158 mem_cgroup_is_descendant(to, memcg);
1160 spin_unlock(&mc.lock);
1164 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1166 if (mc.moving_task && current != mc.moving_task) {
1167 if (mem_cgroup_under_move(memcg)) {
1169 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1170 /* moving charge context might have finished. */
1173 finish_wait(&mc.waitq, &wait);
1180 #define K(x) ((x) << (PAGE_SHIFT-10))
1182 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1183 * @memcg: The memory cgroup that went over limit
1184 * @p: Task that is going to be killed
1186 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1189 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1191 /* oom_info_lock ensures that parallel ooms do not interleave */
1192 static DEFINE_MUTEX(oom_info_lock);
1193 struct mem_cgroup *iter;
1196 mutex_lock(&oom_info_lock);
1200 pr_info("Task in ");
1201 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1202 pr_cont(" killed as a result of limit of ");
1204 pr_info("Memory limit reached of cgroup ");
1207 pr_cont_cgroup_path(memcg->css.cgroup);
1212 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1213 K((u64)page_counter_read(&memcg->memory)),
1214 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1215 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1216 K((u64)page_counter_read(&memcg->memsw)),
1217 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1218 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1219 K((u64)page_counter_read(&memcg->kmem)),
1220 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1222 for_each_mem_cgroup_tree(iter, memcg) {
1223 pr_info("Memory cgroup stats for ");
1224 pr_cont_cgroup_path(iter->css.cgroup);
1227 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1228 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1230 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1231 K(mem_cgroup_read_stat(iter, i)));
1234 for (i = 0; i < NR_LRU_LISTS; i++)
1235 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1236 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1240 mutex_unlock(&oom_info_lock);
1244 * This function returns the number of memcg under hierarchy tree. Returns
1245 * 1(self count) if no children.
1247 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1250 struct mem_cgroup *iter;
1252 for_each_mem_cgroup_tree(iter, memcg)
1258 * Return the memory (and swap, if configured) limit for a memcg.
1260 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1262 unsigned long limit;
1264 limit = memcg->memory.limit;
1265 if (mem_cgroup_swappiness(memcg)) {
1266 unsigned long memsw_limit;
1268 memsw_limit = memcg->memsw.limit;
1269 limit = min(limit + total_swap_pages, memsw_limit);
1274 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1277 struct oom_control oc = {
1280 .gfp_mask = gfp_mask,
1283 struct mem_cgroup *iter;
1284 unsigned long chosen_points = 0;
1285 unsigned long totalpages;
1286 unsigned int points = 0;
1287 struct task_struct *chosen = NULL;
1289 mutex_lock(&oom_lock);
1292 * If current has a pending SIGKILL or is exiting, then automatically
1293 * select it. The goal is to allow it to allocate so that it may
1294 * quickly exit and free its memory.
1296 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1297 mark_oom_victim(current);
1301 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1302 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1303 for_each_mem_cgroup_tree(iter, memcg) {
1304 struct css_task_iter it;
1305 struct task_struct *task;
1307 css_task_iter_start(&iter->css, &it);
1308 while ((task = css_task_iter_next(&it))) {
1309 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1310 case OOM_SCAN_SELECT:
1312 put_task_struct(chosen);
1314 chosen_points = ULONG_MAX;
1315 get_task_struct(chosen);
1317 case OOM_SCAN_CONTINUE:
1319 case OOM_SCAN_ABORT:
1320 css_task_iter_end(&it);
1321 mem_cgroup_iter_break(memcg, iter);
1323 put_task_struct(chosen);
1328 points = oom_badness(task, memcg, NULL, totalpages);
1329 if (!points || points < chosen_points)
1331 /* Prefer thread group leaders for display purposes */
1332 if (points == chosen_points &&
1333 thread_group_leader(chosen))
1337 put_task_struct(chosen);
1339 chosen_points = points;
1340 get_task_struct(chosen);
1342 css_task_iter_end(&it);
1346 points = chosen_points * 1000 / totalpages;
1347 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1348 "Memory cgroup out of memory");
1351 mutex_unlock(&oom_lock);
1354 #if MAX_NUMNODES > 1
1357 * test_mem_cgroup_node_reclaimable
1358 * @memcg: the target memcg
1359 * @nid: the node ID to be checked.
1360 * @noswap : specify true here if the user wants flle only information.
1362 * This function returns whether the specified memcg contains any
1363 * reclaimable pages on a node. Returns true if there are any reclaimable
1364 * pages in the node.
1366 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1367 int nid, bool noswap)
1369 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1371 if (noswap || !total_swap_pages)
1373 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1380 * Always updating the nodemask is not very good - even if we have an empty
1381 * list or the wrong list here, we can start from some node and traverse all
1382 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1385 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1389 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1390 * pagein/pageout changes since the last update.
1392 if (!atomic_read(&memcg->numainfo_events))
1394 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1397 /* make a nodemask where this memcg uses memory from */
1398 memcg->scan_nodes = node_states[N_MEMORY];
1400 for_each_node_mask(nid, node_states[N_MEMORY]) {
1402 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1403 node_clear(nid, memcg->scan_nodes);
1406 atomic_set(&memcg->numainfo_events, 0);
1407 atomic_set(&memcg->numainfo_updating, 0);
1411 * Selecting a node where we start reclaim from. Because what we need is just
1412 * reducing usage counter, start from anywhere is O,K. Considering
1413 * memory reclaim from current node, there are pros. and cons.
1415 * Freeing memory from current node means freeing memory from a node which
1416 * we'll use or we've used. So, it may make LRU bad. And if several threads
1417 * hit limits, it will see a contention on a node. But freeing from remote
1418 * node means more costs for memory reclaim because of memory latency.
1420 * Now, we use round-robin. Better algorithm is welcomed.
1422 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1426 mem_cgroup_may_update_nodemask(memcg);
1427 node = memcg->last_scanned_node;
1429 node = next_node(node, memcg->scan_nodes);
1430 if (node == MAX_NUMNODES)
1431 node = first_node(memcg->scan_nodes);
1433 * We call this when we hit limit, not when pages are added to LRU.
1434 * No LRU may hold pages because all pages are UNEVICTABLE or
1435 * memcg is too small and all pages are not on LRU. In that case,
1436 * we use curret node.
1438 if (unlikely(node == MAX_NUMNODES))
1439 node = numa_node_id();
1441 memcg->last_scanned_node = node;
1445 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1451 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1454 unsigned long *total_scanned)
1456 struct mem_cgroup *victim = NULL;
1459 unsigned long excess;
1460 unsigned long nr_scanned;
1461 struct mem_cgroup_reclaim_cookie reclaim = {
1466 excess = soft_limit_excess(root_memcg);
1469 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1474 * If we have not been able to reclaim
1475 * anything, it might because there are
1476 * no reclaimable pages under this hierarchy
1481 * We want to do more targeted reclaim.
1482 * excess >> 2 is not to excessive so as to
1483 * reclaim too much, nor too less that we keep
1484 * coming back to reclaim from this cgroup
1486 if (total >= (excess >> 2) ||
1487 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1492 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1494 *total_scanned += nr_scanned;
1495 if (!soft_limit_excess(root_memcg))
1498 mem_cgroup_iter_break(root_memcg, victim);
1502 #ifdef CONFIG_LOCKDEP
1503 static struct lockdep_map memcg_oom_lock_dep_map = {
1504 .name = "memcg_oom_lock",
1508 static DEFINE_SPINLOCK(memcg_oom_lock);
1511 * Check OOM-Killer is already running under our hierarchy.
1512 * If someone is running, return false.
1514 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1516 struct mem_cgroup *iter, *failed = NULL;
1518 spin_lock(&memcg_oom_lock);
1520 for_each_mem_cgroup_tree(iter, memcg) {
1521 if (iter->oom_lock) {
1523 * this subtree of our hierarchy is already locked
1524 * so we cannot give a lock.
1527 mem_cgroup_iter_break(memcg, iter);
1530 iter->oom_lock = true;
1535 * OK, we failed to lock the whole subtree so we have
1536 * to clean up what we set up to the failing subtree
1538 for_each_mem_cgroup_tree(iter, memcg) {
1539 if (iter == failed) {
1540 mem_cgroup_iter_break(memcg, iter);
1543 iter->oom_lock = false;
1546 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1548 spin_unlock(&memcg_oom_lock);
1553 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1555 struct mem_cgroup *iter;
1557 spin_lock(&memcg_oom_lock);
1558 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1559 for_each_mem_cgroup_tree(iter, memcg)
1560 iter->oom_lock = false;
1561 spin_unlock(&memcg_oom_lock);
1564 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1566 struct mem_cgroup *iter;
1568 spin_lock(&memcg_oom_lock);
1569 for_each_mem_cgroup_tree(iter, memcg)
1571 spin_unlock(&memcg_oom_lock);
1574 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1576 struct mem_cgroup *iter;
1579 * When a new child is created while the hierarchy is under oom,
1580 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1582 spin_lock(&memcg_oom_lock);
1583 for_each_mem_cgroup_tree(iter, memcg)
1584 if (iter->under_oom > 0)
1586 spin_unlock(&memcg_oom_lock);
1589 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1591 struct oom_wait_info {
1592 struct mem_cgroup *memcg;
1596 static int memcg_oom_wake_function(wait_queue_t *wait,
1597 unsigned mode, int sync, void *arg)
1599 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1600 struct mem_cgroup *oom_wait_memcg;
1601 struct oom_wait_info *oom_wait_info;
1603 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1604 oom_wait_memcg = oom_wait_info->memcg;
1606 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1607 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1609 return autoremove_wake_function(wait, mode, sync, arg);
1612 static void memcg_oom_recover(struct mem_cgroup *memcg)
1615 * For the following lockless ->under_oom test, the only required
1616 * guarantee is that it must see the state asserted by an OOM when
1617 * this function is called as a result of userland actions
1618 * triggered by the notification of the OOM. This is trivially
1619 * achieved by invoking mem_cgroup_mark_under_oom() before
1620 * triggering notification.
1622 if (memcg && memcg->under_oom)
1623 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1626 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1628 if (!current->memcg_oom.may_oom)
1631 * We are in the middle of the charge context here, so we
1632 * don't want to block when potentially sitting on a callstack
1633 * that holds all kinds of filesystem and mm locks.
1635 * Also, the caller may handle a failed allocation gracefully
1636 * (like optional page cache readahead) and so an OOM killer
1637 * invocation might not even be necessary.
1639 * That's why we don't do anything here except remember the
1640 * OOM context and then deal with it at the end of the page
1641 * fault when the stack is unwound, the locks are released,
1642 * and when we know whether the fault was overall successful.
1644 css_get(&memcg->css);
1645 current->memcg_oom.memcg = memcg;
1646 current->memcg_oom.gfp_mask = mask;
1647 current->memcg_oom.order = order;
1651 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1652 * @handle: actually kill/wait or just clean up the OOM state
1654 * This has to be called at the end of a page fault if the memcg OOM
1655 * handler was enabled.
1657 * Memcg supports userspace OOM handling where failed allocations must
1658 * sleep on a waitqueue until the userspace task resolves the
1659 * situation. Sleeping directly in the charge context with all kinds
1660 * of locks held is not a good idea, instead we remember an OOM state
1661 * in the task and mem_cgroup_oom_synchronize() has to be called at
1662 * the end of the page fault to complete the OOM handling.
1664 * Returns %true if an ongoing memcg OOM situation was detected and
1665 * completed, %false otherwise.
1667 bool mem_cgroup_oom_synchronize(bool handle)
1669 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1670 struct oom_wait_info owait;
1673 /* OOM is global, do not handle */
1677 if (!handle || oom_killer_disabled)
1680 owait.memcg = memcg;
1681 owait.wait.flags = 0;
1682 owait.wait.func = memcg_oom_wake_function;
1683 owait.wait.private = current;
1684 INIT_LIST_HEAD(&owait.wait.task_list);
1686 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1687 mem_cgroup_mark_under_oom(memcg);
1689 locked = mem_cgroup_oom_trylock(memcg);
1692 mem_cgroup_oom_notify(memcg);
1694 if (locked && !memcg->oom_kill_disable) {
1695 mem_cgroup_unmark_under_oom(memcg);
1696 finish_wait(&memcg_oom_waitq, &owait.wait);
1697 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1698 current->memcg_oom.order);
1701 mem_cgroup_unmark_under_oom(memcg);
1702 finish_wait(&memcg_oom_waitq, &owait.wait);
1706 mem_cgroup_oom_unlock(memcg);
1708 * There is no guarantee that an OOM-lock contender
1709 * sees the wakeups triggered by the OOM kill
1710 * uncharges. Wake any sleepers explicitely.
