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>
65 #include <linux/tracehook.h>
69 #include <net/tcp_memcontrol.h>
72 #include <asm/uaccess.h>
74 #include <trace/events/vmscan.h>
76 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
77 EXPORT_SYMBOL(memory_cgrp_subsys);
79 struct mem_cgroup *root_mem_cgroup __read_mostly;
81 #define MEM_CGROUP_RECLAIM_RETRIES 5
83 /* Whether the swap controller is active */
84 #ifdef CONFIG_MEMCG_SWAP
85 int do_swap_account __read_mostly;
87 #define do_swap_account 0
90 /* Whether legacy memory+swap accounting is active */
91 static bool do_memsw_account(void)
93 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
96 static const char * const mem_cgroup_stat_names[] = {
106 static const char * const mem_cgroup_events_names[] = {
113 static const char * const mem_cgroup_lru_names[] = {
121 #define THRESHOLDS_EVENTS_TARGET 128
122 #define SOFTLIMIT_EVENTS_TARGET 1024
123 #define NUMAINFO_EVENTS_TARGET 1024
126 * Cgroups above their limits are maintained in a RB-Tree, independent of
127 * their hierarchy representation
130 struct mem_cgroup_tree_per_zone {
131 struct rb_root rb_root;
135 struct mem_cgroup_tree_per_node {
136 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
139 struct mem_cgroup_tree {
140 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
143 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
146 struct mem_cgroup_eventfd_list {
147 struct list_head list;
148 struct eventfd_ctx *eventfd;
152 * cgroup_event represents events which userspace want to receive.
154 struct mem_cgroup_event {
156 * memcg which the event belongs to.
158 struct mem_cgroup *memcg;
160 * eventfd to signal userspace about the event.
162 struct eventfd_ctx *eventfd;
164 * Each of these stored in a list by the cgroup.
166 struct list_head list;
168 * register_event() callback will be used to add new userspace
169 * waiter for changes related to this event. Use eventfd_signal()
170 * on eventfd to send notification to userspace.
172 int (*register_event)(struct mem_cgroup *memcg,
173 struct eventfd_ctx *eventfd, const char *args);
175 * unregister_event() callback will be called when userspace closes
176 * the eventfd or on cgroup removing. This callback must be set,
177 * if you want provide notification functionality.
179 void (*unregister_event)(struct mem_cgroup *memcg,
180 struct eventfd_ctx *eventfd);
182 * All fields below needed to unregister event when
183 * userspace closes eventfd.
186 wait_queue_head_t *wqh;
188 struct work_struct remove;
191 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
192 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
194 /* Stuffs for move charges at task migration. */
196 * Types of charges to be moved.
198 #define MOVE_ANON 0x1U
199 #define MOVE_FILE 0x2U
200 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
202 /* "mc" and its members are protected by cgroup_mutex */
203 static struct move_charge_struct {
204 spinlock_t lock; /* for from, to */
205 struct mem_cgroup *from;
206 struct mem_cgroup *to;
208 unsigned long precharge;
209 unsigned long moved_charge;
210 unsigned long moved_swap;
211 struct task_struct *moving_task; /* a task moving charges */
212 wait_queue_head_t waitq; /* a waitq for other context */
214 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
215 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
219 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
220 * limit reclaim to prevent infinite loops, if they ever occur.
222 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
223 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
226 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
227 MEM_CGROUP_CHARGE_TYPE_ANON,
228 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
229 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
233 /* for encoding cft->private value on file */
241 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
242 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
243 #define MEMFILE_ATTR(val) ((val) & 0xffff)
244 /* Used for OOM nofiier */
245 #define OOM_CONTROL (0)
248 * The memcg_create_mutex will be held whenever a new cgroup is created.
249 * As a consequence, any change that needs to protect against new child cgroups
250 * appearing has to hold it as well.
252 static DEFINE_MUTEX(memcg_create_mutex);
254 /* Some nice accessors for the vmpressure. */
255 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
258 memcg = root_mem_cgroup;
259 return &memcg->vmpressure;
262 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
264 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
267 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
269 return (memcg == root_mem_cgroup);
273 * We restrict the id in the range of [1, 65535], so it can fit into
276 #define MEM_CGROUP_ID_MAX USHRT_MAX
278 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
280 return memcg->css.id;
284 * A helper function to get mem_cgroup from ID. must be called under
285 * rcu_read_lock(). The caller is responsible for calling
286 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
287 * refcnt from swap can be called against removed memcg.)
289 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
291 struct cgroup_subsys_state *css;
293 css = css_from_id(id, &memory_cgrp_subsys);
294 return mem_cgroup_from_css(css);
297 /* Writing them here to avoid exposing memcg's inner layout */
298 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
300 struct static_key memcg_sockets_enabled_key;
301 EXPORT_SYMBOL(memcg_sockets_enabled_key);
303 void sock_update_memcg(struct sock *sk)
305 struct mem_cgroup *memcg;
307 /* Socket cloning can throw us here with sk_cgrp already
308 * filled. It won't however, necessarily happen from
309 * process context. So the test for root memcg given
310 * the current task's memcg won't help us in this case.
312 * Respecting the original socket's memcg is a better
313 * decision in this case.
316 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
317 css_get(&sk->sk_memcg->css);
322 memcg = mem_cgroup_from_task(current);
323 if (memcg != root_mem_cgroup &&
324 memcg->tcp_mem.active &&
325 css_tryget_online(&memcg->css))
326 sk->sk_memcg = memcg;
329 EXPORT_SYMBOL(sock_update_memcg);
331 void sock_release_memcg(struct sock *sk)
333 WARN_ON(!sk->sk_memcg);
334 css_put(&sk->sk_memcg->css);
338 * mem_cgroup_charge_skmem - charge socket memory
339 * @memcg: memcg to charge
340 * @nr_pages: number of pages to charge
342 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
343 * @memcg's configured limit, %false if the charge had to be forced.
345 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
347 struct page_counter *counter;
349 if (page_counter_try_charge(&memcg->tcp_mem.memory_allocated,
350 nr_pages, &counter)) {
351 memcg->tcp_mem.memory_pressure = 0;
354 page_counter_charge(&memcg->tcp_mem.memory_allocated, nr_pages);
355 memcg->tcp_mem.memory_pressure = 1;
360 * mem_cgroup_uncharge_skmem - uncharge socket memory
361 * @memcg - memcg to uncharge
362 * @nr_pages - number of pages to uncharge
364 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
366 page_counter_uncharge(&memcg->tcp_mem.memory_allocated, nr_pages);
371 #ifdef CONFIG_MEMCG_KMEM
373 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
374 * The main reason for not using cgroup id for this:
375 * this works better in sparse environments, where we have a lot of memcgs,
376 * but only a few kmem-limited. Or also, if we have, for instance, 200
377 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
378 * 200 entry array for that.
380 * The current size of the caches array is stored in memcg_nr_cache_ids. It
381 * will double each time we have to increase it.
383 static DEFINE_IDA(memcg_cache_ida);
384 int memcg_nr_cache_ids;
386 /* Protects memcg_nr_cache_ids */
387 static DECLARE_RWSEM(memcg_cache_ids_sem);
389 void memcg_get_cache_ids(void)
391 down_read(&memcg_cache_ids_sem);
394 void memcg_put_cache_ids(void)
396 up_read(&memcg_cache_ids_sem);
400 * MIN_SIZE is different than 1, because we would like to avoid going through
401 * the alloc/free process all the time. In a small machine, 4 kmem-limited
402 * cgroups is a reasonable guess. In the future, it could be a parameter or
403 * tunable, but that is strictly not necessary.
405 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
406 * this constant directly from cgroup, but it is understandable that this is
407 * better kept as an internal representation in cgroup.c. In any case, the
408 * cgrp_id space is not getting any smaller, and we don't have to necessarily
409 * increase ours as well if it increases.
411 #define MEMCG_CACHES_MIN_SIZE 4
412 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
415 * A lot of the calls to the cache allocation functions are expected to be
416 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
417 * conditional to this static branch, we'll have to allow modules that does
418 * kmem_cache_alloc and the such to see this symbol as well
420 struct static_key memcg_kmem_enabled_key;
421 EXPORT_SYMBOL(memcg_kmem_enabled_key);
423 #endif /* CONFIG_MEMCG_KMEM */
425 static struct mem_cgroup_per_zone *
426 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
428 int nid = zone_to_nid(zone);
429 int zid = zone_idx(zone);
431 return &memcg->nodeinfo[nid]->zoneinfo[zid];
435 * mem_cgroup_css_from_page - css of the memcg associated with a page
436 * @page: page of interest
438 * If memcg is bound to the default hierarchy, css of the memcg associated
439 * with @page is returned. The returned css remains associated with @page
440 * until it is released.
442 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
445 * XXX: The above description of behavior on the default hierarchy isn't
446 * strictly true yet as replace_page_cache_page() can modify the
447 * association before @page is released even on the default hierarchy;
448 * however, the current and planned usages don't mix the the two functions
449 * and replace_page_cache_page() will soon be updated to make the invariant
452 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
454 struct mem_cgroup *memcg;
458 memcg = page->mem_cgroup;
460 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
461 memcg = root_mem_cgroup;
468 * page_cgroup_ino - return inode number of the memcg a page is charged to
471 * Look up the closest online ancestor of the memory cgroup @page is charged to
472 * and return its inode number or 0 if @page is not charged to any cgroup. It
473 * is safe to call this function without holding a reference to @page.
475 * Note, this function is inherently racy, because there is nothing to prevent
476 * the cgroup inode from getting torn down and potentially reallocated a moment
477 * after page_cgroup_ino() returns, so it only should be used by callers that
478 * do not care (such as procfs interfaces).
480 ino_t page_cgroup_ino(struct page *page)
482 struct mem_cgroup *memcg;
483 unsigned long ino = 0;
486 memcg = READ_ONCE(page->mem_cgroup);
487 while (memcg && !(memcg->css.flags & CSS_ONLINE))
488 memcg = parent_mem_cgroup(memcg);
490 ino = cgroup_ino(memcg->css.cgroup);
495 static struct mem_cgroup_per_zone *
496 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
498 int nid = page_to_nid(page);
499 int zid = page_zonenum(page);
501 return &memcg->nodeinfo[nid]->zoneinfo[zid];
504 static struct mem_cgroup_tree_per_zone *
505 soft_limit_tree_node_zone(int nid, int zid)
507 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
510 static struct mem_cgroup_tree_per_zone *
511 soft_limit_tree_from_page(struct page *page)
513 int nid = page_to_nid(page);
514 int zid = page_zonenum(page);
516 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
519 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
520 struct mem_cgroup_tree_per_zone *mctz,
521 unsigned long new_usage_in_excess)
523 struct rb_node **p = &mctz->rb_root.rb_node;
524 struct rb_node *parent = NULL;
525 struct mem_cgroup_per_zone *mz_node;
530 mz->usage_in_excess = new_usage_in_excess;
531 if (!mz->usage_in_excess)
535 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
537 if (mz->usage_in_excess < mz_node->usage_in_excess)
540 * We can't avoid mem cgroups that are over their soft
541 * limit by the same amount
543 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
546 rb_link_node(&mz->tree_node, parent, p);
547 rb_insert_color(&mz->tree_node, &mctz->rb_root);
551 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
552 struct mem_cgroup_tree_per_zone *mctz)
556 rb_erase(&mz->tree_node, &mctz->rb_root);
560 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
561 struct mem_cgroup_tree_per_zone *mctz)
565 spin_lock_irqsave(&mctz->lock, flags);
566 __mem_cgroup_remove_exceeded(mz, mctz);
567 spin_unlock_irqrestore(&mctz->lock, flags);
570 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
572 unsigned long nr_pages = page_counter_read(&memcg->memory);
573 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
574 unsigned long excess = 0;
576 if (nr_pages > soft_limit)
577 excess = nr_pages - soft_limit;
582 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
584 unsigned long excess;
585 struct mem_cgroup_per_zone *mz;
586 struct mem_cgroup_tree_per_zone *mctz;
588 mctz = soft_limit_tree_from_page(page);
590 * Necessary to update all ancestors when hierarchy is used.
591 * because their event counter is not touched.
593 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
594 mz = mem_cgroup_page_zoneinfo(memcg, page);
595 excess = soft_limit_excess(memcg);
597 * We have to update the tree if mz is on RB-tree or
598 * mem is over its softlimit.
600 if (excess || mz->on_tree) {
603 spin_lock_irqsave(&mctz->lock, flags);
604 /* if on-tree, remove it */
606 __mem_cgroup_remove_exceeded(mz, mctz);
608 * Insert again. mz->usage_in_excess will be updated.
609 * If excess is 0, no tree ops.
611 __mem_cgroup_insert_exceeded(mz, mctz, excess);
612 spin_unlock_irqrestore(&mctz->lock, flags);
617 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
619 struct mem_cgroup_tree_per_zone *mctz;
620 struct mem_cgroup_per_zone *mz;
624 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
625 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
626 mctz = soft_limit_tree_node_zone(nid, zid);
627 mem_cgroup_remove_exceeded(mz, mctz);
632 static struct mem_cgroup_per_zone *
633 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
635 struct rb_node *rightmost = NULL;
636 struct mem_cgroup_per_zone *mz;
640 rightmost = rb_last(&mctz->rb_root);
642 goto done; /* Nothing to reclaim from */
644 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
646 * Remove the node now but someone else can add it back,
647 * we will to add it back at the end of reclaim to its correct
648 * position in the tree.
650 __mem_cgroup_remove_exceeded(mz, mctz);
651 if (!soft_limit_excess(mz->memcg) ||
652 !css_tryget_online(&mz->memcg->css))
658 static struct mem_cgroup_per_zone *
659 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
661 struct mem_cgroup_per_zone *mz;
663 spin_lock_irq(&mctz->lock);
664 mz = __mem_cgroup_largest_soft_limit_node(mctz);
665 spin_unlock_irq(&mctz->lock);
670 * Return page count for single (non recursive) @memcg.
672 * Implementation Note: reading percpu statistics for memcg.
674 * Both of vmstat[] and percpu_counter has threshold and do periodic
675 * synchronization to implement "quick" read. There are trade-off between
676 * reading cost and precision of value. Then, we may have a chance to implement
677 * a periodic synchronization of counter in memcg's counter.
679 * But this _read() function is used for user interface now. The user accounts
680 * memory usage by memory cgroup and he _always_ requires exact value because
681 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
682 * have to visit all online cpus and make sum. So, for now, unnecessary
683 * synchronization is not implemented. (just implemented for cpu hotplug)
685 * If there are kernel internal actions which can make use of some not-exact
686 * value, and reading all cpu value can be performance bottleneck in some
687 * common workload, threshold and synchronization as vmstat[] should be
691 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
696 /* Per-cpu values can be negative, use a signed accumulator */
697 for_each_possible_cpu(cpu)
698 val += per_cpu(memcg->stat->count[idx], cpu);
700 * Summing races with updates, so val may be negative. Avoid exposing
701 * transient negative values.
708 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
709 enum mem_cgroup_events_index idx)
711 unsigned long val = 0;
714 for_each_possible_cpu(cpu)
715 val += per_cpu(memcg->stat->events[idx], cpu);
719 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
724 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
725 * counted as CACHE even if it's on ANON LRU.
