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 mem_cgroup *from;
200 struct mem_cgroup *to;
202 unsigned long precharge;
203 unsigned long moved_charge;
204 unsigned long moved_swap;
205 struct task_struct *moving_task; /* a task moving charges */
206 wait_queue_head_t waitq; /* a waitq for other context */
208 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
209 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
213 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
214 * limit reclaim to prevent infinite loops, if they ever occur.
216 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
217 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
220 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
221 MEM_CGROUP_CHARGE_TYPE_ANON,
222 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
223 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
227 /* for encoding cft->private value on file */
235 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
236 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
237 #define MEMFILE_ATTR(val) ((val) & 0xffff)
238 /* Used for OOM nofiier */
239 #define OOM_CONTROL (0)
242 * The memcg_create_mutex will be held whenever a new cgroup is created.
243 * As a consequence, any change that needs to protect against new child cgroups
244 * appearing has to hold it as well.
246 static DEFINE_MUTEX(memcg_create_mutex);
248 /* Some nice accessors for the vmpressure. */
249 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
252 memcg = root_mem_cgroup;
253 return &memcg->vmpressure;
256 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
258 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
261 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
263 return (memcg == root_mem_cgroup);
267 * We restrict the id in the range of [1, 65535], so it can fit into
270 #define MEM_CGROUP_ID_MAX USHRT_MAX
272 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
274 return memcg->css.id;
278 * A helper function to get mem_cgroup from ID. must be called under
279 * rcu_read_lock(). The caller is responsible for calling
280 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
281 * refcnt from swap can be called against removed memcg.)
283 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
285 struct cgroup_subsys_state *css;
287 css = css_from_id(id, &memory_cgrp_subsys);
288 return mem_cgroup_from_css(css);
291 /* Writing them here to avoid exposing memcg's inner layout */
292 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
294 void sock_update_memcg(struct sock *sk)
296 if (mem_cgroup_sockets_enabled) {
297 struct mem_cgroup *memcg;
298 struct cg_proto *cg_proto;
300 BUG_ON(!sk->sk_prot->proto_cgroup);
302 /* Socket cloning can throw us here with sk_cgrp already
303 * filled. It won't however, necessarily happen from
304 * process context. So the test for root memcg given
305 * the current task's memcg won't help us in this case.
307 * Respecting the original socket's memcg is a better
308 * decision in this case.
311 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
312 css_get(&sk->sk_cgrp->memcg->css);
317 memcg = mem_cgroup_from_task(current);
318 cg_proto = sk->sk_prot->proto_cgroup(memcg);
319 if (cg_proto && test_bit(MEMCG_SOCK_ACTIVE, &cg_proto->flags) &&
320 css_tryget_online(&memcg->css)) {
321 sk->sk_cgrp = cg_proto;
326 EXPORT_SYMBOL(sock_update_memcg);
328 void sock_release_memcg(struct sock *sk)
330 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
331 struct mem_cgroup *memcg;
332 WARN_ON(!sk->sk_cgrp->memcg);
333 memcg = sk->sk_cgrp->memcg;
334 css_put(&sk->sk_cgrp->memcg->css);
338 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
340 if (!memcg || mem_cgroup_is_root(memcg))
343 return &memcg->tcp_mem;
345 EXPORT_SYMBOL(tcp_proto_cgroup);
349 #ifdef CONFIG_MEMCG_KMEM
351 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida);
362 int memcg_nr_cache_ids;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 struct static_key memcg_kmem_enabled_key;
399 EXPORT_SYMBOL(memcg_kmem_enabled_key);
401 #endif /* CONFIG_MEMCG_KMEM */
403 static struct mem_cgroup_per_zone *
404 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
406 int nid = zone_to_nid(zone);
407 int zid = zone_idx(zone);
409 return &memcg->nodeinfo[nid]->zoneinfo[zid];
413 * mem_cgroup_css_from_page - css of the memcg associated with a page
414 * @page: page of interest
416 * If memcg is bound to the default hierarchy, css of the memcg associated
417 * with @page is returned. The returned css remains associated with @page
418 * until it is released.
420 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
423 * XXX: The above description of behavior on the default hierarchy isn't
424 * strictly true yet as replace_page_cache_page() can modify the
425 * association before @page is released even on the default hierarchy;
426 * however, the current and planned usages don't mix the the two functions
427 * and replace_page_cache_page() will soon be updated to make the invariant
430 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
432 struct mem_cgroup *memcg;
436 memcg = page->mem_cgroup;
438 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
439 memcg = root_mem_cgroup;
446 * page_cgroup_ino - return inode number of the memcg a page is charged to
449 * Look up the closest online ancestor of the memory cgroup @page is charged to
450 * and return its inode number or 0 if @page is not charged to any cgroup. It
451 * is safe to call this function without holding a reference to @page.
453 * Note, this function is inherently racy, because there is nothing to prevent
454 * the cgroup inode from getting torn down and potentially reallocated a moment
455 * after page_cgroup_ino() returns, so it only should be used by callers that
456 * do not care (such as procfs interfaces).
