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 #define MEM_CGROUP_RECLAIM_RETRIES 5
80 static struct mem_cgroup *root_mem_cgroup __read_mostly;
81 struct cgroup_subsys_state *mem_cgroup_root_css __read_mostly;
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 static const char * const mem_cgroup_stat_names[] = {
100 static const char * const mem_cgroup_events_names[] = {
107 static const char * const mem_cgroup_lru_names[] = {
115 #define THRESHOLDS_EVENTS_TARGET 128
116 #define SOFTLIMIT_EVENTS_TARGET 1024
117 #define NUMAINFO_EVENTS_TARGET 1024
120 * Cgroups above their limits are maintained in a RB-Tree, independent of
121 * their hierarchy representation
124 struct mem_cgroup_tree_per_zone {
125 struct rb_root rb_root;
129 struct mem_cgroup_tree_per_node {
130 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
133 struct mem_cgroup_tree {
134 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
137 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
140 struct mem_cgroup_eventfd_list {
141 struct list_head list;
142 struct eventfd_ctx *eventfd;
146 * cgroup_event represents events which userspace want to receive.
148 struct mem_cgroup_event {
150 * memcg which the event belongs to.
152 struct mem_cgroup *memcg;
154 * eventfd to signal userspace about the event.
156 struct eventfd_ctx *eventfd;
158 * Each of these stored in a list by the cgroup.
160 struct list_head list;
162 * register_event() callback will be used to add new userspace
163 * waiter for changes related to this event. Use eventfd_signal()
164 * on eventfd to send notification to userspace.
166 int (*register_event)(struct mem_cgroup *memcg,
167 struct eventfd_ctx *eventfd, const char *args);
169 * unregister_event() callback will be called when userspace closes
170 * the eventfd or on cgroup removing. This callback must be set,
171 * if you want provide notification functionality.
173 void (*unregister_event)(struct mem_cgroup *memcg,
174 struct eventfd_ctx *eventfd);
176 * All fields below needed to unregister event when
177 * userspace closes eventfd.
180 wait_queue_head_t *wqh;
182 struct work_struct remove;
185 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
186 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
188 /* Stuffs for move charges at task migration. */
190 * Types of charges to be moved.
192 #define MOVE_ANON 0x1U
193 #define MOVE_FILE 0x2U
194 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
196 /* "mc" and its members are protected by cgroup_mutex */
197 static struct move_charge_struct {
198 spinlock_t lock; /* for from, to */
199 struct mm_struct *mm;
200 struct mem_cgroup *from;
201 struct mem_cgroup *to;
203 unsigned long precharge;
204 unsigned long moved_charge;
205 unsigned long moved_swap;
206 struct task_struct *moving_task; /* a task moving charges */
207 wait_queue_head_t waitq; /* a waitq for other context */
209 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
210 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
214 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
215 * limit reclaim to prevent infinite loops, if they ever occur.
217 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
218 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
221 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
222 MEM_CGROUP_CHARGE_TYPE_ANON,
223 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
224 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
228 /* for encoding cft->private value on file */
236 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
237 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
238 #define MEMFILE_ATTR(val) ((val) & 0xffff)
239 /* Used for OOM nofiier */
240 #define OOM_CONTROL (0)
243 * The memcg_create_mutex will be held whenever a new cgroup is created.
244 * As a consequence, any change that needs to protect against new child cgroups
245 * appearing has to hold it as well.
247 static DEFINE_MUTEX(memcg_create_mutex);
249 /* Some nice accessors for the vmpressure. */
250 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
253 memcg = root_mem_cgroup;
254 return &memcg->vmpressure;
257 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
259 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
262 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
264 return (memcg == root_mem_cgroup);
268 * We restrict the id in the range of [1, 65535], so it can fit into
271 #define MEM_CGROUP_ID_MAX USHRT_MAX
273 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
275 return memcg->css.id;
279 * A helper function to get mem_cgroup from ID. must be called under
280 * rcu_read_lock(). The caller is responsible for calling
281 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
282 * refcnt from swap can be called against removed memcg.)
284 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
286 struct cgroup_subsys_state *css;
288 css = css_from_id(id, &memory_cgrp_subsys);
289 return mem_cgroup_from_css(css);
292 /* Writing them here to avoid exposing memcg's inner layout */
293 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
295 void sock_update_memcg(struct sock *sk)
297 if (mem_cgroup_sockets_enabled) {
298 struct mem_cgroup *memcg;
299 struct cg_proto *cg_proto;
301 BUG_ON(!sk->sk_prot->proto_cgroup);
303 /* Socket cloning can throw us here with sk_cgrp already
304 * filled. It won't however, necessarily happen from
305 * process context. So the test for root memcg given
306 * the current task's memcg won't help us in this case.
308 * Respecting the original socket's memcg is a better
309 * decision in this case.
312 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
313 css_get(&sk->sk_cgrp->memcg->css);
318 memcg = mem_cgroup_from_task(current);
319 cg_proto = sk->sk_prot->proto_cgroup(memcg);
320 if (cg_proto && test_bit(MEMCG_SOCK_ACTIVE, &cg_proto->flags) &&
321 css_tryget_online(&memcg->css)) {
322 sk->sk_cgrp = cg_proto;
327 EXPORT_SYMBOL(sock_update_memcg);
329 void sock_release_memcg(struct sock *sk)
331 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
332 struct mem_cgroup *memcg;
333 WARN_ON(!sk->sk_cgrp->memcg);
334 memcg = sk->sk_cgrp->memcg;
335 css_put(&sk->sk_cgrp->memcg->css);
339 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
341 if (!memcg || mem_cgroup_is_root(memcg))
344 return &memcg->tcp_mem;
346 EXPORT_SYMBOL(tcp_proto_cgroup);
350 #ifdef CONFIG_MEMCG_KMEM
352 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
353 * The main reason for not using cgroup id for this:
354 * this works better in sparse environments, where we have a lot of memcgs,
355 * but only a few kmem-limited. Or also, if we have, for instance, 200
356 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
357 * 200 entry array for that.
359 * The current size of the caches array is stored in memcg_nr_cache_ids. It
360 * will double each time we have to increase it.
362 static DEFINE_IDA(memcg_cache_ida);
363 int memcg_nr_cache_ids;
365 /* Protects memcg_nr_cache_ids */
366 static DECLARE_RWSEM(memcg_cache_ids_sem);
368 void memcg_get_cache_ids(void)
370 down_read(&memcg_cache_ids_sem);
373 void memcg_put_cache_ids(void)
375 up_read(&memcg_cache_ids_sem);
379 * MIN_SIZE is different than 1, because we would like to avoid going through
380 * the alloc/free process all the time. In a small machine, 4 kmem-limited
381 * cgroups is a reasonable guess. In the future, it could be a parameter or
382 * tunable, but that is strictly not necessary.
384 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
385 * this constant directly from cgroup, but it is understandable that this is
386 * better kept as an internal representation in cgroup.c. In any case, the
387 * cgrp_id space is not getting any smaller, and we don't have to necessarily
388 * increase ours as well if it increases.
390 #define MEMCG_CACHES_MIN_SIZE 4
391 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
394 * A lot of the calls to the cache allocation functions are expected to be
395 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
396 * conditional to this static branch, we'll have to allow modules that does
397 * kmem_cache_alloc and the such to see this symbol as well
399 struct static_key memcg_kmem_enabled_key;
400 EXPORT_SYMBOL(memcg_kmem_enabled_key);
402 #endif /* CONFIG_MEMCG_KMEM */
404 static struct mem_cgroup_per_zone *
405 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
407 int nid = zone_to_nid(zone);
408 int zid = zone_idx(zone);
410 return &memcg->nodeinfo[nid]->zoneinfo[zid];
414 * mem_cgroup_css_from_page - css of the memcg associated with a page
415 * @page: page of interest
417 * If memcg is bound to the default hierarchy, css of the memcg associated
418 * with @page is returned. The returned css remains associated with @page
419 * until it is released.
421 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
424 * XXX: The above description of behavior on the default hierarchy isn't
425 * strictly true yet as replace_page_cache_page() can modify the
426 * association before @page is released even on the default hierarchy;
427 * however, the current and planned usages don't mix the the two functions
428 * and replace_page_cache_page() will soon be updated to make the invariant
431 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
433 struct mem_cgroup *memcg;
437 memcg = page->mem_cgroup;
439 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
440 memcg = root_mem_cgroup;
447 * page_cgroup_ino - return inode number of the memcg a page is charged to
450 * Look up the closest online ancestor of the memory cgroup @page is charged to
451 * and return its inode number or 0 if @page is not charged to any cgroup. It
452 * is safe to call this function without holding a reference to @page.
454 * Note, this function is inherently racy, because there is nothing to prevent
455 * the cgroup inode from getting torn down and potentially reallocated a moment
456 * after page_cgroup_ino() returns, so it only should be used by callers that
457 * do not care (such as procfs interfaces).
459 ino_t page_cgroup_ino(struct page *page)
461 struct mem_cgroup *memcg;
462 unsigned long ino = 0;
465 memcg = READ_ONCE(page->mem_cgroup);
466 while (memcg && !(memcg->css.flags & CSS_ONLINE))
467 memcg = parent_mem_cgroup(memcg);
469 ino = cgroup_ino(memcg->css.cgroup);
474 static struct mem_cgroup_per_zone *
475 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
477 int nid = page_to_nid(page);
478 int zid = page_zonenum(page);
480 return &memcg->nodeinfo[nid]->zoneinfo[zid];
483 static struct mem_cgroup_tree_per_zone *
484 soft_limit_tree_node_zone(int nid, int zid)
486 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
489 static struct mem_cgroup_tree_per_zone *
490 soft_limit_tree_from_page(struct page *page)
492 int nid = page_to_nid(page);
493 int zid = page_zonenum(page);
495 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
498 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
499 struct mem_cgroup_tree_per_zone *mctz,
500 unsigned long new_usage_in_excess)
502 struct rb_node **p = &mctz->rb_root.rb_node;
503 struct rb_node *parent = NULL;
504 struct mem_cgroup_per_zone *mz_node;
509 mz->usage_in_excess = new_usage_in_excess;
510 if (!mz->usage_in_excess)
514 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
516 if (mz->usage_in_excess < mz_node->usage_in_excess)
519 * We can't avoid mem cgroups that are over their soft
520 * limit by the same amount
522 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
525 rb_link_node(&mz->tree_node, parent, p);
526 rb_insert_color(&mz->tree_node, &mctz->rb_root);
530 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
531 struct mem_cgroup_tree_per_zone *mctz)
535 rb_erase(&mz->tree_node, &mctz->rb_root);
539 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
540 struct mem_cgroup_tree_per_zone *mctz)
544 spin_lock_irqsave(&mctz->lock, flags);
545 __mem_cgroup_remove_exceeded(mz, mctz);
546 spin_unlock_irqrestore(&mctz->lock, flags);
549 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
551 unsigned long nr_pages = page_counter_read(&memcg->memory);
552 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
553 unsigned long excess = 0;
555 if (nr_pages > soft_limit)
556 excess = nr_pages - soft_limit;
561 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
563 unsigned long excess;
564 struct mem_cgroup_per_zone *mz;
565 struct mem_cgroup_tree_per_zone *mctz;
567 mctz = soft_limit_tree_from_page(page);
569 * Necessary to update all ancestors when hierarchy is used.
570 * because their event counter is not touched.
572 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
573 mz = mem_cgroup_page_zoneinfo(memcg, page);
574 excess = soft_limit_excess(memcg);
576 * We have to update the tree if mz is on RB-tree or
577 * mem is over its softlimit.
579 if (excess || mz->on_tree) {
582 spin_lock_irqsave(&mctz->lock, flags);
583 /* if on-tree, remove it */
585 __mem_cgroup_remove_exceeded(mz, mctz);
587 * Insert again. mz->usage_in_excess will be updated.
588 * If excess is 0, no tree ops.
590 __mem_cgroup_insert_exceeded(mz, mctz, excess);
591 spin_unlock_irqrestore(&mctz->lock, flags);
596 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
598 struct mem_cgroup_tree_per_zone *mctz;
599 struct mem_cgroup_per_zone *mz;
603 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
604 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
605 mctz = soft_limit_tree_node_zone(nid, zid);
606 mem_cgroup_remove_exceeded(mz, mctz);
611 static struct mem_cgroup_per_zone *
612 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
614 struct rb_node *rightmost = NULL;
615 struct mem_cgroup_per_zone *mz;
619 rightmost = rb_last(&mctz->rb_root);
621 goto done; /* Nothing to reclaim from */
623 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
625 * Remove the node now but someone else can add it back,
626 * we will to add it back at the end of reclaim to its correct
627 * position in the tree.
629 __mem_cgroup_remove_exceeded(mz, mctz);
630 if (!soft_limit_excess(mz->memcg) ||
631 !css_tryget_online(&mz->memcg->css))
637 static struct mem_cgroup_per_zone *
638 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
640 struct mem_cgroup_per_zone *mz;
642 spin_lock_irq(&mctz->lock);
643 mz = __mem_cgroup_largest_soft_limit_node(mctz);
644 spin_unlock_irq(&mctz->lock);
649 * Return page count for single (non recursive) @memcg.
651 * Implementation Note: reading percpu statistics for memcg.
653 * Both of vmstat[] and percpu_counter has threshold and do periodic
654 * synchronization to implement "quick" read. There are trade-off between
655 * reading cost and precision of value. Then, we may have a chance to implement
656 * a periodic synchronization of counter in memcg's counter.
658 * But this _read() function is used for user interface now. The user accounts
659 * memory usage by memory cgroup and he _always_ requires exact value because
660 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
661 * have to visit all online cpus and make sum. So, for now, unnecessary
662 * synchronization is not implemented. (just implemented for cpu hotplug)
664 * If there are kernel internal actions which can make use of some not-exact
665 * value, and reading all cpu value can be performance bottleneck in some
666 * common workload, threshold and synchronization as vmstat[] should be
670 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
675 /* Per-cpu values can be negative, use a signed accumulator */
676 for_each_possible_cpu(cpu)
677 val += per_cpu(memcg->stat->count[idx], cpu);
679 * Summing races with updates, so val may be negative. Avoid exposing
680 * transient negative values.
687 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
688 enum mem_cgroup_events_index idx)
690 unsigned long val = 0;
693 for_each_possible_cpu(cpu)
694 val += per_cpu(memcg->stat->events[idx], cpu);
698 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
703 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
704 * counted as CACHE even if it's on ANON LRU.
707 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
710 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
713 if (PageTransHuge(page))
714 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
717 /* pagein of a big page is an event. So, ignore page size */
719 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
721 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
722 nr_pages = -nr_pages; /* for event */
725 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
728 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
730 unsigned int lru_mask)
732 unsigned long nr = 0;
735 VM_BUG_ON((unsigned)nid >= nr_node_ids);
737 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
738 struct mem_cgroup_per_zone *mz;
742 if (!(BIT(lru) & lru_mask))
744 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
745 nr += mz->lru_size[lru];
751 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
752 unsigned int lru_mask)
754 unsigned long nr = 0;
757 for_each_node_state(nid, N_MEMORY)
758 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
762 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
763 enum mem_cgroup_events_target target)
765 unsigned long val, next;
767 val = __this_cpu_read(memcg->stat->nr_page_events);
768 next = __this_cpu_read(memcg->stat->targets[target]);
769 /* from time_after() in jiffies.h */
770 if ((long)next - (long)val < 0) {
772 case MEM_CGROUP_TARGET_THRESH:
773 next = val + THRESHOLDS_EVENTS_TARGET;
775 case MEM_CGROUP_TARGET_SOFTLIMIT:
776 next = val + SOFTLIMIT_EVENTS_TARGET;
778 case MEM_CGROUP_TARGET_NUMAINFO:
779 next = val + NUMAINFO_EVENTS_TARGET;
784 __this_cpu_write(memcg->stat->targets[target], next);
791 * Check events in order.