728 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
731 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
734 if (PageTransHuge(page))
735 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
738 /* pagein of a big page is an event. So, ignore page size */
740 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
742 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
743 nr_pages = -nr_pages; /* for event */
746 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
749 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
751 unsigned int lru_mask)
753 unsigned long nr = 0;
756 VM_BUG_ON((unsigned)nid >= nr_node_ids);
758 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
759 struct mem_cgroup_per_zone *mz;
763 if (!(BIT(lru) & lru_mask))
765 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
766 nr += mz->lru_size[lru];
772 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
773 unsigned int lru_mask)
775 unsigned long nr = 0;
778 for_each_node_state(nid, N_MEMORY)
779 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
783 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
784 enum mem_cgroup_events_target target)
786 unsigned long val, next;
788 val = __this_cpu_read(memcg->stat->nr_page_events);
789 next = __this_cpu_read(memcg->stat->targets[target]);
790 /* from time_after() in jiffies.h */
791 if ((long)next - (long)val < 0) {
793 case MEM_CGROUP_TARGET_THRESH:
794 next = val + THRESHOLDS_EVENTS_TARGET;
796 case MEM_CGROUP_TARGET_SOFTLIMIT:
797 next = val + SOFTLIMIT_EVENTS_TARGET;
799 case MEM_CGROUP_TARGET_NUMAINFO:
800 next = val + NUMAINFO_EVENTS_TARGET;
805 __this_cpu_write(memcg->stat->targets[target], next);
812 * Check events in order.
815 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
817 /* threshold event is triggered in finer grain than soft limit */
818 if (unlikely(mem_cgroup_event_ratelimit(memcg,
819 MEM_CGROUP_TARGET_THRESH))) {
821 bool do_numainfo __maybe_unused;
823 do_softlimit = mem_cgroup_event_ratelimit(memcg,
824 MEM_CGROUP_TARGET_SOFTLIMIT);
826 do_numainfo = mem_cgroup_event_ratelimit(memcg,
827 MEM_CGROUP_TARGET_NUMAINFO);
829 mem_cgroup_threshold(memcg);
830 if (unlikely(do_softlimit))
831 mem_cgroup_update_tree(memcg, page);
833 if (unlikely(do_numainfo))
834 atomic_inc(&memcg->numainfo_events);
839 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
842 * mm_update_next_owner() may clear mm->owner to NULL
843 * if it races with swapoff, page migration, etc.
844 * So this can be called with p == NULL.
849 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
851 EXPORT_SYMBOL(mem_cgroup_from_task);
853 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
855 struct mem_cgroup *memcg = NULL;
860 * Page cache insertions can happen withou an
861 * actual mm context, e.g. during disk probing
862 * on boot, loopback IO, acct() writes etc.
865 memcg = root_mem_cgroup;
867 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
868 if (unlikely(!memcg))
869 memcg = root_mem_cgroup;
871 } while (!css_tryget_online(&memcg->css));
877 * mem_cgroup_iter - iterate over memory cgroup hierarchy
878 * @root: hierarchy root
879 * @prev: previously returned memcg, NULL on first invocation
880 * @reclaim: cookie for shared reclaim walks, NULL for full walks
882 * Returns references to children of the hierarchy below @root, or
883 * @root itself, or %NULL after a full round-trip.
885 * Caller must pass the return value in @prev on subsequent
886 * invocations for reference counting, or use mem_cgroup_iter_break()
887 * to cancel a hierarchy walk before the round-trip is complete.
889 * Reclaimers can specify a zone and a priority level in @reclaim to
890 * divide up the memcgs in the hierarchy among all concurrent
891 * reclaimers operating on the same zone and priority.
893 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
894 struct mem_cgroup *prev,
895 struct mem_cgroup_reclaim_cookie *reclaim)
897 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
898 struct cgroup_subsys_state *css = NULL;
899 struct mem_cgroup *memcg = NULL;
900 struct mem_cgroup *pos = NULL;
902 if (mem_cgroup_disabled())
906 root = root_mem_cgroup;
908 if (prev && !reclaim)
911 if (!root->use_hierarchy && root != root_mem_cgroup) {
920 struct mem_cgroup_per_zone *mz;
922 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
923 iter = &mz->iter[reclaim->priority];
925 if (prev && reclaim->generation != iter->generation)
929 pos = READ_ONCE(iter->position);
930 if (!pos || css_tryget(&pos->css))
933 * css reference reached zero, so iter->position will
934 * be cleared by ->css_released. However, we should not
935 * rely on this happening soon, because ->css_released
936 * is called from a work queue, and by busy-waiting we
937 * might block it. So we clear iter->position right
940 (void)cmpxchg(&iter->position, pos, NULL);
948 css = css_next_descendant_pre(css, &root->css);
951 * Reclaimers share the hierarchy walk, and a
952 * new one might jump in right at the end of
953 * the hierarchy - make sure they see at least
954 * one group and restart from the beginning.
962 * Verify the css and acquire a reference. The root
963 * is provided by the caller, so we know it's alive
964 * and kicking, and don't take an extra reference.
966 memcg = mem_cgroup_from_css(css);
968 if (css == &root->css)
971 if (css_tryget(css)) {
973 * Make sure the memcg is initialized:
974 * mem_cgroup_css_online() orders the the
975 * initialization against setting the flag.
977 if (smp_load_acquire(&memcg->initialized))
988 * The position could have already been updated by a competing
989 * thread, so check that the value hasn't changed since we read
990 * it to avoid reclaiming from the same cgroup twice.
992 (void)cmpxchg(&iter->position, pos, memcg);
1000 reclaim->generation = iter->generation;
1006 if (prev && prev != root)
1007 css_put(&prev->css);
1013 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1014 * @root: hierarchy root
1015 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1017 void mem_cgroup_iter_break(struct mem_cgroup *root,
1018 struct mem_cgroup *prev)
1021 root = root_mem_cgroup;
1022 if (prev && prev != root)
1023 css_put(&prev->css);
1026 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1028 struct mem_cgroup *memcg = dead_memcg;
1029 struct mem_cgroup_reclaim_iter *iter;
1030 struct mem_cgroup_per_zone *mz;
1034 while ((memcg = parent_mem_cgroup(memcg))) {
1035 for_each_node(nid) {
1036 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1037 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
1038 for (i = 0; i <= DEF_PRIORITY; i++) {
1039 iter = &mz->iter[i];
1040 cmpxchg(&iter->position,
1049 * Iteration constructs for visiting all cgroups (under a tree). If
1050 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1051 * be used for reference counting.
1053 #define for_each_mem_cgroup_tree(iter, root) \
1054 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1056 iter = mem_cgroup_iter(root, iter, NULL))
1058 #define for_each_mem_cgroup(iter) \
1059 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1061 iter = mem_cgroup_iter(NULL, iter, NULL))
1064 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1065 * @zone: zone of the wanted lruvec
1066 * @memcg: memcg of the wanted lruvec
1068 * Returns the lru list vector holding pages for the given @zone and
1069 * @mem. This can be the global zone lruvec, if the memory controller
1072 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1073 struct mem_cgroup *memcg)
1075 struct mem_cgroup_per_zone *mz;
1076 struct lruvec *lruvec;
1078 if (mem_cgroup_disabled()) {
1079 lruvec = &zone->lruvec;
1083 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1084 lruvec = &mz->lruvec;
1087 * Since a node can be onlined after the mem_cgroup was created,
1088 * we have to be prepared to initialize lruvec->zone here;
1089 * and if offlined then reonlined, we need to reinitialize it.
1091 if (unlikely(lruvec->zone != zone))
1092 lruvec->zone = zone;
1097 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1099 * @zone: zone of the page
1101 * This function is only safe when following the LRU page isolation
1102 * and putback protocol: the LRU lock must be held, and the page must
1103 * either be PageLRU() or the caller must have isolated/allocated it.
1105 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1107 struct mem_cgroup_per_zone *mz;
1108 struct mem_cgroup *memcg;
1109 struct lruvec *lruvec;
1111 if (mem_cgroup_disabled()) {
1112 lruvec = &zone->lruvec;
1116 memcg = page->mem_cgroup;
1118 * Swapcache readahead pages are added to the LRU - and
1119 * possibly migrated - before they are charged.
1122 memcg = root_mem_cgroup;
1124 mz = mem_cgroup_page_zoneinfo(memcg, page);
1125 lruvec = &mz->lruvec;
1128 * Since a node can be onlined after the mem_cgroup was created,
1129 * we have to be prepared to initialize lruvec->zone here;
1130 * and if offlined then reonlined, we need to reinitialize it.
1132 if (unlikely(lruvec->zone != zone))
1133 lruvec->zone = zone;
1138 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1139 * @lruvec: mem_cgroup per zone lru vector
1140 * @lru: index of lru list the page is sitting on
1141 * @nr_pages: positive when adding or negative when removing
1143 * This function must be called when a page is added to or removed from an
1146 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1149 struct mem_cgroup_per_zone *mz;
1150 unsigned long *lru_size;
1152 if (mem_cgroup_disabled())
1155 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1156 lru_size = mz->lru_size + lru;
1157 *lru_size += nr_pages;
1158 VM_BUG_ON((long)(*lru_size) < 0);
1161 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1163 struct mem_cgroup *task_memcg;
1164 struct task_struct *p;
1167 p = find_lock_task_mm(task);
1169 task_memcg = get_mem_cgroup_from_mm(p->mm);
1173 * All threads may have already detached their mm's, but the oom
1174 * killer still needs to detect if they have already been oom
1175 * killed to prevent needlessly killing additional tasks.
1178 task_memcg = mem_cgroup_from_task(task);
1179 css_get(&task_memcg->css);
1182 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1183 css_put(&task_memcg->css);
1187 #define mem_cgroup_from_counter(counter, member) \
1188 container_of(counter, struct mem_cgroup, member)
1191 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1192 * @memcg: the memory cgroup
1194 * Returns the maximum amount of memory @mem can be charged with, in
1197 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1199 unsigned long margin = 0;
1200 unsigned long count;
1201 unsigned long limit;
1203 count = page_counter_read(&memcg->memory);
1204 limit = READ_ONCE(memcg->memory.limit);
1206 margin = limit - count;
1208 if (do_memsw_account()) {
1209 count = page_counter_read(&memcg->memsw);
1210 limit = READ_ONCE(memcg->memsw.limit);
1212 margin = min(margin, limit - count);
1219 * A routine for checking "mem" is under move_account() or not.
1221 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1222 * moving cgroups. This is for waiting at high-memory pressure
1225 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1227 struct mem_cgroup *from;
1228 struct mem_cgroup *to;
1231 * Unlike task_move routines, we access mc.to, mc.from not under
1232 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1234 spin_lock(&mc.lock);
1240 ret = mem_cgroup_is_descendant(from, memcg) ||
1241 mem_cgroup_is_descendant(to, memcg);
1243 spin_unlock(&mc.lock);
1247 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1249 if (mc.moving_task && current != mc.moving_task) {
1250 if (mem_cgroup_under_move(memcg)) {
1252 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1253 /* moving charge context might have finished. */
1256 finish_wait(&mc.waitq, &wait);
1263 #define K(x) ((x) << (PAGE_SHIFT-10))
1265 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1266 * @memcg: The memory cgroup that went over limit
1267 * @p: Task that is going to be killed
1269 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1272 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1274 /* oom_info_lock ensures that parallel ooms do not interleave */
1275 static DEFINE_MUTEX(oom_info_lock);
1276 struct mem_cgroup *iter;
1279 mutex_lock(&oom_info_lock);
1283 pr_info("Task in ");
1284 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1285 pr_cont(" killed as a result of limit of ");
1287 pr_info("Memory limit reached of cgroup ");
1290 pr_cont_cgroup_path(memcg->css.cgroup);
1295 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1296 K((u64)page_counter_read(&memcg->memory)),
1297 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1298 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1299 K((u64)page_counter_read(&memcg->memsw)),
1300 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1301 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1302 K((u64)page_counter_read(&memcg->kmem)),
1303 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1305 for_each_mem_cgroup_tree(iter, memcg) {
1306 pr_info("Memory cgroup stats for ");
1307 pr_cont_cgroup_path(iter->css.cgroup);
1310 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1311 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
1313 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1314 K(mem_cgroup_read_stat(iter, i)));
1317 for (i = 0; i < NR_LRU_LISTS; i++)
1318 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1319 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1323 mutex_unlock(&oom_info_lock);
1327 * This function returns the number of memcg under hierarchy tree. Returns
1328 * 1(self count) if no children.
1330 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1333 struct mem_cgroup *iter;
1335 for_each_mem_cgroup_tree(iter, memcg)
1341 * Return the memory (and swap, if configured) limit for a memcg.
1343 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1345 unsigned long limit;
1347 limit = memcg->memory.limit;
1348 if (mem_cgroup_swappiness(memcg)) {
1349 unsigned long memsw_limit;
1351 memsw_limit = memcg->memsw.limit;
1352 limit = min(limit + total_swap_pages, memsw_limit);
1357 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1360 struct oom_control oc = {
1363 .gfp_mask = gfp_mask,
1366 struct mem_cgroup *iter;
1367 unsigned long chosen_points = 0;
1368 unsigned long totalpages;
1369 unsigned int points = 0;
1370 struct task_struct *chosen = NULL;
1372 mutex_lock(&oom_lock);
1375 * If current has a pending SIGKILL or is exiting, then automatically
1376 * select it. The goal is to allow it to allocate so that it may
1377 * quickly exit and free its memory.
1379 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1380 mark_oom_victim(current);
1384 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1385 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1386 for_each_mem_cgroup_tree(iter, memcg) {
1387 struct css_task_iter it;
1388 struct task_struct *task;
1390 css_task_iter_start(&iter->css, &it);
1391 while ((task = css_task_iter_next(&it))) {
1392 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1393 case OOM_SCAN_SELECT:
1395 put_task_struct(chosen);
1397 chosen_points = ULONG_MAX;
1398 get_task_struct(chosen);
1400 case OOM_SCAN_CONTINUE:
1402 case OOM_SCAN_ABORT:
1403 css_task_iter_end(&it);
1404 mem_cgroup_iter_break(memcg, iter);
1406 put_task_struct(chosen);
1411 points = oom_badness(task, memcg, NULL, totalpages);
1412 if (!points || points < chosen_points)
1414 /* Prefer thread group leaders for display purposes */
1415 if (points == chosen_points &&
1416 thread_group_leader(chosen))
1420 put_task_struct(chosen);
1422 chosen_points = points;
1423 get_task_struct(chosen);
1425 css_task_iter_end(&it);
1429 points = chosen_points * 1000 / totalpages;
1430 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1431 "Memory cgroup out of memory");
1434 mutex_unlock(&oom_lock);
1437 #if MAX_NUMNODES > 1
1440 * test_mem_cgroup_node_reclaimable
1441 * @memcg: the target memcg
1442 * @nid: the node ID to be checked.
1443 * @noswap : specify true here if the user wants flle only information.
1445 * This function returns whether the specified memcg contains any
1446 * reclaimable pages on a node. Returns true if there are any reclaimable
1447 * pages in the node.
1449 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1450 int nid, bool noswap)
1452 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1454 if (noswap || !total_swap_pages)
1456 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1463 * Always updating the nodemask is not very good - even if we have an empty
1464 * list or the wrong list here, we can start from some node and traverse all
1465 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1468 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1472 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1473 * pagein/pageout changes since the last update.
1475 if (!atomic_read(&memcg->numainfo_events))
1477 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1480 /* make a nodemask where this memcg uses memory from */
1481 memcg->scan_nodes = node_states[N_MEMORY];
1483 for_each_node_mask(nid, node_states[N_MEMORY]) {
1485 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1486 node_clear(nid, memcg->scan_nodes);
1489 atomic_set(&memcg->numainfo_events, 0);
1490 atomic_set(&memcg->numainfo_updating, 0);
1494 * Selecting a node where we start reclaim from. Because what we need is just
1495 * reducing usage counter, start from anywhere is O,K. Considering
1496 * memory reclaim from current node, there are pros. and cons.