458 ino_t page_cgroup_ino(struct page *page)
460 struct mem_cgroup *memcg;
461 unsigned long ino = 0;
464 memcg = READ_ONCE(page->mem_cgroup);
465 while (memcg && !(memcg->css.flags & CSS_ONLINE))
466 memcg = parent_mem_cgroup(memcg);
468 ino = cgroup_ino(memcg->css.cgroup);
473 static struct mem_cgroup_per_zone *
474 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
476 int nid = page_to_nid(page);
477 int zid = page_zonenum(page);
479 return &memcg->nodeinfo[nid]->zoneinfo[zid];
482 static struct mem_cgroup_tree_per_zone *
483 soft_limit_tree_node_zone(int nid, int zid)
485 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
488 static struct mem_cgroup_tree_per_zone *
489 soft_limit_tree_from_page(struct page *page)
491 int nid = page_to_nid(page);
492 int zid = page_zonenum(page);
494 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
497 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
498 struct mem_cgroup_tree_per_zone *mctz,
499 unsigned long new_usage_in_excess)
501 struct rb_node **p = &mctz->rb_root.rb_node;
502 struct rb_node *parent = NULL;
503 struct mem_cgroup_per_zone *mz_node;
508 mz->usage_in_excess = new_usage_in_excess;
509 if (!mz->usage_in_excess)
513 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
515 if (mz->usage_in_excess < mz_node->usage_in_excess)
518 * We can't avoid mem cgroups that are over their soft
519 * limit by the same amount
521 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
524 rb_link_node(&mz->tree_node, parent, p);
525 rb_insert_color(&mz->tree_node, &mctz->rb_root);
529 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
530 struct mem_cgroup_tree_per_zone *mctz)
534 rb_erase(&mz->tree_node, &mctz->rb_root);
538 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
539 struct mem_cgroup_tree_per_zone *mctz)
543 spin_lock_irqsave(&mctz->lock, flags);
544 __mem_cgroup_remove_exceeded(mz, mctz);
545 spin_unlock_irqrestore(&mctz->lock, flags);
548 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
550 unsigned long nr_pages = page_counter_read(&memcg->memory);
551 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
552 unsigned long excess = 0;
554 if (nr_pages > soft_limit)
555 excess = nr_pages - soft_limit;
560 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
562 unsigned long excess;
563 struct mem_cgroup_per_zone *mz;
564 struct mem_cgroup_tree_per_zone *mctz;
566 mctz = soft_limit_tree_from_page(page);
568 * Necessary to update all ancestors when hierarchy is used.
569 * because their event counter is not touched.
571 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
572 mz = mem_cgroup_page_zoneinfo(memcg, page);
573 excess = soft_limit_excess(memcg);
575 * We have to update the tree if mz is on RB-tree or
576 * mem is over its softlimit.
578 if (excess || mz->on_tree) {
581 spin_lock_irqsave(&mctz->lock, flags);
582 /* if on-tree, remove it */
584 __mem_cgroup_remove_exceeded(mz, mctz);
586 * Insert again. mz->usage_in_excess will be updated.
587 * If excess is 0, no tree ops.
589 __mem_cgroup_insert_exceeded(mz, mctz, excess);
590 spin_unlock_irqrestore(&mctz->lock, flags);
595 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
597 struct mem_cgroup_tree_per_zone *mctz;
598 struct mem_cgroup_per_zone *mz;
602 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
603 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
604 mctz = soft_limit_tree_node_zone(nid, zid);
605 mem_cgroup_remove_exceeded(mz, mctz);
610 static struct mem_cgroup_per_zone *
611 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
613 struct rb_node *rightmost = NULL;
614 struct mem_cgroup_per_zone *mz;
618 rightmost = rb_last(&mctz->rb_root);
620 goto done; /* Nothing to reclaim from */
622 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
624 * Remove the node now but someone else can add it back,
625 * we will to add it back at the end of reclaim to its correct
626 * position in the tree.
628 __mem_cgroup_remove_exceeded(mz, mctz);
629 if (!soft_limit_excess(mz->memcg) ||
630 !css_tryget_online(&mz->memcg->css))
636 static struct mem_cgroup_per_zone *
637 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
639 struct mem_cgroup_per_zone *mz;
641 spin_lock_irq(&mctz->lock);
642 mz = __mem_cgroup_largest_soft_limit_node(mctz);
643 spin_unlock_irq(&mctz->lock);
648 * Return page count for single (non recursive) @memcg.
650 * Implementation Note: reading percpu statistics for memcg.
652 * Both of vmstat[] and percpu_counter has threshold and do periodic
653 * synchronization to implement "quick" read. There are trade-off between
654 * reading cost and precision of value. Then, we may have a chance to implement
655 * a periodic synchronization of counter in memcg's counter.
657 * But this _read() function is used for user interface now. The user accounts
658 * memory usage by memory cgroup and he _always_ requires exact value because
659 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
660 * have to visit all online cpus and make sum. So, for now, unnecessary
661 * synchronization is not implemented. (just implemented for cpu hotplug)
663 * If there are kernel internal actions which can make use of some not-exact
664 * value, and reading all cpu value can be performance bottleneck in some
665 * common workload, threshold and synchronization as vmstat[] should be
669 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
674 /* Per-cpu values can be negative, use a signed accumulator */
675 for_each_possible_cpu(cpu)
676 val += per_cpu(memcg->stat->count[idx], cpu);
678 * Summing races with updates, so val may be negative. Avoid exposing
679 * transient negative values.
686 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
687 enum mem_cgroup_events_index idx)
689 unsigned long val = 0;
692 for_each_possible_cpu(cpu)
693 val += per_cpu(memcg->stat->events[idx], cpu);
697 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
699 bool compound, int nr_pages)
702 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
703 * counted as CACHE even if it's on ANON LRU.
706 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
709 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
713 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
714 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
718 /* pagein of a big page is an event. So, ignore page size */
720 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
722 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
723 nr_pages = -nr_pages; /* for event */
726 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
729 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
731 unsigned int lru_mask)
733 unsigned long nr = 0;
736 VM_BUG_ON((unsigned)nid >= nr_node_ids);
738 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
739 struct mem_cgroup_per_zone *mz;
743 if (!(BIT(lru) & lru_mask))
745 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
746 nr += mz->lru_size[lru];
752 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
753 unsigned int lru_mask)
755 unsigned long nr = 0;
758 for_each_node_state(nid, N_MEMORY)
759 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
763 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
764 enum mem_cgroup_events_target target)
766 unsigned long val, next;
768 val = __this_cpu_read(memcg->stat->nr_page_events);
769 next = __this_cpu_read(memcg->stat->targets[target]);
770 /* from time_after() in jiffies.h */
771 if ((long)next - (long)val < 0) {
773 case MEM_CGROUP_TARGET_THRESH:
774 next = val + THRESHOLDS_EVENTS_TARGET;
776 case MEM_CGROUP_TARGET_SOFTLIMIT:
777 next = val + SOFTLIMIT_EVENTS_TARGET;
779 case MEM_CGROUP_TARGET_NUMAINFO:
780 next = val + NUMAINFO_EVENTS_TARGET;
785 __this_cpu_write(memcg->stat->targets[target], next);
792 * Check events in order.