1712 memcg_oom_recover(memcg);
1715 current->memcg_oom.memcg = NULL;
1716 css_put(&memcg->css);
1721 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1722 * @page: page that is going to change accounted state
1724 * This function must mark the beginning of an accounted page state
1725 * change to prevent double accounting when the page is concurrently
1726 * being moved to another memcg:
1728 * memcg = mem_cgroup_begin_page_stat(page);
1729 * if (TestClearPageState(page))
1730 * mem_cgroup_update_page_stat(memcg, state, -1);
1731 * mem_cgroup_end_page_stat(memcg);
1733 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1735 struct mem_cgroup *memcg;
1736 unsigned long flags;
1739 * The RCU lock is held throughout the transaction. The fast
1740 * path can get away without acquiring the memcg->move_lock
1741 * because page moving starts with an RCU grace period.
1743 * The RCU lock also protects the memcg from being freed when
1744 * the page state that is going to change is the only thing
1745 * preventing the page from being uncharged.
1746 * E.g. end-writeback clearing PageWriteback(), which allows
1747 * migration to go ahead and uncharge the page before the
1748 * account transaction might be complete.
1752 if (mem_cgroup_disabled())
1755 memcg = page->mem_cgroup;
1756 if (unlikely(!memcg))
1759 if (atomic_read(&memcg->moving_account) <= 0)
1762 spin_lock_irqsave(&memcg->move_lock, flags);
1763 if (memcg != page->mem_cgroup) {
1764 spin_unlock_irqrestore(&memcg->move_lock, flags);
1769 * When charge migration first begins, we can have locked and
1770 * unlocked page stat updates happening concurrently. Track
1771 * the task who has the lock for mem_cgroup_end_page_stat().
1773 memcg->move_lock_task = current;
1774 memcg->move_lock_flags = flags;
1778 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1781 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1782 * @memcg: the memcg that was accounted against
1784 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1786 if (memcg && memcg->move_lock_task == current) {
1787 unsigned long flags = memcg->move_lock_flags;
1789 memcg->move_lock_task = NULL;
1790 memcg->move_lock_flags = 0;
1792 spin_unlock_irqrestore(&memcg->move_lock, flags);
1797 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1800 * size of first charge trial. "32" comes from vmscan.c's magic value.
1801 * TODO: maybe necessary to use big numbers in big irons.
1803 #define CHARGE_BATCH 32U
1804 struct memcg_stock_pcp {
1805 struct mem_cgroup *cached; /* this never be root cgroup */
1806 unsigned int nr_pages;
1807 struct work_struct work;
1808 unsigned long flags;
1809 #define FLUSHING_CACHED_CHARGE 0
1811 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1812 static DEFINE_MUTEX(percpu_charge_mutex);
1815 * consume_stock: Try to consume stocked charge on this cpu.
1816 * @memcg: memcg to consume from.
1817 * @nr_pages: how many pages to charge.
1819 * The charges will only happen if @memcg matches the current cpu's memcg
1820 * stock, and at least @nr_pages are available in that stock. Failure to
1821 * service an allocation will refill the stock.
1823 * returns true if successful, false otherwise.
1825 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1827 struct memcg_stock_pcp *stock;
1830 if (nr_pages > CHARGE_BATCH)
1833 stock = &get_cpu_var(memcg_stock);
1834 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1835 stock->nr_pages -= nr_pages;
1838 put_cpu_var(memcg_stock);
1843 * Returns stocks cached in percpu and reset cached information.
1845 static void drain_stock(struct memcg_stock_pcp *stock)
1847 struct mem_cgroup *old = stock->cached;
1849 if (stock->nr_pages) {
1850 page_counter_uncharge(&old->memory, stock->nr_pages);
1851 if (do_swap_account)
1852 page_counter_uncharge(&old->memsw, stock->nr_pages);
1853 css_put_many(&old->css, stock->nr_pages);
1854 stock->nr_pages = 0;
1856 stock->cached = NULL;
1860 * This must be called under preempt disabled or must be called by
1861 * a thread which is pinned to local cpu.
1863 static void drain_local_stock(struct work_struct *dummy)
1865 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1867 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1871 * Cache charges(val) to local per_cpu area.
1872 * This will be consumed by consume_stock() function, later.
1874 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1876 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1878 if (stock->cached != memcg) { /* reset if necessary */
1880 stock->cached = memcg;
1882 stock->nr_pages += nr_pages;
1883 put_cpu_var(memcg_stock);
1887 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1888 * of the hierarchy under it.
1890 static void drain_all_stock(struct mem_cgroup *root_memcg)
1894 /* If someone's already draining, avoid adding running more workers. */
1895 if (!mutex_trylock(&percpu_charge_mutex))
1897 /* Notify other cpus that system-wide "drain" is running */
1900 for_each_online_cpu(cpu) {
1901 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1902 struct mem_cgroup *memcg;
1904 memcg = stock->cached;
1905 if (!memcg || !stock->nr_pages)
1907 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1909 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1911 drain_local_stock(&stock->work);
1913 schedule_work_on(cpu, &stock->work);
1918 mutex_unlock(&percpu_charge_mutex);
1921 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1922 unsigned long action,
1925 int cpu = (unsigned long)hcpu;
1926 struct memcg_stock_pcp *stock;
1928 if (action == CPU_ONLINE)
1931 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1934 stock = &per_cpu(memcg_stock, cpu);
1939 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1940 unsigned int nr_pages)
1942 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1943 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1944 struct mem_cgroup *mem_over_limit;
1945 struct page_counter *counter;
1946 unsigned long nr_reclaimed;
1947 bool may_swap = true;
1948 bool drained = false;
1951 if (mem_cgroup_is_root(memcg))
1954 if (consume_stock(memcg, nr_pages))
1957 if (!do_swap_account ||
1958 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1959 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
1961 if (do_swap_account)
1962 page_counter_uncharge(&memcg->memsw, batch);
1963 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1965 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1969 if (batch > nr_pages) {
1975 * Unlike in global OOM situations, memcg is not in a physical
1976 * memory shortage. Allow dying and OOM-killed tasks to
1977 * bypass the last charges so that they can exit quickly and
1978 * free their memory.
1980 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1981 fatal_signal_pending(current) ||
1982 current->flags & PF_EXITING))
1985 if (unlikely(task_in_memcg_oom(current)))
1988 if (!(gfp_mask & __GFP_WAIT))
1991 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1993 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1994 gfp_mask, may_swap);
1996 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2000 drain_all_stock(mem_over_limit);
2005 if (gfp_mask & __GFP_NORETRY)
2008 * Even though the limit is exceeded at this point, reclaim
2009 * may have been able to free some pages. Retry the charge
2010 * before killing the task.
2012 * Only for regular pages, though: huge pages are rather
2013 * unlikely to succeed so close to the limit, and we fall back
2014 * to regular pages anyway in case of failure.
2016 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2019 * At task move, charge accounts can be doubly counted. So, it's
2020 * better to wait until the end of task_move if something is going on.
2022 if (mem_cgroup_wait_acct_move(mem_over_limit))
2028 if (gfp_mask & __GFP_NOFAIL)
2031 if (fatal_signal_pending(current))
2034 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2036 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2038 if (!(gfp_mask & __GFP_NOFAIL))
2044 css_get_many(&memcg->css, batch);
2045 if (batch > nr_pages)
2046 refill_stock(memcg, batch - nr_pages);
2047 if (!(gfp_mask & __GFP_WAIT))
2050 * If the hierarchy is above the normal consumption range,
2051 * make the charging task trim their excess contribution.
2054 if (page_counter_read(&memcg->memory) <= memcg->high)
2056 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2057 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2058 } while ((memcg = parent_mem_cgroup(memcg)));
2063 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2065 if (mem_cgroup_is_root(memcg))
2068 page_counter_uncharge(&memcg->memory, nr_pages);
2069 if (do_swap_account)
2070 page_counter_uncharge(&memcg->memsw, nr_pages);
2072 css_put_many(&memcg->css, nr_pages);
2076 * try_get_mem_cgroup_from_page - look up page's memcg association
2079 * Look up, get a css reference, and return the memcg that owns @page.
2081 * The page must be locked to prevent racing with swap-in and page
2082 * cache charges. If coming from an unlocked page table, the caller
2083 * must ensure the page is on the LRU or this can race with charging.
2085 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2087 struct mem_cgroup *memcg;
2091 VM_BUG_ON_PAGE(!PageLocked(page), page);
2093 memcg = page->mem_cgroup;
2095 if (!css_tryget_online(&memcg->css))
2097 } else if (PageSwapCache(page)) {
2098 ent.val = page_private(page);
2099 id = lookup_swap_cgroup_id(ent);
2101 memcg = mem_cgroup_from_id(id);
2102 if (memcg && !css_tryget_online(&memcg->css))
2109 static void lock_page_lru(struct page *page, int *isolated)
2111 struct zone *zone = page_zone(page);
2113 spin_lock_irq(&zone->lru_lock);
2114 if (PageLRU(page)) {
2115 struct lruvec *lruvec;
2117 lruvec = mem_cgroup_page_lruvec(page, zone);
2119 del_page_from_lru_list(page, lruvec, page_lru(page));
2125 static void unlock_page_lru(struct page *page, int isolated)
2127 struct zone *zone = page_zone(page);
2130 struct lruvec *lruvec;
2132 lruvec = mem_cgroup_page_lruvec(page, zone);
2133 VM_BUG_ON_PAGE(PageLRU(page), page);
2135 add_page_to_lru_list(page, lruvec, page_lru(page));
2137 spin_unlock_irq(&zone->lru_lock);
2140 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2145 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2148 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2149 * may already be on some other mem_cgroup's LRU. Take care of it.
2152 lock_page_lru(page, &isolated);
2155 * Nobody should be changing or seriously looking at
2156 * page->mem_cgroup at this point:
2158 * - the page is uncharged
2160 * - the page is off-LRU
2162 * - an anonymous fault has exclusive page access, except for
2163 * a locked page table
2165 * - a page cache insertion, a swapin fault, or a migration
2166 * have the page locked
2168 page->mem_cgroup = memcg;
2171 unlock_page_lru(page, isolated);
2174 #ifdef CONFIG_MEMCG_KMEM
2175 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2176 unsigned long nr_pages)
2178 struct page_counter *counter;
2181 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2185 ret = try_charge(memcg, gfp, nr_pages);
2186 if (ret == -EINTR) {
2188 * try_charge() chose to bypass to root due to OOM kill or
2189 * fatal signal. Since our only options are to either fail
2190 * the allocation or charge it to this cgroup, do it as a
2191 * temporary condition. But we can't fail. From a kmem/slab
2192 * perspective, the cache has already been selected, by
2193 * mem_cgroup_kmem_get_cache(), so it is too late to change
2196 * This condition will only trigger if the task entered
2197 * memcg_charge_kmem in a sane state, but was OOM-killed
2198 * during try_charge() above. Tasks that were already dying
2199 * when the allocation triggers should have been already
2200 * directed to the root cgroup in memcontrol.h
2202 page_counter_charge(&memcg->memory, nr_pages);
2203 if (do_swap_account)
2204 page_counter_charge(&memcg->memsw, nr_pages);
2205 css_get_many(&memcg->css, nr_pages);
2208 page_counter_uncharge(&memcg->kmem, nr_pages);
2213 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2215 page_counter_uncharge(&memcg->memory, nr_pages);
2216 if (do_swap_account)
2217 page_counter_uncharge(&memcg->memsw, nr_pages);
2219 page_counter_uncharge(&memcg->kmem, nr_pages);
2221 css_put_many(&memcg->css, nr_pages);
2224 static int memcg_alloc_cache_id(void)
2229 id = ida_simple_get(&memcg_cache_ida,
2230 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2234 if (id < memcg_nr_cache_ids)
2238 * There's no space for the new id in memcg_caches arrays,
2239 * so we have to grow them.
2241 down_write(&memcg_cache_ids_sem);
2243 size = 2 * (id + 1);
2244 if (size < MEMCG_CACHES_MIN_SIZE)
2245 size = MEMCG_CACHES_MIN_SIZE;
2246 else if (size > MEMCG_CACHES_MAX_SIZE)
2247 size = MEMCG_CACHES_MAX_SIZE;
2249 err = memcg_update_all_caches(size);
2251 err = memcg_update_all_list_lrus(size);
2253 memcg_nr_cache_ids = size;
2255 up_write(&memcg_cache_ids_sem);
2258 ida_simple_remove(&memcg_cache_ida, id);
2264 static void memcg_free_cache_id(int id)
2266 ida_simple_remove(&memcg_cache_ida, id);
2269 struct memcg_kmem_cache_create_work {
2270 struct mem_cgroup *memcg;
2271 struct kmem_cache *cachep;
2272 struct work_struct work;
2275 static void memcg_kmem_cache_create_func(struct work_struct *w)
2277 struct memcg_kmem_cache_create_work *cw =
2278 container_of(w, struct memcg_kmem_cache_create_work, work);
2279 struct mem_cgroup *memcg = cw->memcg;
2280 struct kmem_cache *cachep = cw->cachep;
2282 memcg_create_kmem_cache(memcg, cachep);
2284 css_put(&memcg->css);
2289 * Enqueue the creation of a per-memcg kmem_cache.