1498 * Freeing memory from current node means freeing memory from a node which
1499 * we'll use or we've used. So, it may make LRU bad. And if several threads
1500 * hit limits, it will see a contention on a node. But freeing from remote
1501 * node means more costs for memory reclaim because of memory latency.
1503 * Now, we use round-robin. Better algorithm is welcomed.
1505 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1509 mem_cgroup_may_update_nodemask(memcg);
1510 node = memcg->last_scanned_node;
1512 node = next_node(node, memcg->scan_nodes);
1513 if (node == MAX_NUMNODES)
1514 node = first_node(memcg->scan_nodes);
1516 * We call this when we hit limit, not when pages are added to LRU.
1517 * No LRU may hold pages because all pages are UNEVICTABLE or
1518 * memcg is too small and all pages are not on LRU. In that case,
1519 * we use curret node.
1521 if (unlikely(node == MAX_NUMNODES))
1522 node = numa_node_id();
1524 memcg->last_scanned_node = node;
1528 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1534 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1537 unsigned long *total_scanned)
1539 struct mem_cgroup *victim = NULL;
1542 unsigned long excess;
1543 unsigned long nr_scanned;
1544 struct mem_cgroup_reclaim_cookie reclaim = {
1549 excess = soft_limit_excess(root_memcg);
1552 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1557 * If we have not been able to reclaim
1558 * anything, it might because there are
1559 * no reclaimable pages under this hierarchy
1564 * We want to do more targeted reclaim.
1565 * excess >> 2 is not to excessive so as to
1566 * reclaim too much, nor too less that we keep
1567 * coming back to reclaim from this cgroup
1569 if (total >= (excess >> 2) ||
1570 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1575 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1577 *total_scanned += nr_scanned;
1578 if (!soft_limit_excess(root_memcg))
1581 mem_cgroup_iter_break(root_memcg, victim);
1585 #ifdef CONFIG_LOCKDEP
1586 static struct lockdep_map memcg_oom_lock_dep_map = {
1587 .name = "memcg_oom_lock",
1591 static DEFINE_SPINLOCK(memcg_oom_lock);
1594 * Check OOM-Killer is already running under our hierarchy.
1595 * If someone is running, return false.
1597 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1599 struct mem_cgroup *iter, *failed = NULL;
1601 spin_lock(&memcg_oom_lock);
1603 for_each_mem_cgroup_tree(iter, memcg) {
1604 if (iter->oom_lock) {
1606 * this subtree of our hierarchy is already locked
1607 * so we cannot give a lock.
1610 mem_cgroup_iter_break(memcg, iter);
1613 iter->oom_lock = true;
1618 * OK, we failed to lock the whole subtree so we have
1619 * to clean up what we set up to the failing subtree
1621 for_each_mem_cgroup_tree(iter, memcg) {
1622 if (iter == failed) {
1623 mem_cgroup_iter_break(memcg, iter);
1626 iter->oom_lock = false;
1629 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1631 spin_unlock(&memcg_oom_lock);
1636 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1638 struct mem_cgroup *iter;
1640 spin_lock(&memcg_oom_lock);
1641 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1642 for_each_mem_cgroup_tree(iter, memcg)
1643 iter->oom_lock = false;
1644 spin_unlock(&memcg_oom_lock);
1647 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1649 struct mem_cgroup *iter;
1651 spin_lock(&memcg_oom_lock);
1652 for_each_mem_cgroup_tree(iter, memcg)
1654 spin_unlock(&memcg_oom_lock);
1657 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1659 struct mem_cgroup *iter;
1662 * When a new child is created while the hierarchy is under oom,
1663 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1665 spin_lock(&memcg_oom_lock);
1666 for_each_mem_cgroup_tree(iter, memcg)
1667 if (iter->under_oom > 0)
1669 spin_unlock(&memcg_oom_lock);
1672 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1674 struct oom_wait_info {
1675 struct mem_cgroup *memcg;
1679 static int memcg_oom_wake_function(wait_queue_t *wait,
1680 unsigned mode, int sync, void *arg)
1682 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1683 struct mem_cgroup *oom_wait_memcg;
1684 struct oom_wait_info *oom_wait_info;
1686 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1687 oom_wait_memcg = oom_wait_info->memcg;
1689 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1690 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1692 return autoremove_wake_function(wait, mode, sync, arg);
1695 static void memcg_oom_recover(struct mem_cgroup *memcg)
1698 * For the following lockless ->under_oom test, the only required
1699 * guarantee is that it must see the state asserted by an OOM when
1700 * this function is called as a result of userland actions
1701 * triggered by the notification of the OOM. This is trivially
1702 * achieved by invoking mem_cgroup_mark_under_oom() before
1703 * triggering notification.
1705 if (memcg && memcg->under_oom)
1706 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1709 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1711 if (!current->memcg_may_oom)
1714 * We are in the middle of the charge context here, so we
1715 * don't want to block when potentially sitting on a callstack
1716 * that holds all kinds of filesystem and mm locks.
1718 * Also, the caller may handle a failed allocation gracefully
1719 * (like optional page cache readahead) and so an OOM killer
1720 * invocation might not even be necessary.
1722 * That's why we don't do anything here except remember the
1723 * OOM context and then deal with it at the end of the page
1724 * fault when the stack is unwound, the locks are released,
1725 * and when we know whether the fault was overall successful.
1727 css_get(&memcg->css);
1728 current->memcg_in_oom = memcg;
1729 current->memcg_oom_gfp_mask = mask;
1730 current->memcg_oom_order = order;
1734 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1735 * @handle: actually kill/wait or just clean up the OOM state
1737 * This has to be called at the end of a page fault if the memcg OOM
1738 * handler was enabled.
1740 * Memcg supports userspace OOM handling where failed allocations must
1741 * sleep on a waitqueue until the userspace task resolves the
1742 * situation. Sleeping directly in the charge context with all kinds
1743 * of locks held is not a good idea, instead we remember an OOM state
1744 * in the task and mem_cgroup_oom_synchronize() has to be called at
1745 * the end of the page fault to complete the OOM handling.
1747 * Returns %true if an ongoing memcg OOM situation was detected and
1748 * completed, %false otherwise.
1750 bool mem_cgroup_oom_synchronize(bool handle)
1752 struct mem_cgroup *memcg = current->memcg_in_oom;
1753 struct oom_wait_info owait;
1756 /* OOM is global, do not handle */
1760 if (!handle || oom_killer_disabled)
1763 owait.memcg = memcg;
1764 owait.wait.flags = 0;
1765 owait.wait.func = memcg_oom_wake_function;
1766 owait.wait.private = current;
1767 INIT_LIST_HEAD(&owait.wait.task_list);
1769 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1770 mem_cgroup_mark_under_oom(memcg);
1772 locked = mem_cgroup_oom_trylock(memcg);
1775 mem_cgroup_oom_notify(memcg);
1777 if (locked && !memcg->oom_kill_disable) {
1778 mem_cgroup_unmark_under_oom(memcg);
1779 finish_wait(&memcg_oom_waitq, &owait.wait);
1780 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1781 current->memcg_oom_order);
1784 mem_cgroup_unmark_under_oom(memcg);
1785 finish_wait(&memcg_oom_waitq, &owait.wait);
1789 mem_cgroup_oom_unlock(memcg);
1791 * There is no guarantee that an OOM-lock contender
1792 * sees the wakeups triggered by the OOM kill
1793 * uncharges. Wake any sleepers explicitely.
1795 memcg_oom_recover(memcg);
1798 current->memcg_in_oom = NULL;
1799 css_put(&memcg->css);
1804 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1805 * @page: page that is going to change accounted state
1807 * This function must mark the beginning of an accounted page state
1808 * change to prevent double accounting when the page is concurrently
1809 * being moved to another memcg:
1811 * memcg = mem_cgroup_begin_page_stat(page);
1812 * if (TestClearPageState(page))
1813 * mem_cgroup_update_page_stat(memcg, state, -1);
1814 * mem_cgroup_end_page_stat(memcg);
1816 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1818 struct mem_cgroup *memcg;
1819 unsigned long flags;
1822 * The RCU lock is held throughout the transaction. The fast
1823 * path can get away without acquiring the memcg->move_lock
1824 * because page moving starts with an RCU grace period.
1826 * The RCU lock also protects the memcg from being freed when
1827 * the page state that is going to change is the only thing
1828 * preventing the page from being uncharged.
1829 * E.g. end-writeback clearing PageWriteback(), which allows
1830 * migration to go ahead and uncharge the page before the
1831 * account transaction might be complete.
1835 if (mem_cgroup_disabled())
1838 memcg = page->mem_cgroup;
1839 if (unlikely(!memcg))
1842 if (atomic_read(&memcg->moving_account) <= 0)
1845 spin_lock_irqsave(&memcg->move_lock, flags);
1846 if (memcg != page->mem_cgroup) {
1847 spin_unlock_irqrestore(&memcg->move_lock, flags);
1852 * When charge migration first begins, we can have locked and
1853 * unlocked page stat updates happening concurrently. Track
1854 * the task who has the lock for mem_cgroup_end_page_stat().
1856 memcg->move_lock_task = current;
1857 memcg->move_lock_flags = flags;
1861 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1864 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1865 * @memcg: the memcg that was accounted against
1867 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1869 if (memcg && memcg->move_lock_task == current) {
1870 unsigned long flags = memcg->move_lock_flags;
1872 memcg->move_lock_task = NULL;
1873 memcg->move_lock_flags = 0;
1875 spin_unlock_irqrestore(&memcg->move_lock, flags);
1880 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1883 * size of first charge trial. "32" comes from vmscan.c's magic value.
1884 * TODO: maybe necessary to use big numbers in big irons.
1886 #define CHARGE_BATCH 32U
1887 struct memcg_stock_pcp {
1888 struct mem_cgroup *cached; /* this never be root cgroup */
1889 unsigned int nr_pages;
1890 struct work_struct work;
1891 unsigned long flags;
1892 #define FLUSHING_CACHED_CHARGE 0
1894 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1895 static DEFINE_MUTEX(percpu_charge_mutex);
1898 * consume_stock: Try to consume stocked charge on this cpu.
1899 * @memcg: memcg to consume from.
1900 * @nr_pages: how many pages to charge.
1902 * The charges will only happen if @memcg matches the current cpu's memcg
1903 * stock, and at least @nr_pages are available in that stock. Failure to
1904 * service an allocation will refill the stock.
1906 * returns true if successful, false otherwise.
1908 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1910 struct memcg_stock_pcp *stock;
1913 if (nr_pages > CHARGE_BATCH)
1916 stock = &get_cpu_var(memcg_stock);
1917 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1918 stock->nr_pages -= nr_pages;
1921 put_cpu_var(memcg_stock);
1926 * Returns stocks cached in percpu and reset cached information.
1928 static void drain_stock(struct memcg_stock_pcp *stock)
1930 struct mem_cgroup *old = stock->cached;
1932 if (stock->nr_pages) {
1933 page_counter_uncharge(&old->memory, stock->nr_pages);
1934 if (do_memsw_account())
1935 page_counter_uncharge(&old->memsw, stock->nr_pages);
1936 css_put_many(&old->css, stock->nr_pages);
1937 stock->nr_pages = 0;
1939 stock->cached = NULL;
1943 * This must be called under preempt disabled or must be called by
1944 * a thread which is pinned to local cpu.
1946 static void drain_local_stock(struct work_struct *dummy)
1948 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1950 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1954 * Cache charges(val) to local per_cpu area.
1955 * This will be consumed by consume_stock() function, later.
1957 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1959 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1961 if (stock->cached != memcg) { /* reset if necessary */
1963 stock->cached = memcg;
1965 stock->nr_pages += nr_pages;
1966 put_cpu_var(memcg_stock);
1970 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1971 * of the hierarchy under it.
1973 static void drain_all_stock(struct mem_cgroup *root_memcg)
1977 /* If someone's already draining, avoid adding running more workers. */
1978 if (!mutex_trylock(&percpu_charge_mutex))
1980 /* Notify other cpus that system-wide "drain" is running */
1983 for_each_online_cpu(cpu) {
1984 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1985 struct mem_cgroup *memcg;
1987 memcg = stock->cached;
1988 if (!memcg || !stock->nr_pages)
1990 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1992 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1994 drain_local_stock(&stock->work);
1996 schedule_work_on(cpu, &stock->work);
2001 mutex_unlock(&percpu_charge_mutex);
2004 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2005 unsigned long action,
2008 int cpu = (unsigned long)hcpu;
2009 struct memcg_stock_pcp *stock;
2011 if (action == CPU_ONLINE)
2014 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2017 stock = &per_cpu(memcg_stock, cpu);
2023 * Scheduled by try_charge() to be executed from the userland return path
2024 * and reclaims memory over the high limit.
2026 void mem_cgroup_handle_over_high(void)
2028 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2029 struct mem_cgroup *memcg, *pos;
2031 if (likely(!nr_pages))
2034 pos = memcg = get_mem_cgroup_from_mm(current->mm);
2037 if (page_counter_read(&pos->memory) <= pos->high)
2039 mem_cgroup_events(pos, MEMCG_HIGH, 1);
2040 try_to_free_mem_cgroup_pages(pos, nr_pages, GFP_KERNEL, true);
2041 } while ((pos = parent_mem_cgroup(pos)));
2043 css_put(&memcg->css);
2044 current->memcg_nr_pages_over_high = 0;
2047 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2048 unsigned int nr_pages)
2050 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2051 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2052 struct mem_cgroup *mem_over_limit;
2053 struct page_counter *counter;
2054 unsigned long nr_reclaimed;
2055 bool may_swap = true;
2056 bool drained = false;
2058 if (mem_cgroup_is_root(memcg))
2061 if (consume_stock(memcg, nr_pages))
2064 if (!do_memsw_account() ||
2065 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2066 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2068 if (do_memsw_account())
2069 page_counter_uncharge(&memcg->memsw, batch);
2070 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2072 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2076 if (batch > nr_pages) {
2082 * Unlike in global OOM situations, memcg is not in a physical
2083 * memory shortage. Allow dying and OOM-killed tasks to
2084 * bypass the last charges so that they can exit quickly and
2085 * free their memory.
2087 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2088 fatal_signal_pending(current) ||
2089 current->flags & PF_EXITING))
2092 if (unlikely(task_in_memcg_oom(current)))
2095 if (!gfpflags_allow_blocking(gfp_mask))
2098 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2100 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2101 gfp_mask, may_swap);
2103 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2107 drain_all_stock(mem_over_limit);
2112 if (gfp_mask & __GFP_NORETRY)
2115 * Even though the limit is exceeded at this point, reclaim
2116 * may have been able to free some pages. Retry the charge
2117 * before killing the task.
2119 * Only for regular pages, though: huge pages are rather
2120 * unlikely to succeed so close to the limit, and we fall back
2121 * to regular pages anyway in case of failure.
2123 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2126 * At task move, charge accounts can be doubly counted. So, it's
2127 * better to wait until the end of task_move if something is going on.
2129 if (mem_cgroup_wait_acct_move(mem_over_limit))
2135 if (gfp_mask & __GFP_NOFAIL)
2138 if (fatal_signal_pending(current))
2141 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2143 mem_cgroup_oom(mem_over_limit, gfp_mask,
2144 get_order(nr_pages * PAGE_SIZE));
2146 if (!(gfp_mask & __GFP_NOFAIL))
2150 * The allocation either can't fail or will lead to more memory
2151 * being freed very soon. Allow memory usage go over the limit
2152 * temporarily by force charging it.