795 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
797 /* threshold event is triggered in finer grain than soft limit */
798 if (unlikely(mem_cgroup_event_ratelimit(memcg,
799 MEM_CGROUP_TARGET_THRESH))) {
801 bool do_numainfo __maybe_unused;
803 do_softlimit = mem_cgroup_event_ratelimit(memcg,
804 MEM_CGROUP_TARGET_SOFTLIMIT);
806 do_numainfo = mem_cgroup_event_ratelimit(memcg,
807 MEM_CGROUP_TARGET_NUMAINFO);
809 mem_cgroup_threshold(memcg);
810 if (unlikely(do_softlimit))
811 mem_cgroup_update_tree(memcg, page);
813 if (unlikely(do_numainfo))
814 atomic_inc(&memcg->numainfo_events);
819 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
822 * mm_update_next_owner() may clear mm->owner to NULL
823 * if it races with swapoff, page migration, etc.
824 * So this can be called with p == NULL.
829 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
831 EXPORT_SYMBOL(mem_cgroup_from_task);
833 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
835 struct mem_cgroup *memcg = NULL;
840 * Page cache insertions can happen withou an
841 * actual mm context, e.g. during disk probing
842 * on boot, loopback IO, acct() writes etc.
845 memcg = root_mem_cgroup;
847 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
848 if (unlikely(!memcg))
849 memcg = root_mem_cgroup;
851 } while (!css_tryget_online(&memcg->css));
857 * mem_cgroup_iter - iterate over memory cgroup hierarchy
858 * @root: hierarchy root
859 * @prev: previously returned memcg, NULL on first invocation
860 * @reclaim: cookie for shared reclaim walks, NULL for full walks
862 * Returns references to children of the hierarchy below @root, or
863 * @root itself, or %NULL after a full round-trip.
865 * Caller must pass the return value in @prev on subsequent
866 * invocations for reference counting, or use mem_cgroup_iter_break()
867 * to cancel a hierarchy walk before the round-trip is complete.
869 * Reclaimers can specify a zone and a priority level in @reclaim to
870 * divide up the memcgs in the hierarchy among all concurrent
871 * reclaimers operating on the same zone and priority.
873 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
874 struct mem_cgroup *prev,
875 struct mem_cgroup_reclaim_cookie *reclaim)
877 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
878 struct cgroup_subsys_state *css = NULL;
879 struct mem_cgroup *memcg = NULL;
880 struct mem_cgroup *pos = NULL;
882 if (mem_cgroup_disabled())
886 root = root_mem_cgroup;
888 if (prev && !reclaim)
891 if (!root->use_hierarchy && root != root_mem_cgroup) {
900 struct mem_cgroup_per_zone *mz;
902 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
903 iter = &mz->iter[reclaim->priority];
905 if (prev && reclaim->generation != iter->generation)
909 pos = READ_ONCE(iter->position);
911 * A racing update may change the position and
912 * put the last reference, hence css_tryget(),
913 * or retry to see the updated position.
915 } while (pos && !css_tryget(&pos->css));
922 css = css_next_descendant_pre(css, &root->css);
925 * Reclaimers share the hierarchy walk, and a
926 * new one might jump in right at the end of
927 * the hierarchy - make sure they see at least
928 * one group and restart from the beginning.
936 * Verify the css and acquire a reference. The root
937 * is provided by the caller, so we know it's alive
938 * and kicking, and don't take an extra reference.
940 memcg = mem_cgroup_from_css(css);
942 if (css == &root->css)
945 if (css_tryget(css)) {
947 * Make sure the memcg is initialized:
948 * mem_cgroup_css_online() orders the the
949 * initialization against setting the flag.
951 if (smp_load_acquire(&memcg->initialized))
961 if (cmpxchg(&iter->position, pos, memcg) == pos) {
963 css_get(&memcg->css);
969 * pairs with css_tryget when dereferencing iter->position
978 reclaim->generation = iter->generation;
984 if (prev && prev != root)
991 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
992 * @root: hierarchy root
993 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
995 void mem_cgroup_iter_break(struct mem_cgroup *root,
996 struct mem_cgroup *prev)
999 root = root_mem_cgroup;
1000 if (prev && prev != root)
1001 css_put(&prev->css);
1005 * Iteration constructs for visiting all cgroups (under a tree). If
1006 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1007 * be used for reference counting.
1009 #define for_each_mem_cgroup_tree(iter, root) \
1010 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1012 iter = mem_cgroup_iter(root, iter, NULL))
1014 #define for_each_mem_cgroup(iter) \
1015 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1017 iter = mem_cgroup_iter(NULL, iter, NULL))
1020 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1021 * @zone: zone of the wanted lruvec
1022 * @memcg: memcg of the wanted lruvec
1024 * Returns the lru list vector holding pages for the given @zone and
1025 * @mem. This can be the global zone lruvec, if the memory controller
1028 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1029 struct mem_cgroup *memcg)
1031 struct mem_cgroup_per_zone *mz;
1032 struct lruvec *lruvec;
1034 if (mem_cgroup_disabled()) {
1035 lruvec = &zone->lruvec;
1039 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1040 lruvec = &mz->lruvec;
1043 * Since a node can be onlined after the mem_cgroup was created,
1044 * we have to be prepared to initialize lruvec->zone here;
1045 * and if offlined then reonlined, we need to reinitialize it.