2291 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2292 struct kmem_cache *cachep)
2294 struct memcg_kmem_cache_create_work *cw;
2296 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2300 css_get(&memcg->css);
2303 cw->cachep = cachep;
2304 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2306 schedule_work(&cw->work);
2309 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2310 struct kmem_cache *cachep)
2313 * We need to stop accounting when we kmalloc, because if the
2314 * corresponding kmalloc cache is not yet created, the first allocation
2315 * in __memcg_schedule_kmem_cache_create will recurse.
2317 * However, it is better to enclose the whole function. Depending on
2318 * the debugging options enabled, INIT_WORK(), for instance, can
2319 * trigger an allocation. This too, will make us recurse. Because at
2320 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2321 * the safest choice is to do it like this, wrapping the whole function.
2323 current->memcg_kmem_skip_account = 1;
2324 __memcg_schedule_kmem_cache_create(memcg, cachep);
2325 current->memcg_kmem_skip_account = 0;
2329 * Return the kmem_cache we're supposed to use for a slab allocation.
2330 * We try to use the current memcg's version of the cache.
2332 * If the cache does not exist yet, if we are the first user of it,
2333 * we either create it immediately, if possible, or create it asynchronously
2335 * In the latter case, we will let the current allocation go through with
2336 * the original cache.
2338 * Can't be called in interrupt context or from kernel threads.
2339 * This function needs to be called with rcu_read_lock() held.
2341 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2343 struct mem_cgroup *memcg;
2344 struct kmem_cache *memcg_cachep;
2347 VM_BUG_ON(!is_root_cache(cachep));
2349 if (current->memcg_kmem_skip_account)
2352 memcg = get_mem_cgroup_from_mm(current->mm);
2353 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2357 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2358 if (likely(memcg_cachep))
2359 return memcg_cachep;
2362 * If we are in a safe context (can wait, and not in interrupt
2363 * context), we could be be predictable and return right away.
2364 * This would guarantee that the allocation being performed
2365 * already belongs in the new cache.
2367 * However, there are some clashes that can arrive from locking.
2368 * For instance, because we acquire the slab_mutex while doing
2369 * memcg_create_kmem_cache, this means no further allocation
2370 * could happen with the slab_mutex held. So it's better to
2373 memcg_schedule_kmem_cache_create(memcg, cachep);
2375 css_put(&memcg->css);
2379 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2381 if (!is_root_cache(cachep))
2382 css_put(&cachep->memcg_params.memcg->css);
2386 * We need to verify if the allocation against current->mm->owner's memcg is
2387 * possible for the given order. But the page is not allocated yet, so we'll
2388 * need a further commit step to do the final arrangements.
2390 * It is possible for the task to switch cgroups in this mean time, so at
2391 * commit time, we can't rely on task conversion any longer. We'll then use
2392 * the handle argument to return to the caller which cgroup we should commit
2393 * against. We could also return the memcg directly and avoid the pointer
2394 * passing, but a boolean return value gives better semantics considering
2395 * the compiled-out case as well.
2397 * Returning true means the allocation is possible.
2400 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2402 struct mem_cgroup *memcg;
2407 memcg = get_mem_cgroup_from_mm(current->mm);
2409 if (!memcg_kmem_is_active(memcg)) {
2410 css_put(&memcg->css);
2414 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2418 css_put(&memcg->css);
2422 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2425 VM_BUG_ON(mem_cgroup_is_root(memcg));
2427 /* The page allocation failed. Revert */
2429 memcg_uncharge_kmem(memcg, 1 << order);
2432 page->mem_cgroup = memcg;
2435 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2437 struct mem_cgroup *memcg = page->mem_cgroup;
2442 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2444 memcg_uncharge_kmem(memcg, 1 << order);
2445 page->mem_cgroup = NULL;
2448 struct mem_cgroup *__mem_cgroup_from_kmem(void *ptr)
2450 struct mem_cgroup *memcg = NULL;
2451 struct kmem_cache *cachep;
2454 page = virt_to_head_page(ptr);
2455 if (PageSlab(page)) {
2456 cachep = page->slab_cache;
2457 if (!is_root_cache(cachep))
2458 memcg = cachep->memcg_params.memcg;
2460 /* page allocated by alloc_kmem_pages */
2461 memcg = page->mem_cgroup;
2465 #endif /* CONFIG_MEMCG_KMEM */
2467 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2470 * Because tail pages are not marked as "used", set it. We're under
2471 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2472 * charge/uncharge will be never happen and move_account() is done under
2473 * compound_lock(), so we don't have to take care of races.
2475 void mem_cgroup_split_huge_fixup(struct page *head)
2479 if (mem_cgroup_disabled())
2482 for (i = 1; i < HPAGE_PMD_NR; i++)
2483 head[i].mem_cgroup = head->mem_cgroup;
2485 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2488 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2490 #ifdef CONFIG_MEMCG_SWAP
2491 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2494 int val = (charge) ? 1 : -1;
2495 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2499 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2500 * @entry: swap entry to be moved
2501 * @from: mem_cgroup which the entry is moved from
2502 * @to: mem_cgroup which the entry is moved to
2504 * It succeeds only when the swap_cgroup's record for this entry is the same
2505 * as the mem_cgroup's id of @from.
2507 * Returns 0 on success, -EINVAL on failure.
2509 * The caller must have charged to @to, IOW, called page_counter_charge() about
2510 * both res and memsw, and called css_get().
2512 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2513 struct mem_cgroup *from, struct mem_cgroup *to)
2515 unsigned short old_id, new_id;
2517 old_id = mem_cgroup_id(from);
2518 new_id = mem_cgroup_id(to);
2520 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2521 mem_cgroup_swap_statistics(from, false);
2522 mem_cgroup_swap_statistics(to, true);
2528 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2529 struct mem_cgroup *from, struct mem_cgroup *to)
2535 static DEFINE_MUTEX(memcg_limit_mutex);
2537 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2538 unsigned long limit)
2540 unsigned long curusage;
2541 unsigned long oldusage;
2542 bool enlarge = false;
2547 * For keeping hierarchical_reclaim simple, how long we should retry
2548 * is depends on callers. We set our retry-count to be function
2549 * of # of children which we should visit in this loop.
2551 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2552 mem_cgroup_count_children(memcg);
2554 oldusage = page_counter_read(&memcg->memory);
2557 if (signal_pending(current)) {
2562 mutex_lock(&memcg_limit_mutex);
2563 if (limit > memcg->memsw.limit) {
2564 mutex_unlock(&memcg_limit_mutex);
2568 if (limit > memcg->memory.limit)
2570 ret = page_counter_limit(&memcg->memory, limit);
2571 mutex_unlock(&memcg_limit_mutex);
2576 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2578 curusage = page_counter_read(&memcg->memory);
2579 /* Usage is reduced ? */
2580 if (curusage >= oldusage)
2583 oldusage = curusage;
2584 } while (retry_count);
2586 if (!ret && enlarge)
2587 memcg_oom_recover(memcg);
2592 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2593 unsigned long limit)
2595 unsigned long curusage;
2596 unsigned long oldusage;
2597 bool enlarge = false;
2601 /* see mem_cgroup_resize_res_limit */
2602 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2603 mem_cgroup_count_children(memcg);
2605 oldusage = page_counter_read(&memcg->memsw);
2608 if (signal_pending(current)) {
2613 mutex_lock(&memcg_limit_mutex);
2614 if (limit < memcg->memory.limit) {
2615 mutex_unlock(&memcg_limit_mutex);
2619 if (limit > memcg->memsw.limit)
2621 ret = page_counter_limit(&memcg->memsw, limit);
2622 mutex_unlock(&memcg_limit_mutex);
2627 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2629 curusage = page_counter_read(&memcg->memsw);
2630 /* Usage is reduced ? */
2631 if (curusage >= oldusage)
2634 oldusage = curusage;
2635 } while (retry_count);
2637 if (!ret && enlarge)
2638 memcg_oom_recover(memcg);
2643 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2645 unsigned long *total_scanned)
2647 unsigned long nr_reclaimed = 0;
2648 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2649 unsigned long reclaimed;
2651 struct mem_cgroup_tree_per_zone *mctz;
2652 unsigned long excess;
2653 unsigned long nr_scanned;
2658 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2660 * This loop can run a while, specially if mem_cgroup's continuously
2661 * keep exceeding their soft limit and putting the system under
2668 mz = mem_cgroup_largest_soft_limit_node(mctz);
2673 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2674 gfp_mask, &nr_scanned);
2675 nr_reclaimed += reclaimed;
2676 *total_scanned += nr_scanned;
2677 spin_lock_irq(&mctz->lock);
2678 __mem_cgroup_remove_exceeded(mz, mctz);
2681 * If we failed to reclaim anything from this memory cgroup
2682 * it is time to move on to the next cgroup
2686 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2688 excess = soft_limit_excess(mz->memcg);
2690 * One school of thought says that we should not add
2691 * back the node to the tree if reclaim returns 0.
2692 * But our reclaim could return 0, simply because due
2693 * to priority we are exposing a smaller subset of
2694 * memory to reclaim from. Consider this as a longer
2697 /* If excess == 0, no tree ops */
2698 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2699 spin_unlock_irq(&mctz->lock);
2700 css_put(&mz->memcg->css);
2703 * Could not reclaim anything and there are no more
2704 * mem cgroups to try or we seem to be looping without
2705 * reclaiming anything.
2707 if (!nr_reclaimed &&
2709 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2711 } while (!nr_reclaimed);
2713 css_put(&next_mz->memcg->css);
2714 return nr_reclaimed;
2718 * Test whether @memcg has children, dead or alive. Note that this
2719 * function doesn't care whether @memcg has use_hierarchy enabled and
2720 * returns %true if there are child csses according to the cgroup
2721 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2723 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2728 * The lock does not prevent addition or deletion of children, but
2729 * it prevents a new child from being initialized based on this
2730 * parent in css_online(), so it's enough to decide whether
2731 * hierarchically inherited attributes can still be changed or not.
2733 lockdep_assert_held(&memcg_create_mutex);
2736 ret = css_next_child(NULL, &memcg->css);
2742 * Reclaims as many pages from the given memcg as possible and moves
2743 * the rest to the parent.
2745 * Caller is responsible for holding css reference for memcg.
2747 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2749 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2751 /* we call try-to-free pages for make this cgroup empty */
2752 lru_add_drain_all();
2753 /* try to free all pages in this cgroup */
2754 while (nr_retries && page_counter_read(&memcg->memory)) {
2757 if (signal_pending(current))
2760 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2764 /* maybe some writeback is necessary */
2765 congestion_wait(BLK_RW_ASYNC, HZ/10);
2773 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2774 char *buf, size_t nbytes,
2777 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2779 if (mem_cgroup_is_root(memcg))
2781 return mem_cgroup_force_empty(memcg) ?: nbytes;
2784 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2787 return mem_cgroup_from_css(css)->use_hierarchy;
2790 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2791 struct cftype *cft, u64 val)
2794 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2795 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2797 mutex_lock(&memcg_create_mutex);
2799 if (memcg->use_hierarchy == val)
2803 * If parent's use_hierarchy is set, we can't make any modifications
2804 * in the child subtrees. If it is unset, then the change can
2805 * occur, provided the current cgroup has no children.
2807 * For the root cgroup, parent_mem is NULL, we allow value to be
2808 * set if there are no children.