794 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
796 /* threshold event is triggered in finer grain than soft limit */
797 if (unlikely(mem_cgroup_event_ratelimit(memcg,
798 MEM_CGROUP_TARGET_THRESH))) {
800 bool do_numainfo __maybe_unused;
802 do_softlimit = mem_cgroup_event_ratelimit(memcg,
803 MEM_CGROUP_TARGET_SOFTLIMIT);
805 do_numainfo = mem_cgroup_event_ratelimit(memcg,
806 MEM_CGROUP_TARGET_NUMAINFO);
808 mem_cgroup_threshold(memcg);
809 if (unlikely(do_softlimit))
810 mem_cgroup_update_tree(memcg, page);
812 if (unlikely(do_numainfo))
813 atomic_inc(&memcg->numainfo_events);
818 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
821 * mm_update_next_owner() may clear mm->owner to NULL
822 * if it races with swapoff, page migration, etc.
823 * So this can be called with p == NULL.
828 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
830 EXPORT_SYMBOL(mem_cgroup_from_task);
832 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
834 struct mem_cgroup *memcg = NULL;
839 * Page cache insertions can happen withou an
840 * actual mm context, e.g. during disk probing
841 * on boot, loopback IO, acct() writes etc.
844 memcg = root_mem_cgroup;
846 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
847 if (unlikely(!memcg))
848 memcg = root_mem_cgroup;
850 } while (!css_tryget_online(&memcg->css));
856 * mem_cgroup_iter - iterate over memory cgroup hierarchy
857 * @root: hierarchy root
858 * @prev: previously returned memcg, NULL on first invocation
859 * @reclaim: cookie for shared reclaim walks, NULL for full walks
861 * Returns references to children of the hierarchy below @root, or
862 * @root itself, or %NULL after a full round-trip.
864 * Caller must pass the return value in @prev on subsequent
865 * invocations for reference counting, or use mem_cgroup_iter_break()
866 * to cancel a hierarchy walk before the round-trip is complete.
868 * Reclaimers can specify a zone and a priority level in @reclaim to
869 * divide up the memcgs in the hierarchy among all concurrent
870 * reclaimers operating on the same zone and priority.
872 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
873 struct mem_cgroup *prev,
874 struct mem_cgroup_reclaim_cookie *reclaim)
876 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
877 struct cgroup_subsys_state *css = NULL;
878 struct mem_cgroup *memcg = NULL;
879 struct mem_cgroup *pos = NULL;
881 if (mem_cgroup_disabled())
885 root = root_mem_cgroup;
887 if (prev && !reclaim)
890 if (!root->use_hierarchy && root != root_mem_cgroup) {
899 struct mem_cgroup_per_zone *mz;
901 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
902 iter = &mz->iter[reclaim->priority];
904 if (prev && reclaim->generation != iter->generation)
908 pos = READ_ONCE(iter->position);
909 if (!pos || css_tryget(&pos->css))
912 * css reference reached zero, so iter->position will
913 * be cleared by ->css_released. However, we should not
914 * rely on this happening soon, because ->css_released
915 * is called from a work queue, and by busy-waiting we
916 * might block it. So we clear iter->position right
919 (void)cmpxchg(&iter->position, pos, NULL);
927 css = css_next_descendant_pre(css, &root->css);
930 * Reclaimers share the hierarchy walk, and a
931 * new one might jump in right at the end of
932 * the hierarchy - make sure they see at least
933 * one group and restart from the beginning.
941 * Verify the css and acquire a reference. The root
942 * is provided by the caller, so we know it's alive
943 * and kicking, and don't take an extra reference.
945 memcg = mem_cgroup_from_css(css);
947 if (css == &root->css)
950 if (css_tryget(css)) {
952 * Make sure the memcg is initialized:
953 * mem_cgroup_css_online() orders the the
954 * initialization against setting the flag.
956 if (smp_load_acquire(&memcg->initialized))
967 * The position could have already been updated by a competing
968 * thread, so check that the value hasn't changed since we read
969 * it to avoid reclaiming from the same cgroup twice.
971 (void)cmpxchg(&iter->position, pos, memcg);
979 reclaim->generation = iter->generation;
985 if (prev && prev != root)
992 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
993 * @root: hierarchy root
994 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
996 void mem_cgroup_iter_break(struct mem_cgroup *root,
997 struct mem_cgroup *prev)
1000 root = root_mem_cgroup;
1001 if (prev && prev != root)
1002 css_put(&prev->css);
1005 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1007 struct mem_cgroup *memcg = dead_memcg;
1008 struct mem_cgroup_reclaim_iter *iter;
1009 struct mem_cgroup_per_zone *mz;
1013 while ((memcg = parent_mem_cgroup(memcg))) {
1014 for_each_node(nid) {
1015 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1016 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
1017 for (i = 0; i <= DEF_PRIORITY; i++) {
1018 iter = &mz->iter[i];
1019 cmpxchg(&iter->position,
1028 * Iteration constructs for visiting all cgroups (under a tree). If
1029 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1030 * be used for reference counting.
1032 #define for_each_mem_cgroup_tree(iter, root) \
1033 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1035 iter = mem_cgroup_iter(root, iter, NULL))
1037 #define for_each_mem_cgroup(iter) \
1038 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1040 iter = mem_cgroup_iter(NULL, iter, NULL))
1043 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1044 * @zone: zone of the wanted lruvec
1045 * @memcg: memcg of the wanted lruvec
1047 * Returns the lru list vector holding pages for the given @zone and
1048 * @mem. This can be the global zone lruvec, if the memory controller
1051 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1052 struct mem_cgroup *memcg)
1054 struct mem_cgroup_per_zone *mz;
1055 struct lruvec *lruvec;
1057 if (mem_cgroup_disabled()) {
1058 lruvec = &zone->lruvec;
1062 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1063 lruvec = &mz->lruvec;
1066 * Since a node can be onlined after the mem_cgroup was created,
1067 * we have to be prepared to initialize lruvec->zone here;
1068 * and if offlined then reonlined, we need to reinitialize it.
1070 if (unlikely(lruvec->zone != zone))
1071 lruvec->zone = zone;
1076 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1078 * @zone: zone of the page
1080 * This function is only safe when following the LRU page isolation
1081 * and putback protocol: the LRU lock must be held, and the page must
1082 * either be PageLRU() or the caller must have isolated/allocated it.
1084 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1086 struct mem_cgroup_per_zone *mz;
1087 struct mem_cgroup *memcg;
1088 struct lruvec *lruvec;
1090 if (mem_cgroup_disabled()) {
1091 lruvec = &zone->lruvec;
1095 memcg = page->mem_cgroup;
1097 * Swapcache readahead pages are added to the LRU - and
1098 * possibly migrated - before they are charged.
1101 memcg = root_mem_cgroup;
1103 mz = mem_cgroup_page_zoneinfo(memcg, page);
1104 lruvec = &mz->lruvec;
1107 * Since a node can be onlined after the mem_cgroup was created,
1108 * we have to be prepared to initialize lruvec->zone here;
1109 * and if offlined then reonlined, we need to reinitialize it.
1111 if (unlikely(lruvec->zone != zone))
1112 lruvec->zone = zone;
1117 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1118 * @lruvec: mem_cgroup per zone lru vector
1119 * @lru: index of lru list the page is sitting on
1120 * @nr_pages: positive when adding or negative when removing
1122 * This function must be called when a page is added to or removed from an
1125 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1128 struct mem_cgroup_per_zone *mz;
1129 unsigned long *lru_size;
1131 if (mem_cgroup_disabled())
1134 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1135 lru_size = mz->lru_size + lru;
1136 *lru_size += nr_pages;
1137 VM_BUG_ON((long)(*lru_size) < 0);
1140 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1142 struct mem_cgroup *task_memcg;
1143 struct task_struct *p;
1146 p = find_lock_task_mm(task);
1148 task_memcg = get_mem_cgroup_from_mm(p->mm);
1152 * All threads may have already detached their mm's, but the oom
1153 * killer still needs to detect if they have already been oom
1154 * killed to prevent needlessly killing additional tasks.
1157 task_memcg = mem_cgroup_from_task(task);
1158 css_get(&task_memcg->css);
1161 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1162 css_put(&task_memcg->css);
1166 #define mem_cgroup_from_counter(counter, member) \
1167 container_of(counter, struct mem_cgroup, member)
1170 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1171 * @memcg: the memory cgroup
1173 * Returns the maximum amount of memory @mem can be charged with, in
1176 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1178 unsigned long margin = 0;
1179 unsigned long count;
1180 unsigned long limit;
1182 count = page_counter_read(&memcg->memory);
1183 limit = READ_ONCE(memcg->memory.limit);
1185 margin = limit - count;
1187 if (do_swap_account) {
1188 count = page_counter_read(&memcg->memsw);
1189 limit = READ_ONCE(memcg->memsw.limit);
1191 margin = min(margin, limit - count);
1198 * A routine for checking "mem" is under move_account() or not.
1200 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1201 * moving cgroups. This is for waiting at high-memory pressure
1204 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1206 struct mem_cgroup *from;
1207 struct mem_cgroup *to;
1210 * Unlike task_move routines, we access mc.to, mc.from not under
1211 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1213 spin_lock(&mc.lock);
1219 ret = mem_cgroup_is_descendant(from, memcg) ||
1220 mem_cgroup_is_descendant(to, memcg);
1222 spin_unlock(&mc.lock);
1226 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1228 if (mc.moving_task && current != mc.moving_task) {
1229 if (mem_cgroup_under_move(memcg)) {
1231 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1232 /* moving charge context might have finished. */
1235 finish_wait(&mc.waitq, &wait);
1242 #define K(x) ((x) << (PAGE_SHIFT-10))
1244 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1245 * @memcg: The memory cgroup that went over limit
1246 * @p: Task that is going to be killed
1248 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1251 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1253 /* oom_info_lock ensures that parallel ooms do not interleave */
1254 static DEFINE_MUTEX(oom_info_lock);
1255 struct mem_cgroup *iter;
1258 mutex_lock(&oom_info_lock);
1262 pr_info("Task in ");
1263 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1264 pr_cont(" killed as a result of limit of ");
1266 pr_info("Memory limit reached of cgroup ");
1269 pr_cont_cgroup_path(memcg->css.cgroup);
1274 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1275 K((u64)page_counter_read(&memcg->memory)),
1276 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1277 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1278 K((u64)page_counter_read(&memcg->memsw)),
1279 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1280 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1281 K((u64)page_counter_read(&memcg->kmem)),
1282 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1284 for_each_mem_cgroup_tree(iter, memcg) {
1285 pr_info("Memory cgroup stats for ");
1286 pr_cont_cgroup_path(iter->css.cgroup);
1289 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1290 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1292 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1293 K(mem_cgroup_read_stat(iter, i)));
1296 for (i = 0; i < NR_LRU_LISTS; i++)
1297 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1298 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1302 mutex_unlock(&oom_info_lock);
1306 * This function returns the number of memcg under hierarchy tree. Returns
1307 * 1(self count) if no children.
1309 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1312 struct mem_cgroup *iter;
1314 for_each_mem_cgroup_tree(iter, memcg)
1320 * Return the memory (and swap, if configured) limit for a memcg.
1322 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1324 unsigned long limit;
1326 limit = memcg->memory.limit;
1327 if (mem_cgroup_swappiness(memcg)) {
1328 unsigned long memsw_limit;
1330 memsw_limit = memcg->memsw.limit;
1331 limit = min(limit + total_swap_pages, memsw_limit);
1336 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1339 struct oom_control oc = {
1342 .gfp_mask = gfp_mask,
1345 struct mem_cgroup *iter;
1346 unsigned long chosen_points = 0;
1347 unsigned long totalpages;
1348 unsigned int points = 0;
1349 struct task_struct *chosen = NULL;
1351 mutex_lock(&oom_lock);
1354 * If current has a pending SIGKILL or is exiting, then automatically
1355 * select it. The goal is to allow it to allocate so that it may
1356 * quickly exit and free its memory.
1358 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1359 mark_oom_victim(current);
1363 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1364 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1365 for_each_mem_cgroup_tree(iter, memcg) {
1366 struct css_task_iter it;
1367 struct task_struct *task;
1369 css_task_iter_start(&iter->css, &it);
1370 while ((task = css_task_iter_next(&it))) {
1371 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1372 case OOM_SCAN_SELECT:
1374 put_task_struct(chosen);
1376 chosen_points = ULONG_MAX;
1377 get_task_struct(chosen);
1379 case OOM_SCAN_CONTINUE:
1381 case OOM_SCAN_ABORT:
1382 css_task_iter_end(&it);
1383 mem_cgroup_iter_break(memcg, iter);
1385 put_task_struct(chosen);
1390 points = oom_badness(task, memcg, NULL, totalpages);
1391 if (!points || points < chosen_points)
1393 /* Prefer thread group leaders for display purposes */
1394 if (points == chosen_points &&
1395 thread_group_leader(chosen))
1399 put_task_struct(chosen);
1401 chosen_points = points;
1402 get_task_struct(chosen);
1404 css_task_iter_end(&it);
1408 points = chosen_points * 1000 / totalpages;
1409 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1410 "Memory cgroup out of memory");
1413 mutex_unlock(&oom_lock);
1417 #if MAX_NUMNODES > 1
1420 * test_mem_cgroup_node_reclaimable
1421 * @memcg: the target memcg
1422 * @nid: the node ID to be checked.
1423 * @noswap : specify true here if the user wants flle only information.
1425 * This function returns whether the specified memcg contains any
1426 * reclaimable pages on a node. Returns true if there are any reclaimable
1427 * pages in the node.
1429 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1430 int nid, bool noswap)
1432 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1434 if (noswap || !total_swap_pages)
1436 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1443 * Always updating the nodemask is not very good - even if we have an empty
1444 * list or the wrong list here, we can start from some node and traverse all
1445 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1448 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1452 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1453 * pagein/pageout changes since the last update.
1455 if (!atomic_read(&memcg->numainfo_events))
1457 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1460 /* make a nodemask where this memcg uses memory from */
1461 memcg->scan_nodes = node_states[N_MEMORY];
1463 for_each_node_mask(nid, node_states[N_MEMORY]) {
1465 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1466 node_clear(nid, memcg->scan_nodes);
1469 atomic_set(&memcg->numainfo_events, 0);
1470 atomic_set(&memcg->numainfo_updating, 0);
1474 * Selecting a node where we start reclaim from. Because what we need is just
1475 * reducing usage counter, start from anywhere is O,K. Considering
1476 * memory reclaim from current node, there are pros. and cons.
1478 * Freeing memory from current node means freeing memory from a node which
1479 * we'll use or we've used. So, it may make LRU bad. And if several threads
1480 * hit limits, it will see a contention on a node. But freeing from remote
1481 * node means more costs for memory reclaim because of memory latency.
1483 * Now, we use round-robin. Better algorithm is welcomed.
1485 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1489 mem_cgroup_may_update_nodemask(memcg);
1490 node = memcg->last_scanned_node;
1492 node = next_node(node, memcg->scan_nodes);
1493 if (node == MAX_NUMNODES)
1494 node = first_node(memcg->scan_nodes);
1496 * We call this when we hit limit, not when pages are added to LRU.