2154 page_counter_charge(&memcg->memory, nr_pages);
2155 if (do_memsw_account())
2156 page_counter_charge(&memcg->memsw, nr_pages);
2157 css_get_many(&memcg->css, nr_pages);
2162 css_get_many(&memcg->css, batch);
2163 if (batch > nr_pages)
2164 refill_stock(memcg, batch - nr_pages);
2167 * If the hierarchy is above the normal consumption range, schedule
2168 * reclaim on returning to userland. We can perform reclaim here
2169 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2170 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2171 * not recorded as it most likely matches current's and won't
2172 * change in the meantime. As high limit is checked again before
2173 * reclaim, the cost of mismatch is negligible.
2176 if (page_counter_read(&memcg->memory) > memcg->high) {
2177 current->memcg_nr_pages_over_high += batch;
2178 set_notify_resume(current);
2181 } while ((memcg = parent_mem_cgroup(memcg)));
2186 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2188 if (mem_cgroup_is_root(memcg))
2191 page_counter_uncharge(&memcg->memory, nr_pages);
2192 if (do_memsw_account())
2193 page_counter_uncharge(&memcg->memsw, nr_pages);
2195 css_put_many(&memcg->css, nr_pages);
2198 static void lock_page_lru(struct page *page, int *isolated)
2200 struct zone *zone = page_zone(page);
2202 spin_lock_irq(&zone->lru_lock);
2203 if (PageLRU(page)) {
2204 struct lruvec *lruvec;
2206 lruvec = mem_cgroup_page_lruvec(page, zone);
2208 del_page_from_lru_list(page, lruvec, page_lru(page));
2214 static void unlock_page_lru(struct page *page, int isolated)
2216 struct zone *zone = page_zone(page);
2219 struct lruvec *lruvec;
2221 lruvec = mem_cgroup_page_lruvec(page, zone);
2222 VM_BUG_ON_PAGE(PageLRU(page), page);
2224 add_page_to_lru_list(page, lruvec, page_lru(page));
2226 spin_unlock_irq(&zone->lru_lock);
2229 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2234 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2237 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2238 * may already be on some other mem_cgroup's LRU. Take care of it.
2241 lock_page_lru(page, &isolated);
2244 * Nobody should be changing or seriously looking at
2245 * page->mem_cgroup at this point:
2247 * - the page is uncharged
2249 * - the page is off-LRU
2251 * - an anonymous fault has exclusive page access, except for
2252 * a locked page table
2254 * - a page cache insertion, a swapin fault, or a migration
2255 * have the page locked
2257 page->mem_cgroup = memcg;
2260 unlock_page_lru(page, isolated);
2263 #ifdef CONFIG_MEMCG_KMEM
2264 static int memcg_alloc_cache_id(void)
2269 id = ida_simple_get(&memcg_cache_ida,
2270 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2274 if (id < memcg_nr_cache_ids)
2278 * There's no space for the new id in memcg_caches arrays,
2279 * so we have to grow them.
2281 down_write(&memcg_cache_ids_sem);
2283 size = 2 * (id + 1);
2284 if (size < MEMCG_CACHES_MIN_SIZE)
2285 size = MEMCG_CACHES_MIN_SIZE;
2286 else if (size > MEMCG_CACHES_MAX_SIZE)
2287 size = MEMCG_CACHES_MAX_SIZE;
2289 err = memcg_update_all_caches(size);
2291 err = memcg_update_all_list_lrus(size);
2293 memcg_nr_cache_ids = size;
2295 up_write(&memcg_cache_ids_sem);
2298 ida_simple_remove(&memcg_cache_ida, id);
2304 static void memcg_free_cache_id(int id)
2306 ida_simple_remove(&memcg_cache_ida, id);
2309 struct memcg_kmem_cache_create_work {
2310 struct mem_cgroup *memcg;
2311 struct kmem_cache *cachep;
2312 struct work_struct work;
2315 static void memcg_kmem_cache_create_func(struct work_struct *w)
2317 struct memcg_kmem_cache_create_work *cw =
2318 container_of(w, struct memcg_kmem_cache_create_work, work);
2319 struct mem_cgroup *memcg = cw->memcg;
2320 struct kmem_cache *cachep = cw->cachep;
2322 memcg_create_kmem_cache(memcg, cachep);
2324 css_put(&memcg->css);
2329 * Enqueue the creation of a per-memcg kmem_cache.
2331 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2332 struct kmem_cache *cachep)
2334 struct memcg_kmem_cache_create_work *cw;
2336 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2340 css_get(&memcg->css);
2343 cw->cachep = cachep;
2344 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2346 schedule_work(&cw->work);
2349 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2350 struct kmem_cache *cachep)
2353 * We need to stop accounting when we kmalloc, because if the
2354 * corresponding kmalloc cache is not yet created, the first allocation
2355 * in __memcg_schedule_kmem_cache_create will recurse.
2357 * However, it is better to enclose the whole function. Depending on
2358 * the debugging options enabled, INIT_WORK(), for instance, can
2359 * trigger an allocation. This too, will make us recurse. Because at
2360 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2361 * the safest choice is to do it like this, wrapping the whole function.
2363 current->memcg_kmem_skip_account = 1;
2364 __memcg_schedule_kmem_cache_create(memcg, cachep);
2365 current->memcg_kmem_skip_account = 0;
2369 * Return the kmem_cache we're supposed to use for a slab allocation.
2370 * We try to use the current memcg's version of the cache.
2372 * If the cache does not exist yet, if we are the first user of it,
2373 * we either create it immediately, if possible, or create it asynchronously
2375 * In the latter case, we will let the current allocation go through with
2376 * the original cache.
2378 * Can't be called in interrupt context or from kernel threads.
2379 * This function needs to be called with rcu_read_lock() held.
2381 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp)
2383 struct mem_cgroup *memcg;
2384 struct kmem_cache *memcg_cachep;
2387 VM_BUG_ON(!is_root_cache(cachep));
2389 if (cachep->flags & SLAB_ACCOUNT)
2390 gfp |= __GFP_ACCOUNT;
2392 if (!(gfp & __GFP_ACCOUNT))
2395 if (current->memcg_kmem_skip_account)
2398 memcg = get_mem_cgroup_from_mm(current->mm);
2399 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2403 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2404 if (likely(memcg_cachep))
2405 return memcg_cachep;
2408 * If we are in a safe context (can wait, and not in interrupt
2409 * context), we could be be predictable and return right away.
2410 * This would guarantee that the allocation being performed
2411 * already belongs in the new cache.
2413 * However, there are some clashes that can arrive from locking.
2414 * For instance, because we acquire the slab_mutex while doing
2415 * memcg_create_kmem_cache, this means no further allocation
2416 * could happen with the slab_mutex held. So it's better to
2419 memcg_schedule_kmem_cache_create(memcg, cachep);
2421 css_put(&memcg->css);
2425 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2427 if (!is_root_cache(cachep))
2428 css_put(&cachep->memcg_params.memcg->css);
2431 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2432 struct mem_cgroup *memcg)
2434 unsigned int nr_pages = 1 << order;
2435 struct page_counter *counter;
2438 if (!memcg_kmem_is_active(memcg))
2441 if (!page_counter_try_charge(&memcg->kmem, nr_pages, &counter))
2444 ret = try_charge(memcg, gfp, nr_pages);
2446 page_counter_uncharge(&memcg->kmem, nr_pages);
2450 page->mem_cgroup = memcg;
2455 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2457 struct mem_cgroup *memcg;
2460 memcg = get_mem_cgroup_from_mm(current->mm);
2461 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2462 css_put(&memcg->css);
2466 void __memcg_kmem_uncharge(struct page *page, int order)
2468 struct mem_cgroup *memcg = page->mem_cgroup;
2469 unsigned int nr_pages = 1 << order;
2474 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2476 page_counter_uncharge(&memcg->kmem, nr_pages);
2477 page_counter_uncharge(&memcg->memory, nr_pages);
2478 if (do_memsw_account())
2479 page_counter_uncharge(&memcg->memsw, nr_pages);
2481 page->mem_cgroup = NULL;
2482 css_put_many(&memcg->css, nr_pages);
2484 #endif /* CONFIG_MEMCG_KMEM */
2486 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2489 * Because tail pages are not marked as "used", set it. We're under
2490 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2491 * charge/uncharge will be never happen and move_account() is done under
2492 * compound_lock(), so we don't have to take care of races.
2494 void mem_cgroup_split_huge_fixup(struct page *head)
2498 if (mem_cgroup_disabled())
2501 for (i = 1; i < HPAGE_PMD_NR; i++)
2502 head[i].mem_cgroup = head->mem_cgroup;
2504 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2507 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2509 #ifdef CONFIG_MEMCG_SWAP
2510 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2513 int val = (charge) ? 1 : -1;
2514 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2518 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2519 * @entry: swap entry to be moved
2520 * @from: mem_cgroup which the entry is moved from
2521 * @to: mem_cgroup which the entry is moved to
2523 * It succeeds only when the swap_cgroup's record for this entry is the same
2524 * as the mem_cgroup's id of @from.
2526 * Returns 0 on success, -EINVAL on failure.
2528 * The caller must have charged to @to, IOW, called page_counter_charge() about
2529 * both res and memsw, and called css_get().
2531 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2532 struct mem_cgroup *from, struct mem_cgroup *to)
2534 unsigned short old_id, new_id;
2536 old_id = mem_cgroup_id(from);
2537 new_id = mem_cgroup_id(to);
2539 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2540 mem_cgroup_swap_statistics(from, false);
2541 mem_cgroup_swap_statistics(to, true);
2547 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2548 struct mem_cgroup *from, struct mem_cgroup *to)
2554 static DEFINE_MUTEX(memcg_limit_mutex);
2556 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2557 unsigned long limit)
2559 unsigned long curusage;
2560 unsigned long oldusage;
2561 bool enlarge = false;
2566 * For keeping hierarchical_reclaim simple, how long we should retry
2567 * is depends on callers. We set our retry-count to be function
2568 * of # of children which we should visit in this loop.
2570 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2571 mem_cgroup_count_children(memcg);
2573 oldusage = page_counter_read(&memcg->memory);
2576 if (signal_pending(current)) {
2581 mutex_lock(&memcg_limit_mutex);
2582 if (limit > memcg->memsw.limit) {
2583 mutex_unlock(&memcg_limit_mutex);
2587 if (limit > memcg->memory.limit)
2589 ret = page_counter_limit(&memcg->memory, limit);
2590 mutex_unlock(&memcg_limit_mutex);
2595 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2597 curusage = page_counter_read(&memcg->memory);
2598 /* Usage is reduced ? */
2599 if (curusage >= oldusage)
2602 oldusage = curusage;
2603 } while (retry_count);
2605 if (!ret && enlarge)
2606 memcg_oom_recover(memcg);
2611 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2612 unsigned long limit)
2614 unsigned long curusage;
2615 unsigned long oldusage;
2616 bool enlarge = false;
2620 /* see mem_cgroup_resize_res_limit */
2621 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2622 mem_cgroup_count_children(memcg);
2624 oldusage = page_counter_read(&memcg->memsw);
2627 if (signal_pending(current)) {
2632 mutex_lock(&memcg_limit_mutex);
2633 if (limit < memcg->memory.limit) {
2634 mutex_unlock(&memcg_limit_mutex);
2638 if (limit > memcg->memsw.limit)
2640 ret = page_counter_limit(&memcg->memsw, limit);
2641 mutex_unlock(&memcg_limit_mutex);
2646 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2648 curusage = page_counter_read(&memcg->memsw);
2649 /* Usage is reduced ? */
2650 if (curusage >= oldusage)
2653 oldusage = curusage;
2654 } while (retry_count);
2656 if (!ret && enlarge)
2657 memcg_oom_recover(memcg);
2662 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2664 unsigned long *total_scanned)
2666 unsigned long nr_reclaimed = 0;
2667 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2668 unsigned long reclaimed;
2670 struct mem_cgroup_tree_per_zone *mctz;
2671 unsigned long excess;
2672 unsigned long nr_scanned;
2677 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2679 * This loop can run a while, specially if mem_cgroup's continuously
2680 * keep exceeding their soft limit and putting the system under
2687 mz = mem_cgroup_largest_soft_limit_node(mctz);
2692 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2693 gfp_mask, &nr_scanned);
2694 nr_reclaimed += reclaimed;
2695 *total_scanned += nr_scanned;
2696 spin_lock_irq(&mctz->lock);
2697 __mem_cgroup_remove_exceeded(mz, mctz);
2700 * If we failed to reclaim anything from this memory cgroup
2701 * it is time to move on to the next cgroup
2705 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2707 excess = soft_limit_excess(mz->memcg);
2709 * One school of thought says that we should not add
2710 * back the node to the tree if reclaim returns 0.
2711 * But our reclaim could return 0, simply because due
2712 * to priority we are exposing a smaller subset of
2713 * memory to reclaim from. Consider this as a longer
2716 /* If excess == 0, no tree ops */
2717 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2718 spin_unlock_irq(&mctz->lock);
2719 css_put(&mz->memcg->css);
2722 * Could not reclaim anything and there are no more
2723 * mem cgroups to try or we seem to be looping without
2724 * reclaiming anything.
2726 if (!nr_reclaimed &&
2728 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2730 } while (!nr_reclaimed);
2732 css_put(&next_mz->memcg->css);
2733 return nr_reclaimed;
2737 * Test whether @memcg has children, dead or alive. Note that this
2738 * function doesn't care whether @memcg has use_hierarchy enabled and
2739 * returns %true if there are child csses according to the cgroup
2740 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2742 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2747 * The lock does not prevent addition or deletion of children, but
2748 * it prevents a new child from being initialized based on this
2749 * parent in css_online(), so it's enough to decide whether
2750 * hierarchically inherited attributes can still be changed or not.
2752 lockdep_assert_held(&memcg_create_mutex);
2755 ret = css_next_child(NULL, &memcg->css);
2761 * Reclaims as many pages from the given memcg as possible and moves
2762 * the rest to the parent.
2764 * Caller is responsible for holding css reference for memcg.
2766 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2768 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2770 /* we call try-to-free pages for make this cgroup empty */
2771 lru_add_drain_all();
2772 /* try to free all pages in this cgroup */
2773 while (nr_retries && page_counter_read(&memcg->memory)) {
2776 if (signal_pending(current))
2779 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2783 /* maybe some writeback is necessary */
2784 congestion_wait(BLK_RW_ASYNC, HZ/10);
2792 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2793 char *buf, size_t nbytes,
2796 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2798 if (mem_cgroup_is_root(memcg))
2800 return mem_cgroup_force_empty(memcg) ?: nbytes;
2803 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2806 return mem_cgroup_from_css(css)->use_hierarchy;
2809 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2810 struct cftype *cft, u64 val)
2813 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2814 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2816 mutex_lock(&memcg_create_mutex);
2818 if (memcg->use_hierarchy == val)
2822 * If parent's use_hierarchy is set, we can't make any modifications
2823 * in the child subtrees. If it is unset, then the change can
2824 * occur, provided the current cgroup has no children.
2826 * For the root cgroup, parent_mem is NULL, we allow value to be
2827 * set if there are no children.