2810 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2811 (val == 1 || val == 0)) {
2812 if (!memcg_has_children(memcg))
2813 memcg->use_hierarchy = val;
2820 mutex_unlock(&memcg_create_mutex);
2825 static unsigned long tree_stat(struct mem_cgroup *memcg,
2826 enum mem_cgroup_stat_index idx)
2828 struct mem_cgroup *iter;
2831 /* Per-cpu values can be negative, use a signed accumulator */
2832 for_each_mem_cgroup_tree(iter, memcg)
2833 val += mem_cgroup_read_stat(iter, idx);
2835 if (val < 0) /* race ? */
2840 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2844 if (mem_cgroup_is_root(memcg)) {
2845 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2846 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2848 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2851 val = page_counter_read(&memcg->memory);
2853 val = page_counter_read(&memcg->memsw);
2855 return val << PAGE_SHIFT;
2866 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2869 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2870 struct page_counter *counter;
2872 switch (MEMFILE_TYPE(cft->private)) {
2874 counter = &memcg->memory;
2877 counter = &memcg->memsw;
2880 counter = &memcg->kmem;
2886 switch (MEMFILE_ATTR(cft->private)) {
2888 if (counter == &memcg->memory)
2889 return mem_cgroup_usage(memcg, false);
2890 if (counter == &memcg->memsw)
2891 return mem_cgroup_usage(memcg, true);
2892 return (u64)page_counter_read(counter) * PAGE_SIZE;
2894 return (u64)counter->limit * PAGE_SIZE;
2896 return (u64)counter->watermark * PAGE_SIZE;
2898 return counter->failcnt;
2899 case RES_SOFT_LIMIT:
2900 return (u64)memcg->soft_limit * PAGE_SIZE;
2906 #ifdef CONFIG_MEMCG_KMEM
2907 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2908 unsigned long nr_pages)
2913 BUG_ON(memcg->kmemcg_id >= 0);
2914 BUG_ON(memcg->kmem_acct_activated);
2915 BUG_ON(memcg->kmem_acct_active);
2918 * For simplicity, we won't allow this to be disabled. It also can't
2919 * be changed if the cgroup has children already, or if tasks had
2922 * If tasks join before we set the limit, a person looking at
2923 * kmem.usage_in_bytes will have no way to determine when it took
2924 * place, which makes the value quite meaningless.
2926 * After it first became limited, changes in the value of the limit are
2927 * of course permitted.
2929 mutex_lock(&memcg_create_mutex);
2930 if (cgroup_has_tasks(memcg->css.cgroup) ||
2931 (memcg->use_hierarchy && memcg_has_children(memcg)))
2933 mutex_unlock(&memcg_create_mutex);
2937 memcg_id = memcg_alloc_cache_id();
2944 * We couldn't have accounted to this cgroup, because it hasn't got
2945 * activated yet, so this should succeed.
2947 err = page_counter_limit(&memcg->kmem, nr_pages);
2950 static_key_slow_inc(&memcg_kmem_enabled_key);
2952 * A memory cgroup is considered kmem-active as soon as it gets
2953 * kmemcg_id. Setting the id after enabling static branching will
2954 * guarantee no one starts accounting before all call sites are
2957 memcg->kmemcg_id = memcg_id;
2958 memcg->kmem_acct_activated = true;
2959 memcg->kmem_acct_active = true;
2964 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2965 unsigned long limit)
2969 mutex_lock(&memcg_limit_mutex);
2970 if (!memcg_kmem_is_active(memcg))
2971 ret = memcg_activate_kmem(memcg, limit);
2973 ret = page_counter_limit(&memcg->kmem, limit);
2974 mutex_unlock(&memcg_limit_mutex);
2978 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2981 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2986 mutex_lock(&memcg_limit_mutex);
2988 * If the parent cgroup is not kmem-active now, it cannot be activated
2989 * after this point, because it has at least one child already.
2991 if (memcg_kmem_is_active(parent))
2992 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
2993 mutex_unlock(&memcg_limit_mutex);
2997 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2998 unsigned long limit)
3002 #endif /* CONFIG_MEMCG_KMEM */
3005 * The user of this function is...
3008 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3009 char *buf, size_t nbytes, loff_t off)
3011 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3012 unsigned long nr_pages;
3015 buf = strstrip(buf);
3016 ret = page_counter_memparse(buf, "-1", &nr_pages);
3020 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3022 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3026 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3028 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3031 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3034 ret = memcg_update_kmem_limit(memcg, nr_pages);
3038 case RES_SOFT_LIMIT:
3039 memcg->soft_limit = nr_pages;
3043 return ret ?: nbytes;
3046 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3047 size_t nbytes, loff_t off)
3049 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3050 struct page_counter *counter;
3052 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3054 counter = &memcg->memory;
3057 counter = &memcg->memsw;
3060 counter = &memcg->kmem;
3066 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3068 page_counter_reset_watermark(counter);
3071 counter->failcnt = 0;
3080 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3083 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3087 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3088 struct cftype *cft, u64 val)
3090 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3092 if (val & ~MOVE_MASK)
3096 * No kind of locking is needed in here, because ->can_attach() will
3097 * check this value once in the beginning of the process, and then carry
3098 * on with stale data. This means that changes to this value will only
3099 * affect task migrations starting after the change.
3101 memcg->move_charge_at_immigrate = val;
3105 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3106 struct cftype *cft, u64 val)
3113 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3117 unsigned int lru_mask;
3120 static const struct numa_stat stats[] = {
3121 { "total", LRU_ALL },
3122 { "file", LRU_ALL_FILE },
3123 { "anon", LRU_ALL_ANON },
3124 { "unevictable", BIT(LRU_UNEVICTABLE) },
3126 const struct numa_stat *stat;
3129 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3131 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3132 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3133 seq_printf(m, "%s=%lu", stat->name, nr);
3134 for_each_node_state(nid, N_MEMORY) {
3135 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3137 seq_printf(m, " N%d=%lu", nid, nr);
3142 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3143 struct mem_cgroup *iter;
3146 for_each_mem_cgroup_tree(iter, memcg)
3147 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3148 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3149 for_each_node_state(nid, N_MEMORY) {
3151 for_each_mem_cgroup_tree(iter, memcg)
3152 nr += mem_cgroup_node_nr_lru_pages(
3153 iter, nid, stat->lru_mask);
3154 seq_printf(m, " N%d=%lu", nid, nr);
3161 #endif /* CONFIG_NUMA */
3163 static int memcg_stat_show(struct seq_file *m, void *v)
3165 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3166 unsigned long memory, memsw;
3167 struct mem_cgroup *mi;
3170 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3171 MEM_CGROUP_STAT_NSTATS);
3172 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3173 MEM_CGROUP_EVENTS_NSTATS);
3174 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3176 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3177 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3179 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3180 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3183 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3184 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3185 mem_cgroup_read_events(memcg, i));
3187 for (i = 0; i < NR_LRU_LISTS; i++)
3188 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3189 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3191 /* Hierarchical information */
3192 memory = memsw = PAGE_COUNTER_MAX;
3193 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3194 memory = min(memory, mi->memory.limit);
3195 memsw = min(memsw, mi->memsw.limit);
3197 seq_printf(m, "hierarchical_memory_limit %llu\n",
3198 (u64)memory * PAGE_SIZE);
3199 if (do_swap_account)
3200 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3201 (u64)memsw * PAGE_SIZE);
3203 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3206 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3208 for_each_mem_cgroup_tree(mi, memcg)
3209 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3210 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3213 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3214 unsigned long long val = 0;
3216 for_each_mem_cgroup_tree(mi, memcg)
3217 val += mem_cgroup_read_events(mi, i);
3218 seq_printf(m, "total_%s %llu\n",
3219 mem_cgroup_events_names[i], val);
3222 for (i = 0; i < NR_LRU_LISTS; i++) {
3223 unsigned long long val = 0;
3225 for_each_mem_cgroup_tree(mi, memcg)
3226 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3227 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3230 #ifdef CONFIG_DEBUG_VM
3233 struct mem_cgroup_per_zone *mz;
3234 struct zone_reclaim_stat *rstat;
3235 unsigned long recent_rotated[2] = {0, 0};
3236 unsigned long recent_scanned[2] = {0, 0};
3238 for_each_online_node(nid)
3239 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3240 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3241 rstat = &mz->lruvec.reclaim_stat;
3243 recent_rotated[0] += rstat->recent_rotated[0];
3244 recent_rotated[1] += rstat->recent_rotated[1];
3245 recent_scanned[0] += rstat->recent_scanned[0];
3246 recent_scanned[1] += rstat->recent_scanned[1];
3248 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3249 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3250 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3251 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3258 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3261 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3263 return mem_cgroup_swappiness(memcg);
3266 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3267 struct cftype *cft, u64 val)
3269 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3275 memcg->swappiness = val;
3277 vm_swappiness = val;
3282 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3284 struct mem_cgroup_threshold_ary *t;
3285 unsigned long usage;
3290 t = rcu_dereference(memcg->thresholds.primary);
3292 t = rcu_dereference(memcg->memsw_thresholds.primary);
3297 usage = mem_cgroup_usage(memcg, swap);
3300 * current_threshold points to threshold just below or equal to usage.
3301 * If it's not true, a threshold was crossed after last
3302 * call of __mem_cgroup_threshold().
3304 i = t->current_threshold;
3307 * Iterate backward over array of thresholds starting from
3308 * current_threshold and check if a threshold is crossed.
3309 * If none of thresholds below usage is crossed, we read
3310 * only one element of the array here.
3312 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3313 eventfd_signal(t->entries[i].eventfd, 1);
3315 /* i = current_threshold + 1 */
3319 * Iterate forward over array of thresholds starting from
3320 * current_threshold+1 and check if a threshold is crossed.
3321 * If none of thresholds above usage is crossed, we read
3322 * only one element of the array here.
3324 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3325 eventfd_signal(t->entries[i].eventfd, 1);
3327 /* Update current_threshold */
3328 t->current_threshold = i - 1;
3333 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3336 __mem_cgroup_threshold(memcg, false);
3337 if (do_swap_account)
3338 __mem_cgroup_threshold(memcg, true);
3340 memcg = parent_mem_cgroup(memcg);
3344 static int compare_thresholds(const void *a, const void *b)
3346 const struct mem_cgroup_threshold *_a = a;
3347 const struct mem_cgroup_threshold *_b = b;
3349 if (_a->threshold > _b->threshold)
3352 if (_a->threshold < _b->threshold)
3358 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3360 struct mem_cgroup_eventfd_list *ev;
3362 spin_lock(&memcg_oom_lock);
3364 list_for_each_entry(ev, &memcg->oom_notify, list)
3365 eventfd_signal(ev->eventfd, 1);
3367 spin_unlock(&memcg_oom_lock);
3371 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3373 struct mem_cgroup *iter;
3375 for_each_mem_cgroup_tree(iter, memcg)
3376 mem_cgroup_oom_notify_cb(iter);
3379 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3380 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3382 struct mem_cgroup_thresholds *thresholds;
3383 struct mem_cgroup_threshold_ary *new;
3384 unsigned long threshold;
3385 unsigned long usage;
3388 ret = page_counter_memparse(args, "-1", &threshold);
3392 mutex_lock(&memcg->thresholds_lock);
3395 thresholds = &memcg->thresholds;
3396 usage = mem_cgroup_usage(memcg, false);
3397 } else if (type == _MEMSWAP) {
3398 thresholds = &memcg->memsw_thresholds;
3399 usage = mem_cgroup_usage(memcg, true);
3403 /* Check if a threshold crossed before adding a new one */
3404 if (thresholds->primary)
3405 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3407 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3409 /* Allocate memory for new array of thresholds */
3410 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3418 /* Copy thresholds (if any) to new array */
3419 if (thresholds->primary) {
3420 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3421 sizeof(struct mem_cgroup_threshold));
3424 /* Add new threshold */
3425 new->entries[size - 1].eventfd = eventfd;
3426 new->entries[size - 1].threshold = threshold;
3428 /* Sort thresholds. Registering of new threshold isn't time-critical */
3429 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3430 compare_thresholds, NULL);
3432 /* Find current threshold */
3433 new->current_threshold = -1;
3434 for (i = 0; i < size; i++) {
3435 if (new->entries[i].threshold <= usage) {
3437 * new->current_threshold will not be used until
3438 * rcu_assign_pointer(), so it's safe to increment
3441 ++new->current_threshold;
3446 /* Free old spare buffer and save old primary buffer as spare */
3447 kfree(thresholds->spare);
3448 thresholds->spare = thresholds->primary;
3450 rcu_assign_pointer(thresholds->primary, new);
3452 /* To be sure that nobody uses thresholds */
3456 mutex_unlock(&memcg->thresholds_lock);
3461 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3462 struct eventfd_ctx *eventfd, const char *args)
3464 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3467 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3468 struct eventfd_ctx *eventfd, const char *args)
3470 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3473 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3474 struct eventfd_ctx *eventfd, enum res_type type)
3476 struct mem_cgroup_thresholds *thresholds;
3477 struct mem_cgroup_threshold_ary *new;
3478 unsigned long usage;
3481 mutex_lock(&memcg->thresholds_lock);
3484 thresholds = &memcg->thresholds;
3485 usage = mem_cgroup_usage(memcg, false);
3486 } else if (type == _MEMSWAP) {
3487 thresholds = &memcg->memsw_thresholds;
3488 usage = mem_cgroup_usage(memcg, true);
3492 if (!thresholds->primary)
3495 /* Check if a threshold crossed before removing */
3496 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3498 /* Calculate new number of threshold */
3500 for (i = 0; i < thresholds->primary->size; i++) {
3501 if (thresholds->primary->entries[i].eventfd != eventfd)
3505 new = thresholds->spare;
3507 /* Set thresholds array to NULL if we don't have thresholds */
3516 /* Copy thresholds and find current threshold */
3517 new->current_threshold = -1;
3518 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3519 if (thresholds->primary->entries[i].eventfd == eventfd)
3522 new->entries[j] = thresholds->primary->entries[i];
3523 if (new->entries[j].threshold <= usage) {
3525 * new->current_threshold will not be used
3526 * until rcu_assign_pointer(), so it's safe to increment
3529 ++new->current_threshold;
3535 /* Swap primary and spare array */
3536 thresholds->spare = thresholds->primary;
3537 /* If all events are unregistered, free the spare array */
3539 kfree(thresholds->spare);
3540 thresholds->spare = NULL;
3543 rcu_assign_pointer(thresholds->primary, new);
3545 /* To be sure that nobody uses thresholds */
3548 mutex_unlock(&memcg->thresholds_lock);
3551 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3552 struct eventfd_ctx *eventfd)
3554 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3557 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3558 struct eventfd_ctx *eventfd)
3560 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3563 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3564 struct eventfd_ctx *eventfd, const char *args)
3566 struct mem_cgroup_eventfd_list *event;
3568 event = kmalloc(sizeof(*event), GFP_KERNEL);
3572 spin_lock(&memcg_oom_lock);
3574 event->eventfd = eventfd;
3575 list_add(&event->list, &memcg->oom_notify);
3577 /* already in OOM ? */
3578 if (memcg->under_oom)
3579 eventfd_signal(eventfd, 1);
3580 spin_unlock(&memcg_oom_lock);
3585 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3586 struct eventfd_ctx *eventfd)
3588 struct mem_cgroup_eventfd_list *ev, *tmp;
3590 spin_lock(&memcg_oom_lock);
3592 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3593 if (ev->eventfd == eventfd) {
3594 list_del(&ev->list);
3599 spin_unlock(&memcg_oom_lock);
3602 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3604 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3606 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3607 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3611 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3612 struct cftype *cft, u64 val)
3614 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3616 /* cannot set to root cgroup and only 0 and 1 are allowed */
3617 if (!css->parent || !((val == 0) || (val == 1)))
3620 memcg->oom_kill_disable = val;
3622 memcg_oom_recover(memcg);
3627 #ifdef CONFIG_MEMCG_KMEM
3628 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3632 ret = memcg_propagate_kmem(memcg);
3636 return mem_cgroup_sockets_init(memcg, ss);
3639 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3641 struct cgroup_subsys_state *css;
3642 struct mem_cgroup *parent, *child;
3645 if (!memcg->kmem_acct_active)
3649 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3650 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3651 * guarantees no cache will be created for this cgroup after we are
3652 * done (see memcg_create_kmem_cache()).