1047 if (unlikely(lruvec->zone != zone))
1048 lruvec->zone = zone;
1053 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1055 * @zone: zone of the page
1057 * This function is only safe when following the LRU page isolation
1058 * and putback protocol: the LRU lock must be held, and the page must
1059 * either be PageLRU() or the caller must have isolated/allocated it.
1061 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1063 struct mem_cgroup_per_zone *mz;
1064 struct mem_cgroup *memcg;
1065 struct lruvec *lruvec;
1067 if (mem_cgroup_disabled()) {
1068 lruvec = &zone->lruvec;
1072 memcg = page->mem_cgroup;
1074 * Swapcache readahead pages are added to the LRU - and
1075 * possibly migrated - before they are charged.
1078 memcg = root_mem_cgroup;
1080 mz = mem_cgroup_page_zoneinfo(memcg, page);
1081 lruvec = &mz->lruvec;
1084 * Since a node can be onlined after the mem_cgroup was created,
1085 * we have to be prepared to initialize lruvec->zone here;
1086 * and if offlined then reonlined, we need to reinitialize it.
1088 if (unlikely(lruvec->zone != zone))
1089 lruvec->zone = zone;
1094 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1095 * @lruvec: mem_cgroup per zone lru vector
1096 * @lru: index of lru list the page is sitting on
1097 * @nr_pages: positive when adding or negative when removing
1099 * This function must be called when a page is added to or removed from an
1102 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1105 struct mem_cgroup_per_zone *mz;
1106 unsigned long *lru_size;
1108 if (mem_cgroup_disabled())
1111 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1112 lru_size = mz->lru_size + lru;
1113 *lru_size += nr_pages;
1114 VM_BUG_ON((long)(*lru_size) < 0);
1117 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1119 struct mem_cgroup *task_memcg;
1120 struct task_struct *p;
1123 p = find_lock_task_mm(task);
1125 task_memcg = get_mem_cgroup_from_mm(p->mm);
1129 * All threads may have already detached their mm's, but the oom
1130 * killer still needs to detect if they have already been oom
1131 * killed to prevent needlessly killing additional tasks.
1134 task_memcg = mem_cgroup_from_task(task);
1135 css_get(&task_memcg->css);
1138 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1139 css_put(&task_memcg->css);
1143 #define mem_cgroup_from_counter(counter, member) \
1144 container_of(counter, struct mem_cgroup, member)
1147 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1148 * @memcg: the memory cgroup
1150 * Returns the maximum amount of memory @mem can be charged with, in
1153 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1155 unsigned long margin = 0;
1156 unsigned long count;
1157 unsigned long limit;
1159 count = page_counter_read(&memcg->memory);
1160 limit = READ_ONCE(memcg->memory.limit);
1162 margin = limit - count;
1164 if (do_swap_account) {
1165 count = page_counter_read(&memcg->memsw);
1166 limit = READ_ONCE(memcg->memsw.limit);
1168 margin = min(margin, limit - count);
1175 * A routine for checking "mem" is under move_account() or not.
1177 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1178 * moving cgroups. This is for waiting at high-memory pressure
1181 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1183 struct mem_cgroup *from;
1184 struct mem_cgroup *to;
1187 * Unlike task_move routines, we access mc.to, mc.from not under
1188 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1190 spin_lock(&mc.lock);
1196 ret = mem_cgroup_is_descendant(from, memcg) ||
1197 mem_cgroup_is_descendant(to, memcg);
1199 spin_unlock(&mc.lock);
1203 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1205 if (mc.moving_task && current != mc.moving_task) {
1206 if (mem_cgroup_under_move(memcg)) {
1208 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1209 /* moving charge context might have finished. */
1212 finish_wait(&mc.waitq, &wait);
1219 #define K(x) ((x) << (PAGE_SHIFT-10))
1221 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1222 * @memcg: The memory cgroup that went over limit
1223 * @p: Task that is going to be killed
1225 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1228 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1230 /* oom_info_lock ensures that parallel ooms do not interleave */
1231 static DEFINE_MUTEX(oom_info_lock);
1232 struct mem_cgroup *iter;
1235 mutex_lock(&oom_info_lock);
1239 pr_info("Task in ");
1240 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1241 pr_cont(" killed as a result of limit of ");
1243 pr_info("Memory limit reached of cgroup ");
1246 pr_cont_cgroup_path(memcg->css.cgroup);
1251 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1252 K((u64)page_counter_read(&memcg->memory)),
1253 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1254 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1255 K((u64)page_counter_read(&memcg->memsw)),
1256 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1257 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1258 K((u64)page_counter_read(&memcg->kmem)),
1259 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1261 for_each_mem_cgroup_tree(iter, memcg) {
1262 pr_info("Memory cgroup stats for ");
1263 pr_cont_cgroup_path(iter->css.cgroup);
1266 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1267 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1269 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1270 K(mem_cgroup_read_stat(iter, i)));
1273 for (i = 0; i < NR_LRU_LISTS; i++)
1274 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1275 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1279 mutex_unlock(&oom_info_lock);
1283 * This function returns the number of memcg under hierarchy tree. Returns
1284 * 1(self count) if no children.
1286 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1289 struct mem_cgroup *iter;
1291 for_each_mem_cgroup_tree(iter, memcg)
1297 * Return the memory (and swap, if configured) limit for a memcg.