2829 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2830 (val == 1 || val == 0)) {
2831 if (!memcg_has_children(memcg))
2832 memcg->use_hierarchy = val;
2839 mutex_unlock(&memcg_create_mutex);
2844 static unsigned long tree_stat(struct mem_cgroup *memcg,
2845 enum mem_cgroup_stat_index idx)
2847 struct mem_cgroup *iter;
2848 unsigned long val = 0;
2850 for_each_mem_cgroup_tree(iter, memcg)
2851 val += mem_cgroup_read_stat(iter, idx);
2856 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2860 if (mem_cgroup_is_root(memcg)) {
2861 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2862 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2864 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2867 val = page_counter_read(&memcg->memory);
2869 val = page_counter_read(&memcg->memsw);
2882 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2885 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2886 struct page_counter *counter;
2888 switch (MEMFILE_TYPE(cft->private)) {
2890 counter = &memcg->memory;
2893 counter = &memcg->memsw;
2896 counter = &memcg->kmem;
2902 switch (MEMFILE_ATTR(cft->private)) {
2904 if (counter == &memcg->memory)
2905 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2906 if (counter == &memcg->memsw)
2907 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2908 return (u64)page_counter_read(counter) * PAGE_SIZE;
2910 return (u64)counter->limit * PAGE_SIZE;
2912 return (u64)counter->watermark * PAGE_SIZE;
2914 return counter->failcnt;
2915 case RES_SOFT_LIMIT:
2916 return (u64)memcg->soft_limit * PAGE_SIZE;
2922 #ifdef CONFIG_MEMCG_KMEM
2923 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2924 unsigned long nr_pages)
2929 BUG_ON(memcg->kmemcg_id >= 0);
2930 BUG_ON(memcg->kmem_acct_activated);
2931 BUG_ON(memcg->kmem_acct_active);
2934 * For simplicity, we won't allow this to be disabled. It also can't
2935 * be changed if the cgroup has children already, or if tasks had
2938 * If tasks join before we set the limit, a person looking at
2939 * kmem.usage_in_bytes will have no way to determine when it took
2940 * place, which makes the value quite meaningless.
2942 * After it first became limited, changes in the value of the limit are
2943 * of course permitted.
2945 mutex_lock(&memcg_create_mutex);
2946 if (cgroup_is_populated(memcg->css.cgroup) ||
2947 (memcg->use_hierarchy && memcg_has_children(memcg)))
2949 mutex_unlock(&memcg_create_mutex);
2953 memcg_id = memcg_alloc_cache_id();
2960 * We couldn't have accounted to this cgroup, because it hasn't got
2961 * activated yet, so this should succeed.
2963 err = page_counter_limit(&memcg->kmem, nr_pages);
2966 static_key_slow_inc(&memcg_kmem_enabled_key);
2968 * A memory cgroup is considered kmem-active as soon as it gets
2969 * kmemcg_id. Setting the id after enabling static branching will
2970 * guarantee no one starts accounting before all call sites are
2973 memcg->kmemcg_id = memcg_id;
2974 memcg->kmem_acct_activated = true;
2975 memcg->kmem_acct_active = true;
2980 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2981 unsigned long limit)
2985 mutex_lock(&memcg_limit_mutex);
2986 if (!memcg_kmem_is_active(memcg))
2987 ret = memcg_activate_kmem(memcg, limit);
2989 ret = page_counter_limit(&memcg->kmem, limit);
2990 mutex_unlock(&memcg_limit_mutex);
2994 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2997 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3002 mutex_lock(&memcg_limit_mutex);
3004 * If the parent cgroup is not kmem-active now, it cannot be activated
3005 * after this point, because it has at least one child already.
3007 if (memcg_kmem_is_active(parent))
3008 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3009 mutex_unlock(&memcg_limit_mutex);
3013 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3014 unsigned long limit)
3018 #endif /* CONFIG_MEMCG_KMEM */
3021 * The user of this function is...
3024 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3025 char *buf, size_t nbytes, loff_t off)
3027 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3028 unsigned long nr_pages;
3031 buf = strstrip(buf);
3032 ret = page_counter_memparse(buf, "-1", &nr_pages);
3036 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3038 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3042 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3044 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3047 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3050 ret = memcg_update_kmem_limit(memcg, nr_pages);
3054 case RES_SOFT_LIMIT:
3055 memcg->soft_limit = nr_pages;
3059 return ret ?: nbytes;
3062 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3063 size_t nbytes, loff_t off)
3065 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3066 struct page_counter *counter;
3068 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3070 counter = &memcg->memory;
3073 counter = &memcg->memsw;
3076 counter = &memcg->kmem;
3082 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3084 page_counter_reset_watermark(counter);
3087 counter->failcnt = 0;
3096 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3099 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3103 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3104 struct cftype *cft, u64 val)
3106 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3108 if (val & ~MOVE_MASK)
3112 * No kind of locking is needed in here, because ->can_attach() will
3113 * check this value once in the beginning of the process, and then carry
3114 * on with stale data. This means that changes to this value will only
3115 * affect task migrations starting after the change.
3117 memcg->move_charge_at_immigrate = val;
3121 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3122 struct cftype *cft, u64 val)
3129 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3133 unsigned int lru_mask;
3136 static const struct numa_stat stats[] = {
3137 { "total", LRU_ALL },
3138 { "file", LRU_ALL_FILE },
3139 { "anon", LRU_ALL_ANON },
3140 { "unevictable", BIT(LRU_UNEVICTABLE) },
3142 const struct numa_stat *stat;
3145 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3147 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3148 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3149 seq_printf(m, "%s=%lu", stat->name, nr);
3150 for_each_node_state(nid, N_MEMORY) {
3151 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3153 seq_printf(m, " N%d=%lu", nid, nr);
3158 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3159 struct mem_cgroup *iter;
3162 for_each_mem_cgroup_tree(iter, memcg)
3163 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3164 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3165 for_each_node_state(nid, N_MEMORY) {
3167 for_each_mem_cgroup_tree(iter, memcg)
3168 nr += mem_cgroup_node_nr_lru_pages(
3169 iter, nid, stat->lru_mask);
3170 seq_printf(m, " N%d=%lu", nid, nr);
3177 #endif /* CONFIG_NUMA */
3179 static int memcg_stat_show(struct seq_file *m, void *v)
3181 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3182 unsigned long memory, memsw;
3183 struct mem_cgroup *mi;
3186 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3187 MEM_CGROUP_STAT_NSTATS);
3188 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3189 MEM_CGROUP_EVENTS_NSTATS);
3190 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3192 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3193 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3195 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3196 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3199 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3200 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3201 mem_cgroup_read_events(memcg, i));
3203 for (i = 0; i < NR_LRU_LISTS; i++)
3204 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3205 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3207 /* Hierarchical information */
3208 memory = memsw = PAGE_COUNTER_MAX;
3209 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3210 memory = min(memory, mi->memory.limit);
3211 memsw = min(memsw, mi->memsw.limit);
3213 seq_printf(m, "hierarchical_memory_limit %llu\n",
3214 (u64)memory * PAGE_SIZE);
3215 if (do_memsw_account())
3216 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3217 (u64)memsw * PAGE_SIZE);
3219 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3220 unsigned long long val = 0;
3222 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3224 for_each_mem_cgroup_tree(mi, memcg)
3225 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3226 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3229 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3230 unsigned long long val = 0;
3232 for_each_mem_cgroup_tree(mi, memcg)
3233 val += mem_cgroup_read_events(mi, i);
3234 seq_printf(m, "total_%s %llu\n",
3235 mem_cgroup_events_names[i], val);
3238 for (i = 0; i < NR_LRU_LISTS; i++) {
3239 unsigned long long val = 0;
3241 for_each_mem_cgroup_tree(mi, memcg)
3242 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3243 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3246 #ifdef CONFIG_DEBUG_VM
3249 struct mem_cgroup_per_zone *mz;
3250 struct zone_reclaim_stat *rstat;
3251 unsigned long recent_rotated[2] = {0, 0};
3252 unsigned long recent_scanned[2] = {0, 0};
3254 for_each_online_node(nid)
3255 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3256 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3257 rstat = &mz->lruvec.reclaim_stat;
3259 recent_rotated[0] += rstat->recent_rotated[0];
3260 recent_rotated[1] += rstat->recent_rotated[1];
3261 recent_scanned[0] += rstat->recent_scanned[0];
3262 recent_scanned[1] += rstat->recent_scanned[1];
3264 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3265 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3266 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3267 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3274 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3277 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3279 return mem_cgroup_swappiness(memcg);
3282 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3283 struct cftype *cft, u64 val)
3285 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3291 memcg->swappiness = val;
3293 vm_swappiness = val;
3298 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3300 struct mem_cgroup_threshold_ary *t;
3301 unsigned long usage;
3306 t = rcu_dereference(memcg->thresholds.primary);
3308 t = rcu_dereference(memcg->memsw_thresholds.primary);
3313 usage = mem_cgroup_usage(memcg, swap);
3316 * current_threshold points to threshold just below or equal to usage.
3317 * If it's not true, a threshold was crossed after last
3318 * call of __mem_cgroup_threshold().
3320 i = t->current_threshold;
3323 * Iterate backward over array of thresholds starting from
3324 * current_threshold and check if a threshold is crossed.
3325 * If none of thresholds below usage is crossed, we read
3326 * only one element of the array here.
3328 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3329 eventfd_signal(t->entries[i].eventfd, 1);
3331 /* i = current_threshold + 1 */
3335 * Iterate forward over array of thresholds starting from
3336 * current_threshold+1 and check if a threshold is crossed.
3337 * If none of thresholds above usage is crossed, we read
3338 * only one element of the array here.
3340 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3341 eventfd_signal(t->entries[i].eventfd, 1);
3343 /* Update current_threshold */
3344 t->current_threshold = i - 1;
3349 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3352 __mem_cgroup_threshold(memcg, false);
3353 if (do_memsw_account())
3354 __mem_cgroup_threshold(memcg, true);
3356 memcg = parent_mem_cgroup(memcg);
3360 static int compare_thresholds(const void *a, const void *b)
3362 const struct mem_cgroup_threshold *_a = a;
3363 const struct mem_cgroup_threshold *_b = b;
3365 if (_a->threshold > _b->threshold)
3368 if (_a->threshold < _b->threshold)
3374 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3376 struct mem_cgroup_eventfd_list *ev;
3378 spin_lock(&memcg_oom_lock);
3380 list_for_each_entry(ev, &memcg->oom_notify, list)
3381 eventfd_signal(ev->eventfd, 1);
3383 spin_unlock(&memcg_oom_lock);
3387 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3389 struct mem_cgroup *iter;
3391 for_each_mem_cgroup_tree(iter, memcg)
3392 mem_cgroup_oom_notify_cb(iter);
3395 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3396 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3398 struct mem_cgroup_thresholds *thresholds;
3399 struct mem_cgroup_threshold_ary *new;
3400 unsigned long threshold;
3401 unsigned long usage;
3404 ret = page_counter_memparse(args, "-1", &threshold);
3408 mutex_lock(&memcg->thresholds_lock);
3411 thresholds = &memcg->thresholds;
3412 usage = mem_cgroup_usage(memcg, false);
3413 } else if (type == _MEMSWAP) {
3414 thresholds = &memcg->memsw_thresholds;
3415 usage = mem_cgroup_usage(memcg, true);
3419 /* Check if a threshold crossed before adding a new one */
3420 if (thresholds->primary)
3421 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3423 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3425 /* Allocate memory for new array of thresholds */
3426 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3434 /* Copy thresholds (if any) to new array */
3435 if (thresholds->primary) {
3436 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3437 sizeof(struct mem_cgroup_threshold));
3440 /* Add new threshold */
3441 new->entries[size - 1].eventfd = eventfd;
3442 new->entries[size - 1].threshold = threshold;
3444 /* Sort thresholds. Registering of new threshold isn't time-critical */
3445 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3446 compare_thresholds, NULL);
3448 /* Find current threshold */
3449 new->current_threshold = -1;
3450 for (i = 0; i < size; i++) {
3451 if (new->entries[i].threshold <= usage) {
3453 * new->current_threshold will not be used until
3454 * rcu_assign_pointer(), so it's safe to increment
3457 ++new->current_threshold;
3462 /* Free old spare buffer and save old primary buffer as spare */
3463 kfree(thresholds->spare);
3464 thresholds->spare = thresholds->primary;
3466 rcu_assign_pointer(thresholds->primary, new);
3468 /* To be sure that nobody uses thresholds */
3472 mutex_unlock(&memcg->thresholds_lock);
3477 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3478 struct eventfd_ctx *eventfd, const char *args)
3480 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3483 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3484 struct eventfd_ctx *eventfd, const char *args)
3486 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3489 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3490 struct eventfd_ctx *eventfd, enum res_type type)
3492 struct mem_cgroup_thresholds *thresholds;
3493 struct mem_cgroup_threshold_ary *new;
3494 unsigned long usage;
3497 mutex_lock(&memcg->thresholds_lock);
3500 thresholds = &memcg->thresholds;
3501 usage = mem_cgroup_usage(memcg, false);
3502 } else if (type == _MEMSWAP) {
3503 thresholds = &memcg->memsw_thresholds;
3504 usage = mem_cgroup_usage(memcg, true);
3508 if (!thresholds->primary)
3511 /* Check if a threshold crossed before removing */
3512 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3514 /* Calculate new number of threshold */
3516 for (i = 0; i < thresholds->primary->size; i++) {
3517 if (thresholds->primary->entries[i].eventfd != eventfd)
3521 new = thresholds->spare;
3523 /* Set thresholds array to NULL if we don't have thresholds */
3532 /* Copy thresholds and find current threshold */
3533 new->current_threshold = -1;
3534 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3535 if (thresholds->primary->entries[i].eventfd == eventfd)
3538 new->entries[j] = thresholds->primary->entries[i];
3539 if (new->entries[j].threshold <= usage) {
3541 * new->current_threshold will not be used
3542 * until rcu_assign_pointer(), so it's safe to increment
3545 ++new->current_threshold;
3551 /* Swap primary and spare array */
3552 thresholds->spare = thresholds->primary;
3553 /* If all events are unregistered, free the spare array */
3555 kfree(thresholds->spare);
3556 thresholds->spare = NULL;
3559 rcu_assign_pointer(thresholds->primary, new);
3561 /* To be sure that nobody uses thresholds */
3564 mutex_unlock(&memcg->thresholds_lock);
3567 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3568 struct eventfd_ctx *eventfd)
3570 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3573 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3574 struct eventfd_ctx *eventfd)
3576 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3579 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3580 struct eventfd_ctx *eventfd, const char *args)
3582 struct mem_cgroup_eventfd_list *event;
3584 event = kmalloc(sizeof(*event), GFP_KERNEL);
3588 spin_lock(&memcg_oom_lock);
3590 event->eventfd = eventfd;
3591 list_add(&event->list, &memcg->oom_notify);
3593 /* already in OOM ? */
3594 if (memcg->under_oom)
3595 eventfd_signal(eventfd, 1);
3596 spin_unlock(&memcg_oom_lock);
3601 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3602 struct eventfd_ctx *eventfd)
3604 struct mem_cgroup_eventfd_list *ev, *tmp;
3606 spin_lock(&memcg_oom_lock);
3608 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3609 if (ev->eventfd == eventfd) {
3610 list_del(&ev->list);
3615 spin_unlock(&memcg_oom_lock);
3618 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3620 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3622 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3623 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3627 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3628 struct cftype *cft, u64 val)
3630 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3632 /* cannot set to root cgroup and only 0 and 1 are allowed */
3633 if (!css->parent || !((val == 0) || (val == 1)))
3636 memcg->oom_kill_disable = val;
3638 memcg_oom_recover(memcg);
3643 #ifdef CONFIG_MEMCG_KMEM
3644 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3648 ret = memcg_propagate_kmem(memcg);
3652 return tcp_init_cgroup(memcg, ss);
3655 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3657 struct cgroup_subsys_state *css;
3658 struct mem_cgroup *parent, *child;
3661 if (!memcg->kmem_acct_active)
3665 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3666 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3667 * guarantees no cache will be created for this cgroup after we are
3668 * done (see memcg_create_kmem_cache()).