3654 memcg->kmem_acct_active = false;
3656 memcg_deactivate_kmem_caches(memcg);
3658 kmemcg_id = memcg->kmemcg_id;
3659 BUG_ON(kmemcg_id < 0);
3661 parent = parent_mem_cgroup(memcg);
3663 parent = root_mem_cgroup;
3666 * Change kmemcg_id of this cgroup and all its descendants to the
3667 * parent's id, and then move all entries from this cgroup's list_lrus
3668 * to ones of the parent. After we have finished, all list_lrus
3669 * corresponding to this cgroup are guaranteed to remain empty. The
3670 * ordering is imposed by list_lru_node->lock taken by
3671 * memcg_drain_all_list_lrus().
3673 css_for_each_descendant_pre(css, &memcg->css) {
3674 child = mem_cgroup_from_css(css);
3675 BUG_ON(child->kmemcg_id != kmemcg_id);
3676 child->kmemcg_id = parent->kmemcg_id;
3677 if (!memcg->use_hierarchy)
3680 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3682 memcg_free_cache_id(kmemcg_id);
3685 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3687 if (memcg->kmem_acct_activated) {
3688 memcg_destroy_kmem_caches(memcg);
3689 static_key_slow_dec(&memcg_kmem_enabled_key);
3690 WARN_ON(page_counter_read(&memcg->kmem));
3692 mem_cgroup_sockets_destroy(memcg);
3695 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3700 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3704 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3709 #ifdef CONFIG_CGROUP_WRITEBACK
3711 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3713 return &memcg->cgwb_list;
3716 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3718 return wb_domain_init(&memcg->cgwb_domain, gfp);
3721 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3723 wb_domain_exit(&memcg->cgwb_domain);
3726 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3728 wb_domain_size_changed(&memcg->cgwb_domain);
3731 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3733 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3735 if (!memcg->css.parent)
3738 return &memcg->cgwb_domain;
3742 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3743 * @wb: bdi_writeback in question
3744 * @pavail: out parameter for number of available pages
3745 * @pdirty: out parameter for number of dirty pages
3746 * @pwriteback: out parameter for number of pages under writeback
3748 * Determine the numbers of available, dirty, and writeback pages in @wb's
3749 * memcg. Dirty and writeback are self-explanatory. Available is a bit
3752 * A memcg's headroom is "min(max, high) - used". The available memory is
3753 * calculated as the lowest headroom of itself and the ancestors plus the
3754 * number of pages already being used for file pages. Note that this
3755 * doesn't consider the actual amount of available memory in the system.
3756 * The caller should further cap *@pavail accordingly.
3758 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pavail,
3759 unsigned long *pdirty, unsigned long *pwriteback)
3761 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3762 struct mem_cgroup *parent;
3763 unsigned long head_room = PAGE_COUNTER_MAX;
3764 unsigned long file_pages;
3766 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3768 /* this should eventually include NR_UNSTABLE_NFS */
3769 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3771 file_pages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3772 (1 << LRU_ACTIVE_FILE));
3773 while ((parent = parent_mem_cgroup(memcg))) {
3774 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3775 unsigned long used = page_counter_read(&memcg->memory);
3777 head_room = min(head_room, ceiling - min(ceiling, used));
3781 *pavail = file_pages + head_room;
3784 #else /* CONFIG_CGROUP_WRITEBACK */
3786 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3791 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3795 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3799 #endif /* CONFIG_CGROUP_WRITEBACK */
3802 * DO NOT USE IN NEW FILES.
3804 * "cgroup.event_control" implementation.
3806 * This is way over-engineered. It tries to support fully configurable
3807 * events for each user. Such level of flexibility is completely
3808 * unnecessary especially in the light of the planned unified hierarchy.
3810 * Please deprecate this and replace with something simpler if at all
3815 * Unregister event and free resources.
3817 * Gets called from workqueue.
3819 static void memcg_event_remove(struct work_struct *work)
3821 struct mem_cgroup_event *event =
3822 container_of(work, struct mem_cgroup_event, remove);
3823 struct mem_cgroup *memcg = event->memcg;
3825 remove_wait_queue(event->wqh, &event->wait);
3827 event->unregister_event(memcg, event->eventfd);
3829 /* Notify userspace the event is going away. */
3830 eventfd_signal(event->eventfd, 1);
3832 eventfd_ctx_put(event->eventfd);
3834 css_put(&memcg->css);
3838 * Gets called on POLLHUP on eventfd when user closes it.
3840 * Called with wqh->lock held and interrupts disabled.
3842 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3843 int sync, void *key)
3845 struct mem_cgroup_event *event =
3846 container_of(wait, struct mem_cgroup_event, wait);
3847 struct mem_cgroup *memcg = event->memcg;
3848 unsigned long flags = (unsigned long)key;
3850 if (flags & POLLHUP) {
3852 * If the event has been detached at cgroup removal, we
3853 * can simply return knowing the other side will cleanup
3856 * We can't race against event freeing since the other
3857 * side will require wqh->lock via remove_wait_queue(),
3860 spin_lock(&memcg->event_list_lock);
3861 if (!list_empty(&event->list)) {
3862 list_del_init(&event->list);
3864 * We are in atomic context, but cgroup_event_remove()
3865 * may sleep, so we have to call it in workqueue.
3867 schedule_work(&event->remove);
3869 spin_unlock(&memcg->event_list_lock);
3875 static void memcg_event_ptable_queue_proc(struct file *file,
3876 wait_queue_head_t *wqh, poll_table *pt)
3878 struct mem_cgroup_event *event =
3879 container_of(pt, struct mem_cgroup_event, pt);
3882 add_wait_queue(wqh, &event->wait);
3886 * DO NOT USE IN NEW FILES.
3888 * Parse input and register new cgroup event handler.
3890 * Input must be in format '<event_fd> <control_fd> <args>'.
3891 * Interpretation of args is defined by control file implementation.
3893 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3894 char *buf, size_t nbytes, loff_t off)
3896 struct cgroup_subsys_state *css = of_css(of);
3897 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3898 struct mem_cgroup_event *event;
3899 struct cgroup_subsys_state *cfile_css;
3900 unsigned int efd, cfd;
3907 buf = strstrip(buf);
3909 efd = simple_strtoul(buf, &endp, 10);
3914 cfd = simple_strtoul(buf, &endp, 10);
3915 if ((*endp != ' ') && (*endp != '\0'))
3919 event = kzalloc(sizeof(*event), GFP_KERNEL);
3923 event->memcg = memcg;
3924 INIT_LIST_HEAD(&event->list);
3925 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3926 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3927 INIT_WORK(&event->remove, memcg_event_remove);
3935 event->eventfd = eventfd_ctx_fileget(efile.file);
3936 if (IS_ERR(event->eventfd)) {
3937 ret = PTR_ERR(event->eventfd);
3944 goto out_put_eventfd;
3947 /* the process need read permission on control file */
3948 /* AV: shouldn't we check that it's been opened for read instead? */
3949 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3954 * Determine the event callbacks and set them in @event. This used
3955 * to be done via struct cftype but cgroup core no longer knows
3956 * about these events. The following is crude but the whole thing
3957 * is for compatibility anyway.
3959 * DO NOT ADD NEW FILES.
3961 name = cfile.file->f_path.dentry->d_name.name;
3963 if (!strcmp(name, "memory.usage_in_bytes")) {
3964 event->register_event = mem_cgroup_usage_register_event;
3965 event->unregister_event = mem_cgroup_usage_unregister_event;
3966 } else if (!strcmp(name, "memory.oom_control")) {
3967 event->register_event = mem_cgroup_oom_register_event;
3968 event->unregister_event = mem_cgroup_oom_unregister_event;
3969 } else if (!strcmp(name, "memory.pressure_level")) {
3970 event->register_event = vmpressure_register_event;
3971 event->unregister_event = vmpressure_unregister_event;
3972 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3973 event->register_event = memsw_cgroup_usage_register_event;
3974 event->unregister_event = memsw_cgroup_usage_unregister_event;
3981 * Verify @cfile should belong to @css. Also, remaining events are
3982 * automatically removed on cgroup destruction but the removal is
3983 * asynchronous, so take an extra ref on @css.