1497 * No LRU may hold pages because all pages are UNEVICTABLE or
1498 * memcg is too small and all pages are not on LRU. In that case,
1499 * we use curret node.
1501 if (unlikely(node == MAX_NUMNODES))
1502 node = numa_node_id();
1504 memcg->last_scanned_node = node;
1508 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1514 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1517 unsigned long *total_scanned)
1519 struct mem_cgroup *victim = NULL;
1522 unsigned long excess;
1523 unsigned long nr_scanned;
1524 struct mem_cgroup_reclaim_cookie reclaim = {
1529 excess = soft_limit_excess(root_memcg);
1532 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1537 * If we have not been able to reclaim
1538 * anything, it might because there are
1539 * no reclaimable pages under this hierarchy
1544 * We want to do more targeted reclaim.
1545 * excess >> 2 is not to excessive so as to
1546 * reclaim too much, nor too less that we keep
1547 * coming back to reclaim from this cgroup
1549 if (total >= (excess >> 2) ||
1550 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1555 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1557 *total_scanned += nr_scanned;
1558 if (!soft_limit_excess(root_memcg))
1561 mem_cgroup_iter_break(root_memcg, victim);
1565 #ifdef CONFIG_LOCKDEP
1566 static struct lockdep_map memcg_oom_lock_dep_map = {
1567 .name = "memcg_oom_lock",
1571 static DEFINE_SPINLOCK(memcg_oom_lock);
1574 * Check OOM-Killer is already running under our hierarchy.
1575 * If someone is running, return false.
1577 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1579 struct mem_cgroup *iter, *failed = NULL;
1581 spin_lock(&memcg_oom_lock);
1583 for_each_mem_cgroup_tree(iter, memcg) {
1584 if (iter->oom_lock) {
1586 * this subtree of our hierarchy is already locked
1587 * so we cannot give a lock.
1590 mem_cgroup_iter_break(memcg, iter);
1593 iter->oom_lock = true;
1598 * OK, we failed to lock the whole subtree so we have
1599 * to clean up what we set up to the failing subtree
1601 for_each_mem_cgroup_tree(iter, memcg) {
1602 if (iter == failed) {
1603 mem_cgroup_iter_break(memcg, iter);
1606 iter->oom_lock = false;
1609 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1611 spin_unlock(&memcg_oom_lock);
1616 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1618 struct mem_cgroup *iter;
1620 spin_lock(&memcg_oom_lock);
1621 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1622 for_each_mem_cgroup_tree(iter, memcg)
1623 iter->oom_lock = false;
1624 spin_unlock(&memcg_oom_lock);
1627 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1629 struct mem_cgroup *iter;
1631 spin_lock(&memcg_oom_lock);
1632 for_each_mem_cgroup_tree(iter, memcg)
1634 spin_unlock(&memcg_oom_lock);
1637 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1639 struct mem_cgroup *iter;
1642 * When a new child is created while the hierarchy is under oom,
1643 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1645 spin_lock(&memcg_oom_lock);
1646 for_each_mem_cgroup_tree(iter, memcg)
1647 if (iter->under_oom > 0)
1649 spin_unlock(&memcg_oom_lock);
1652 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1654 struct oom_wait_info {
1655 struct mem_cgroup *memcg;
1659 static int memcg_oom_wake_function(wait_queue_t *wait,
1660 unsigned mode, int sync, void *arg)
1662 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1663 struct mem_cgroup *oom_wait_memcg;
1664 struct oom_wait_info *oom_wait_info;
1666 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1667 oom_wait_memcg = oom_wait_info->memcg;
1669 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1670 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1672 return autoremove_wake_function(wait, mode, sync, arg);
1675 static void memcg_oom_recover(struct mem_cgroup *memcg)
1678 * For the following lockless ->under_oom test, the only required
1679 * guarantee is that it must see the state asserted by an OOM when
1680 * this function is called as a result of userland actions
1681 * triggered by the notification of the OOM. This is trivially
1682 * achieved by invoking mem_cgroup_mark_under_oom() before
1683 * triggering notification.
1685 if (memcg && memcg->under_oom)
1686 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1689 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1691 if (!current->memcg_may_oom)
1694 * We are in the middle of the charge context here, so we
1695 * don't want to block when potentially sitting on a callstack
1696 * that holds all kinds of filesystem and mm locks.
1698 * Also, the caller may handle a failed allocation gracefully
1699 * (like optional page cache readahead) and so an OOM killer
1700 * invocation might not even be necessary.
1702 * That's why we don't do anything here except remember the
1703 * OOM context and then deal with it at the end of the page
1704 * fault when the stack is unwound, the locks are released,
1705 * and when we know whether the fault was overall successful.
1707 css_get(&memcg->css);
1708 current->memcg_in_oom = memcg;
1709 current->memcg_oom_gfp_mask = mask;
1710 current->memcg_oom_order = order;
1714 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1715 * @handle: actually kill/wait or just clean up the OOM state
1717 * This has to be called at the end of a page fault if the memcg OOM
1718 * handler was enabled.
1720 * Memcg supports userspace OOM handling where failed allocations must
1721 * sleep on a waitqueue until the userspace task resolves the
1722 * situation. Sleeping directly in the charge context with all kinds
1723 * of locks held is not a good idea, instead we remember an OOM state
1724 * in the task and mem_cgroup_oom_synchronize() has to be called at
1725 * the end of the page fault to complete the OOM handling.
1727 * Returns %true if an ongoing memcg OOM situation was detected and
1728 * completed, %false otherwise.
1730 bool mem_cgroup_oom_synchronize(bool handle)
1732 struct mem_cgroup *memcg = current->memcg_in_oom;
1733 struct oom_wait_info owait;
1736 /* OOM is global, do not handle */
1740 if (!handle || oom_killer_disabled)
1743 owait.memcg = memcg;
1744 owait.wait.flags = 0;
1745 owait.wait.func = memcg_oom_wake_function;
1746 owait.wait.private = current;
1747 INIT_LIST_HEAD(&owait.wait.task_list);
1749 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1750 mem_cgroup_mark_under_oom(memcg);
1752 locked = mem_cgroup_oom_trylock(memcg);
1755 mem_cgroup_oom_notify(memcg);
1757 if (locked && !memcg->oom_kill_disable) {
1758 mem_cgroup_unmark_under_oom(memcg);
1759 finish_wait(&memcg_oom_waitq, &owait.wait);
1760 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1761 current->memcg_oom_order);
1764 mem_cgroup_unmark_under_oom(memcg);
1765 finish_wait(&memcg_oom_waitq, &owait.wait);
1769 mem_cgroup_oom_unlock(memcg);
1771 * There is no guarantee that an OOM-lock contender
1772 * sees the wakeups triggered by the OOM kill
1773 * uncharges. Wake any sleepers explicitely.
1775 memcg_oom_recover(memcg);
1778 current->memcg_in_oom = NULL;
1779 css_put(&memcg->css);
1784 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1785 * @page: page that is going to change accounted state
1787 * This function must mark the beginning of an accounted page state
1788 * change to prevent double accounting when the page is concurrently
1789 * being moved to another memcg:
1791 * memcg = mem_cgroup_begin_page_stat(page);
1792 * if (TestClearPageState(page))
1793 * mem_cgroup_update_page_stat(memcg, state, -1);
1794 * mem_cgroup_end_page_stat(memcg);
1796 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1798 struct mem_cgroup *memcg;
1799 unsigned long flags;
1802 * The RCU lock is held throughout the transaction. The fast
1803 * path can get away without acquiring the memcg->move_lock
1804 * because page moving starts with an RCU grace period.
1806 * The RCU lock also protects the memcg from being freed when
1807 * the page state that is going to change is the only thing
1808 * preventing the page from being uncharged.
1809 * E.g. end-writeback clearing PageWriteback(), which allows
1810 * migration to go ahead and uncharge the page before the
1811 * account transaction might be complete.
1815 if (mem_cgroup_disabled())
1818 memcg = page->mem_cgroup;
1819 if (unlikely(!memcg))
1822 if (atomic_read(&memcg->moving_account) <= 0)
1825 spin_lock_irqsave(&memcg->move_lock, flags);
1826 if (memcg != page->mem_cgroup) {
1827 spin_unlock_irqrestore(&memcg->move_lock, flags);
1832 * When charge migration first begins, we can have locked and
1833 * unlocked page stat updates happening concurrently. Track
1834 * the task who has the lock for mem_cgroup_end_page_stat().
1836 memcg->move_lock_task = current;
1837 memcg->move_lock_flags = flags;
1841 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1844 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1845 * @memcg: the memcg that was accounted against
1847 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1849 if (memcg && memcg->move_lock_task == current) {
1850 unsigned long flags = memcg->move_lock_flags;
1852 memcg->move_lock_task = NULL;
1853 memcg->move_lock_flags = 0;
1855 spin_unlock_irqrestore(&memcg->move_lock, flags);
1860 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1863 * size of first charge trial. "32" comes from vmscan.c's magic value.
1864 * TODO: maybe necessary to use big numbers in big irons.
1866 #define CHARGE_BATCH 32U
1867 struct memcg_stock_pcp {
1868 struct mem_cgroup *cached; /* this never be root cgroup */
1869 unsigned int nr_pages;
1870 struct work_struct work;
1871 unsigned long flags;
1872 #define FLUSHING_CACHED_CHARGE 0
1874 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1875 static DEFINE_MUTEX(percpu_charge_mutex);
1878 * consume_stock: Try to consume stocked charge on this cpu.
1879 * @memcg: memcg to consume from.
1880 * @nr_pages: how many pages to charge.
1882 * The charges will only happen if @memcg matches the current cpu's memcg
1883 * stock, and at least @nr_pages are available in that stock. Failure to
1884 * service an allocation will refill the stock.
1886 * returns true if successful, false otherwise.
1888 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1890 struct memcg_stock_pcp *stock;
1893 if (nr_pages > CHARGE_BATCH)
1896 stock = &get_cpu_var(memcg_stock);
1897 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1898 stock->nr_pages -= nr_pages;
1901 put_cpu_var(memcg_stock);
1906 * Returns stocks cached in percpu and reset cached information.
1908 static void drain_stock(struct memcg_stock_pcp *stock)
1910 struct mem_cgroup *old = stock->cached;
1912 if (stock->nr_pages) {
1913 page_counter_uncharge(&old->memory, stock->nr_pages);
1914 if (do_swap_account)
1915 page_counter_uncharge(&old->memsw, stock->nr_pages);
1916 css_put_many(&old->css, stock->nr_pages);
1917 stock->nr_pages = 0;
1919 stock->cached = NULL;
1923 * This must be called under preempt disabled or must be called by
1924 * a thread which is pinned to local cpu.
1926 static void drain_local_stock(struct work_struct *dummy)
1928 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1930 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1934 * Cache charges(val) to local per_cpu area.
1935 * This will be consumed by consume_stock() function, later.
1937 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1939 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1941 if (stock->cached != memcg) { /* reset if necessary */
1943 stock->cached = memcg;
1945 stock->nr_pages += nr_pages;
1946 put_cpu_var(memcg_stock);
1950 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1951 * of the hierarchy under it.
1953 static void drain_all_stock(struct mem_cgroup *root_memcg)
1957 /* If someone's already draining, avoid adding running more workers. */
1958 if (!mutex_trylock(&percpu_charge_mutex))
1960 /* Notify other cpus that system-wide "drain" is running */
1963 for_each_online_cpu(cpu) {
1964 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1965 struct mem_cgroup *memcg;
1967 memcg = stock->cached;
1968 if (!memcg || !stock->nr_pages)
1970 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1972 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1974 drain_local_stock(&stock->work);
1976 schedule_work_on(cpu, &stock->work);
1981 mutex_unlock(&percpu_charge_mutex);
1984 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1985 unsigned long action,
1988 int cpu = (unsigned long)hcpu;
1989 struct memcg_stock_pcp *stock;
1991 if (action == CPU_ONLINE)
1994 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1997 stock = &per_cpu(memcg_stock, cpu);
2003 * Scheduled by try_charge() to be executed from the userland return path
2004 * and reclaims memory over the high limit.
2006 void mem_cgroup_handle_over_high(void)
2008 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2009 struct mem_cgroup *memcg, *pos;
2011 if (likely(!nr_pages))
2014 pos = memcg = get_mem_cgroup_from_mm(current->mm);
2017 if (page_counter_read(&pos->memory) <= pos->high)
2019 mem_cgroup_events(pos, MEMCG_HIGH, 1);
2020 try_to_free_mem_cgroup_pages(pos, nr_pages, GFP_KERNEL, true);
2021 } while ((pos = parent_mem_cgroup(pos)));
2023 css_put(&memcg->css);
2024 current->memcg_nr_pages_over_high = 0;
2027 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2028 unsigned int nr_pages)
2030 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2031 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2032 struct mem_cgroup *mem_over_limit;
2033 struct page_counter *counter;
2034 unsigned long nr_reclaimed;
2035 bool may_swap = true;
2036 bool drained = false;
2038 if (mem_cgroup_is_root(memcg))
2041 if (consume_stock(memcg, nr_pages))
2044 if (!do_swap_account ||
2045 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2046 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2048 if (do_swap_account)
2049 page_counter_uncharge(&memcg->memsw, batch);
2050 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2052 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2056 if (batch > nr_pages) {
2062 * Unlike in global OOM situations, memcg is not in a physical
2063 * memory shortage. Allow dying and OOM-killed tasks to
2064 * bypass the last charges so that they can exit quickly and
2065 * free their memory.
2067 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2068 fatal_signal_pending(current) ||
2069 current->flags & PF_EXITING))
2072 if (unlikely(task_in_memcg_oom(current)))
2075 if (!gfpflags_allow_blocking(gfp_mask))
2078 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2080 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2081 gfp_mask, may_swap);
2083 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2087 drain_all_stock(mem_over_limit);
2092 if (gfp_mask & __GFP_NORETRY)
2095 * Even though the limit is exceeded at this point, reclaim
2096 * may have been able to free some pages. Retry the charge
2097 * before killing the task.
2099 * Only for regular pages, though: huge pages are rather
2100 * unlikely to succeed so close to the limit, and we fall back
2101 * to regular pages anyway in case of failure.
2103 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2106 * At task move, charge accounts can be doubly counted. So, it's
2107 * better to wait until the end of task_move if something is going on.
2109 if (mem_cgroup_wait_acct_move(mem_over_limit))
2115 if (gfp_mask & __GFP_NOFAIL)
2118 if (fatal_signal_pending(current))
2121 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2123 mem_cgroup_oom(mem_over_limit, gfp_mask,
2124 get_order(nr_pages * PAGE_SIZE));
2126 if (!(gfp_mask & __GFP_NOFAIL))
2130 * The allocation either can't fail or will lead to more memory
2131 * being freed very soon. Allow memory usage go over the limit
2132 * temporarily by force charging it.
2134 page_counter_charge(&memcg->memory, nr_pages);
2135 if (do_swap_account)
2136 page_counter_charge(&memcg->memsw, nr_pages);
2137 css_get_many(&memcg->css, nr_pages);
2142 css_get_many(&memcg->css, batch);
2143 if (batch > nr_pages)
2144 refill_stock(memcg, batch - nr_pages);
2147 * If the hierarchy is above the normal consumption range, schedule
2148 * reclaim on returning to userland. We can perform reclaim here
2149 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2150 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2151 * not recorded as it most likely matches current's and won't
2152 * change in the meantime. As high limit is checked again before
2153 * reclaim, the cost of mismatch is negligible.