3670 memcg->kmem_acct_active = false;
3672 memcg_deactivate_kmem_caches(memcg);
3674 kmemcg_id = memcg->kmemcg_id;
3675 BUG_ON(kmemcg_id < 0);
3677 parent = parent_mem_cgroup(memcg);
3679 parent = root_mem_cgroup;
3682 * Change kmemcg_id of this cgroup and all its descendants to the
3683 * parent's id, and then move all entries from this cgroup's list_lrus
3684 * to ones of the parent. After we have finished, all list_lrus
3685 * corresponding to this cgroup are guaranteed to remain empty. The
3686 * ordering is imposed by list_lru_node->lock taken by
3687 * memcg_drain_all_list_lrus().
3689 css_for_each_descendant_pre(css, &memcg->css) {
3690 child = mem_cgroup_from_css(css);
3691 BUG_ON(child->kmemcg_id != kmemcg_id);
3692 child->kmemcg_id = parent->kmemcg_id;
3693 if (!memcg->use_hierarchy)
3696 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3698 memcg_free_cache_id(kmemcg_id);
3701 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3703 if (memcg->kmem_acct_activated) {
3704 memcg_destroy_kmem_caches(memcg);
3705 static_key_slow_dec(&memcg_kmem_enabled_key);
3706 WARN_ON(page_counter_read(&memcg->kmem));
3708 tcp_destroy_cgroup(memcg);
3711 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3716 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3720 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3725 #ifdef CONFIG_CGROUP_WRITEBACK
3727 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3729 return &memcg->cgwb_list;
3732 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3734 return wb_domain_init(&memcg->cgwb_domain, gfp);
3737 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3739 wb_domain_exit(&memcg->cgwb_domain);
3742 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3744 wb_domain_size_changed(&memcg->cgwb_domain);
3747 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3749 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3751 if (!memcg->css.parent)
3754 return &memcg->cgwb_domain;
3758 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3759 * @wb: bdi_writeback in question
3760 * @pfilepages: out parameter for number of file pages
3761 * @pheadroom: out parameter for number of allocatable pages according to memcg
3762 * @pdirty: out parameter for number of dirty pages
3763 * @pwriteback: out parameter for number of pages under writeback
3765 * Determine the numbers of file, headroom, dirty, and writeback pages in
3766 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3767 * is a bit more involved.
3769 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3770 * headroom is calculated as the lowest headroom of itself and the
3771 * ancestors. Note that this doesn't consider the actual amount of
3772 * available memory in the system. The caller should further cap
3773 * *@pheadroom accordingly.
3775 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3776 unsigned long *pheadroom, unsigned long *pdirty,
3777 unsigned long *pwriteback)
3779 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3780 struct mem_cgroup *parent;
3782 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3784 /* this should eventually include NR_UNSTABLE_NFS */
3785 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3786 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3787 (1 << LRU_ACTIVE_FILE));
3788 *pheadroom = PAGE_COUNTER_MAX;
3790 while ((parent = parent_mem_cgroup(memcg))) {
3791 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3792 unsigned long used = page_counter_read(&memcg->memory);
3794 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3799 #else /* CONFIG_CGROUP_WRITEBACK */
3801 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3806 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3810 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3814 #endif /* CONFIG_CGROUP_WRITEBACK */
3817 * DO NOT USE IN NEW FILES.
3819 * "cgroup.event_control" implementation.
3821 * This is way over-engineered. It tries to support fully configurable
3822 * events for each user. Such level of flexibility is completely
3823 * unnecessary especially in the light of the planned unified hierarchy.
3825 * Please deprecate this and replace with something simpler if at all
3830 * Unregister event and free resources.
3832 * Gets called from workqueue.
3834 static void memcg_event_remove(struct work_struct *work)
3836 struct mem_cgroup_event *event =
3837 container_of(work, struct mem_cgroup_event, remove);
3838 struct mem_cgroup *memcg = event->memcg;
3840 remove_wait_queue(event->wqh, &event->wait);
3842 event->unregister_event(memcg, event->eventfd);
3844 /* Notify userspace the event is going away. */
3845 eventfd_signal(event->eventfd, 1);
3847 eventfd_ctx_put(event->eventfd);
3849 css_put(&memcg->css);
3853 * Gets called on POLLHUP on eventfd when user closes it.
3855 * Called with wqh->lock held and interrupts disabled.
3857 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3858 int sync, void *key)
3860 struct mem_cgroup_event *event =
3861 container_of(wait, struct mem_cgroup_event, wait);
3862 struct mem_cgroup *memcg = event->memcg;
3863 unsigned long flags = (unsigned long)key;
3865 if (flags & POLLHUP) {
3867 * If the event has been detached at cgroup removal, we
3868 * can simply return knowing the other side will cleanup
3871 * We can't race against event freeing since the other
3872 * side will require wqh->lock via remove_wait_queue(),
3875 spin_lock(&memcg->event_list_lock);
3876 if (!list_empty(&event->list)) {
3877 list_del_init(&event->list);
3879 * We are in atomic context, but cgroup_event_remove()
3880 * may sleep, so we have to call it in workqueue.
3882 schedule_work(&event->remove);
3884 spin_unlock(&memcg->event_list_lock);
3890 static void memcg_event_ptable_queue_proc(struct file *file,
3891 wait_queue_head_t *wqh, poll_table *pt)
3893 struct mem_cgroup_event *event =
3894 container_of(pt, struct mem_cgroup_event, pt);
3897 add_wait_queue(wqh, &event->wait);
3901 * DO NOT USE IN NEW FILES.
3903 * Parse input and register new cgroup event handler.
3905 * Input must be in format '<event_fd> <control_fd> <args>'.
3906 * Interpretation of args is defined by control file implementation.
3908 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3909 char *buf, size_t nbytes, loff_t off)
3911 struct cgroup_subsys_state *css = of_css(of);
3912 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3913 struct mem_cgroup_event *event;
3914 struct cgroup_subsys_state *cfile_css;
3915 unsigned int efd, cfd;
3922 buf = strstrip(buf);
3924 efd = simple_strtoul(buf, &endp, 10);
3929 cfd = simple_strtoul(buf, &endp, 10);
3930 if ((*endp != ' ') && (*endp != '\0'))
3934 event = kzalloc(sizeof(*event), GFP_KERNEL);
3938 event->memcg = memcg;
3939 INIT_LIST_HEAD(&event->list);
3940 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3941 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3942 INIT_WORK(&event->remove, memcg_event_remove);
3950 event->eventfd = eventfd_ctx_fileget(efile.file);
3951 if (IS_ERR(event->eventfd)) {
3952 ret = PTR_ERR(event->eventfd);
3959 goto out_put_eventfd;
3962 /* the process need read permission on control file */
3963 /* AV: shouldn't we check that it's been opened for read instead? */
3964 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3969 * Determine the event callbacks and set them in @event. This used
3970 * to be done via struct cftype but cgroup core no longer knows
3971 * about these events. The following is crude but the whole thing
3972 * is for compatibility anyway.
3974 * DO NOT ADD NEW FILES.
3976 name = cfile.file->f_path.dentry->d_name.name;
3978 if (!strcmp(name, "memory.usage_in_bytes")) {
3979 event->register_event = mem_cgroup_usage_register_event;
3980 event->unregister_event = mem_cgroup_usage_unregister_event;
3981 } else if (!strcmp(name, "memory.oom_control")) {
3982 event->register_event = mem_cgroup_oom_register_event;
3983 event->unregister_event = mem_cgroup_oom_unregister_event;
3984 } else if (!strcmp(name, "memory.pressure_level")) {
3985 event->register_event = vmpressure_register_event;
3986 event->unregister_event = vmpressure_unregister_event;
3987 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3988 event->register_event = memsw_cgroup_usage_register_event;
3989 event->unregister_event = memsw_cgroup_usage_unregister_event;
3996 * Verify @cfile should belong to @css. Also, remaining events are
3997 * automatically removed on cgroup destruction but the removal is
3998 * asynchronous, so take an extra ref on @css.
4000 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4001 &memory_cgrp_subsys);
4003 if (IS_ERR(cfile_css))
4005 if (cfile_css != css) {
4010 ret = event->register_event(memcg, event->eventfd, buf);
4014 efile.file->f_op->poll(efile.file, &event->pt);
4016 spin_lock(&memcg->event_list_lock);
4017 list_add(&event->list, &memcg->event_list);
4018 spin_unlock(&memcg->event_list_lock);
4030 eventfd_ctx_put(event->eventfd);
4039 static struct cftype mem_cgroup_legacy_files[] = {
4041 .name = "usage_in_bytes",
4042 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4043 .read_u64 = mem_cgroup_read_u64,
4046 .name = "max_usage_in_bytes",
4047 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4048 .write = mem_cgroup_reset,
4049 .read_u64 = mem_cgroup_read_u64,
4052 .name = "limit_in_bytes",
4053 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4054 .write = mem_cgroup_write,
4055 .read_u64 = mem_cgroup_read_u64,
4058 .name = "soft_limit_in_bytes",
4059 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4060 .write = mem_cgroup_write,
4061 .read_u64 = mem_cgroup_read_u64,
4065 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4066 .write = mem_cgroup_reset,
4067 .read_u64 = mem_cgroup_read_u64,
4071 .seq_show = memcg_stat_show,
4074 .name = "force_empty",
4075 .write = mem_cgroup_force_empty_write,
4078 .name = "use_hierarchy",
4079 .write_u64 = mem_cgroup_hierarchy_write,
4080 .read_u64 = mem_cgroup_hierarchy_read,
4083 .name = "cgroup.event_control", /* XXX: for compat */
4084 .write = memcg_write_event_control,
4085 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4088 .name = "swappiness",
4089 .read_u64 = mem_cgroup_swappiness_read,
4090 .write_u64 = mem_cgroup_swappiness_write,
4093 .name = "move_charge_at_immigrate",
4094 .read_u64 = mem_cgroup_move_charge_read,
4095 .write_u64 = mem_cgroup_move_charge_write,
4098 .name = "oom_control",
4099 .seq_show = mem_cgroup_oom_control_read,
4100 .write_u64 = mem_cgroup_oom_control_write,
4101 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4104 .name = "pressure_level",
4108 .name = "numa_stat",
4109 .seq_show = memcg_numa_stat_show,
4112 #ifdef CONFIG_MEMCG_KMEM
4114 .name = "kmem.limit_in_bytes",
4115 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4116 .write = mem_cgroup_write,
4117 .read_u64 = mem_cgroup_read_u64,
4120 .name = "kmem.usage_in_bytes",
4121 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4122 .read_u64 = mem_cgroup_read_u64,
4125 .name = "kmem.failcnt",
4126 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4127 .write = mem_cgroup_reset,
4128 .read_u64 = mem_cgroup_read_u64,
4131 .name = "kmem.max_usage_in_bytes",
4132 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4133 .write = mem_cgroup_reset,
4134 .read_u64 = mem_cgroup_read_u64,
4136 #ifdef CONFIG_SLABINFO
4138 .name = "kmem.slabinfo",
4139 .seq_start = slab_start,
4140 .seq_next = slab_next,
4141 .seq_stop = slab_stop,
4142 .seq_show = memcg_slab_show,
4146 { }, /* terminate */
4149 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4151 struct mem_cgroup_per_node *pn;
4152 struct mem_cgroup_per_zone *mz;
4153 int zone, tmp = node;
4155 * This routine is called against possible nodes.
4156 * But it's BUG to call kmalloc() against offline node.
4158 * TODO: this routine can waste much memory for nodes which will
4159 * never be onlined. It's better to use memory hotplug callback
4162 if (!node_state(node, N_NORMAL_MEMORY))
4164 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4168 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4169 mz = &pn->zoneinfo[zone];
4170 lruvec_init(&mz->lruvec);
4171 mz->usage_in_excess = 0;
4172 mz->on_tree = false;
4175 memcg->nodeinfo[node] = pn;
4179 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4181 kfree(memcg->nodeinfo[node]);
4184 static struct mem_cgroup *mem_cgroup_alloc(void)
4186 struct mem_cgroup *memcg;
4189 size = sizeof(struct mem_cgroup);
4190 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4192 memcg = kzalloc(size, GFP_KERNEL);
4196 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4200 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4206 free_percpu(memcg->stat);
4213 * At destroying mem_cgroup, references from swap_cgroup can remain.
4214 * (scanning all at force_empty is too costly...)
4216 * Instead of clearing all references at force_empty, we remember
4217 * the number of reference from swap_cgroup and free mem_cgroup when
4218 * it goes down to 0.
4220 * Removal of cgroup itself succeeds regardless of refs from swap.
4223 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4227 mem_cgroup_remove_from_trees(memcg);
4230 free_mem_cgroup_per_zone_info(memcg, node);
4232 free_percpu(memcg->stat);
4233 memcg_wb_domain_exit(memcg);
4238 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4240 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4242 if (!memcg->memory.parent)
4244 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4246 EXPORT_SYMBOL(parent_mem_cgroup);
4248 static struct cgroup_subsys_state * __ref
4249 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4251 struct mem_cgroup *memcg;
4252 long error = -ENOMEM;
4255 memcg = mem_cgroup_alloc();
4257 return ERR_PTR(error);
4260 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4264 if (parent_css == NULL) {
4265 root_mem_cgroup = memcg;
4266 page_counter_init(&memcg->memory, NULL);
4267 memcg->high = PAGE_COUNTER_MAX;
4268 memcg->soft_limit = PAGE_COUNTER_MAX;
4269 page_counter_init(&memcg->memsw, NULL);
4270 page_counter_init(&memcg->kmem, NULL);
4273 memcg->last_scanned_node = MAX_NUMNODES;
4274 INIT_LIST_HEAD(&memcg->oom_notify);
4275 memcg->move_charge_at_immigrate = 0;
4276 mutex_init(&memcg->thresholds_lock);
4277 spin_lock_init(&memcg->move_lock);
4278 vmpressure_init(&memcg->vmpressure);
4279 INIT_LIST_HEAD(&memcg->event_list);
4280 spin_lock_init(&memcg->event_list_lock);
4281 #ifdef CONFIG_MEMCG_KMEM
4282 memcg->kmemcg_id = -1;
4284 #ifdef CONFIG_CGROUP_WRITEBACK
4285 INIT_LIST_HEAD(&memcg->cgwb_list);
4290 __mem_cgroup_free(memcg);
4291 return ERR_PTR(error);
4295 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4297 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4298 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4301 if (css->id > MEM_CGROUP_ID_MAX)
4307 mutex_lock(&memcg_create_mutex);
4309 memcg->use_hierarchy = parent->use_hierarchy;
4310 memcg->oom_kill_disable = parent->oom_kill_disable;
4311 memcg->swappiness = mem_cgroup_swappiness(parent);
4313 if (parent->use_hierarchy) {
4314 page_counter_init(&memcg->memory, &parent->memory);
4315 memcg->high = PAGE_COUNTER_MAX;
4316 memcg->soft_limit = PAGE_COUNTER_MAX;
4317 page_counter_init(&memcg->memsw, &parent->memsw);
4318 page_counter_init(&memcg->kmem, &parent->kmem);
4321 * No need to take a reference to the parent because cgroup
4322 * core guarantees its existence.
4325 page_counter_init(&memcg->memory, NULL);
4326 memcg->high = PAGE_COUNTER_MAX;
4327 memcg->soft_limit = PAGE_COUNTER_MAX;
4328 page_counter_init(&memcg->memsw, NULL);
4329 page_counter_init(&memcg->kmem, NULL);
4331 * Deeper hierachy with use_hierarchy == false doesn't make
4332 * much sense so let cgroup subsystem know about this
4333 * unfortunate state in our controller.