3985 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3986 &memory_cgrp_subsys);
3988 if (IS_ERR(cfile_css))
3990 if (cfile_css != css) {
3995 ret = event->register_event(memcg, event->eventfd, buf);
3999 efile.file->f_op->poll(efile.file, &event->pt);
4001 spin_lock(&memcg->event_list_lock);
4002 list_add(&event->list, &memcg->event_list);
4003 spin_unlock(&memcg->event_list_lock);
4015 eventfd_ctx_put(event->eventfd);
4024 static struct cftype mem_cgroup_legacy_files[] = {
4026 .name = "usage_in_bytes",
4027 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4028 .read_u64 = mem_cgroup_read_u64,
4031 .name = "max_usage_in_bytes",
4032 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4033 .write = mem_cgroup_reset,
4034 .read_u64 = mem_cgroup_read_u64,
4037 .name = "limit_in_bytes",
4038 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4039 .write = mem_cgroup_write,
4040 .read_u64 = mem_cgroup_read_u64,
4043 .name = "soft_limit_in_bytes",
4044 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4045 .write = mem_cgroup_write,
4046 .read_u64 = mem_cgroup_read_u64,
4050 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4051 .write = mem_cgroup_reset,
4052 .read_u64 = mem_cgroup_read_u64,
4056 .seq_show = memcg_stat_show,
4059 .name = "force_empty",
4060 .write = mem_cgroup_force_empty_write,
4063 .name = "use_hierarchy",
4064 .write_u64 = mem_cgroup_hierarchy_write,
4065 .read_u64 = mem_cgroup_hierarchy_read,
4068 .name = "cgroup.event_control", /* XXX: for compat */
4069 .write = memcg_write_event_control,
4070 .flags = CFTYPE_NO_PREFIX,
4074 .name = "swappiness",
4075 .read_u64 = mem_cgroup_swappiness_read,
4076 .write_u64 = mem_cgroup_swappiness_write,
4079 .name = "move_charge_at_immigrate",
4080 .read_u64 = mem_cgroup_move_charge_read,
4081 .write_u64 = mem_cgroup_move_charge_write,
4084 .name = "oom_control",
4085 .seq_show = mem_cgroup_oom_control_read,
4086 .write_u64 = mem_cgroup_oom_control_write,
4087 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4090 .name = "pressure_level",
4094 .name = "numa_stat",
4095 .seq_show = memcg_numa_stat_show,
4098 #ifdef CONFIG_MEMCG_KMEM
4100 .name = "kmem.limit_in_bytes",
4101 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4102 .write = mem_cgroup_write,
4103 .read_u64 = mem_cgroup_read_u64,
4106 .name = "kmem.usage_in_bytes",
4107 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4108 .read_u64 = mem_cgroup_read_u64,
4111 .name = "kmem.failcnt",
4112 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4113 .write = mem_cgroup_reset,
4114 .read_u64 = mem_cgroup_read_u64,
4117 .name = "kmem.max_usage_in_bytes",
4118 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4119 .write = mem_cgroup_reset,
4120 .read_u64 = mem_cgroup_read_u64,
4122 #ifdef CONFIG_SLABINFO
4124 .name = "kmem.slabinfo",
4125 .seq_start = slab_start,
4126 .seq_next = slab_next,
4127 .seq_stop = slab_stop,
4128 .seq_show = memcg_slab_show,
4132 { }, /* terminate */
4135 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4137 struct mem_cgroup_per_node *pn;
4138 struct mem_cgroup_per_zone *mz;
4139 int zone, tmp = node;
4141 * This routine is called against possible nodes.
4142 * But it's BUG to call kmalloc() against offline node.
4144 * TODO: this routine can waste much memory for nodes which will
4145 * never be onlined. It's better to use memory hotplug callback
4148 if (!node_state(node, N_NORMAL_MEMORY))
4150 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4154 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4155 mz = &pn->zoneinfo[zone];
4156 lruvec_init(&mz->lruvec);
4157 mz->usage_in_excess = 0;
4158 mz->on_tree = false;
4161 memcg->nodeinfo[node] = pn;
4165 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4167 kfree(memcg->nodeinfo[node]);
4170 static struct mem_cgroup *mem_cgroup_alloc(void)
4172 struct mem_cgroup *memcg;
4175 size = sizeof(struct mem_cgroup);
4176 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4178 memcg = kzalloc(size, GFP_KERNEL);
4182 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4186 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4189 spin_lock_init(&memcg->pcp_counter_lock);
4193 free_percpu(memcg->stat);
4200 * At destroying mem_cgroup, references from swap_cgroup can remain.
4201 * (scanning all at force_empty is too costly...)
4203 * Instead of clearing all references at force_empty, we remember
4204 * the number of reference from swap_cgroup and free mem_cgroup when
4205 * it goes down to 0.
4207 * Removal of cgroup itself succeeds regardless of refs from swap.
4210 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4214 mem_cgroup_remove_from_trees(memcg);
4217 free_mem_cgroup_per_zone_info(memcg, node);
4219 free_percpu(memcg->stat);
4220 memcg_wb_domain_exit(memcg);
4225 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4227 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4229 if (!memcg->memory.parent)
4231 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4233 EXPORT_SYMBOL(parent_mem_cgroup);
4235 static struct cgroup_subsys_state * __ref
4236 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4238 struct mem_cgroup *memcg;
4239 long error = -ENOMEM;
4242 memcg = mem_cgroup_alloc();
4244 return ERR_PTR(error);
4247 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4251 if (parent_css == NULL) {
4252 root_mem_cgroup = memcg;
4253 mem_cgroup_root_css = &memcg->css;
4254 page_counter_init(&memcg->memory, NULL);
4255 memcg->high = PAGE_COUNTER_MAX;
4256 memcg->soft_limit = PAGE_COUNTER_MAX;
4257 page_counter_init(&memcg->memsw, NULL);
4258 page_counter_init(&memcg->kmem, NULL);
4261 memcg->last_scanned_node = MAX_NUMNODES;
4262 INIT_LIST_HEAD(&memcg->oom_notify);
4263 memcg->move_charge_at_immigrate = 0;
4264 mutex_init(&memcg->thresholds_lock);
4265 spin_lock_init(&memcg->move_lock);
4266 vmpressure_init(&memcg->vmpressure);
4267 INIT_LIST_HEAD(&memcg->event_list);
4268 spin_lock_init(&memcg->event_list_lock);
4269 #ifdef CONFIG_MEMCG_KMEM
4270 memcg->kmemcg_id = -1;
4272 #ifdef CONFIG_CGROUP_WRITEBACK
4273 INIT_LIST_HEAD(&memcg->cgwb_list);
4278 __mem_cgroup_free(memcg);
4279 return ERR_PTR(error);
4283 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4285 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4286 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4289 if (css->id > MEM_CGROUP_ID_MAX)
4295 mutex_lock(&memcg_create_mutex);
4297 memcg->use_hierarchy = parent->use_hierarchy;
4298 memcg->oom_kill_disable = parent->oom_kill_disable;
4299 memcg->swappiness = mem_cgroup_swappiness(parent);
4301 if (parent->use_hierarchy) {
4302 page_counter_init(&memcg->memory, &parent->memory);
4303 memcg->high = PAGE_COUNTER_MAX;
4304 memcg->soft_limit = PAGE_COUNTER_MAX;
4305 page_counter_init(&memcg->memsw, &parent->memsw);
4306 page_counter_init(&memcg->kmem, &parent->kmem);
4309 * No need to take a reference to the parent because cgroup
4310 * core guarantees its existence.
4313 page_counter_init(&memcg->memory, NULL);
4314 memcg->high = PAGE_COUNTER_MAX;
4315 memcg->soft_limit = PAGE_COUNTER_MAX;
4316 page_counter_init(&memcg->memsw, NULL);
4317 page_counter_init(&memcg->kmem, NULL);
4319 * Deeper hierachy with use_hierarchy == false doesn't make
4320 * much sense so let cgroup subsystem know about this
4321 * unfortunate state in our controller.
4323 if (parent != root_mem_cgroup)
4324 memory_cgrp_subsys.broken_hierarchy = true;
4326 mutex_unlock(&memcg_create_mutex);
4328 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4333 * Make sure the memcg is initialized: mem_cgroup_iter()
4334 * orders reading memcg->initialized against its callers
4335 * reading the memcg members.
4337 smp_store_release(&memcg->initialized, 1);
4342 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4344 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4345 struct mem_cgroup_event *event, *tmp;
4348 * Unregister events and notify userspace.
4349 * Notify userspace about cgroup removing only after rmdir of cgroup
4350 * directory to avoid race between userspace and kernelspace.
4352 spin_lock(&memcg->event_list_lock);
4353 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4354 list_del_init(&event->list);
4355 schedule_work(&event->remove);
4357 spin_unlock(&memcg->event_list_lock);
4359 vmpressure_cleanup(&memcg->vmpressure);
4361 memcg_deactivate_kmem(memcg);
4363 wb_memcg_offline(memcg);
4366 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4368 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4370 memcg_destroy_kmem(memcg);
4371 __mem_cgroup_free(memcg);
4375 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4376 * @css: the target css
4378 * Reset the states of the mem_cgroup associated with @css. This is
4379 * invoked when the userland requests disabling on the default hierarchy
4380 * but the memcg is pinned through dependency. The memcg should stop
4381 * applying policies and should revert to the vanilla state as it may be
4382 * made visible again.
4384 * The current implementation only resets the essential configurations.
4385 * This needs to be expanded to cover all the visible parts.
4387 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4389 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4391 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4392 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4393 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4395 memcg->high = PAGE_COUNTER_MAX;
4396 memcg->soft_limit = PAGE_COUNTER_MAX;
4397 memcg_wb_domain_size_changed(memcg);
4401 /* Handlers for move charge at task migration. */
4402 static int mem_cgroup_do_precharge(unsigned long count)
4406 /* Try a single bulk charge without reclaim first */
4407 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4409 mc.precharge += count;
4412 if (ret == -EINTR) {
4413 cancel_charge(root_mem_cgroup, count);
4417 /* Try charges one by one with reclaim */
4419 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4421 * In case of failure, any residual charges against
4422 * mc.to will be dropped by mem_cgroup_clear_mc()
4423 * later on. However, cancel any charges that are
4424 * bypassed to root right away or they'll be lost.
4427 cancel_charge(root_mem_cgroup, 1);
4437 * get_mctgt_type - get target type of moving charge
4438 * @vma: the vma the pte to be checked belongs
4439 * @addr: the address corresponding to the pte to be checked
4440 * @ptent: the pte to be checked
4441 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4444 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4445 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4446 * move charge. if @target is not NULL, the page is stored in target->page
4447 * with extra refcnt got(Callers should handle it).
4448 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4449 * target for charge migration. if @target is not NULL, the entry is stored
4452 * Called with pte lock held.
4459 enum mc_target_type {
4465 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4466 unsigned long addr, pte_t ptent)
4468 struct page *page = vm_normal_page(vma, addr, ptent);
4470 if (!page || !page_mapped(page))
4472 if (PageAnon(page)) {
4473 if (!(mc.flags & MOVE_ANON))
4476 if (!(mc.flags & MOVE_FILE))
4479 if (!get_page_unless_zero(page))
4486 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4487 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4489 struct page *page = NULL;
4490 swp_entry_t ent = pte_to_swp_entry(ptent);
4492 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4495 * Because lookup_swap_cache() updates some statistics counter,
4496 * we call find_get_page() with swapper_space directly.
4498 page = find_get_page(swap_address_space(ent), ent.val);
4499 if (do_swap_account)
4500 entry->val = ent.val;
4505 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4506 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4512 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4513 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4515 struct page *page = NULL;
4516 struct address_space *mapping;
4519 if (!vma->vm_file) /* anonymous vma */
4521 if (!(mc.flags & MOVE_FILE))
4524 mapping = vma->vm_file->f_mapping;
4525 pgoff = linear_page_index(vma, addr);
4527 /* page is moved even if it's not RSS of this task(page-faulted). */
4529 /* shmem/tmpfs may report page out on swap: account for that too. */
4530 if (shmem_mapping(mapping)) {
4531 page = find_get_entry(mapping, pgoff);
4532 if (radix_tree_exceptional_entry(page)) {
4533 swp_entry_t swp = radix_to_swp_entry(page);
4534 if (do_swap_account)
4536 page = find_get_page(swap_address_space(swp), swp.val);
4539 page = find_get_page(mapping, pgoff);
4541 page = find_get_page(mapping, pgoff);
4547 * mem_cgroup_move_account - move account of the page
4549 * @nr_pages: number of regular pages (>1 for huge pages)
4550 * @from: mem_cgroup which the page is moved from.
4551 * @to: mem_cgroup which the page is moved to. @from != @to.
4553 * The caller must confirm following.
4554 * - page is not on LRU (isolate_page() is useful.)
4555 * - compound_lock is held when nr_pages > 1
4557 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4560 static int mem_cgroup_move_account(struct page *page,
4561 unsigned int nr_pages,
4562 struct mem_cgroup *from,
4563 struct mem_cgroup *to)
4565 unsigned long flags;
4569 VM_BUG_ON(from == to);
4570 VM_BUG_ON_PAGE(PageLRU(page), page);
4572 * The page is isolated from LRU. So, collapse function
4573 * will not handle this page. But page splitting can happen.
4574 * Do this check under compound_page_lock(). The caller should
4578 if (nr_pages > 1 && !PageTransHuge(page))
4582 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4583 * of its source page while we change it: page migration takes
4584 * both pages off the LRU, but page cache replacement doesn't.
4586 if (!trylock_page(page))
4590 if (page->mem_cgroup != from)
4593 anon = PageAnon(page);
4595 spin_lock_irqsave(&from->move_lock, flags);
4597 if (!anon && page_mapped(page)) {
4598 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4600 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4605 * move_lock grabbed above and caller set from->moving_account, so
4606 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4607 * So mapping should be stable for dirty pages.