2156 if (page_counter_read(&memcg->memory) > memcg->high) {
2157 current->memcg_nr_pages_over_high += batch;
2158 set_notify_resume(current);
2161 } while ((memcg = parent_mem_cgroup(memcg)));
2166 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2168 if (mem_cgroup_is_root(memcg))
2171 page_counter_uncharge(&memcg->memory, nr_pages);
2172 if (do_swap_account)
2173 page_counter_uncharge(&memcg->memsw, nr_pages);
2175 css_put_many(&memcg->css, nr_pages);
2178 static void lock_page_lru(struct page *page, int *isolated)
2180 struct zone *zone = page_zone(page);
2182 spin_lock_irq(&zone->lru_lock);
2183 if (PageLRU(page)) {
2184 struct lruvec *lruvec;
2186 lruvec = mem_cgroup_page_lruvec(page, zone);
2188 del_page_from_lru_list(page, lruvec, page_lru(page));
2194 static void unlock_page_lru(struct page *page, int isolated)
2196 struct zone *zone = page_zone(page);
2199 struct lruvec *lruvec;
2201 lruvec = mem_cgroup_page_lruvec(page, zone);
2202 VM_BUG_ON_PAGE(PageLRU(page), page);
2204 add_page_to_lru_list(page, lruvec, page_lru(page));
2206 spin_unlock_irq(&zone->lru_lock);
2209 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2214 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2217 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2218 * may already be on some other mem_cgroup's LRU. Take care of it.
2221 lock_page_lru(page, &isolated);
2224 * Nobody should be changing or seriously looking at
2225 * page->mem_cgroup at this point:
2227 * - the page is uncharged
2229 * - the page is off-LRU
2231 * - an anonymous fault has exclusive page access, except for
2232 * a locked page table
2234 * - a page cache insertion, a swapin fault, or a migration
2235 * have the page locked
2237 page->mem_cgroup = memcg;
2240 unlock_page_lru(page, isolated);
2243 #ifdef CONFIG_MEMCG_KMEM
2244 static int memcg_alloc_cache_id(void)
2249 id = ida_simple_get(&memcg_cache_ida,
2250 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2254 if (id < memcg_nr_cache_ids)
2258 * There's no space for the new id in memcg_caches arrays,
2259 * so we have to grow them.
2261 down_write(&memcg_cache_ids_sem);
2263 size = 2 * (id + 1);
2264 if (size < MEMCG_CACHES_MIN_SIZE)
2265 size = MEMCG_CACHES_MIN_SIZE;
2266 else if (size > MEMCG_CACHES_MAX_SIZE)
2267 size = MEMCG_CACHES_MAX_SIZE;
2269 err = memcg_update_all_caches(size);
2271 err = memcg_update_all_list_lrus(size);
2273 memcg_nr_cache_ids = size;
2275 up_write(&memcg_cache_ids_sem);
2278 ida_simple_remove(&memcg_cache_ida, id);
2284 static void memcg_free_cache_id(int id)
2286 ida_simple_remove(&memcg_cache_ida, id);
2289 struct memcg_kmem_cache_create_work {
2290 struct mem_cgroup *memcg;
2291 struct kmem_cache *cachep;
2292 struct work_struct work;
2295 static void memcg_kmem_cache_create_func(struct work_struct *w)
2297 struct memcg_kmem_cache_create_work *cw =
2298 container_of(w, struct memcg_kmem_cache_create_work, work);
2299 struct mem_cgroup *memcg = cw->memcg;
2300 struct kmem_cache *cachep = cw->cachep;
2302 memcg_create_kmem_cache(memcg, cachep);
2304 css_put(&memcg->css);
2309 * Enqueue the creation of a per-memcg kmem_cache.
2311 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2312 struct kmem_cache *cachep)
2314 struct memcg_kmem_cache_create_work *cw;
2316 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2320 css_get(&memcg->css);
2323 cw->cachep = cachep;
2324 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2326 schedule_work(&cw->work);
2329 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2330 struct kmem_cache *cachep)
2333 * We need to stop accounting when we kmalloc, because if the
2334 * corresponding kmalloc cache is not yet created, the first allocation
2335 * in __memcg_schedule_kmem_cache_create will recurse.
2337 * However, it is better to enclose the whole function. Depending on
2338 * the debugging options enabled, INIT_WORK(), for instance, can
2339 * trigger an allocation. This too, will make us recurse. Because at
2340 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2341 * the safest choice is to do it like this, wrapping the whole function.
2343 current->memcg_kmem_skip_account = 1;
2344 __memcg_schedule_kmem_cache_create(memcg, cachep);
2345 current->memcg_kmem_skip_account = 0;
2349 * Return the kmem_cache we're supposed to use for a slab allocation.
2350 * We try to use the current memcg's version of the cache.
2352 * If the cache does not exist yet, if we are the first user of it,
2353 * we either create it immediately, if possible, or create it asynchronously
2355 * In the latter case, we will let the current allocation go through with
2356 * the original cache.
2358 * Can't be called in interrupt context or from kernel threads.
2359 * This function needs to be called with rcu_read_lock() held.
2361 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2363 struct mem_cgroup *memcg;
2364 struct kmem_cache *memcg_cachep;
2367 VM_BUG_ON(!is_root_cache(cachep));
2369 if (current->memcg_kmem_skip_account)
2372 memcg = get_mem_cgroup_from_mm(current->mm);
2373 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2377 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2378 if (likely(memcg_cachep))
2379 return memcg_cachep;
2382 * If we are in a safe context (can wait, and not in interrupt
2383 * context), we could be be predictable and return right away.
2384 * This would guarantee that the allocation being performed
2385 * already belongs in the new cache.
2387 * However, there are some clashes that can arrive from locking.
2388 * For instance, because we acquire the slab_mutex while doing
2389 * memcg_create_kmem_cache, this means no further allocation
2390 * could happen with the slab_mutex held. So it's better to
2393 memcg_schedule_kmem_cache_create(memcg, cachep);
2395 css_put(&memcg->css);
2399 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2401 if (!is_root_cache(cachep))
2402 css_put(&cachep->memcg_params.memcg->css);
2405 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2406 struct mem_cgroup *memcg)
2408 unsigned int nr_pages = 1 << order;
2409 struct page_counter *counter;
2412 if (!memcg_kmem_is_active(memcg))
2415 if (!page_counter_try_charge(&memcg->kmem, nr_pages, &counter))
2418 ret = try_charge(memcg, gfp, nr_pages);
2420 page_counter_uncharge(&memcg->kmem, nr_pages);
2424 page->mem_cgroup = memcg;
2429 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2431 struct mem_cgroup *memcg;
2434 memcg = get_mem_cgroup_from_mm(current->mm);
2435 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2436 css_put(&memcg->css);
2440 void __memcg_kmem_uncharge(struct page *page, int order)
2442 struct mem_cgroup *memcg = page->mem_cgroup;
2443 unsigned int nr_pages = 1 << order;
2448 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2450 page_counter_uncharge(&memcg->kmem, nr_pages);
2451 page_counter_uncharge(&memcg->memory, nr_pages);
2452 if (do_swap_account)
2453 page_counter_uncharge(&memcg->memsw, nr_pages);
2455 page->mem_cgroup = NULL;
2456 css_put_many(&memcg->css, nr_pages);
2458 #endif /* CONFIG_MEMCG_KMEM */
2460 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2463 * Because tail pages are not marked as "used", set it. We're under
2464 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2465 * charge/uncharge will be never happen and move_account() is done under
2466 * compound_lock(), so we don't have to take care of races.
2468 void mem_cgroup_split_huge_fixup(struct page *head)
2472 if (mem_cgroup_disabled())
2475 for (i = 1; i < HPAGE_PMD_NR; i++)
2476 head[i].mem_cgroup = head->mem_cgroup;
2478 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2481 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2483 #ifdef CONFIG_MEMCG_SWAP
2484 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2487 int val = (charge) ? 1 : -1;
2488 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2492 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2493 * @entry: swap entry to be moved
2494 * @from: mem_cgroup which the entry is moved from
2495 * @to: mem_cgroup which the entry is moved to
2497 * It succeeds only when the swap_cgroup's record for this entry is the same
2498 * as the mem_cgroup's id of @from.
2500 * Returns 0 on success, -EINVAL on failure.
2502 * The caller must have charged to @to, IOW, called page_counter_charge() about
2503 * both res and memsw, and called css_get().
2505 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2506 struct mem_cgroup *from, struct mem_cgroup *to)
2508 unsigned short old_id, new_id;
2510 old_id = mem_cgroup_id(from);
2511 new_id = mem_cgroup_id(to);
2513 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2514 mem_cgroup_swap_statistics(from, false);
2515 mem_cgroup_swap_statistics(to, true);
2521 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2522 struct mem_cgroup *from, struct mem_cgroup *to)
2528 static DEFINE_MUTEX(memcg_limit_mutex);
2530 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2531 unsigned long limit)
2533 unsigned long curusage;
2534 unsigned long oldusage;
2535 bool enlarge = false;
2540 * For keeping hierarchical_reclaim simple, how long we should retry
2541 * is depends on callers. We set our retry-count to be function
2542 * of # of children which we should visit in this loop.
2544 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2545 mem_cgroup_count_children(memcg);
2547 oldusage = page_counter_read(&memcg->memory);
2550 if (signal_pending(current)) {
2555 mutex_lock(&memcg_limit_mutex);
2556 if (limit > memcg->memsw.limit) {
2557 mutex_unlock(&memcg_limit_mutex);
2561 if (limit > memcg->memory.limit)
2563 ret = page_counter_limit(&memcg->memory, limit);
2564 mutex_unlock(&memcg_limit_mutex);
2569 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2571 curusage = page_counter_read(&memcg->memory);
2572 /* Usage is reduced ? */
2573 if (curusage >= oldusage)
2576 oldusage = curusage;
2577 } while (retry_count);
2579 if (!ret && enlarge)
2580 memcg_oom_recover(memcg);
2585 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2586 unsigned long limit)
2588 unsigned long curusage;
2589 unsigned long oldusage;
2590 bool enlarge = false;
2594 /* see mem_cgroup_resize_res_limit */
2595 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2596 mem_cgroup_count_children(memcg);
2598 oldusage = page_counter_read(&memcg->memsw);
2601 if (signal_pending(current)) {
2606 mutex_lock(&memcg_limit_mutex);
2607 if (limit < memcg->memory.limit) {
2608 mutex_unlock(&memcg_limit_mutex);
2612 if (limit > memcg->memsw.limit)
2614 ret = page_counter_limit(&memcg->memsw, limit);
2615 mutex_unlock(&memcg_limit_mutex);
2620 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2622 curusage = page_counter_read(&memcg->memsw);
2623 /* Usage is reduced ? */
2624 if (curusage >= oldusage)
2627 oldusage = curusage;
2628 } while (retry_count);
2630 if (!ret && enlarge)
2631 memcg_oom_recover(memcg);
2636 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2638 unsigned long *total_scanned)
2640 unsigned long nr_reclaimed = 0;
2641 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2642 unsigned long reclaimed;
2644 struct mem_cgroup_tree_per_zone *mctz;
2645 unsigned long excess;
2646 unsigned long nr_scanned;
2651 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2653 * This loop can run a while, specially if mem_cgroup's continuously
2654 * keep exceeding their soft limit and putting the system under
2661 mz = mem_cgroup_largest_soft_limit_node(mctz);
2666 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2667 gfp_mask, &nr_scanned);
2668 nr_reclaimed += reclaimed;
2669 *total_scanned += nr_scanned;
2670 spin_lock_irq(&mctz->lock);
2671 __mem_cgroup_remove_exceeded(mz, mctz);
2674 * If we failed to reclaim anything from this memory cgroup
2675 * it is time to move on to the next cgroup
2679 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2681 excess = soft_limit_excess(mz->memcg);
2683 * One school of thought says that we should not add
2684 * back the node to the tree if reclaim returns 0.
2685 * But our reclaim could return 0, simply because due
2686 * to priority we are exposing a smaller subset of
2687 * memory to reclaim from. Consider this as a longer
2690 /* If excess == 0, no tree ops */
2691 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2692 spin_unlock_irq(&mctz->lock);
2693 css_put(&mz->memcg->css);
2696 * Could not reclaim anything and there are no more
2697 * mem cgroups to try or we seem to be looping without
2698 * reclaiming anything.
2700 if (!nr_reclaimed &&
2702 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2704 } while (!nr_reclaimed);
2706 css_put(&next_mz->memcg->css);
2707 return nr_reclaimed;
2711 * Test whether @memcg has children, dead or alive. Note that this
2712 * function doesn't care whether @memcg has use_hierarchy enabled and
2713 * returns %true if there are child csses according to the cgroup
2714 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2716 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2721 * The lock does not prevent addition or deletion of children, but
2722 * it prevents a new child from being initialized based on this
2723 * parent in css_online(), so it's enough to decide whether
2724 * hierarchically inherited attributes can still be changed or not.
2726 lockdep_assert_held(&memcg_create_mutex);
2729 ret = css_next_child(NULL, &memcg->css);
2735 * Reclaims as many pages from the given memcg as possible and moves
2736 * the rest to the parent.
2738 * Caller is responsible for holding css reference for memcg.
2740 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2742 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2744 /* we call try-to-free pages for make this cgroup empty */
2745 lru_add_drain_all();
2746 /* try to free all pages in this cgroup */
2747 while (nr_retries && page_counter_read(&memcg->memory)) {
2750 if (signal_pending(current))
2753 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2757 /* maybe some writeback is necessary */
2758 congestion_wait(BLK_RW_ASYNC, HZ/10);
2766 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2767 char *buf, size_t nbytes,
2770 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2772 if (mem_cgroup_is_root(memcg))
2774 return mem_cgroup_force_empty(memcg) ?: nbytes;
2777 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2780 return mem_cgroup_from_css(css)->use_hierarchy;
2783 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2784 struct cftype *cft, u64 val)
2787 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2788 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2790 mutex_lock(&memcg_create_mutex);
2792 if (memcg->use_hierarchy == val)
2796 * If parent's use_hierarchy is set, we can't make any modifications
2797 * in the child subtrees. If it is unset, then the change can
2798 * occur, provided the current cgroup has no children.
2800 * For the root cgroup, parent_mem is NULL, we allow value to be
2801 * set if there are no children.
2803 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2804 (val == 1 || val == 0)) {
2805 if (!memcg_has_children(memcg))
2806 memcg->use_hierarchy = val;
2813 mutex_unlock(&memcg_create_mutex);
2818 static unsigned long tree_stat(struct mem_cgroup *memcg,
2819 enum mem_cgroup_stat_index idx)
2821 struct mem_cgroup *iter;
2822 unsigned long val = 0;
2824 for_each_mem_cgroup_tree(iter, memcg)
2825 val += mem_cgroup_read_stat(iter, idx);
2830 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2834 if (mem_cgroup_is_root(memcg)) {
2835 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2836 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2838 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2841 val = page_counter_read(&memcg->memory);
2843 val = page_counter_read(&memcg->memsw);
2856 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2859 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2860 struct page_counter *counter;
2862 switch (MEMFILE_TYPE(cft->private)) {
2864 counter = &memcg->memory;
2867 counter = &memcg->memsw;
2870 counter = &memcg->kmem;
2876 switch (MEMFILE_ATTR(cft->private)) {
2878 if (counter == &memcg->memory)
2879 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2880 if (counter == &memcg->memsw)
2881 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2882 return (u64)page_counter_read(counter) * PAGE_SIZE;
2884 return (u64)counter->limit * PAGE_SIZE;
2886 return (u64)counter->watermark * PAGE_SIZE;
2888 return counter->failcnt;
2889 case RES_SOFT_LIMIT:
2890 return (u64)memcg->soft_limit * PAGE_SIZE;
2896 #ifdef CONFIG_MEMCG_KMEM
2897 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2898 unsigned long nr_pages)
2903 BUG_ON(memcg->kmemcg_id >= 0);
2904 BUG_ON(memcg->kmem_acct_activated);
2905 BUG_ON(memcg->kmem_acct_active);
2908 * For simplicity, we won't allow this to be disabled. It also can't
2909 * be changed if the cgroup has children already, or if tasks had
2912 * If tasks join before we set the limit, a person looking at
2913 * kmem.usage_in_bytes will have no way to determine when it took
2914 * place, which makes the value quite meaningless.