4335 if (parent != root_mem_cgroup)
4336 memory_cgrp_subsys.broken_hierarchy = true;
4338 mutex_unlock(&memcg_create_mutex);
4340 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4345 * Make sure the memcg is initialized: mem_cgroup_iter()
4346 * orders reading memcg->initialized against its callers
4347 * reading the memcg members.
4349 smp_store_release(&memcg->initialized, 1);
4354 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4356 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4357 struct mem_cgroup_event *event, *tmp;
4360 * Unregister events and notify userspace.
4361 * Notify userspace about cgroup removing only after rmdir of cgroup
4362 * directory to avoid race between userspace and kernelspace.
4364 spin_lock(&memcg->event_list_lock);
4365 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4366 list_del_init(&event->list);
4367 schedule_work(&event->remove);
4369 spin_unlock(&memcg->event_list_lock);
4371 vmpressure_cleanup(&memcg->vmpressure);
4373 memcg_deactivate_kmem(memcg);
4375 wb_memcg_offline(memcg);
4378 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4380 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4382 invalidate_reclaim_iterators(memcg);
4385 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4387 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4389 memcg_destroy_kmem(memcg);
4390 __mem_cgroup_free(memcg);
4394 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4395 * @css: the target css
4397 * Reset the states of the mem_cgroup associated with @css. This is
4398 * invoked when the userland requests disabling on the default hierarchy
4399 * but the memcg is pinned through dependency. The memcg should stop
4400 * applying policies and should revert to the vanilla state as it may be
4401 * made visible again.
4403 * The current implementation only resets the essential configurations.
4404 * This needs to be expanded to cover all the visible parts.
4406 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4408 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4410 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4411 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4412 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4414 memcg->high = PAGE_COUNTER_MAX;
4415 memcg->soft_limit = PAGE_COUNTER_MAX;
4416 memcg_wb_domain_size_changed(memcg);
4420 /* Handlers for move charge at task migration. */
4421 static int mem_cgroup_do_precharge(unsigned long count)
4425 /* Try a single bulk charge without reclaim first, kswapd may wake */
4426 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4428 mc.precharge += count;
4432 /* Try charges one by one with reclaim */
4434 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4444 * get_mctgt_type - get target type of moving charge
4445 * @vma: the vma the pte to be checked belongs
4446 * @addr: the address corresponding to the pte to be checked
4447 * @ptent: the pte to be checked
4448 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4451 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4452 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4453 * move charge. if @target is not NULL, the page is stored in target->page
4454 * with extra refcnt got(Callers should handle it).
4455 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4456 * target for charge migration. if @target is not NULL, the entry is stored
4459 * Called with pte lock held.
4466 enum mc_target_type {
4472 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4473 unsigned long addr, pte_t ptent)
4475 struct page *page = vm_normal_page(vma, addr, ptent);
4477 if (!page || !page_mapped(page))
4479 if (PageAnon(page)) {
4480 if (!(mc.flags & MOVE_ANON))
4483 if (!(mc.flags & MOVE_FILE))
4486 if (!get_page_unless_zero(page))
4493 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4494 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4496 struct page *page = NULL;
4497 swp_entry_t ent = pte_to_swp_entry(ptent);
4499 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4502 * Because lookup_swap_cache() updates some statistics counter,
4503 * we call find_get_page() with swapper_space directly.
4505 page = find_get_page(swap_address_space(ent), ent.val);
4506 if (do_memsw_account())
4507 entry->val = ent.val;
4512 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4513 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4519 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4520 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4522 struct page *page = NULL;
4523 struct address_space *mapping;
4526 if (!vma->vm_file) /* anonymous vma */
4528 if (!(mc.flags & MOVE_FILE))
4531 mapping = vma->vm_file->f_mapping;
4532 pgoff = linear_page_index(vma, addr);
4534 /* page is moved even if it's not RSS of this task(page-faulted). */
4536 /* shmem/tmpfs may report page out on swap: account for that too. */
4537 if (shmem_mapping(mapping)) {
4538 page = find_get_entry(mapping, pgoff);
4539 if (radix_tree_exceptional_entry(page)) {
4540 swp_entry_t swp = radix_to_swp_entry(page);
4541 if (do_memsw_account())
4543 page = find_get_page(swap_address_space(swp), swp.val);
4546 page = find_get_page(mapping, pgoff);
4548 page = find_get_page(mapping, pgoff);
4554 * mem_cgroup_move_account - move account of the page
4556 * @nr_pages: number of regular pages (>1 for huge pages)
4557 * @from: mem_cgroup which the page is moved from.
4558 * @to: mem_cgroup which the page is moved to. @from != @to.
4560 * The caller must confirm following.
4561 * - page is not on LRU (isolate_page() is useful.)
4562 * - compound_lock is held when nr_pages > 1
4564 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4567 static int mem_cgroup_move_account(struct page *page,
4568 unsigned int nr_pages,
4569 struct mem_cgroup *from,
4570 struct mem_cgroup *to)
4572 unsigned long flags;
4576 VM_BUG_ON(from == to);
4577 VM_BUG_ON_PAGE(PageLRU(page), page);
4579 * The page is isolated from LRU. So, collapse function
4580 * will not handle this page. But page splitting can happen.
4581 * Do this check under compound_page_lock(). The caller should
4585 if (nr_pages > 1 && !PageTransHuge(page))
4589 * Prevent mem_cgroup_replace_page() from looking at
4590 * page->mem_cgroup of its source page while we change it.
4592 if (!trylock_page(page))
4596 if (page->mem_cgroup != from)
4599 anon = PageAnon(page);
4601 spin_lock_irqsave(&from->move_lock, flags);
4603 if (!anon && page_mapped(page)) {
4604 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4606 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4611 * move_lock grabbed above and caller set from->moving_account, so
4612 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4613 * So mapping should be stable for dirty pages.
4615 if (!anon && PageDirty(page)) {
4616 struct address_space *mapping = page_mapping(page);
4618 if (mapping_cap_account_dirty(mapping)) {
4619 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4621 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4626 if (PageWriteback(page)) {
4627 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4629 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4634 * It is safe to change page->mem_cgroup here because the page
4635 * is referenced, charged, and isolated - we can't race with
4636 * uncharging, charging, migration, or LRU putback.
4639 /* caller should have done css_get */
4640 page->mem_cgroup = to;
4641 spin_unlock_irqrestore(&from->move_lock, flags);
4645 local_irq_disable();
4646 mem_cgroup_charge_statistics(to, page, nr_pages);
4647 memcg_check_events(to, page);
4648 mem_cgroup_charge_statistics(from, page, -nr_pages);
4649 memcg_check_events(from, page);
4657 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4658 unsigned long addr, pte_t ptent, union mc_target *target)
4660 struct page *page = NULL;
4661 enum mc_target_type ret = MC_TARGET_NONE;
4662 swp_entry_t ent = { .val = 0 };
4664 if (pte_present(ptent))
4665 page = mc_handle_present_pte(vma, addr, ptent);
4666 else if (is_swap_pte(ptent))
4667 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4668 else if (pte_none(ptent))
4669 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4671 if (!page && !ent.val)
4675 * Do only loose check w/o serialization.
4676 * mem_cgroup_move_account() checks the page is valid or
4677 * not under LRU exclusion.
4679 if (page->mem_cgroup == mc.from) {
4680 ret = MC_TARGET_PAGE;
4682 target->page = page;
4684 if (!ret || !target)
4687 /* There is a swap entry and a page doesn't exist or isn't charged */
4688 if (ent.val && !ret &&
4689 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4690 ret = MC_TARGET_SWAP;
4697 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4699 * We don't consider swapping or file mapped pages because THP does not
4700 * support them for now.
4701 * Caller should make sure that pmd_trans_huge(pmd) is true.
4703 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4704 unsigned long addr, pmd_t pmd, union mc_target *target)
4706 struct page *page = NULL;
4707 enum mc_target_type ret = MC_TARGET_NONE;
4709 page = pmd_page(pmd);
4710 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4711 if (!(mc.flags & MOVE_ANON))
4713 if (page->mem_cgroup == mc.from) {
4714 ret = MC_TARGET_PAGE;
4717 target->page = page;
4723 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4724 unsigned long addr, pmd_t pmd, union mc_target *target)
4726 return MC_TARGET_NONE;
4730 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4731 unsigned long addr, unsigned long end,
4732 struct mm_walk *walk)
4734 struct vm_area_struct *vma = walk->vma;
4738 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4739 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4740 mc.precharge += HPAGE_PMD_NR;
4745 if (pmd_trans_unstable(pmd))
4747 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4748 for (; addr != end; pte++, addr += PAGE_SIZE)
4749 if (get_mctgt_type(vma, addr, *pte, NULL))
4750 mc.precharge++; /* increment precharge temporarily */
4751 pte_unmap_unlock(pte - 1, ptl);
4757 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4759 unsigned long precharge;
4761 struct mm_walk mem_cgroup_count_precharge_walk = {
4762 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4765 down_read(&mm->mmap_sem);
4766 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4767 up_read(&mm->mmap_sem);
4769 precharge = mc.precharge;
4775 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4777 unsigned long precharge = mem_cgroup_count_precharge(mm);
4779 VM_BUG_ON(mc.moving_task);
4780 mc.moving_task = current;
4781 return mem_cgroup_do_precharge(precharge);
4784 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4785 static void __mem_cgroup_clear_mc(void)
4787 struct mem_cgroup *from = mc.from;
4788 struct mem_cgroup *to = mc.to;
4790 /* we must uncharge all the leftover precharges from mc.to */
4792 cancel_charge(mc.to, mc.precharge);
4796 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4797 * we must uncharge here.
4799 if (mc.moved_charge) {
4800 cancel_charge(mc.from, mc.moved_charge);
4801 mc.moved_charge = 0;
4803 /* we must fixup refcnts and charges */
4804 if (mc.moved_swap) {
4805 /* uncharge swap account from the old cgroup */
4806 if (!mem_cgroup_is_root(mc.from))
4807 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4810 * we charged both to->memory and to->memsw, so we
4811 * should uncharge to->memory.
4813 if (!mem_cgroup_is_root(mc.to))
4814 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4816 css_put_many(&mc.from->css, mc.moved_swap);
4818 /* we've already done css_get(mc.to) */
4821 memcg_oom_recover(from);
4822 memcg_oom_recover(to);
4823 wake_up_all(&mc.waitq);
4826 static void mem_cgroup_clear_mc(void)
4829 * we must clear moving_task before waking up waiters at the end of
4832 mc.moving_task = NULL;
4833 __mem_cgroup_clear_mc();
4834 spin_lock(&mc.lock);
4837 spin_unlock(&mc.lock);
4840 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4842 struct cgroup_subsys_state *css;
4843 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4844 struct mem_cgroup *from;
4845 struct task_struct *leader, *p;
4846 struct mm_struct *mm;
4847 unsigned long move_flags;
4850 /* charge immigration isn't supported on the default hierarchy */
4851 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4855 * Multi-process migrations only happen on the default hierarchy
4856 * where charge immigration is not used. Perform charge
4857 * immigration if @tset contains a leader and whine if there are
4861 cgroup_taskset_for_each_leader(leader, css, tset) {
4864 memcg = mem_cgroup_from_css(css);
4870 * We are now commited to this value whatever it is. Changes in this
4871 * tunable will only affect upcoming migrations, not the current one.
4872 * So we need to save it, and keep it going.
4874 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4878 from = mem_cgroup_from_task(p);
4880 VM_BUG_ON(from == memcg);
4882 mm = get_task_mm(p);
4885 /* We move charges only when we move a owner of the mm */
4886 if (mm->owner == p) {
4889 VM_BUG_ON(mc.precharge);
4890 VM_BUG_ON(mc.moved_charge);
4891 VM_BUG_ON(mc.moved_swap);
4893 spin_lock(&mc.lock);
4896 mc.flags = move_flags;
4897 spin_unlock(&mc.lock);
4898 /* We set mc.moving_task later */
4900 ret = mem_cgroup_precharge_mc(mm);
4902 mem_cgroup_clear_mc();
4908 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4911 mem_cgroup_clear_mc();
4914 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4915 unsigned long addr, unsigned long end,
4916 struct mm_walk *walk)
4919 struct vm_area_struct *vma = walk->vma;
4922 enum mc_target_type target_type;
4923 union mc_target target;
4927 * We don't take compound_lock() here but no race with splitting thp
4929 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4930 * under splitting, which means there's no concurrent thp split,
4931 * - if another thread runs into split_huge_page() just after we
4932 * entered this if-block, the thread must wait for page table lock
4933 * to be unlocked in __split_huge_page_splitting(), where the main
4934 * part of thp split is not executed yet.
4936 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4937 if (mc.precharge < HPAGE_PMD_NR) {
4941 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4942 if (target_type == MC_TARGET_PAGE) {
4944 if (!isolate_lru_page(page)) {
4945 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
4947 mc.precharge -= HPAGE_PMD_NR;
4948 mc.moved_charge += HPAGE_PMD_NR;
4950 putback_lru_page(page);
4958 if (pmd_trans_unstable(pmd))
4961 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4962 for (; addr != end; addr += PAGE_SIZE) {
4963 pte_t ptent = *(pte++);
4969 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4970 case MC_TARGET_PAGE:
4972 if (isolate_lru_page(page))
4974 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
4976 /* we uncharge from mc.from later. */
4979 putback_lru_page(page);
4980 put: /* get_mctgt_type() gets the page */
4983 case MC_TARGET_SWAP:
4985 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4987 /* we fixup refcnts and charges later. */
4995 pte_unmap_unlock(pte - 1, ptl);
5000 * We have consumed all precharges we got in can_attach().
5001 * We try charge one by one, but don't do any additional
5002 * charges to mc.to if we have failed in charge once in attach()
5005 ret = mem_cgroup_do_precharge(1);
5013 static void mem_cgroup_move_charge(struct mm_struct *mm)
5015 struct mm_walk mem_cgroup_move_charge_walk = {
5016 .pmd_entry = mem_cgroup_move_charge_pte_range,
5020 lru_add_drain_all();
5022 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5023 * move_lock while we're moving its pages to another memcg.
5024 * Then wait for already started RCU-only updates to finish.
5026 atomic_inc(&mc.from->moving_account);
5029 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5031 * Someone who are holding the mmap_sem might be waiting in
5032 * waitq. So we cancel all extra charges, wake up all waiters,
5033 * and retry. Because we cancel precharges, we might not be able
5034 * to move enough charges, but moving charge is a best-effort
5035 * feature anyway, so it wouldn't be a big problem.
5037 __mem_cgroup_clear_mc();
5042 * When we have consumed all precharges and failed in doing
5043 * additional charge, the page walk just aborts.
5045 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5046 up_read(&mm->mmap_sem);
5047 atomic_dec(&mc.from->moving_account);
5050 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5052 struct cgroup_subsys_state *css;
5053 struct task_struct *p = cgroup_taskset_first(tset, &css);
5054 struct mm_struct *mm = get_task_mm(p);
5058 mem_cgroup_move_charge(mm);
5062 mem_cgroup_clear_mc();
5064 #else /* !CONFIG_MMU */
5065 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5069 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5072 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5078 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5079 * to verify whether we're attached to the default hierarchy on each mount
5082 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5085 * use_hierarchy is forced on the default hierarchy. cgroup core
5086 * guarantees that @root doesn't have any children, so turning it
5087 * on for the root memcg is enough.