1299 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1301 unsigned long limit;
1303 limit = memcg->memory.limit;
1304 if (mem_cgroup_swappiness(memcg)) {
1305 unsigned long memsw_limit;
1307 memsw_limit = memcg->memsw.limit;
1308 limit = min(limit + total_swap_pages, memsw_limit);
1313 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1316 struct oom_control oc = {
1319 .gfp_mask = gfp_mask,
1322 struct mem_cgroup *iter;
1323 unsigned long chosen_points = 0;
1324 unsigned long totalpages;
1325 unsigned int points = 0;
1326 struct task_struct *chosen = NULL;
1328 mutex_lock(&oom_lock);
1331 * If current has a pending SIGKILL or is exiting, then automatically
1332 * select it. The goal is to allow it to allocate so that it may
1333 * quickly exit and free its memory.
1335 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1336 mark_oom_victim(current);
1340 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1341 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1342 for_each_mem_cgroup_tree(iter, memcg) {
1343 struct css_task_iter it;
1344 struct task_struct *task;
1346 css_task_iter_start(&iter->css, &it);
1347 while ((task = css_task_iter_next(&it))) {
1348 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1349 case OOM_SCAN_SELECT:
1351 put_task_struct(chosen);
1353 chosen_points = ULONG_MAX;
1354 get_task_struct(chosen);
1356 case OOM_SCAN_CONTINUE:
1358 case OOM_SCAN_ABORT:
1359 css_task_iter_end(&it);
1360 mem_cgroup_iter_break(memcg, iter);
1362 put_task_struct(chosen);
1367 points = oom_badness(task, memcg, NULL, totalpages);
1368 if (!points || points < chosen_points)
1370 /* Prefer thread group leaders for display purposes */
1371 if (points == chosen_points &&
1372 thread_group_leader(chosen))
1376 put_task_struct(chosen);
1378 chosen_points = points;
1379 get_task_struct(chosen);
1381 css_task_iter_end(&it);
1385 points = chosen_points * 1000 / totalpages;
1386 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1387 "Memory cgroup out of memory");
1390 mutex_unlock(&oom_lock);
1393 #if MAX_NUMNODES > 1
1396 * test_mem_cgroup_node_reclaimable
1397 * @memcg: the target memcg
1398 * @nid: the node ID to be checked.
1399 * @noswap : specify true here if the user wants flle only information.
1401 * This function returns whether the specified memcg contains any
1402 * reclaimable pages on a node. Returns true if there are any reclaimable
1403 * pages in the node.
1405 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1406 int nid, bool noswap)
1408 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1410 if (noswap || !total_swap_pages)
1412 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1419 * Always updating the nodemask is not very good - even if we have an empty
1420 * list or the wrong list here, we can start from some node and traverse all
1421 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1424 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1428 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1429 * pagein/pageout changes since the last update.
1431 if (!atomic_read(&memcg->numainfo_events))
1433 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1436 /* make a nodemask where this memcg uses memory from */
1437 memcg->scan_nodes = node_states[N_MEMORY];
1439 for_each_node_mask(nid, node_states[N_MEMORY]) {
1441 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1442 node_clear(nid, memcg->scan_nodes);
1445 atomic_set(&memcg->numainfo_events, 0);
1446 atomic_set(&memcg->numainfo_updating, 0);
1450 * Selecting a node where we start reclaim from. Because what we need is just
1451 * reducing usage counter, start from anywhere is O,K. Considering
1452 * memory reclaim from current node, there are pros. and cons.
1454 * Freeing memory from current node means freeing memory from a node which
1455 * we'll use or we've used. So, it may make LRU bad. And if several threads
1456 * hit limits, it will see a contention on a node. But freeing from remote
1457 * node means more costs for memory reclaim because of memory latency.
1459 * Now, we use round-robin. Better algorithm is welcomed.
1461 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1465 mem_cgroup_may_update_nodemask(memcg);
1466 node = memcg->last_scanned_node;
1468 node = next_node(node, memcg->scan_nodes);
1469 if (node == MAX_NUMNODES)
1470 node = first_node(memcg->scan_nodes);
1472 * We call this when we hit limit, not when pages are added to LRU.
1473 * No LRU may hold pages because all pages are UNEVICTABLE or
1474 * memcg is too small and all pages are not on LRU. In that case,
1475 * we use curret node.
1477 if (unlikely(node == MAX_NUMNODES))
1478 node = numa_node_id();
1480 memcg->last_scanned_node = node;
1484 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1490 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1493 unsigned long *total_scanned)
1495 struct mem_cgroup *victim = NULL;
1498 unsigned long excess;
1499 unsigned long nr_scanned;
1500 struct mem_cgroup_reclaim_cookie reclaim = {
1505 excess = soft_limit_excess(root_memcg);
1508 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1513 * If we have not been able to reclaim
1514 * anything, it might because there are
1515 * no reclaimable pages under this hierarchy
1520 * We want to do more targeted reclaim.
1521 * excess >> 2 is not to excessive so as to
1522 * reclaim too much, nor too less that we keep
1523 * coming back to reclaim from this cgroup
1525 if (total >= (excess >> 2) ||
1526 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1531 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1533 *total_scanned += nr_scanned;
1534 if (!soft_limit_excess(root_memcg))
1537 mem_cgroup_iter_break(root_memcg, victim);
1541 #ifdef CONFIG_LOCKDEP
1542 static struct lockdep_map memcg_oom_lock_dep_map = {
1543 .name = "memcg_oom_lock",
1547 static DEFINE_SPINLOCK(memcg_oom_lock);
1550 * Check OOM-Killer is already running under our hierarchy.
1551 * If someone is running, return false.
1553 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1555 struct mem_cgroup *iter, *failed = NULL;
1557 spin_lock(&memcg_oom_lock);
1559 for_each_mem_cgroup_tree(iter, memcg) {
1560 if (iter->oom_lock) {
1562 * this subtree of our hierarchy is already locked
1563 * so we cannot give a lock.