2916 * After it first became limited, changes in the value of the limit are
2917 * of course permitted.
2919 mutex_lock(&memcg_create_mutex);
2920 if (cgroup_is_populated(memcg->css.cgroup) ||
2921 (memcg->use_hierarchy && memcg_has_children(memcg)))
2923 mutex_unlock(&memcg_create_mutex);
2927 memcg_id = memcg_alloc_cache_id();
2934 * We couldn't have accounted to this cgroup, because it hasn't got
2935 * activated yet, so this should succeed.
2937 err = page_counter_limit(&memcg->kmem, nr_pages);
2940 static_key_slow_inc(&memcg_kmem_enabled_key);
2942 * A memory cgroup is considered kmem-active as soon as it gets
2943 * kmemcg_id. Setting the id after enabling static branching will
2944 * guarantee no one starts accounting before all call sites are
2947 memcg->kmemcg_id = memcg_id;
2948 memcg->kmem_acct_activated = true;
2949 memcg->kmem_acct_active = true;
2954 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2955 unsigned long limit)
2959 mutex_lock(&memcg_limit_mutex);
2960 if (!memcg_kmem_is_active(memcg))
2961 ret = memcg_activate_kmem(memcg, limit);
2963 ret = page_counter_limit(&memcg->kmem, limit);
2964 mutex_unlock(&memcg_limit_mutex);
2968 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2971 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2976 mutex_lock(&memcg_limit_mutex);
2978 * If the parent cgroup is not kmem-active now, it cannot be activated
2979 * after this point, because it has at least one child already.
2981 if (memcg_kmem_is_active(parent))
2982 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
2983 mutex_unlock(&memcg_limit_mutex);
2987 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2988 unsigned long limit)
2992 #endif /* CONFIG_MEMCG_KMEM */
2995 * The user of this function is...
2998 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2999 char *buf, size_t nbytes, loff_t off)
3001 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3002 unsigned long nr_pages;
3005 buf = strstrip(buf);
3006 ret = page_counter_memparse(buf, "-1", &nr_pages);
3010 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3012 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3016 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3018 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3021 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3024 ret = memcg_update_kmem_limit(memcg, nr_pages);
3028 case RES_SOFT_LIMIT:
3029 memcg->soft_limit = nr_pages;
3033 return ret ?: nbytes;
3036 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3037 size_t nbytes, loff_t off)
3039 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3040 struct page_counter *counter;
3042 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3044 counter = &memcg->memory;
3047 counter = &memcg->memsw;
3050 counter = &memcg->kmem;
3056 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3058 page_counter_reset_watermark(counter);
3061 counter->failcnt = 0;
3070 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3073 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3077 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3078 struct cftype *cft, u64 val)
3080 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3082 if (val & ~MOVE_MASK)
3086 * No kind of locking is needed in here, because ->can_attach() will
3087 * check this value once in the beginning of the process, and then carry
3088 * on with stale data. This means that changes to this value will only
3089 * affect task migrations starting after the change.
3091 memcg->move_charge_at_immigrate = val;
3095 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3096 struct cftype *cft, u64 val)
3103 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3107 unsigned int lru_mask;
3110 static const struct numa_stat stats[] = {
3111 { "total", LRU_ALL },
3112 { "file", LRU_ALL_FILE },
3113 { "anon", LRU_ALL_ANON },
3114 { "unevictable", BIT(LRU_UNEVICTABLE) },
3116 const struct numa_stat *stat;
3119 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3121 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3122 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3123 seq_printf(m, "%s=%lu", stat->name, nr);
3124 for_each_node_state(nid, N_MEMORY) {
3125 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3127 seq_printf(m, " N%d=%lu", nid, nr);
3132 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3133 struct mem_cgroup *iter;
3136 for_each_mem_cgroup_tree(iter, memcg)
3137 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3138 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3139 for_each_node_state(nid, N_MEMORY) {
3141 for_each_mem_cgroup_tree(iter, memcg)
3142 nr += mem_cgroup_node_nr_lru_pages(
3143 iter, nid, stat->lru_mask);
3144 seq_printf(m, " N%d=%lu", nid, nr);
3151 #endif /* CONFIG_NUMA */
3153 static int memcg_stat_show(struct seq_file *m, void *v)
3155 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3156 unsigned long memory, memsw;
3157 struct mem_cgroup *mi;
3160 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3161 MEM_CGROUP_STAT_NSTATS);
3162 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3163 MEM_CGROUP_EVENTS_NSTATS);
3164 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3166 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3167 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3169 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3170 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3173 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3174 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3175 mem_cgroup_read_events(memcg, i));
3177 for (i = 0; i < NR_LRU_LISTS; i++)
3178 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3179 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3181 /* Hierarchical information */
3182 memory = memsw = PAGE_COUNTER_MAX;
3183 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3184 memory = min(memory, mi->memory.limit);
3185 memsw = min(memsw, mi->memsw.limit);
3187 seq_printf(m, "hierarchical_memory_limit %llu\n",
3188 (u64)memory * PAGE_SIZE);
3189 if (do_swap_account)
3190 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3191 (u64)memsw * PAGE_SIZE);
3193 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3194 unsigned long long val = 0;
3196 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3198 for_each_mem_cgroup_tree(mi, memcg)
3199 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3200 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3203 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3204 unsigned long long val = 0;
3206 for_each_mem_cgroup_tree(mi, memcg)
3207 val += mem_cgroup_read_events(mi, i);
3208 seq_printf(m, "total_%s %llu\n",
3209 mem_cgroup_events_names[i], val);
3212 for (i = 0; i < NR_LRU_LISTS; i++) {
3213 unsigned long long val = 0;
3215 for_each_mem_cgroup_tree(mi, memcg)
3216 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3217 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3220 #ifdef CONFIG_DEBUG_VM
3223 struct mem_cgroup_per_zone *mz;
3224 struct zone_reclaim_stat *rstat;
3225 unsigned long recent_rotated[2] = {0, 0};
3226 unsigned long recent_scanned[2] = {0, 0};
3228 for_each_online_node(nid)
3229 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3230 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3231 rstat = &mz->lruvec.reclaim_stat;
3233 recent_rotated[0] += rstat->recent_rotated[0];
3234 recent_rotated[1] += rstat->recent_rotated[1];
3235 recent_scanned[0] += rstat->recent_scanned[0];
3236 recent_scanned[1] += rstat->recent_scanned[1];
3238 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3239 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3240 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3241 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3248 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3251 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3253 return mem_cgroup_swappiness(memcg);
3256 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3257 struct cftype *cft, u64 val)
3259 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3265 memcg->swappiness = val;
3267 vm_swappiness = val;
3272 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3274 struct mem_cgroup_threshold_ary *t;
3275 unsigned long usage;
3280 t = rcu_dereference(memcg->thresholds.primary);
3282 t = rcu_dereference(memcg->memsw_thresholds.primary);
3287 usage = mem_cgroup_usage(memcg, swap);
3290 * current_threshold points to threshold just below or equal to usage.
3291 * If it's not true, a threshold was crossed after last
3292 * call of __mem_cgroup_threshold().
3294 i = t->current_threshold;
3297 * Iterate backward over array of thresholds starting from
3298 * current_threshold and check if a threshold is crossed.
3299 * If none of thresholds below usage is crossed, we read
3300 * only one element of the array here.
3302 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3303 eventfd_signal(t->entries[i].eventfd, 1);
3305 /* i = current_threshold + 1 */
3309 * Iterate forward over array of thresholds starting from
3310 * current_threshold+1 and check if a threshold is crossed.
3311 * If none of thresholds above usage is crossed, we read
3312 * only one element of the array here.
3314 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3315 eventfd_signal(t->entries[i].eventfd, 1);
3317 /* Update current_threshold */
3318 t->current_threshold = i - 1;
3323 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3326 __mem_cgroup_threshold(memcg, false);
3327 if (do_swap_account)
3328 __mem_cgroup_threshold(memcg, true);
3330 memcg = parent_mem_cgroup(memcg);
3334 static int compare_thresholds(const void *a, const void *b)
3336 const struct mem_cgroup_threshold *_a = a;
3337 const struct mem_cgroup_threshold *_b = b;
3339 if (_a->threshold > _b->threshold)
3342 if (_a->threshold < _b->threshold)
3348 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3350 struct mem_cgroup_eventfd_list *ev;
3352 spin_lock(&memcg_oom_lock);
3354 list_for_each_entry(ev, &memcg->oom_notify, list)
3355 eventfd_signal(ev->eventfd, 1);
3357 spin_unlock(&memcg_oom_lock);
3361 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3363 struct mem_cgroup *iter;
3365 for_each_mem_cgroup_tree(iter, memcg)
3366 mem_cgroup_oom_notify_cb(iter);
3369 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3370 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3372 struct mem_cgroup_thresholds *thresholds;
3373 struct mem_cgroup_threshold_ary *new;
3374 unsigned long threshold;
3375 unsigned long usage;
3378 ret = page_counter_memparse(args, "-1", &threshold);
3382 mutex_lock(&memcg->thresholds_lock);
3385 thresholds = &memcg->thresholds;
3386 usage = mem_cgroup_usage(memcg, false);
3387 } else if (type == _MEMSWAP) {
3388 thresholds = &memcg->memsw_thresholds;
3389 usage = mem_cgroup_usage(memcg, true);
3393 /* Check if a threshold crossed before adding a new one */
3394 if (thresholds->primary)
3395 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3397 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3399 /* Allocate memory for new array of thresholds */
3400 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3408 /* Copy thresholds (if any) to new array */
3409 if (thresholds->primary) {
3410 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3411 sizeof(struct mem_cgroup_threshold));
3414 /* Add new threshold */
3415 new->entries[size - 1].eventfd = eventfd;
3416 new->entries[size - 1].threshold = threshold;
3418 /* Sort thresholds. Registering of new threshold isn't time-critical */
3419 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3420 compare_thresholds, NULL);
3422 /* Find current threshold */
3423 new->current_threshold = -1;
3424 for (i = 0; i < size; i++) {
3425 if (new->entries[i].threshold <= usage) {
3427 * new->current_threshold will not be used until
3428 * rcu_assign_pointer(), so it's safe to increment
3431 ++new->current_threshold;
3436 /* Free old spare buffer and save old primary buffer as spare */
3437 kfree(thresholds->spare);
3438 thresholds->spare = thresholds->primary;
3440 rcu_assign_pointer(thresholds->primary, new);
3442 /* To be sure that nobody uses thresholds */
3446 mutex_unlock(&memcg->thresholds_lock);
3451 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3452 struct eventfd_ctx *eventfd, const char *args)
3454 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3457 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3458 struct eventfd_ctx *eventfd, const char *args)
3460 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3463 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3464 struct eventfd_ctx *eventfd, enum res_type type)
3466 struct mem_cgroup_thresholds *thresholds;
3467 struct mem_cgroup_threshold_ary *new;
3468 unsigned long usage;
3471 mutex_lock(&memcg->thresholds_lock);
3474 thresholds = &memcg->thresholds;
3475 usage = mem_cgroup_usage(memcg, false);
3476 } else if (type == _MEMSWAP) {
3477 thresholds = &memcg->memsw_thresholds;
3478 usage = mem_cgroup_usage(memcg, true);
3482 if (!thresholds->primary)
3485 /* Check if a threshold crossed before removing */
3486 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3488 /* Calculate new number of threshold */
3490 for (i = 0; i < thresholds->primary->size; i++) {
3491 if (thresholds->primary->entries[i].eventfd != eventfd)
3495 new = thresholds->spare;
3497 /* Set thresholds array to NULL if we don't have thresholds */
3506 /* Copy thresholds and find current threshold */
3507 new->current_threshold = -1;
3508 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3509 if (thresholds->primary->entries[i].eventfd == eventfd)
3512 new->entries[j] = thresholds->primary->entries[i];
3513 if (new->entries[j].threshold <= usage) {
3515 * new->current_threshold will not be used
3516 * until rcu_assign_pointer(), so it's safe to increment
3519 ++new->current_threshold;
3525 /* Swap primary and spare array */
3526 thresholds->spare = thresholds->primary;
3528 rcu_assign_pointer(thresholds->primary, new);
3530 /* To be sure that nobody uses thresholds */
3533 /* If all events are unregistered, free the spare array */
3535 kfree(thresholds->spare);
3536 thresholds->spare = NULL;
3539 mutex_unlock(&memcg->thresholds_lock);
3542 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3543 struct eventfd_ctx *eventfd)
3545 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3548 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3549 struct eventfd_ctx *eventfd)
3551 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3554 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3555 struct eventfd_ctx *eventfd, const char *args)
3557 struct mem_cgroup_eventfd_list *event;
3559 event = kmalloc(sizeof(*event), GFP_KERNEL);
3563 spin_lock(&memcg_oom_lock);
3565 event->eventfd = eventfd;
3566 list_add(&event->list, &memcg->oom_notify);
3568 /* already in OOM ? */
3569 if (memcg->under_oom)
3570 eventfd_signal(eventfd, 1);
3571 spin_unlock(&memcg_oom_lock);
3576 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3577 struct eventfd_ctx *eventfd)
3579 struct mem_cgroup_eventfd_list *ev, *tmp;
3581 spin_lock(&memcg_oom_lock);
3583 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3584 if (ev->eventfd == eventfd) {
3585 list_del(&ev->list);
3590 spin_unlock(&memcg_oom_lock);
3593 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3595 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3597 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3598 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3602 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3603 struct cftype *cft, u64 val)
3605 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3607 /* cannot set to root cgroup and only 0 and 1 are allowed */
3608 if (!css->parent || !((val == 0) || (val == 1)))
3611 memcg->oom_kill_disable = val;
3613 memcg_oom_recover(memcg);
3618 #ifdef CONFIG_MEMCG_KMEM
3619 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3623 ret = memcg_propagate_kmem(memcg);
3627 return mem_cgroup_sockets_init(memcg, ss);
3630 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3632 struct cgroup_subsys_state *css;
3633 struct mem_cgroup *parent, *child;
3636 if (!memcg->kmem_acct_active)
3640 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3641 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3642 * guarantees no cache will be created for this cgroup after we are
3643 * done (see memcg_create_kmem_cache()).
3645 memcg->kmem_acct_active = false;
3647 memcg_deactivate_kmem_caches(memcg);
3649 kmemcg_id = memcg->kmemcg_id;
3650 BUG_ON(kmemcg_id < 0);
3652 parent = parent_mem_cgroup(memcg);
3654 parent = root_mem_cgroup;
3657 * Change kmemcg_id of this cgroup and all its descendants to the
3658 * parent's id, and then move all entries from this cgroup's list_lrus
3659 * to ones of the parent. After we have finished, all list_lrus
3660 * corresponding to this cgroup are guaranteed to remain empty. The
3661 * ordering is imposed by list_lru_node->lock taken by
3662 * memcg_drain_all_list_lrus().