5089 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5090 root_mem_cgroup->use_hierarchy = true;
5092 root_mem_cgroup->use_hierarchy = false;
5095 static u64 memory_current_read(struct cgroup_subsys_state *css,
5098 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5100 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5103 static int memory_low_show(struct seq_file *m, void *v)
5105 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5106 unsigned long low = READ_ONCE(memcg->low);
5108 if (low == PAGE_COUNTER_MAX)
5109 seq_puts(m, "max\n");
5111 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5116 static ssize_t memory_low_write(struct kernfs_open_file *of,
5117 char *buf, size_t nbytes, loff_t off)
5119 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5123 buf = strstrip(buf);
5124 err = page_counter_memparse(buf, "max", &low);
5133 static int memory_high_show(struct seq_file *m, void *v)
5135 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5136 unsigned long high = READ_ONCE(memcg->high);
5138 if (high == PAGE_COUNTER_MAX)
5139 seq_puts(m, "max\n");
5141 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5146 static ssize_t memory_high_write(struct kernfs_open_file *of,
5147 char *buf, size_t nbytes, loff_t off)
5149 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5153 buf = strstrip(buf);
5154 err = page_counter_memparse(buf, "max", &high);
5160 memcg_wb_domain_size_changed(memcg);
5164 static int memory_max_show(struct seq_file *m, void *v)
5166 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5167 unsigned long max = READ_ONCE(memcg->memory.limit);
5169 if (max == PAGE_COUNTER_MAX)
5170 seq_puts(m, "max\n");
5172 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5177 static ssize_t memory_max_write(struct kernfs_open_file *of,
5178 char *buf, size_t nbytes, loff_t off)
5180 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5184 buf = strstrip(buf);
5185 err = page_counter_memparse(buf, "max", &max);
5189 err = mem_cgroup_resize_limit(memcg, max);
5193 memcg_wb_domain_size_changed(memcg);
5197 static int memory_events_show(struct seq_file *m, void *v)
5199 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5201 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5202 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5203 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5204 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5209 static struct cftype memory_files[] = {
5212 .flags = CFTYPE_NOT_ON_ROOT,
5213 .read_u64 = memory_current_read,
5217 .flags = CFTYPE_NOT_ON_ROOT,
5218 .seq_show = memory_low_show,
5219 .write = memory_low_write,
5223 .flags = CFTYPE_NOT_ON_ROOT,
5224 .seq_show = memory_high_show,
5225 .write = memory_high_write,
5229 .flags = CFTYPE_NOT_ON_ROOT,
5230 .seq_show = memory_max_show,
5231 .write = memory_max_write,
5235 .flags = CFTYPE_NOT_ON_ROOT,
5236 .file_offset = offsetof(struct mem_cgroup, events_file),
5237 .seq_show = memory_events_show,
5242 struct cgroup_subsys memory_cgrp_subsys = {
5243 .css_alloc = mem_cgroup_css_alloc,
5244 .css_online = mem_cgroup_css_online,
5245 .css_offline = mem_cgroup_css_offline,
5246 .css_released = mem_cgroup_css_released,
5247 .css_free = mem_cgroup_css_free,
5248 .css_reset = mem_cgroup_css_reset,
5249 .can_attach = mem_cgroup_can_attach,
5250 .cancel_attach = mem_cgroup_cancel_attach,
5251 .attach = mem_cgroup_move_task,
5252 .bind = mem_cgroup_bind,
5253 .dfl_cftypes = memory_files,
5254 .legacy_cftypes = mem_cgroup_legacy_files,
5259 * mem_cgroup_low - check if memory consumption is below the normal range
5260 * @root: the highest ancestor to consider
5261 * @memcg: the memory cgroup to check
5263 * Returns %true if memory consumption of @memcg, and that of all
5264 * configurable ancestors up to @root, is below the normal range.
5266 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5268 if (mem_cgroup_disabled())
5272 * The toplevel group doesn't have a configurable range, so
5273 * it's never low when looked at directly, and it is not
5274 * considered an ancestor when assessing the hierarchy.
5277 if (memcg == root_mem_cgroup)
5280 if (page_counter_read(&memcg->memory) >= memcg->low)
5283 while (memcg != root) {
5284 memcg = parent_mem_cgroup(memcg);
5286 if (memcg == root_mem_cgroup)
5289 if (page_counter_read(&memcg->memory) >= memcg->low)
5296 * mem_cgroup_try_charge - try charging a page
5297 * @page: page to charge
5298 * @mm: mm context of the victim
5299 * @gfp_mask: reclaim mode
5300 * @memcgp: charged memcg return
5302 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5303 * pages according to @gfp_mask if necessary.
5305 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5306 * Otherwise, an error code is returned.
5308 * After page->mapping has been set up, the caller must finalize the
5309 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5310 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5312 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5313 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5315 struct mem_cgroup *memcg = NULL;
5316 unsigned int nr_pages = 1;
5319 if (mem_cgroup_disabled())
5322 if (PageSwapCache(page)) {
5324 * Every swap fault against a single page tries to charge the
5325 * page, bail as early as possible. shmem_unuse() encounters
5326 * already charged pages, too. The USED bit is protected by
5327 * the page lock, which serializes swap cache removal, which
5328 * in turn serializes uncharging.
5330 VM_BUG_ON_PAGE(!PageLocked(page), page);
5331 if (page->mem_cgroup)
5334 if (do_memsw_account()) {
5335 swp_entry_t ent = { .val = page_private(page), };
5336 unsigned short id = lookup_swap_cgroup_id(ent);
5339 memcg = mem_cgroup_from_id(id);
5340 if (memcg && !css_tryget_online(&memcg->css))
5346 if (PageTransHuge(page)) {
5347 nr_pages <<= compound_order(page);
5348 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5352 memcg = get_mem_cgroup_from_mm(mm);
5354 ret = try_charge(memcg, gfp_mask, nr_pages);
5356 css_put(&memcg->css);
5363 * mem_cgroup_commit_charge - commit a page charge
5364 * @page: page to charge
5365 * @memcg: memcg to charge the page to
5366 * @lrucare: page might be on LRU already
5368 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5369 * after page->mapping has been set up. This must happen atomically
5370 * as part of the page instantiation, i.e. under the page table lock
5371 * for anonymous pages, under the page lock for page and swap cache.
5373 * In addition, the page must not be on the LRU during the commit, to
5374 * prevent racing with task migration. If it might be, use @lrucare.
5376 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5378 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5381 unsigned int nr_pages = 1;
5383 VM_BUG_ON_PAGE(!page->mapping, page);
5384 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5386 if (mem_cgroup_disabled())
5389 * Swap faults will attempt to charge the same page multiple
5390 * times. But reuse_swap_page() might have removed the page
5391 * from swapcache already, so we can't check PageSwapCache().
5396 commit_charge(page, memcg, lrucare);
5398 if (PageTransHuge(page)) {
5399 nr_pages <<= compound_order(page);
5400 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5403 local_irq_disable();
5404 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5405 memcg_check_events(memcg, page);
5408 if (do_memsw_account() && PageSwapCache(page)) {
5409 swp_entry_t entry = { .val = page_private(page) };
5411 * The swap entry might not get freed for a long time,
5412 * let's not wait for it. The page already received a
5413 * memory+swap charge, drop the swap entry duplicate.
5415 mem_cgroup_uncharge_swap(entry);
5420 * mem_cgroup_cancel_charge - cancel a page charge
5421 * @page: page to charge
5422 * @memcg: memcg to charge the page to
5424 * Cancel a charge transaction started by mem_cgroup_try_charge().
5426 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5428 unsigned int nr_pages = 1;
5430 if (mem_cgroup_disabled())
5433 * Swap faults will attempt to charge the same page multiple
5434 * times. But reuse_swap_page() might have removed the page
5435 * from swapcache already, so we can't check PageSwapCache().
5440 if (PageTransHuge(page)) {
5441 nr_pages <<= compound_order(page);
5442 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5445 cancel_charge(memcg, nr_pages);
5448 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5449 unsigned long nr_anon, unsigned long nr_file,
5450 unsigned long nr_huge, struct page *dummy_page)
5452 unsigned long nr_pages = nr_anon + nr_file;
5453 unsigned long flags;
5455 if (!mem_cgroup_is_root(memcg)) {
5456 page_counter_uncharge(&memcg->memory, nr_pages);
5457 if (do_memsw_account())
5458 page_counter_uncharge(&memcg->memsw, nr_pages);
5459 memcg_oom_recover(memcg);
5462 local_irq_save(flags);
5463 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5464 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5465 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5466 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5467 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5468 memcg_check_events(memcg, dummy_page);
5469 local_irq_restore(flags);
5471 if (!mem_cgroup_is_root(memcg))
5472 css_put_many(&memcg->css, nr_pages);
5475 static void uncharge_list(struct list_head *page_list)
5477 struct mem_cgroup *memcg = NULL;
5478 unsigned long nr_anon = 0;
5479 unsigned long nr_file = 0;
5480 unsigned long nr_huge = 0;
5481 unsigned long pgpgout = 0;
5482 struct list_head *next;
5485 next = page_list->next;
5487 unsigned int nr_pages = 1;
5489 page = list_entry(next, struct page, lru);
5490 next = page->lru.next;
5492 VM_BUG_ON_PAGE(PageLRU(page), page);
5493 VM_BUG_ON_PAGE(page_count(page), page);
5495 if (!page->mem_cgroup)
5499 * Nobody should be changing or seriously looking at
5500 * page->mem_cgroup at this point, we have fully
5501 * exclusive access to the page.
5504 if (memcg != page->mem_cgroup) {
5506 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5508 pgpgout = nr_anon = nr_file = nr_huge = 0;
5510 memcg = page->mem_cgroup;
5513 if (PageTransHuge(page)) {
5514 nr_pages <<= compound_order(page);
5515 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5516 nr_huge += nr_pages;
5520 nr_anon += nr_pages;
5522 nr_file += nr_pages;
5524 page->mem_cgroup = NULL;
5527 } while (next != page_list);
5530 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5535 * mem_cgroup_uncharge - uncharge a page
5536 * @page: page to uncharge
5538 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5539 * mem_cgroup_commit_charge().
5541 void mem_cgroup_uncharge(struct page *page)
5543 if (mem_cgroup_disabled())
5546 /* Don't touch page->lru of any random page, pre-check: */
5547 if (!page->mem_cgroup)
5550 INIT_LIST_HEAD(&page->lru);
5551 uncharge_list(&page->lru);
5555 * mem_cgroup_uncharge_list - uncharge a list of page
5556 * @page_list: list of pages to uncharge
5558 * Uncharge a list of pages previously charged with
5559 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5561 void mem_cgroup_uncharge_list(struct list_head *page_list)
5563 if (mem_cgroup_disabled())
5566 if (!list_empty(page_list))
5567 uncharge_list(page_list);
5571 * mem_cgroup_replace_page - migrate a charge to another page
5572 * @oldpage: currently charged page
5573 * @newpage: page to transfer the charge to
5575 * Migrate the charge from @oldpage to @newpage.
5577 * Both pages must be locked, @newpage->mapping must be set up.
5578 * Either or both pages might be on the LRU already.
5580 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5582 struct mem_cgroup *memcg;
5585 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5586 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5587 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5588 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5591 if (mem_cgroup_disabled())
5594 /* Page cache replacement: new page already charged? */
5595 if (newpage->mem_cgroup)
5598 /* Swapcache readahead pages can get replaced before being charged */
5599 memcg = oldpage->mem_cgroup;
5603 lock_page_lru(oldpage, &isolated);
5604 oldpage->mem_cgroup = NULL;
5605 unlock_page_lru(oldpage, isolated);
5607 commit_charge(newpage, memcg, true);
5611 * subsys_initcall() for memory controller.
5613 * Some parts like hotcpu_notifier() have to be initialized from this context
5614 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5615 * everything that doesn't depend on a specific mem_cgroup structure should
5616 * be initialized from here.
5618 static int __init mem_cgroup_init(void)
5622 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5624 for_each_possible_cpu(cpu)
5625 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5628 for_each_node(node) {
5629 struct mem_cgroup_tree_per_node *rtpn;
5632 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5633 node_online(node) ? node : NUMA_NO_NODE);
5635 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5636 struct mem_cgroup_tree_per_zone *rtpz;
5638 rtpz = &rtpn->rb_tree_per_zone[zone];
5639 rtpz->rb_root = RB_ROOT;
5640 spin_lock_init(&rtpz->lock);
5642 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5647 subsys_initcall(mem_cgroup_init);
5649 #ifdef CONFIG_MEMCG_SWAP
5651 * mem_cgroup_swapout - transfer a memsw charge to swap
5652 * @page: page whose memsw charge to transfer
5653 * @entry: swap entry to move the charge to
5655 * Transfer the memsw charge of @page to @entry.
5657 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5659 struct mem_cgroup *memcg;
5660 unsigned short oldid;
5662 VM_BUG_ON_PAGE(PageLRU(page), page);
5663 VM_BUG_ON_PAGE(page_count(page), page);
5665 if (!do_memsw_account())
5668 memcg = page->mem_cgroup;
5670 /* Readahead page, never charged */
5674 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5675 VM_BUG_ON_PAGE(oldid, page);
5676 mem_cgroup_swap_statistics(memcg, true);
5678 page->mem_cgroup = NULL;
5680 if (!mem_cgroup_is_root(memcg))
5681 page_counter_uncharge(&memcg->memory, 1);
5684 * Interrupts should be disabled here because the caller holds the
5685 * mapping->tree_lock lock which is taken with interrupts-off. It is
5686 * important here to have the interrupts disabled because it is the
5687 * only synchronisation we have for udpating the per-CPU variables.
5689 VM_BUG_ON(!irqs_disabled());
5690 mem_cgroup_charge_statistics(memcg, page, -1);
5691 memcg_check_events(memcg, page);
5695 * mem_cgroup_uncharge_swap - uncharge a swap entry
5696 * @entry: swap entry to uncharge
5698 * Drop the memsw charge associated with @entry.
5700 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5702 struct mem_cgroup *memcg;
5705 if (!do_memsw_account())
5708 id = swap_cgroup_record(entry, 0);
5710 memcg = mem_cgroup_from_id(id);
5712 if (!mem_cgroup_is_root(memcg))
5713 page_counter_uncharge(&memcg->memsw, 1);
5714 mem_cgroup_swap_statistics(memcg, false);
5715 css_put(&memcg->css);
5720 /* for remember boot option*/
5721 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5722 static int really_do_swap_account __initdata = 1;
5724 static int really_do_swap_account __initdata;
5727 static int __init enable_swap_account(char *s)
5729 if (!strcmp(s, "1"))
5730 really_do_swap_account = 1;
5731 else if (!strcmp(s, "0"))
5732 really_do_swap_account = 0;
5735 __setup("swapaccount=", enable_swap_account);
5737 static struct cftype memsw_cgroup_files[] = {
5739 .name = "memsw.usage_in_bytes",
5740 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5741 .read_u64 = mem_cgroup_read_u64,
5744 .name = "memsw.max_usage_in_bytes",
5745 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5746 .write = mem_cgroup_reset,
5747 .read_u64 = mem_cgroup_read_u64,
5750 .name = "memsw.limit_in_bytes",
5751 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5752 .write = mem_cgroup_write,
5753 .read_u64 = mem_cgroup_read_u64,
5756 .name = "memsw.failcnt",
5757 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5758 .write = mem_cgroup_reset,
5759 .read_u64 = mem_cgroup_read_u64,
5761 { }, /* terminate */
5764 static int __init mem_cgroup_swap_init(void)
5766 if (!mem_cgroup_disabled() && really_do_swap_account) {
5767 do_swap_account = 1;
5768 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5769 memsw_cgroup_files));
5773 subsys_initcall(mem_cgroup_swap_init);
5775 #endif /* CONFIG_MEMCG_SWAP */