4609 if (!anon && PageDirty(page)) {
4610 struct address_space *mapping = page_mapping(page);
4612 if (mapping_cap_account_dirty(mapping)) {
4613 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4615 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4620 if (PageWriteback(page)) {
4621 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4623 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4628 * It is safe to change page->mem_cgroup here because the page
4629 * is referenced, charged, and isolated - we can't race with
4630 * uncharging, charging, migration, or LRU putback.
4633 /* caller should have done css_get */
4634 page->mem_cgroup = to;
4635 spin_unlock_irqrestore(&from->move_lock, flags);
4639 local_irq_disable();
4640 mem_cgroup_charge_statistics(to, page, nr_pages);
4641 memcg_check_events(to, page);
4642 mem_cgroup_charge_statistics(from, page, -nr_pages);
4643 memcg_check_events(from, page);
4651 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4652 unsigned long addr, pte_t ptent, union mc_target *target)
4654 struct page *page = NULL;
4655 enum mc_target_type ret = MC_TARGET_NONE;
4656 swp_entry_t ent = { .val = 0 };
4658 if (pte_present(ptent))
4659 page = mc_handle_present_pte(vma, addr, ptent);
4660 else if (is_swap_pte(ptent))
4661 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4662 else if (pte_none(ptent))
4663 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4665 if (!page && !ent.val)
4669 * Do only loose check w/o serialization.
4670 * mem_cgroup_move_account() checks the page is valid or
4671 * not under LRU exclusion.
4673 if (page->mem_cgroup == mc.from) {
4674 ret = MC_TARGET_PAGE;
4676 target->page = page;
4678 if (!ret || !target)
4681 /* There is a swap entry and a page doesn't exist or isn't charged */
4682 if (ent.val && !ret &&
4683 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4684 ret = MC_TARGET_SWAP;
4691 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4693 * We don't consider swapping or file mapped pages because THP does not
4694 * support them for now.
4695 * Caller should make sure that pmd_trans_huge(pmd) is true.
4697 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4698 unsigned long addr, pmd_t pmd, union mc_target *target)
4700 struct page *page = NULL;
4701 enum mc_target_type ret = MC_TARGET_NONE;
4703 page = pmd_page(pmd);
4704 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4705 if (!(mc.flags & MOVE_ANON))
4707 if (page->mem_cgroup == mc.from) {
4708 ret = MC_TARGET_PAGE;
4711 target->page = page;
4717 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4718 unsigned long addr, pmd_t pmd, union mc_target *target)
4720 return MC_TARGET_NONE;
4724 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4725 unsigned long addr, unsigned long end,
4726 struct mm_walk *walk)
4728 struct vm_area_struct *vma = walk->vma;
4732 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4733 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4734 mc.precharge += HPAGE_PMD_NR;
4739 if (pmd_trans_unstable(pmd))
4741 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4742 for (; addr != end; pte++, addr += PAGE_SIZE)
4743 if (get_mctgt_type(vma, addr, *pte, NULL))
4744 mc.precharge++; /* increment precharge temporarily */
4745 pte_unmap_unlock(pte - 1, ptl);
4751 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4753 unsigned long precharge;
4755 struct mm_walk mem_cgroup_count_precharge_walk = {
4756 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4759 down_read(&mm->mmap_sem);
4760 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4761 up_read(&mm->mmap_sem);
4763 precharge = mc.precharge;
4769 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4771 unsigned long precharge = mem_cgroup_count_precharge(mm);
4773 VM_BUG_ON(mc.moving_task);
4774 mc.moving_task = current;
4775 return mem_cgroup_do_precharge(precharge);
4778 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4779 static void __mem_cgroup_clear_mc(void)
4781 struct mem_cgroup *from = mc.from;
4782 struct mem_cgroup *to = mc.to;
4784 /* we must uncharge all the leftover precharges from mc.to */
4786 cancel_charge(mc.to, mc.precharge);
4790 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4791 * we must uncharge here.
4793 if (mc.moved_charge) {
4794 cancel_charge(mc.from, mc.moved_charge);
4795 mc.moved_charge = 0;
4797 /* we must fixup refcnts and charges */
4798 if (mc.moved_swap) {
4799 /* uncharge swap account from the old cgroup */
4800 if (!mem_cgroup_is_root(mc.from))
4801 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4804 * we charged both to->memory and to->memsw, so we
4805 * should uncharge to->memory.
4807 if (!mem_cgroup_is_root(mc.to))
4808 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4810 css_put_many(&mc.from->css, mc.moved_swap);
4812 /* we've already done css_get(mc.to) */
4815 memcg_oom_recover(from);
4816 memcg_oom_recover(to);
4817 wake_up_all(&mc.waitq);
4820 static void mem_cgroup_clear_mc(void)
4823 * we must clear moving_task before waking up waiters at the end of
4826 mc.moving_task = NULL;
4827 __mem_cgroup_clear_mc();
4828 spin_lock(&mc.lock);
4831 spin_unlock(&mc.lock);
4834 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
4835 struct cgroup_taskset *tset)
4837 struct task_struct *p = cgroup_taskset_first(tset);
4839 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4840 unsigned long move_flags;
4843 * We are now commited to this value whatever it is. Changes in this
4844 * tunable will only affect upcoming migrations, not the current one.
4845 * So we need to save it, and keep it going.
4847 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4849 struct mm_struct *mm;
4850 struct mem_cgroup *from = mem_cgroup_from_task(p);
4852 VM_BUG_ON(from == memcg);
4854 mm = get_task_mm(p);
4857 /* We move charges only when we move a owner of the mm */
4858 if (mm->owner == p) {
4861 VM_BUG_ON(mc.precharge);
4862 VM_BUG_ON(mc.moved_charge);
4863 VM_BUG_ON(mc.moved_swap);
4865 spin_lock(&mc.lock);
4868 mc.flags = move_flags;
4869 spin_unlock(&mc.lock);
4870 /* We set mc.moving_task later */
4872 ret = mem_cgroup_precharge_mc(mm);
4874 mem_cgroup_clear_mc();
4881 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
4882 struct cgroup_taskset *tset)
4885 mem_cgroup_clear_mc();
4888 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4889 unsigned long addr, unsigned long end,
4890 struct mm_walk *walk)
4893 struct vm_area_struct *vma = walk->vma;
4896 enum mc_target_type target_type;
4897 union mc_target target;
4901 * We don't take compound_lock() here but no race with splitting thp
4903 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4904 * under splitting, which means there's no concurrent thp split,
4905 * - if another thread runs into split_huge_page() just after we
4906 * entered this if-block, the thread must wait for page table lock
4907 * to be unlocked in __split_huge_page_splitting(), where the main
4908 * part of thp split is not executed yet.
4910 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4911 if (mc.precharge < HPAGE_PMD_NR) {
4915 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4916 if (target_type == MC_TARGET_PAGE) {
4918 if (!isolate_lru_page(page)) {
4919 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
4921 mc.precharge -= HPAGE_PMD_NR;
4922 mc.moved_charge += HPAGE_PMD_NR;
4924 putback_lru_page(page);
4932 if (pmd_trans_unstable(pmd))
4935 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4936 for (; addr != end; addr += PAGE_SIZE) {
4937 pte_t ptent = *(pte++);
4943 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4944 case MC_TARGET_PAGE:
4946 if (isolate_lru_page(page))
4948 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
4950 /* we uncharge from mc.from later. */
4953 putback_lru_page(page);
4954 put: /* get_mctgt_type() gets the page */
4957 case MC_TARGET_SWAP:
4959 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4961 /* we fixup refcnts and charges later. */
4969 pte_unmap_unlock(pte - 1, ptl);
4974 * We have consumed all precharges we got in can_attach().
4975 * We try charge one by one, but don't do any additional
4976 * charges to mc.to if we have failed in charge once in attach()
4979 ret = mem_cgroup_do_precharge(1);
4987 static void mem_cgroup_move_charge(struct mm_struct *mm)
4989 struct mm_walk mem_cgroup_move_charge_walk = {
4990 .pmd_entry = mem_cgroup_move_charge_pte_range,
4994 lru_add_drain_all();
4996 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4997 * move_lock while we're moving its pages to another memcg.
4998 * Then wait for already started RCU-only updates to finish.
5000 atomic_inc(&mc.from->moving_account);
5003 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5005 * Someone who are holding the mmap_sem might be waiting in
5006 * waitq. So we cancel all extra charges, wake up all waiters,
5007 * and retry. Because we cancel precharges, we might not be able
5008 * to move enough charges, but moving charge is a best-effort
5009 * feature anyway, so it wouldn't be a big problem.
5011 __mem_cgroup_clear_mc();
5016 * When we have consumed all precharges and failed in doing
5017 * additional charge, the page walk just aborts.
5019 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5020 up_read(&mm->mmap_sem);
5021 atomic_dec(&mc.from->moving_account);
5024 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5025 struct cgroup_taskset *tset)
5027 struct task_struct *p = cgroup_taskset_first(tset);
5028 struct mm_struct *mm = get_task_mm(p);
5032 mem_cgroup_move_charge(mm);
5036 mem_cgroup_clear_mc();
5038 #else /* !CONFIG_MMU */
5039 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5040 struct cgroup_taskset *tset)
5044 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5045 struct cgroup_taskset *tset)
5048 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5049 struct cgroup_taskset *tset)
5055 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5056 * to verify whether we're attached to the default hierarchy on each mount
5059 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5062 * use_hierarchy is forced on the default hierarchy. cgroup core
5063 * guarantees that @root doesn't have any children, so turning it
5064 * on for the root memcg is enough.
5066 if (cgroup_on_dfl(root_css->cgroup))
5067 root_mem_cgroup->use_hierarchy = true;
5069 root_mem_cgroup->use_hierarchy = false;
5072 static u64 memory_current_read(struct cgroup_subsys_state *css,
5075 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5078 static int memory_low_show(struct seq_file *m, void *v)
5080 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5081 unsigned long low = READ_ONCE(memcg->low);
5083 if (low == PAGE_COUNTER_MAX)
5084 seq_puts(m, "max\n");
5086 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5091 static ssize_t memory_low_write(struct kernfs_open_file *of,
5092 char *buf, size_t nbytes, loff_t off)
5094 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5098 buf = strstrip(buf);
5099 err = page_counter_memparse(buf, "max", &low);
5108 static int memory_high_show(struct seq_file *m, void *v)
5110 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5111 unsigned long high = READ_ONCE(memcg->high);
5113 if (high == PAGE_COUNTER_MAX)
5114 seq_puts(m, "max\n");
5116 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5121 static ssize_t memory_high_write(struct kernfs_open_file *of,
5122 char *buf, size_t nbytes, loff_t off)
5124 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5128 buf = strstrip(buf);
5129 err = page_counter_memparse(buf, "max", &high);
5135 memcg_wb_domain_size_changed(memcg);
5139 static int memory_max_show(struct seq_file *m, void *v)
5141 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5142 unsigned long max = READ_ONCE(memcg->memory.limit);
5144 if (max == PAGE_COUNTER_MAX)
5145 seq_puts(m, "max\n");
5147 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5152 static ssize_t memory_max_write(struct kernfs_open_file *of,
5153 char *buf, size_t nbytes, loff_t off)
5155 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5159 buf = strstrip(buf);
5160 err = page_counter_memparse(buf, "max", &max);
5164 err = mem_cgroup_resize_limit(memcg, max);
5168 memcg_wb_domain_size_changed(memcg);
5172 static int memory_events_show(struct seq_file *m, void *v)
5174 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5176 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5177 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5178 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5179 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5184 static struct cftype memory_files[] = {
5187 .read_u64 = memory_current_read,
5191 .flags = CFTYPE_NOT_ON_ROOT,
5192 .seq_show = memory_low_show,
5193 .write = memory_low_write,
5197 .flags = CFTYPE_NOT_ON_ROOT,
5198 .seq_show = memory_high_show,
5199 .write = memory_high_write,
5203 .flags = CFTYPE_NOT_ON_ROOT,
5204 .seq_show = memory_max_show,
5205 .write = memory_max_write,
5209 .flags = CFTYPE_NOT_ON_ROOT,
5210 .seq_show = memory_events_show,
5215 struct cgroup_subsys memory_cgrp_subsys = {
5216 .css_alloc = mem_cgroup_css_alloc,
5217 .css_online = mem_cgroup_css_online,
5218 .css_offline = mem_cgroup_css_offline,
5219 .css_free = mem_cgroup_css_free,
5220 .css_reset = mem_cgroup_css_reset,
5221 .can_attach = mem_cgroup_can_attach,
5222 .cancel_attach = mem_cgroup_cancel_attach,
5223 .attach = mem_cgroup_move_task,
5224 .bind = mem_cgroup_bind,
5225 .dfl_cftypes = memory_files,
5226 .legacy_cftypes = mem_cgroup_legacy_files,
5231 * mem_cgroup_low - check if memory consumption is below the normal range
5232 * @root: the highest ancestor to consider
5233 * @memcg: the memory cgroup to check
5235 * Returns %true if memory consumption of @memcg, and that of all
5236 * configurable ancestors up to @root, is below the normal range.