3664 css_for_each_descendant_pre(css, &memcg->css) {
3665 child = mem_cgroup_from_css(css);
3666 BUG_ON(child->kmemcg_id != kmemcg_id);
3667 child->kmemcg_id = parent->kmemcg_id;
3668 if (!memcg->use_hierarchy)
3671 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3673 memcg_free_cache_id(kmemcg_id);
3676 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3678 if (memcg->kmem_acct_activated) {
3679 memcg_destroy_kmem_caches(memcg);
3680 static_key_slow_dec(&memcg_kmem_enabled_key);
3681 WARN_ON(page_counter_read(&memcg->kmem));
3683 mem_cgroup_sockets_destroy(memcg);
3686 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3691 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3695 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3700 #ifdef CONFIG_CGROUP_WRITEBACK
3702 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3704 return &memcg->cgwb_list;
3707 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3709 return wb_domain_init(&memcg->cgwb_domain, gfp);
3712 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3714 wb_domain_exit(&memcg->cgwb_domain);
3717 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3719 wb_domain_size_changed(&memcg->cgwb_domain);
3722 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3724 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3726 if (!memcg->css.parent)
3729 return &memcg->cgwb_domain;
3733 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3734 * @wb: bdi_writeback in question
3735 * @pfilepages: out parameter for number of file pages
3736 * @pheadroom: out parameter for number of allocatable pages according to memcg
3737 * @pdirty: out parameter for number of dirty pages
3738 * @pwriteback: out parameter for number of pages under writeback
3740 * Determine the numbers of file, headroom, dirty, and writeback pages in
3741 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3742 * is a bit more involved.
3744 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3745 * headroom is calculated as the lowest headroom of itself and the
3746 * ancestors. Note that this doesn't consider the actual amount of
3747 * available memory in the system. The caller should further cap
3748 * *@pheadroom accordingly.
3750 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3751 unsigned long *pheadroom, unsigned long *pdirty,
3752 unsigned long *pwriteback)
3754 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3755 struct mem_cgroup *parent;
3757 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3759 /* this should eventually include NR_UNSTABLE_NFS */
3760 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3761 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3762 (1 << LRU_ACTIVE_FILE));
3763 *pheadroom = PAGE_COUNTER_MAX;
3765 while ((parent = parent_mem_cgroup(memcg))) {
3766 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3767 unsigned long used = page_counter_read(&memcg->memory);
3769 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3774 #else /* CONFIG_CGROUP_WRITEBACK */
3776 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3781 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3785 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3789 #endif /* CONFIG_CGROUP_WRITEBACK */
3792 * DO NOT USE IN NEW FILES.
3794 * "cgroup.event_control" implementation.
3796 * This is way over-engineered. It tries to support fully configurable
3797 * events for each user. Such level of flexibility is completely
3798 * unnecessary especially in the light of the planned unified hierarchy.
3800 * Please deprecate this and replace with something simpler if at all
3805 * Unregister event and free resources.
3807 * Gets called from workqueue.
3809 static void memcg_event_remove(struct work_struct *work)
3811 struct mem_cgroup_event *event =
3812 container_of(work, struct mem_cgroup_event, remove);
3813 struct mem_cgroup *memcg = event->memcg;
3815 remove_wait_queue(event->wqh, &event->wait);
3817 event->unregister_event(memcg, event->eventfd);
3819 /* Notify userspace the event is going away. */
3820 eventfd_signal(event->eventfd, 1);
3822 eventfd_ctx_put(event->eventfd);
3824 css_put(&memcg->css);
3828 * Gets called on POLLHUP on eventfd when user closes it.
3830 * Called with wqh->lock held and interrupts disabled.
3832 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3833 int sync, void *key)
3835 struct mem_cgroup_event *event =
3836 container_of(wait, struct mem_cgroup_event, wait);
3837 struct mem_cgroup *memcg = event->memcg;
3838 unsigned long flags = (unsigned long)key;
3840 if (flags & POLLHUP) {
3842 * If the event has been detached at cgroup removal, we
3843 * can simply return knowing the other side will cleanup
3846 * We can't race against event freeing since the other
3847 * side will require wqh->lock via remove_wait_queue(),
3850 spin_lock(&memcg->event_list_lock);
3851 if (!list_empty(&event->list)) {
3852 list_del_init(&event->list);
3854 * We are in atomic context, but cgroup_event_remove()
3855 * may sleep, so we have to call it in workqueue.
3857 schedule_work(&event->remove);
3859 spin_unlock(&memcg->event_list_lock);
3865 static void memcg_event_ptable_queue_proc(struct file *file,
3866 wait_queue_head_t *wqh, poll_table *pt)
3868 struct mem_cgroup_event *event =
3869 container_of(pt, struct mem_cgroup_event, pt);
3872 add_wait_queue(wqh, &event->wait);
3876 * DO NOT USE IN NEW FILES.
3878 * Parse input and register new cgroup event handler.
3880 * Input must be in format '<event_fd> <control_fd> <args>'.
3881 * Interpretation of args is defined by control file implementation.
3883 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3884 char *buf, size_t nbytes, loff_t off)
3886 struct cgroup_subsys_state *css = of_css(of);
3887 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3888 struct mem_cgroup_event *event;
3889 struct cgroup_subsys_state *cfile_css;
3890 unsigned int efd, cfd;
3897 buf = strstrip(buf);
3899 efd = simple_strtoul(buf, &endp, 10);
3904 cfd = simple_strtoul(buf, &endp, 10);
3905 if ((*endp != ' ') && (*endp != '\0'))
3909 event = kzalloc(sizeof(*event), GFP_KERNEL);
3913 event->memcg = memcg;
3914 INIT_LIST_HEAD(&event->list);
3915 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3916 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3917 INIT_WORK(&event->remove, memcg_event_remove);
3925 event->eventfd = eventfd_ctx_fileget(efile.file);
3926 if (IS_ERR(event->eventfd)) {
3927 ret = PTR_ERR(event->eventfd);
3934 goto out_put_eventfd;
3937 /* the process need read permission on control file */
3938 /* AV: shouldn't we check that it's been opened for read instead? */
3939 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3944 * Determine the event callbacks and set them in @event. This used
3945 * to be done via struct cftype but cgroup core no longer knows
3946 * about these events. The following is crude but the whole thing
3947 * is for compatibility anyway.
3949 * DO NOT ADD NEW FILES.
3951 name = cfile.file->f_path.dentry->d_name.name;
3953 if (!strcmp(name, "memory.usage_in_bytes")) {
3954 event->register_event = mem_cgroup_usage_register_event;
3955 event->unregister_event = mem_cgroup_usage_unregister_event;
3956 } else if (!strcmp(name, "memory.oom_control")) {
3957 event->register_event = mem_cgroup_oom_register_event;
3958 event->unregister_event = mem_cgroup_oom_unregister_event;
3959 } else if (!strcmp(name, "memory.pressure_level")) {
3960 event->register_event = vmpressure_register_event;
3961 event->unregister_event = vmpressure_unregister_event;
3962 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3963 event->register_event = memsw_cgroup_usage_register_event;
3964 event->unregister_event = memsw_cgroup_usage_unregister_event;
3971 * Verify @cfile should belong to @css. Also, remaining events are
3972 * automatically removed on cgroup destruction but the removal is
3973 * asynchronous, so take an extra ref on @css.
3975 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3976 &memory_cgrp_subsys);
3978 if (IS_ERR(cfile_css))
3980 if (cfile_css != css) {
3985 ret = event->register_event(memcg, event->eventfd, buf);
3989 efile.file->f_op->poll(efile.file, &event->pt);
3991 spin_lock(&memcg->event_list_lock);
3992 list_add(&event->list, &memcg->event_list);
3993 spin_unlock(&memcg->event_list_lock);
4005 eventfd_ctx_put(event->eventfd);
4014 static struct cftype mem_cgroup_legacy_files[] = {
4016 .name = "usage_in_bytes",
4017 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4018 .read_u64 = mem_cgroup_read_u64,
4021 .name = "max_usage_in_bytes",
4022 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4023 .write = mem_cgroup_reset,
4024 .read_u64 = mem_cgroup_read_u64,
4027 .name = "limit_in_bytes",
4028 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4029 .write = mem_cgroup_write,
4030 .read_u64 = mem_cgroup_read_u64,
4033 .name = "soft_limit_in_bytes",
4034 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4035 .write = mem_cgroup_write,
4036 .read_u64 = mem_cgroup_read_u64,
4040 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4041 .write = mem_cgroup_reset,
4042 .read_u64 = mem_cgroup_read_u64,
4046 .seq_show = memcg_stat_show,
4049 .name = "force_empty",
4050 .write = mem_cgroup_force_empty_write,
4053 .name = "use_hierarchy",
4054 .write_u64 = mem_cgroup_hierarchy_write,
4055 .read_u64 = mem_cgroup_hierarchy_read,
4058 .name = "cgroup.event_control", /* XXX: for compat */
4059 .write = memcg_write_event_control,
4060 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4063 .name = "swappiness",
4064 .read_u64 = mem_cgroup_swappiness_read,
4065 .write_u64 = mem_cgroup_swappiness_write,
4068 .name = "move_charge_at_immigrate",
4069 .read_u64 = mem_cgroup_move_charge_read,
4070 .write_u64 = mem_cgroup_move_charge_write,
4073 .name = "oom_control",
4074 .seq_show = mem_cgroup_oom_control_read,
4075 .write_u64 = mem_cgroup_oom_control_write,
4076 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4079 .name = "pressure_level",
4083 .name = "numa_stat",
4084 .seq_show = memcg_numa_stat_show,
4087 #ifdef CONFIG_MEMCG_KMEM
4089 .name = "kmem.limit_in_bytes",
4090 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4091 .write = mem_cgroup_write,
4092 .read_u64 = mem_cgroup_read_u64,
4095 .name = "kmem.usage_in_bytes",
4096 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4097 .read_u64 = mem_cgroup_read_u64,
4100 .name = "kmem.failcnt",
4101 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4102 .write = mem_cgroup_reset,
4103 .read_u64 = mem_cgroup_read_u64,
4106 .name = "kmem.max_usage_in_bytes",
4107 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4108 .write = mem_cgroup_reset,
4109 .read_u64 = mem_cgroup_read_u64,
4111 #ifdef CONFIG_SLABINFO
4113 .name = "kmem.slabinfo",
4114 .seq_start = slab_start,
4115 .seq_next = slab_next,
4116 .seq_stop = slab_stop,
4117 .seq_show = memcg_slab_show,
4121 { }, /* terminate */
4124 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4126 struct mem_cgroup_per_node *pn;
4127 struct mem_cgroup_per_zone *mz;
4128 int zone, tmp = node;
4130 * This routine is called against possible nodes.
4131 * But it's BUG to call kmalloc() against offline node.
4133 * TODO: this routine can waste much memory for nodes which will
4134 * never be onlined. It's better to use memory hotplug callback
4137 if (!node_state(node, N_NORMAL_MEMORY))
4139 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4143 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4144 mz = &pn->zoneinfo[zone];
4145 lruvec_init(&mz->lruvec);
4146 mz->usage_in_excess = 0;
4147 mz->on_tree = false;
4150 memcg->nodeinfo[node] = pn;
4154 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4156 kfree(memcg->nodeinfo[node]);
4159 static struct mem_cgroup *mem_cgroup_alloc(void)
4161 struct mem_cgroup *memcg;
4164 size = sizeof(struct mem_cgroup);
4165 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4167 memcg = kzalloc(size, GFP_KERNEL);
4171 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4175 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4181 free_percpu(memcg->stat);
4188 * At destroying mem_cgroup, references from swap_cgroup can remain.
4189 * (scanning all at force_empty is too costly...)
4191 * Instead of clearing all references at force_empty, we remember
4192 * the number of reference from swap_cgroup and free mem_cgroup when
4193 * it goes down to 0.
4195 * Removal of cgroup itself succeeds regardless of refs from swap.
4198 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4202 mem_cgroup_remove_from_trees(memcg);
4205 free_mem_cgroup_per_zone_info(memcg, node);
4207 free_percpu(memcg->stat);
4208 memcg_wb_domain_exit(memcg);
4213 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4215 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4217 if (!memcg->memory.parent)
4219 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4221 EXPORT_SYMBOL(parent_mem_cgroup);
4223 static struct cgroup_subsys_state * __ref
4224 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4226 struct mem_cgroup *memcg;
4227 long error = -ENOMEM;
4230 memcg = mem_cgroup_alloc();
4232 return ERR_PTR(error);
4235 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4239 if (parent_css == NULL) {
4240 root_mem_cgroup = memcg;
4241 mem_cgroup_root_css = &memcg->css;
4242 page_counter_init(&memcg->memory, NULL);
4243 memcg->high = PAGE_COUNTER_MAX;
4244 memcg->soft_limit = PAGE_COUNTER_MAX;
4245 page_counter_init(&memcg->memsw, NULL);
4246 page_counter_init(&memcg->kmem, NULL);
4249 memcg->last_scanned_node = MAX_NUMNODES;
4250 INIT_LIST_HEAD(&memcg->oom_notify);
4251 memcg->move_charge_at_immigrate = 0;
4252 mutex_init(&memcg->thresholds_lock);
4253 spin_lock_init(&memcg->move_lock);
4254 vmpressure_init(&memcg->vmpressure);
4255 INIT_LIST_HEAD(&memcg->event_list);
4256 spin_lock_init(&memcg->event_list_lock);
4257 #ifdef CONFIG_MEMCG_KMEM
4258 memcg->kmemcg_id = -1;
4260 #ifdef CONFIG_CGROUP_WRITEBACK
4261 INIT_LIST_HEAD(&memcg->cgwb_list);
4266 __mem_cgroup_free(memcg);
4267 return ERR_PTR(error);
4271 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4273 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4274 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4277 if (css->id > MEM_CGROUP_ID_MAX)
4283 mutex_lock(&memcg_create_mutex);
4285 memcg->use_hierarchy = parent->use_hierarchy;
4286 memcg->oom_kill_disable = parent->oom_kill_disable;
4287 memcg->swappiness = mem_cgroup_swappiness(parent);
4289 if (parent->use_hierarchy) {
4290 page_counter_init(&memcg->memory, &parent->memory);
4291 memcg->high = PAGE_COUNTER_MAX;
4292 memcg->soft_limit = PAGE_COUNTER_MAX;
4293 page_counter_init(&memcg->memsw, &parent->memsw);
4294 page_counter_init(&memcg->kmem, &parent->kmem);
4297 * No need to take a reference to the parent because cgroup
4298 * core guarantees its existence.
4301 page_counter_init(&memcg->memory, NULL);
4302 memcg->high = PAGE_COUNTER_MAX;
4303 memcg->soft_limit = PAGE_COUNTER_MAX;
4304 page_counter_init(&memcg->memsw, NULL);
4305 page_counter_init(&memcg->kmem, NULL);
4307 * Deeper hierachy with use_hierarchy == false doesn't make
4308 * much sense so let cgroup subsystem know about this
4309 * unfortunate state in our controller.
4311 if (parent != root_mem_cgroup)
4312 memory_cgrp_subsys.broken_hierarchy = true;
4314 mutex_unlock(&memcg_create_mutex);
4316 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4321 * Make sure the memcg is initialized: mem_cgroup_iter()
4322 * orders reading memcg->initialized against its callers
4323 * reading the memcg members.
4325 smp_store_release(&memcg->initialized, 1);
4330 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4332 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4333 struct mem_cgroup_event *event, *tmp;
4336 * Unregister events and notify userspace.