1566 mem_cgroup_iter_break(memcg, iter);
1569 iter->oom_lock = true;
1574 * OK, we failed to lock the whole subtree so we have
1575 * to clean up what we set up to the failing subtree
1577 for_each_mem_cgroup_tree(iter, memcg) {
1578 if (iter == failed) {
1579 mem_cgroup_iter_break(memcg, iter);
1582 iter->oom_lock = false;
1585 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1587 spin_unlock(&memcg_oom_lock);
1592 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1594 struct mem_cgroup *iter;
1596 spin_lock(&memcg_oom_lock);
1597 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1598 for_each_mem_cgroup_tree(iter, memcg)
1599 iter->oom_lock = false;
1600 spin_unlock(&memcg_oom_lock);
1603 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1605 struct mem_cgroup *iter;
1607 spin_lock(&memcg_oom_lock);
1608 for_each_mem_cgroup_tree(iter, memcg)
1610 spin_unlock(&memcg_oom_lock);
1613 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1615 struct mem_cgroup *iter;
1618 * When a new child is created while the hierarchy is under oom,
1619 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1621 spin_lock(&memcg_oom_lock);
1622 for_each_mem_cgroup_tree(iter, memcg)
1623 if (iter->under_oom > 0)
1625 spin_unlock(&memcg_oom_lock);
1628 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1630 struct oom_wait_info {
1631 struct mem_cgroup *memcg;
1635 static int memcg_oom_wake_function(wait_queue_t *wait,
1636 unsigned mode, int sync, void *arg)
1638 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1639 struct mem_cgroup *oom_wait_memcg;
1640 struct oom_wait_info *oom_wait_info;
1642 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1643 oom_wait_memcg = oom_wait_info->memcg;
1645 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1646 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1648 return autoremove_wake_function(wait, mode, sync, arg);
1651 static void memcg_oom_recover(struct mem_cgroup *memcg)
1654 * For the following lockless ->under_oom test, the only required
1655 * guarantee is that it must see the state asserted by an OOM when
1656 * this function is called as a result of userland actions
1657 * triggered by the notification of the OOM. This is trivially
1658 * achieved by invoking mem_cgroup_mark_under_oom() before
1659 * triggering notification.
1661 if (memcg && memcg->under_oom)
1662 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1665 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1667 if (!current->memcg_may_oom)
1670 * We are in the middle of the charge context here, so we
1671 * don't want to block when potentially sitting on a callstack
1672 * that holds all kinds of filesystem and mm locks.
1674 * Also, the caller may handle a failed allocation gracefully
1675 * (like optional page cache readahead) and so an OOM killer
1676 * invocation might not even be necessary.
1678 * That's why we don't do anything here except remember the
1679 * OOM context and then deal with it at the end of the page
1680 * fault when the stack is unwound, the locks are released,
1681 * and when we know whether the fault was overall successful.
1683 css_get(&memcg->css);
1684 current->memcg_in_oom = memcg;
1685 current->memcg_oom_gfp_mask = mask;
1686 current->memcg_oom_order = order;
1690 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1691 * @handle: actually kill/wait or just clean up the OOM state
1693 * This has to be called at the end of a page fault if the memcg OOM
1694 * handler was enabled.
1696 * Memcg supports userspace OOM handling where failed allocations must
1697 * sleep on a waitqueue until the userspace task resolves the
1698 * situation. Sleeping directly in the charge context with all kinds
1699 * of locks held is not a good idea, instead we remember an OOM state
1700 * in the task and mem_cgroup_oom_synchronize() has to be called at
1701 * the end of the page fault to complete the OOM handling.
1703 * Returns %true if an ongoing memcg OOM situation was detected and
1704 * completed, %false otherwise.
1706 bool mem_cgroup_oom_synchronize(bool handle)
1708 struct mem_cgroup *memcg = current->memcg_in_oom;
1709 struct oom_wait_info owait;
1712 /* OOM is global, do not handle */
1716 if (!handle || oom_killer_disabled)
1719 owait.memcg = memcg;
1720 owait.wait.flags = 0;
1721 owait.wait.func = memcg_oom_wake_function;
1722 owait.wait.private = current;
1723 INIT_LIST_HEAD(&owait.wait.task_list);
1725 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1726 mem_cgroup_mark_under_oom(memcg);
1728 locked = mem_cgroup_oom_trylock(memcg);
1731 mem_cgroup_oom_notify(memcg);
1733 if (locked && !memcg->oom_kill_disable) {
1734 mem_cgroup_unmark_under_oom(memcg);
1735 finish_wait(&memcg_oom_waitq, &owait.wait);
1736 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1737 current->memcg_oom_order);
1740 mem_cgroup_unmark_under_oom(memcg);
1741 finish_wait(&memcg_oom_waitq, &owait.wait);
1745 mem_cgroup_oom_unlock(memcg);
1747 * There is no guarantee that an OOM-lock contender
1748 * sees the wakeups triggered by the OOM kill
1749 * uncharges. Wake any sleepers explicitely.
1751 memcg_oom_recover(memcg);
1754 current->memcg_in_oom = NULL;
1755 css_put(&memcg->css);
1760 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1761 * @page: page that is going to change accounted state
1763 * This function must mark the beginning of an accounted page state
1764 * change to prevent double accounting when the page is concurrently
1765 * being moved to another memcg:
1767 * memcg = mem_cgroup_begin_page_stat(page);
1768 * if (TestClearPageState(page))
1769 * mem_cgroup_update_page_stat(memcg, state, -1);
1770 * mem_cgroup_end_page_stat(memcg);
1772 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1774 struct mem_cgroup *memcg;
1775 unsigned long flags;
1778 * The RCU lock is held throughout the transaction. The fast
1779 * path can get away without acquiring the memcg->move_lock
1780 * because page moving starts with an RCU grace period.