5238 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5240 if (mem_cgroup_disabled())
5244 * The toplevel group doesn't have a configurable range, so
5245 * it's never low when looked at directly, and it is not
5246 * considered an ancestor when assessing the hierarchy.
5249 if (memcg == root_mem_cgroup)
5252 if (page_counter_read(&memcg->memory) >= memcg->low)
5255 while (memcg != root) {
5256 memcg = parent_mem_cgroup(memcg);
5258 if (memcg == root_mem_cgroup)
5261 if (page_counter_read(&memcg->memory) >= memcg->low)
5268 * mem_cgroup_try_charge - try charging a page
5269 * @page: page to charge
5270 * @mm: mm context of the victim
5271 * @gfp_mask: reclaim mode
5272 * @memcgp: charged memcg return
5274 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5275 * pages according to @gfp_mask if necessary.
5277 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5278 * Otherwise, an error code is returned.
5280 * After page->mapping has been set up, the caller must finalize the
5281 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5282 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5284 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5285 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5287 struct mem_cgroup *memcg = NULL;
5288 unsigned int nr_pages = 1;
5291 if (mem_cgroup_disabled())
5294 if (PageSwapCache(page)) {
5296 * Every swap fault against a single page tries to charge the
5297 * page, bail as early as possible. shmem_unuse() encounters
5298 * already charged pages, too. The USED bit is protected by
5299 * the page lock, which serializes swap cache removal, which
5300 * in turn serializes uncharging.
5302 if (page->mem_cgroup)
5306 if (PageTransHuge(page)) {
5307 nr_pages <<= compound_order(page);
5308 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5311 if (do_swap_account && PageSwapCache(page))
5312 memcg = try_get_mem_cgroup_from_page(page);
5314 memcg = get_mem_cgroup_from_mm(mm);
5316 ret = try_charge(memcg, gfp_mask, nr_pages);
5318 css_put(&memcg->css);
5320 if (ret == -EINTR) {
5321 memcg = root_mem_cgroup;
5330 * mem_cgroup_commit_charge - commit a page charge
5331 * @page: page to charge
5332 * @memcg: memcg to charge the page to
5333 * @lrucare: page might be on LRU already
5335 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5336 * after page->mapping has been set up. This must happen atomically
5337 * as part of the page instantiation, i.e. under the page table lock
5338 * for anonymous pages, under the page lock for page and swap cache.
5340 * In addition, the page must not be on the LRU during the commit, to
5341 * prevent racing with task migration. If it might be, use @lrucare.
5343 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5345 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5348 unsigned int nr_pages = 1;
5350 VM_BUG_ON_PAGE(!page->mapping, page);
5351 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5353 if (mem_cgroup_disabled())
5356 * Swap faults will attempt to charge the same page multiple
5357 * times. But reuse_swap_page() might have removed the page
5358 * from swapcache already, so we can't check PageSwapCache().
5363 commit_charge(page, memcg, lrucare);
5365 if (PageTransHuge(page)) {
5366 nr_pages <<= compound_order(page);
5367 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5370 local_irq_disable();
5371 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5372 memcg_check_events(memcg, page);
5375 if (do_swap_account && PageSwapCache(page)) {
5376 swp_entry_t entry = { .val = page_private(page) };
5378 * The swap entry might not get freed for a long time,
5379 * let's not wait for it. The page already received a
5380 * memory+swap charge, drop the swap entry duplicate.
5382 mem_cgroup_uncharge_swap(entry);
5387 * mem_cgroup_cancel_charge - cancel a page charge
5388 * @page: page to charge
5389 * @memcg: memcg to charge the page to
5391 * Cancel a charge transaction started by mem_cgroup_try_charge().
5393 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5395 unsigned int nr_pages = 1;
5397 if (mem_cgroup_disabled())
5400 * Swap faults will attempt to charge the same page multiple
5401 * times. But reuse_swap_page() might have removed the page
5402 * from swapcache already, so we can't check PageSwapCache().
5407 if (PageTransHuge(page)) {
5408 nr_pages <<= compound_order(page);
5409 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5412 cancel_charge(memcg, nr_pages);
5415 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5416 unsigned long nr_anon, unsigned long nr_file,
5417 unsigned long nr_huge, struct page *dummy_page)
5419 unsigned long nr_pages = nr_anon + nr_file;
5420 unsigned long flags;
5422 if (!mem_cgroup_is_root(memcg)) {
5423 page_counter_uncharge(&memcg->memory, nr_pages);
5424 if (do_swap_account)
5425 page_counter_uncharge(&memcg->memsw, nr_pages);
5426 memcg_oom_recover(memcg);
5429 local_irq_save(flags);
5430 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5431 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5432 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5433 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5434 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5435 memcg_check_events(memcg, dummy_page);
5436 local_irq_restore(flags);
5438 if (!mem_cgroup_is_root(memcg))
5439 css_put_many(&memcg->css, nr_pages);
5442 static void uncharge_list(struct list_head *page_list)
5444 struct mem_cgroup *memcg = NULL;
5445 unsigned long nr_anon = 0;
5446 unsigned long nr_file = 0;
5447 unsigned long nr_huge = 0;
5448 unsigned long pgpgout = 0;
5449 struct list_head *next;
5452 next = page_list->next;
5454 unsigned int nr_pages = 1;
5456 page = list_entry(next, struct page, lru);
5457 next = page->lru.next;
5459 VM_BUG_ON_PAGE(PageLRU(page), page);
5460 VM_BUG_ON_PAGE(page_count(page), page);
5462 if (!page->mem_cgroup)
5466 * Nobody should be changing or seriously looking at
5467 * page->mem_cgroup at this point, we have fully
5468 * exclusive access to the page.
5471 if (memcg != page->mem_cgroup) {
5473 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5475 pgpgout = nr_anon = nr_file = nr_huge = 0;
5477 memcg = page->mem_cgroup;
5480 if (PageTransHuge(page)) {
5481 nr_pages <<= compound_order(page);
5482 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5483 nr_huge += nr_pages;
5487 nr_anon += nr_pages;
5489 nr_file += nr_pages;
5491 page->mem_cgroup = NULL;
5494 } while (next != page_list);
5497 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5502 * mem_cgroup_uncharge - uncharge a page
5503 * @page: page to uncharge
5505 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5506 * mem_cgroup_commit_charge().
5508 void mem_cgroup_uncharge(struct page *page)
5510 if (mem_cgroup_disabled())
5513 /* Don't touch page->lru of any random page, pre-check: */
5514 if (!page->mem_cgroup)
5517 INIT_LIST_HEAD(&page->lru);
5518 uncharge_list(&page->lru);
5522 * mem_cgroup_uncharge_list - uncharge a list of page
5523 * @page_list: list of pages to uncharge
5525 * Uncharge a list of pages previously charged with
5526 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5528 void mem_cgroup_uncharge_list(struct list_head *page_list)
5530 if (mem_cgroup_disabled())
5533 if (!list_empty(page_list))
5534 uncharge_list(page_list);
5538 * mem_cgroup_migrate - migrate a charge to another page
5539 * @oldpage: currently charged page
5540 * @newpage: page to transfer the charge to
5541 * @lrucare: either or both pages might be on the LRU already
5543 * Migrate the charge from @oldpage to @newpage.
5545 * Both pages must be locked, @newpage->mapping must be set up.
5547 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5550 struct mem_cgroup *memcg;
5553 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5554 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5555 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5556 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5557 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5558 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5561 if (mem_cgroup_disabled())
5564 /* Page cache replacement: new page already charged? */
5565 if (newpage->mem_cgroup)
5569 * Swapcache readahead pages can get migrated before being
5570 * charged, and migration from compaction can happen to an
5571 * uncharged page when the PFN walker finds a page that
5572 * reclaim just put back on the LRU but has not released yet.
5574 memcg = oldpage->mem_cgroup;
5579 lock_page_lru(oldpage, &isolated);
5581 oldpage->mem_cgroup = NULL;
5584 unlock_page_lru(oldpage, isolated);
5586 commit_charge(newpage, memcg, lrucare);
5590 * subsys_initcall() for memory controller.
5592 * Some parts like hotcpu_notifier() have to be initialized from this context
5593 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5594 * everything that doesn't depend on a specific mem_cgroup structure should
5595 * be initialized from here.
5597 static int __init mem_cgroup_init(void)
5601 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5603 for_each_possible_cpu(cpu)
5604 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5607 for_each_node(node) {
5608 struct mem_cgroup_tree_per_node *rtpn;
5611 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5612 node_online(node) ? node : NUMA_NO_NODE);
5614 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5615 struct mem_cgroup_tree_per_zone *rtpz;
5617 rtpz = &rtpn->rb_tree_per_zone[zone];
5618 rtpz->rb_root = RB_ROOT;
5619 spin_lock_init(&rtpz->lock);
5621 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5626 subsys_initcall(mem_cgroup_init);
5628 #ifdef CONFIG_MEMCG_SWAP
5630 * mem_cgroup_swapout - transfer a memsw charge to swap
5631 * @page: page whose memsw charge to transfer
5632 * @entry: swap entry to move the charge to
5634 * Transfer the memsw charge of @page to @entry.
5636 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5638 struct mem_cgroup *memcg;
5639 unsigned short oldid;
5641 VM_BUG_ON_PAGE(PageLRU(page), page);
5642 VM_BUG_ON_PAGE(page_count(page), page);
5644 if (!do_swap_account)
5647 memcg = page->mem_cgroup;
5649 /* Readahead page, never charged */
5653 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5654 VM_BUG_ON_PAGE(oldid, page);
5655 mem_cgroup_swap_statistics(memcg, true);
5657 page->mem_cgroup = NULL;
5659 if (!mem_cgroup_is_root(memcg))
5660 page_counter_uncharge(&memcg->memory, 1);
5663 * Interrupts should be disabled here because the caller holds the
5664 * mapping->tree_lock lock which is taken with interrupts-off. It is
5665 * important here to have the interrupts disabled because it is the
5666 * only synchronisation we have for udpating the per-CPU variables.
5668 VM_BUG_ON(!irqs_disabled());
5669 mem_cgroup_charge_statistics(memcg, page, -1);
5670 memcg_check_events(memcg, page);
5674 * mem_cgroup_uncharge_swap - uncharge a swap entry
5675 * @entry: swap entry to uncharge
5677 * Drop the memsw charge associated with @entry.
5679 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5681 struct mem_cgroup *memcg;
5684 if (!do_swap_account)
5687 id = swap_cgroup_record(entry, 0);
5689 memcg = mem_cgroup_from_id(id);
5691 if (!mem_cgroup_is_root(memcg))
5692 page_counter_uncharge(&memcg->memsw, 1);
5693 mem_cgroup_swap_statistics(memcg, false);
5694 css_put(&memcg->css);
5699 /* for remember boot option*/
5700 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5701 static int really_do_swap_account __initdata = 1;
5703 static int really_do_swap_account __initdata;
5706 static int __init enable_swap_account(char *s)
5708 if (!strcmp(s, "1"))
5709 really_do_swap_account = 1;
5710 else if (!strcmp(s, "0"))
5711 really_do_swap_account = 0;
5714 __setup("swapaccount=", enable_swap_account);
5716 static struct cftype memsw_cgroup_files[] = {
5718 .name = "memsw.usage_in_bytes",
5719 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5720 .read_u64 = mem_cgroup_read_u64,
5723 .name = "memsw.max_usage_in_bytes",
5724 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5725 .write = mem_cgroup_reset,
5726 .read_u64 = mem_cgroup_read_u64,
5729 .name = "memsw.limit_in_bytes",
5730 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5731 .write = mem_cgroup_write,
5732 .read_u64 = mem_cgroup_read_u64,
5735 .name = "memsw.failcnt",
5736 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5737 .write = mem_cgroup_reset,
5738 .read_u64 = mem_cgroup_read_u64,
5740 { }, /* terminate */
5743 static int __init mem_cgroup_swap_init(void)
5745 if (!mem_cgroup_disabled() && really_do_swap_account) {
5746 do_swap_account = 1;
5747 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5748 memsw_cgroup_files));
5752 subsys_initcall(mem_cgroup_swap_init);
5754 #endif /* CONFIG_MEMCG_SWAP */