4337 * Notify userspace about cgroup removing only after rmdir of cgroup
4338 * directory to avoid race between userspace and kernelspace.
4340 spin_lock(&memcg->event_list_lock);
4341 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4342 list_del_init(&event->list);
4343 schedule_work(&event->remove);
4345 spin_unlock(&memcg->event_list_lock);
4347 vmpressure_cleanup(&memcg->vmpressure);
4349 memcg_deactivate_kmem(memcg);
4351 wb_memcg_offline(memcg);
4354 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4356 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4358 invalidate_reclaim_iterators(memcg);
4361 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4363 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4365 memcg_destroy_kmem(memcg);
4366 __mem_cgroup_free(memcg);
4370 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4371 * @css: the target css
4373 * Reset the states of the mem_cgroup associated with @css. This is
4374 * invoked when the userland requests disabling on the default hierarchy
4375 * but the memcg is pinned through dependency. The memcg should stop
4376 * applying policies and should revert to the vanilla state as it may be
4377 * made visible again.
4379 * The current implementation only resets the essential configurations.
4380 * This needs to be expanded to cover all the visible parts.
4382 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4384 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4386 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4387 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4388 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4390 memcg->high = PAGE_COUNTER_MAX;
4391 memcg->soft_limit = PAGE_COUNTER_MAX;
4392 memcg_wb_domain_size_changed(memcg);
4396 /* Handlers for move charge at task migration. */
4397 static int mem_cgroup_do_precharge(unsigned long count)
4401 /* Try a single bulk charge without reclaim first, kswapd may wake */
4402 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4404 mc.precharge += count;
4408 /* Try charges one by one with reclaim */
4410 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4420 * get_mctgt_type - get target type of moving charge
4421 * @vma: the vma the pte to be checked belongs
4422 * @addr: the address corresponding to the pte to be checked
4423 * @ptent: the pte to be checked
4424 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4427 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4428 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4429 * move charge. if @target is not NULL, the page is stored in target->page
4430 * with extra refcnt got(Callers should handle it).
4431 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4432 * target for charge migration. if @target is not NULL, the entry is stored
4435 * Called with pte lock held.
4442 enum mc_target_type {
4448 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4449 unsigned long addr, pte_t ptent)
4451 struct page *page = vm_normal_page(vma, addr, ptent);
4453 if (!page || !page_mapped(page))
4455 if (PageAnon(page)) {
4456 if (!(mc.flags & MOVE_ANON))
4459 if (!(mc.flags & MOVE_FILE))
4462 if (!get_page_unless_zero(page))
4469 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4470 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4472 struct page *page = NULL;
4473 swp_entry_t ent = pte_to_swp_entry(ptent);
4475 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4478 * Because lookup_swap_cache() updates some statistics counter,
4479 * we call find_get_page() with swapper_space directly.
4481 page = find_get_page(swap_address_space(ent), ent.val);
4482 if (do_swap_account)
4483 entry->val = ent.val;
4488 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4489 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4495 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4496 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4498 struct page *page = NULL;
4499 struct address_space *mapping;
4502 if (!vma->vm_file) /* anonymous vma */
4504 if (!(mc.flags & MOVE_FILE))
4507 mapping = vma->vm_file->f_mapping;
4508 pgoff = linear_page_index(vma, addr);
4510 /* page is moved even if it's not RSS of this task(page-faulted). */
4512 /* shmem/tmpfs may report page out on swap: account for that too. */
4513 if (shmem_mapping(mapping)) {
4514 page = find_get_entry(mapping, pgoff);
4515 if (radix_tree_exceptional_entry(page)) {
4516 swp_entry_t swp = radix_to_swp_entry(page);
4517 if (do_swap_account)
4519 page = find_get_page(swap_address_space(swp), swp.val);
4522 page = find_get_page(mapping, pgoff);
4524 page = find_get_page(mapping, pgoff);
4530 * mem_cgroup_move_account - move account of the page
4532 * @nr_pages: number of regular pages (>1 for huge pages)
4533 * @from: mem_cgroup which the page is moved from.
4534 * @to: mem_cgroup which the page is moved to. @from != @to.
4536 * The caller must confirm following.
4537 * - page is not on LRU (isolate_page() is useful.)
4538 * - compound_lock is held when nr_pages > 1
4540 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4543 static int mem_cgroup_move_account(struct page *page,
4544 unsigned int nr_pages,
4545 struct mem_cgroup *from,
4546 struct mem_cgroup *to)
4548 unsigned long flags;
4552 VM_BUG_ON(from == to);
4553 VM_BUG_ON_PAGE(PageLRU(page), page);
4555 * The page is isolated from LRU. So, collapse function
4556 * will not handle this page. But page splitting can happen.
4557 * Do this check under compound_page_lock(). The caller should
4561 if (nr_pages > 1 && !PageTransHuge(page))
4565 * Prevent mem_cgroup_replace_page() from looking at
4566 * page->mem_cgroup of its source page while we change it.
4568 if (!trylock_page(page))
4572 if (page->mem_cgroup != from)
4575 anon = PageAnon(page);
4577 spin_lock_irqsave(&from->move_lock, flags);
4579 if (!anon && page_mapped(page)) {
4580 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4582 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4587 * move_lock grabbed above and caller set from->moving_account, so
4588 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4589 * So mapping should be stable for dirty pages.
4591 if (!anon && PageDirty(page)) {
4592 struct address_space *mapping = page_mapping(page);
4594 if (mapping_cap_account_dirty(mapping)) {
4595 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4597 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4602 if (PageWriteback(page)) {
4603 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4605 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4610 * It is safe to change page->mem_cgroup here because the page
4611 * is referenced, charged, and isolated - we can't race with
4612 * uncharging, charging, migration, or LRU putback.
4615 /* caller should have done css_get */
4616 page->mem_cgroup = to;
4617 spin_unlock_irqrestore(&from->move_lock, flags);
4621 local_irq_disable();
4622 mem_cgroup_charge_statistics(to, page, nr_pages);
4623 memcg_check_events(to, page);
4624 mem_cgroup_charge_statistics(from, page, -nr_pages);
4625 memcg_check_events(from, page);
4633 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4634 unsigned long addr, pte_t ptent, union mc_target *target)
4636 struct page *page = NULL;
4637 enum mc_target_type ret = MC_TARGET_NONE;
4638 swp_entry_t ent = { .val = 0 };
4640 if (pte_present(ptent))
4641 page = mc_handle_present_pte(vma, addr, ptent);
4642 else if (is_swap_pte(ptent))
4643 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4644 else if (pte_none(ptent))
4645 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4647 if (!page && !ent.val)
4651 * Do only loose check w/o serialization.
4652 * mem_cgroup_move_account() checks the page is valid or
4653 * not under LRU exclusion.
4655 if (page->mem_cgroup == mc.from) {
4656 ret = MC_TARGET_PAGE;
4658 target->page = page;
4660 if (!ret || !target)
4663 /* There is a swap entry and a page doesn't exist or isn't charged */
4664 if (ent.val && !ret &&
4665 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4666 ret = MC_TARGET_SWAP;
4673 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4675 * We don't consider swapping or file mapped pages because THP does not
4676 * support them for now.
4677 * Caller should make sure that pmd_trans_huge(pmd) is true.
4679 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4680 unsigned long addr, pmd_t pmd, union mc_target *target)
4682 struct page *page = NULL;
4683 enum mc_target_type ret = MC_TARGET_NONE;
4685 page = pmd_page(pmd);
4686 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4687 if (!(mc.flags & MOVE_ANON))
4689 if (page->mem_cgroup == mc.from) {
4690 ret = MC_TARGET_PAGE;
4693 target->page = page;
4699 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4700 unsigned long addr, pmd_t pmd, union mc_target *target)
4702 return MC_TARGET_NONE;
4706 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4707 unsigned long addr, unsigned long end,
4708 struct mm_walk *walk)
4710 struct vm_area_struct *vma = walk->vma;
4714 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4715 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4716 mc.precharge += HPAGE_PMD_NR;
4721 if (pmd_trans_unstable(pmd))
4723 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4724 for (; addr != end; pte++, addr += PAGE_SIZE)
4725 if (get_mctgt_type(vma, addr, *pte, NULL))
4726 mc.precharge++; /* increment precharge temporarily */
4727 pte_unmap_unlock(pte - 1, ptl);
4733 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4735 unsigned long precharge;
4737 struct mm_walk mem_cgroup_count_precharge_walk = {
4738 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4741 down_read(&mm->mmap_sem);
4742 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4743 up_read(&mm->mmap_sem);
4745 precharge = mc.precharge;
4751 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4753 unsigned long precharge = mem_cgroup_count_precharge(mm);
4755 VM_BUG_ON(mc.moving_task);
4756 mc.moving_task = current;
4757 return mem_cgroup_do_precharge(precharge);
4760 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4761 static void __mem_cgroup_clear_mc(void)
4763 struct mem_cgroup *from = mc.from;
4764 struct mem_cgroup *to = mc.to;
4766 /* we must uncharge all the leftover precharges from mc.to */
4768 cancel_charge(mc.to, mc.precharge);
4772 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4773 * we must uncharge here.
4775 if (mc.moved_charge) {
4776 cancel_charge(mc.from, mc.moved_charge);
4777 mc.moved_charge = 0;
4779 /* we must fixup refcnts and charges */
4780 if (mc.moved_swap) {
4781 /* uncharge swap account from the old cgroup */
4782 if (!mem_cgroup_is_root(mc.from))
4783 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4786 * we charged both to->memory and to->memsw, so we
4787 * should uncharge to->memory.
4789 if (!mem_cgroup_is_root(mc.to))
4790 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4792 css_put_many(&mc.from->css, mc.moved_swap);
4794 /* we've already done css_get(mc.to) */
4797 memcg_oom_recover(from);
4798 memcg_oom_recover(to);
4799 wake_up_all(&mc.waitq);
4802 static void mem_cgroup_clear_mc(void)
4804 struct mm_struct *mm = mc.mm;
4807 * we must clear moving_task before waking up waiters at the end of
4810 mc.moving_task = NULL;
4811 __mem_cgroup_clear_mc();
4812 spin_lock(&mc.lock);
4816 spin_unlock(&mc.lock);
4821 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4823 struct cgroup_subsys_state *css;
4824 struct mem_cgroup *memcg;
4825 struct mem_cgroup *from;
4826 struct task_struct *leader, *p;
4827 struct mm_struct *mm;
4828 unsigned long move_flags;
4831 /* charge immigration isn't supported on the default hierarchy */
4832 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4836 * Multi-process migrations only happen on the default hierarchy
4837 * where charge immigration is not used. Perform charge
4838 * immigration if @tset contains a leader and whine if there are
4842 cgroup_taskset_for_each_leader(leader, css, tset) {
4845 memcg = mem_cgroup_from_css(css);
4851 * We are now commited to this value whatever it is. Changes in this
4852 * tunable will only affect upcoming migrations, not the current one.
4853 * So we need to save it, and keep it going.
4855 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4859 from = mem_cgroup_from_task(p);
4861 VM_BUG_ON(from == memcg);
4863 mm = get_task_mm(p);
4866 /* We move charges only when we move a owner of the mm */
4867 if (mm->owner == p) {
4870 VM_BUG_ON(mc.precharge);
4871 VM_BUG_ON(mc.moved_charge);
4872 VM_BUG_ON(mc.moved_swap);
4874 spin_lock(&mc.lock);
4878 mc.flags = move_flags;
4879 spin_unlock(&mc.lock);
4880 /* We set mc.moving_task later */
4882 ret = mem_cgroup_precharge_mc(mm);
4884 mem_cgroup_clear_mc();
4891 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4894 mem_cgroup_clear_mc();
4897 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4898 unsigned long addr, unsigned long end,
4899 struct mm_walk *walk)
4902 struct vm_area_struct *vma = walk->vma;
4905 enum mc_target_type target_type;
4906 union mc_target target;
4910 * We don't take compound_lock() here but no race with splitting thp
4912 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4913 * under splitting, which means there's no concurrent thp split,
4914 * - if another thread runs into split_huge_page() just after we
4915 * entered this if-block, the thread must wait for page table lock
4916 * to be unlocked in __split_huge_page_splitting(), where the main
4917 * part of thp split is not executed yet.
4919 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4920 if (mc.precharge < HPAGE_PMD_NR) {
4924 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4925 if (target_type == MC_TARGET_PAGE) {
4927 if (!isolate_lru_page(page)) {
4928 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
4930 mc.precharge -= HPAGE_PMD_NR;
4931 mc.moved_charge += HPAGE_PMD_NR;
4933 putback_lru_page(page);
4941 if (pmd_trans_unstable(pmd))
4944 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4945 for (; addr != end; addr += PAGE_SIZE) {
4946 pte_t ptent = *(pte++);
4952 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4953 case MC_TARGET_PAGE:
4955 if (isolate_lru_page(page))
4957 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
4959 /* we uncharge from mc.from later. */
4962 putback_lru_page(page);
4963 put: /* get_mctgt_type() gets the page */
4966 case MC_TARGET_SWAP:
4968 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4970 /* we fixup refcnts and charges later. */
4978 pte_unmap_unlock(pte - 1, ptl);
4983 * We have consumed all precharges we got in can_attach().
4984 * We try charge one by one, but don't do any additional
4985 * charges to mc.to if we have failed in charge once in attach()
4988 ret = mem_cgroup_do_precharge(1);
4996 static void mem_cgroup_move_charge(void)
4998 struct mm_walk mem_cgroup_move_charge_walk = {
4999 .pmd_entry = mem_cgroup_move_charge_pte_range,
5003 lru_add_drain_all();
5005 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5006 * move_lock while we're moving its pages to another memcg.
5007 * Then wait for already started RCU-only updates to finish.
5009 atomic_inc(&mc.from->moving_account);
5012 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5014 * Someone who are holding the mmap_sem might be waiting in
5015 * waitq. So we cancel all extra charges, wake up all waiters,
5016 * and retry. Because we cancel precharges, we might not be able
5017 * to move enough charges, but moving charge is a best-effort
5018 * feature anyway, so it wouldn't be a big problem.
5020 __mem_cgroup_clear_mc();
5025 * When we have consumed all precharges and failed in doing
5026 * additional charge, the page walk just aborts.
5028 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5029 up_read(&mc.mm->mmap_sem);
5030 atomic_dec(&mc.from->moving_account);
5033 static void mem_cgroup_move_task(void)
5036 mem_cgroup_move_charge();
5037 mem_cgroup_clear_mc();
5040 #else /* !CONFIG_MMU */
5041 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5045 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5048 static void mem_cgroup_move_task(void)
5054 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5055 * to verify whether we're attached to the default hierarchy on each mount
5058 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5061 * use_hierarchy is forced on the default hierarchy. cgroup core
5062 * guarantees that @root doesn't have any children, so turning it
5063 * on for the root memcg is enough.