1782 * The RCU lock also protects the memcg from being freed when
1783 * the page state that is going to change is the only thing
1784 * preventing the page from being uncharged.
1785 * E.g. end-writeback clearing PageWriteback(), which allows
1786 * migration to go ahead and uncharge the page before the
1787 * account transaction might be complete.
1791 if (mem_cgroup_disabled())
1794 memcg = page->mem_cgroup;
1795 if (unlikely(!memcg))
1798 if (atomic_read(&memcg->moving_account) <= 0)
1801 spin_lock_irqsave(&memcg->move_lock, flags);
1802 if (memcg != page->mem_cgroup) {
1803 spin_unlock_irqrestore(&memcg->move_lock, flags);
1808 * When charge migration first begins, we can have locked and
1809 * unlocked page stat updates happening concurrently. Track
1810 * the task who has the lock for mem_cgroup_end_page_stat().
1812 memcg->move_lock_task = current;
1813 memcg->move_lock_flags = flags;
1817 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1820 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1821 * @memcg: the memcg that was accounted against
1823 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1825 if (memcg && memcg->move_lock_task == current) {
1826 unsigned long flags = memcg->move_lock_flags;
1828 memcg->move_lock_task = NULL;
1829 memcg->move_lock_flags = 0;
1831 spin_unlock_irqrestore(&memcg->move_lock, flags);
1836 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1839 * size of first charge trial. "32" comes from vmscan.c's magic value.
1840 * TODO: maybe necessary to use big numbers in big irons.
1842 #define CHARGE_BATCH 32U
1843 struct memcg_stock_pcp {
1844 struct mem_cgroup *cached; /* this never be root cgroup */
1845 unsigned int nr_pages;
1846 struct work_struct work;
1847 unsigned long flags;
1848 #define FLUSHING_CACHED_CHARGE 0
1850 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1851 static DEFINE_MUTEX(percpu_charge_mutex);
1854 * consume_stock: Try to consume stocked charge on this cpu.
1855 * @memcg: memcg to consume from.
1856 * @nr_pages: how many pages to charge.
1858 * The charges will only happen if @memcg matches the current cpu's memcg
1859 * stock, and at least @nr_pages are available in that stock. Failure to
1860 * service an allocation will refill the stock.
1862 * returns true if successful, false otherwise.
1864 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1866 struct memcg_stock_pcp *stock;
1869 if (nr_pages > CHARGE_BATCH)
1872 stock = &get_cpu_var(memcg_stock);
1873 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1874 stock->nr_pages -= nr_pages;
1877 put_cpu_var(memcg_stock);
1882 * Returns stocks cached in percpu and reset cached information.
1884 static void drain_stock(struct memcg_stock_pcp *stock)
1886 struct mem_cgroup *old = stock->cached;
1888 if (stock->nr_pages) {
1889 page_counter_uncharge(&old->memory, stock->nr_pages);
1890 if (do_swap_account)
1891 page_counter_uncharge(&old->memsw, stock->nr_pages);
1892 css_put_many(&old->css, stock->nr_pages);
1893 stock->nr_pages = 0;
1895 stock->cached = NULL;
1899 * This must be called under preempt disabled or must be called by
1900 * a thread which is pinned to local cpu.
1902 static void drain_local_stock(struct work_struct *dummy)
1904 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1906 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1910 * Cache charges(val) to local per_cpu area.
1911 * This will be consumed by consume_stock() function, later.
1913 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1915 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1917 if (stock->cached != memcg) { /* reset if necessary */
1919 stock->cached = memcg;
1921 stock->nr_pages += nr_pages;
1922 put_cpu_var(memcg_stock);
1926 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1927 * of the hierarchy under it.
1929 static void drain_all_stock(struct mem_cgroup *root_memcg)
1933 /* If someone's already draining, avoid adding running more workers. */
1934 if (!mutex_trylock(&percpu_charge_mutex))
1936 /* Notify other cpus that system-wide "drain" is running */
1939 for_each_online_cpu(cpu) {
1940 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1941 struct mem_cgroup *memcg;
1943 memcg = stock->cached;
1944 if (!memcg || !stock->nr_pages)
1946 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1948 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1950 drain_local_stock(&stock->work);
1952 schedule_work_on(cpu, &stock->work);
1957 mutex_unlock(&percpu_charge_mutex);
1960 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1961 unsigned long action,
1964 int cpu = (unsigned long)hcpu;
1965 struct memcg_stock_pcp *stock;
1967 if (action == CPU_ONLINE)
1970 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1973 stock = &per_cpu(memcg_stock, cpu);
1979 * Scheduled by try_charge() to be executed from the userland return path
1980 * and reclaims memory over the high limit.
1982 void mem_cgroup_handle_over_high(void)
1984 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1985 struct mem_cgroup *memcg, *pos;
1987 if (likely(!nr_pages))
1990 pos = memcg = get_mem_cgroup_from_mm(current->mm);
1993 if (page_counter_read(&pos->memory) <= pos->high)
1995 mem_cgroup_events(pos, MEMCG_HIGH, 1);
1996 try_to_free_mem_cgroup_pages(pos, nr_pages, GFP_KERNEL, true);
1997 } while ((pos = parent_mem_cgroup(pos)));
1999 css_put(&memcg->css);
2000 current->memcg_nr_pages_over_high = 0;
2003 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2004 unsigned int nr_pages)
2006 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2007 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2008 struct mem_cgroup *mem_over_limit;
2009 struct page_counter *counter;
2010 unsigned long nr_reclaimed;
2011 bool may_swap = true;
2012 bool drained = false;
2014 if (mem_cgroup_is_root(memcg))
2017 if (consume_stock(memcg, nr_pages))
2020 if (!do_swap_account ||
2021 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2022 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2024 if (do_swap_account)
2025 page_counter_uncharge(&memcg->memsw, batch);
2026 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2028 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2032 if (batch > nr_pages) {
2038 * Unlike in global OOM situations, memcg is not in a physical
2039 * memory shortage. Allow dying and OOM-killed tasks to
2040 * bypass the last charges so that they can exit quickly and
2041 * free their memory.