5065 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5066 root_mem_cgroup->use_hierarchy = true;
5068 root_mem_cgroup->use_hierarchy = false;
5071 static u64 memory_current_read(struct cgroup_subsys_state *css,
5074 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5076 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5079 static int memory_low_show(struct seq_file *m, void *v)
5081 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5082 unsigned long low = READ_ONCE(memcg->low);
5084 if (low == PAGE_COUNTER_MAX)
5085 seq_puts(m, "max\n");
5087 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5092 static ssize_t memory_low_write(struct kernfs_open_file *of,
5093 char *buf, size_t nbytes, loff_t off)
5095 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5099 buf = strstrip(buf);
5100 err = page_counter_memparse(buf, "max", &low);
5109 static int memory_high_show(struct seq_file *m, void *v)
5111 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5112 unsigned long high = READ_ONCE(memcg->high);
5114 if (high == PAGE_COUNTER_MAX)
5115 seq_puts(m, "max\n");
5117 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5122 static ssize_t memory_high_write(struct kernfs_open_file *of,
5123 char *buf, size_t nbytes, loff_t off)
5125 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5126 unsigned long nr_pages;
5130 buf = strstrip(buf);
5131 err = page_counter_memparse(buf, "max", &high);
5137 nr_pages = page_counter_read(&memcg->memory);
5138 if (nr_pages > high)
5139 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5142 memcg_wb_domain_size_changed(memcg);
5146 static int memory_max_show(struct seq_file *m, void *v)
5148 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5149 unsigned long max = READ_ONCE(memcg->memory.limit);
5151 if (max == PAGE_COUNTER_MAX)
5152 seq_puts(m, "max\n");
5154 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5159 static ssize_t memory_max_write(struct kernfs_open_file *of,
5160 char *buf, size_t nbytes, loff_t off)
5162 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5163 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5164 bool drained = false;
5168 buf = strstrip(buf);
5169 err = page_counter_memparse(buf, "max", &max);
5173 xchg(&memcg->memory.limit, max);
5176 unsigned long nr_pages = page_counter_read(&memcg->memory);
5178 if (nr_pages <= max)
5181 if (signal_pending(current)) {
5187 drain_all_stock(memcg);
5193 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5199 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5200 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5204 memcg_wb_domain_size_changed(memcg);
5208 static int memory_events_show(struct seq_file *m, void *v)
5210 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5212 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5213 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5214 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5215 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5220 static struct cftype memory_files[] = {
5223 .flags = CFTYPE_NOT_ON_ROOT,
5224 .read_u64 = memory_current_read,
5228 .flags = CFTYPE_NOT_ON_ROOT,
5229 .seq_show = memory_low_show,
5230 .write = memory_low_write,
5234 .flags = CFTYPE_NOT_ON_ROOT,
5235 .seq_show = memory_high_show,
5236 .write = memory_high_write,
5240 .flags = CFTYPE_NOT_ON_ROOT,
5241 .seq_show = memory_max_show,
5242 .write = memory_max_write,
5246 .flags = CFTYPE_NOT_ON_ROOT,
5247 .file_offset = offsetof(struct mem_cgroup, events_file),
5248 .seq_show = memory_events_show,
5253 struct cgroup_subsys memory_cgrp_subsys = {
5254 .css_alloc = mem_cgroup_css_alloc,
5255 .css_online = mem_cgroup_css_online,
5256 .css_offline = mem_cgroup_css_offline,
5257 .css_released = mem_cgroup_css_released,
5258 .css_free = mem_cgroup_css_free,
5259 .css_reset = mem_cgroup_css_reset,
5260 .can_attach = mem_cgroup_can_attach,
5261 .cancel_attach = mem_cgroup_cancel_attach,
5262 .post_attach = mem_cgroup_move_task,
5263 .bind = mem_cgroup_bind,
5264 .dfl_cftypes = memory_files,
5265 .legacy_cftypes = mem_cgroup_legacy_files,
5270 * mem_cgroup_low - check if memory consumption is below the normal range
5271 * @root: the highest ancestor to consider
5272 * @memcg: the memory cgroup to check
5274 * Returns %true if memory consumption of @memcg, and that of all
5275 * configurable ancestors up to @root, is below the normal range.
5277 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5279 if (mem_cgroup_disabled())
5283 * The toplevel group doesn't have a configurable range, so
5284 * it's never low when looked at directly, and it is not
5285 * considered an ancestor when assessing the hierarchy.
5288 if (memcg == root_mem_cgroup)
5291 if (page_counter_read(&memcg->memory) >= memcg->low)
5294 while (memcg != root) {
5295 memcg = parent_mem_cgroup(memcg);
5297 if (memcg == root_mem_cgroup)
5300 if (page_counter_read(&memcg->memory) >= memcg->low)
5307 * mem_cgroup_try_charge - try charging a page
5308 * @page: page to charge
5309 * @mm: mm context of the victim
5310 * @gfp_mask: reclaim mode
5311 * @memcgp: charged memcg return
5313 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5314 * pages according to @gfp_mask if necessary.
5316 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5317 * Otherwise, an error code is returned.
5319 * After page->mapping has been set up, the caller must finalize the
5320 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5321 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5323 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5324 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5326 struct mem_cgroup *memcg = NULL;
5327 unsigned int nr_pages = 1;
5330 if (mem_cgroup_disabled())
5333 if (PageSwapCache(page)) {
5335 * Every swap fault against a single page tries to charge the
5336 * page, bail as early as possible. shmem_unuse() encounters
5337 * already charged pages, too. The USED bit is protected by
5338 * the page lock, which serializes swap cache removal, which
5339 * in turn serializes uncharging.
5341 VM_BUG_ON_PAGE(!PageLocked(page), page);
5342 if (page->mem_cgroup)
5345 if (do_swap_account) {
5346 swp_entry_t ent = { .val = page_private(page), };
5347 unsigned short id = lookup_swap_cgroup_id(ent);
5350 memcg = mem_cgroup_from_id(id);
5351 if (memcg && !css_tryget_online(&memcg->css))
5357 if (PageTransHuge(page)) {
5358 nr_pages <<= compound_order(page);
5359 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5363 memcg = get_mem_cgroup_from_mm(mm);
5365 ret = try_charge(memcg, gfp_mask, nr_pages);
5367 css_put(&memcg->css);
5374 * mem_cgroup_commit_charge - commit a page charge
5375 * @page: page to charge
5376 * @memcg: memcg to charge the page to
5377 * @lrucare: page might be on LRU already
5379 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5380 * after page->mapping has been set up. This must happen atomically
5381 * as part of the page instantiation, i.e. under the page table lock
5382 * for anonymous pages, under the page lock for page and swap cache.
5384 * In addition, the page must not be on the LRU during the commit, to
5385 * prevent racing with task migration. If it might be, use @lrucare.
5387 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5389 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5392 unsigned int nr_pages = 1;
5394 VM_BUG_ON_PAGE(!page->mapping, page);
5395 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5397 if (mem_cgroup_disabled())
5400 * Swap faults will attempt to charge the same page multiple
5401 * times. But reuse_swap_page() might have removed the page
5402 * from swapcache already, so we can't check PageSwapCache().
5407 commit_charge(page, memcg, lrucare);
5409 if (PageTransHuge(page)) {
5410 nr_pages <<= compound_order(page);
5411 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5414 local_irq_disable();
5415 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5416 memcg_check_events(memcg, page);
5419 if (do_swap_account && PageSwapCache(page)) {
5420 swp_entry_t entry = { .val = page_private(page) };
5422 * The swap entry might not get freed for a long time,
5423 * let's not wait for it. The page already received a
5424 * memory+swap charge, drop the swap entry duplicate.
5426 mem_cgroup_uncharge_swap(entry);
5431 * mem_cgroup_cancel_charge - cancel a page charge
5432 * @page: page to charge
5433 * @memcg: memcg to charge the page to
5435 * Cancel a charge transaction started by mem_cgroup_try_charge().
5437 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5439 unsigned int nr_pages = 1;
5441 if (mem_cgroup_disabled())
5444 * Swap faults will attempt to charge the same page multiple
5445 * times. But reuse_swap_page() might have removed the page
5446 * from swapcache already, so we can't check PageSwapCache().
5451 if (PageTransHuge(page)) {
5452 nr_pages <<= compound_order(page);
5453 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5456 cancel_charge(memcg, nr_pages);
5459 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5460 unsigned long nr_anon, unsigned long nr_file,
5461 unsigned long nr_huge, struct page *dummy_page)
5463 unsigned long nr_pages = nr_anon + nr_file;
5464 unsigned long flags;
5466 if (!mem_cgroup_is_root(memcg)) {
5467 page_counter_uncharge(&memcg->memory, nr_pages);
5468 if (do_swap_account)
5469 page_counter_uncharge(&memcg->memsw, nr_pages);
5470 memcg_oom_recover(memcg);
5473 local_irq_save(flags);
5474 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5475 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5476 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5477 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5478 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5479 memcg_check_events(memcg, dummy_page);
5480 local_irq_restore(flags);
5482 if (!mem_cgroup_is_root(memcg))
5483 css_put_many(&memcg->css, nr_pages);
5486 static void uncharge_list(struct list_head *page_list)
5488 struct mem_cgroup *memcg = NULL;
5489 unsigned long nr_anon = 0;
5490 unsigned long nr_file = 0;
5491 unsigned long nr_huge = 0;
5492 unsigned long pgpgout = 0;
5493 struct list_head *next;
5496 next = page_list->next;
5498 unsigned int nr_pages = 1;
5500 page = list_entry(next, struct page, lru);
5501 next = page->lru.next;
5503 VM_BUG_ON_PAGE(PageLRU(page), page);
5504 VM_BUG_ON_PAGE(page_count(page), page);
5506 if (!page->mem_cgroup)
5510 * Nobody should be changing or seriously looking at
5511 * page->mem_cgroup at this point, we have fully
5512 * exclusive access to the page.
5515 if (memcg != page->mem_cgroup) {
5517 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5519 pgpgout = nr_anon = nr_file = nr_huge = 0;
5521 memcg = page->mem_cgroup;
5524 if (PageTransHuge(page)) {
5525 nr_pages <<= compound_order(page);
5526 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5527 nr_huge += nr_pages;
5531 nr_anon += nr_pages;
5533 nr_file += nr_pages;
5535 page->mem_cgroup = NULL;
5538 } while (next != page_list);
5541 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5546 * mem_cgroup_uncharge - uncharge a page
5547 * @page: page to uncharge
5549 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5550 * mem_cgroup_commit_charge().
5552 void mem_cgroup_uncharge(struct page *page)
5554 if (mem_cgroup_disabled())
5557 /* Don't touch page->lru of any random page, pre-check: */
5558 if (!page->mem_cgroup)
5561 INIT_LIST_HEAD(&page->lru);
5562 uncharge_list(&page->lru);
5566 * mem_cgroup_uncharge_list - uncharge a list of page
5567 * @page_list: list of pages to uncharge
5569 * Uncharge a list of pages previously charged with
5570 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5572 void mem_cgroup_uncharge_list(struct list_head *page_list)
5574 if (mem_cgroup_disabled())
5577 if (!list_empty(page_list))
5578 uncharge_list(page_list);
5582 * mem_cgroup_replace_page - migrate a charge to another page
5583 * @oldpage: currently charged page
5584 * @newpage: page to transfer the charge to
5586 * Migrate the charge from @oldpage to @newpage.
5588 * Both pages must be locked, @newpage->mapping must be set up.
5589 * Either or both pages might be on the LRU already.
5591 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5593 struct mem_cgroup *memcg;
5596 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5597 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5598 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5599 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5602 if (mem_cgroup_disabled())
5605 /* Page cache replacement: new page already charged? */
5606 if (newpage->mem_cgroup)
5609 /* Swapcache readahead pages can get replaced before being charged */
5610 memcg = oldpage->mem_cgroup;
5614 lock_page_lru(oldpage, &isolated);
5615 oldpage->mem_cgroup = NULL;
5616 unlock_page_lru(oldpage, isolated);
5618 commit_charge(newpage, memcg, true);
5622 * subsys_initcall() for memory controller.
5624 * Some parts like hotcpu_notifier() have to be initialized from this context
5625 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5626 * everything that doesn't depend on a specific mem_cgroup structure should
5627 * be initialized from here.
5629 static int __init mem_cgroup_init(void)
5633 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5635 for_each_possible_cpu(cpu)
5636 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5639 for_each_node(node) {
5640 struct mem_cgroup_tree_per_node *rtpn;
5643 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5644 node_online(node) ? node : NUMA_NO_NODE);
5646 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5647 struct mem_cgroup_tree_per_zone *rtpz;
5649 rtpz = &rtpn->rb_tree_per_zone[zone];
5650 rtpz->rb_root = RB_ROOT;
5651 spin_lock_init(&rtpz->lock);
5653 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5658 subsys_initcall(mem_cgroup_init);
5660 #ifdef CONFIG_MEMCG_SWAP
5662 * mem_cgroup_swapout - transfer a memsw charge to swap
5663 * @page: page whose memsw charge to transfer
5664 * @entry: swap entry to move the charge to
5666 * Transfer the memsw charge of @page to @entry.
5668 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5670 struct mem_cgroup *memcg;
5671 unsigned short oldid;
5673 VM_BUG_ON_PAGE(PageLRU(page), page);
5674 VM_BUG_ON_PAGE(page_count(page), page);
5676 if (!do_swap_account)
5679 memcg = page->mem_cgroup;
5681 /* Readahead page, never charged */
5685 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5686 VM_BUG_ON_PAGE(oldid, page);
5687 mem_cgroup_swap_statistics(memcg, true);
5689 page->mem_cgroup = NULL;
5691 if (!mem_cgroup_is_root(memcg))
5692 page_counter_uncharge(&memcg->memory, 1);
5695 * Interrupts should be disabled here because the caller holds the
5696 * mapping->tree_lock lock which is taken with interrupts-off. It is
5697 * important here to have the interrupts disabled because it is the
5698 * only synchronisation we have for udpating the per-CPU variables.
5700 VM_BUG_ON(!irqs_disabled());
5701 mem_cgroup_charge_statistics(memcg, page, -1);
5702 memcg_check_events(memcg, page);
5706 * mem_cgroup_uncharge_swap - uncharge a swap entry
5707 * @entry: swap entry to uncharge
5709 * Drop the memsw charge associated with @entry.
5711 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5713 struct mem_cgroup *memcg;
5716 if (!do_swap_account)
5719 id = swap_cgroup_record(entry, 0);
5721 memcg = mem_cgroup_from_id(id);
5723 if (!mem_cgroup_is_root(memcg))
5724 page_counter_uncharge(&memcg->memsw, 1);
5725 mem_cgroup_swap_statistics(memcg, false);
5726 css_put(&memcg->css);
5731 /* for remember boot option*/
5732 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5733 static int really_do_swap_account __initdata = 1;
5735 static int really_do_swap_account __initdata;
5738 static int __init enable_swap_account(char *s)
5740 if (!strcmp(s, "1"))
5741 really_do_swap_account = 1;
5742 else if (!strcmp(s, "0"))
5743 really_do_swap_account = 0;
5746 __setup("swapaccount=", enable_swap_account);
5748 static struct cftype memsw_cgroup_files[] = {
5750 .name = "memsw.usage_in_bytes",
5751 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5752 .read_u64 = mem_cgroup_read_u64,
5755 .name = "memsw.max_usage_in_bytes",
5756 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5757 .write = mem_cgroup_reset,
5758 .read_u64 = mem_cgroup_read_u64,
5761 .name = "memsw.limit_in_bytes",
5762 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5763 .write = mem_cgroup_write,
5764 .read_u64 = mem_cgroup_read_u64,
5767 .name = "memsw.failcnt",
5768 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5769 .write = mem_cgroup_reset,
5770 .read_u64 = mem_cgroup_read_u64,
5772 { }, /* terminate */
5775 static int __init mem_cgroup_swap_init(void)
5777 if (!mem_cgroup_disabled() && really_do_swap_account) {
5778 do_swap_account = 1;
5779 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5780 memsw_cgroup_files));
5784 subsys_initcall(mem_cgroup_swap_init);
5786 #endif /* CONFIG_MEMCG_SWAP */