2043 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2044 fatal_signal_pending(current) ||
2045 current->flags & PF_EXITING))
2048 if (unlikely(task_in_memcg_oom(current)))
2051 if (!gfpflags_allow_blocking(gfp_mask))
2054 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2056 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2057 gfp_mask, may_swap);
2059 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2063 drain_all_stock(mem_over_limit);
2068 if (gfp_mask & __GFP_NORETRY)
2071 * Even though the limit is exceeded at this point, reclaim
2072 * may have been able to free some pages. Retry the charge
2073 * before killing the task.
2075 * Only for regular pages, though: huge pages are rather
2076 * unlikely to succeed so close to the limit, and we fall back
2077 * to regular pages anyway in case of failure.
2079 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2082 * At task move, charge accounts can be doubly counted. So, it's
2083 * better to wait until the end of task_move if something is going on.
2085 if (mem_cgroup_wait_acct_move(mem_over_limit))
2091 if (gfp_mask & __GFP_NOFAIL)
2094 if (fatal_signal_pending(current))
2097 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2099 mem_cgroup_oom(mem_over_limit, gfp_mask,
2100 get_order(nr_pages * PAGE_SIZE));
2102 if (!(gfp_mask & __GFP_NOFAIL))
2106 * The allocation either can't fail or will lead to more memory
2107 * being freed very soon. Allow memory usage go over the limit
2108 * temporarily by force charging it.
2110 page_counter_charge(&memcg->memory, nr_pages);
2111 if (do_swap_account)
2112 page_counter_charge(&memcg->memsw, nr_pages);
2113 css_get_many(&memcg->css, nr_pages);
2118 css_get_many(&memcg->css, batch);
2119 if (batch > nr_pages)
2120 refill_stock(memcg, batch - nr_pages);
2123 * If the hierarchy is above the normal consumption range, schedule
2124 * reclaim on returning to userland. We can perform reclaim here
2125 * if __GFP_WAIT but let's always punt for simplicity and so that
2126 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2127 * not recorded as it most likely matches current's and won't
2128 * change in the meantime. As high limit is checked again before
2129 * reclaim, the cost of mismatch is negligible.
2132 if (page_counter_read(&memcg->memory) > memcg->high) {
2133 current->memcg_nr_pages_over_high += nr_pages;
2134 set_notify_resume(current);
2137 } while ((memcg = parent_mem_cgroup(memcg)));
2142 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2144 if (mem_cgroup_is_root(memcg))
2147 page_counter_uncharge(&memcg->memory, nr_pages);
2148 if (do_swap_account)
2149 page_counter_uncharge(&memcg->memsw, nr_pages);
2151 css_put_many(&memcg->css, nr_pages);
2154 static void lock_page_lru(struct page *page, int *isolated)
2156 struct zone *zone = page_zone(page);
2158 spin_lock_irq(&zone->lru_lock);
2159 if (PageLRU(page)) {
2160 struct lruvec *lruvec;
2162 lruvec = mem_cgroup_page_lruvec(page, zone);
2164 del_page_from_lru_list(page, lruvec, page_lru(page));
2170 static void unlock_page_lru(struct page *page, int isolated)
2172 struct zone *zone = page_zone(page);
2175 struct lruvec *lruvec;
2177 lruvec = mem_cgroup_page_lruvec(page, zone);
2178 VM_BUG_ON_PAGE(PageLRU(page), page);
2180 add_page_to_lru_list(page, lruvec, page_lru(page));
2182 spin_unlock_irq(&zone->lru_lock);
2185 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2190 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2193 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2194 * may already be on some other mem_cgroup's LRU. Take care of it.
2197 lock_page_lru(page, &isolated);
2200 * Nobody should be changing or seriously looking at
2201 * page->mem_cgroup at this point:
2203 * - the page is uncharged
2205 * - the page is off-LRU
2207 * - an anonymous fault has exclusive page access, except for
2208 * a locked page table
2210 * - a page cache insertion, a swapin fault, or a migration
2211 * have the page locked
2213 page->mem_cgroup = memcg;
2216 unlock_page_lru(page, isolated);
2219 #ifdef CONFIG_MEMCG_KMEM
2220 static int memcg_alloc_cache_id(void)
2225 id = ida_simple_get(&memcg_cache_ida,
2226 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2230 if (id < memcg_nr_cache_ids)
2234 * There's no space for the new id in memcg_caches arrays,
2235 * so we have to grow them.
2237 down_write(&memcg_cache_ids_sem);
2239 size = 2 * (id + 1);
2240 if (size < MEMCG_CACHES_MIN_SIZE)
2241 size = MEMCG_CACHES_MIN_SIZE;
2242 else if (size > MEMCG_CACHES_MAX_SIZE)
2243 size = MEMCG_CACHES_MAX_SIZE;
2245 err = memcg_update_all_caches(size);
2247 err = memcg_update_all_list_lrus(size);
2249 memcg_nr_cache_ids = size;
2251 up_write(&memcg_cache_ids_sem);
2254 ida_simple_remove(&memcg_cache_ida, id);
2260 static void memcg_free_cache_id(int id)
2262 ida_simple_remove(&memcg_cache_ida, id);
2265 struct memcg_kmem_cache_create_work {
2266 struct mem_cgroup *memcg;
2267 struct kmem_cache *cachep;
2268 struct work_struct work;
2271 static void memcg_kmem_cache_create_func(struct work_struct *w)
2273 struct memcg_kmem_cache_create_work *cw =
2274 container_of(w, struct memcg_kmem_cache_create_work, work);
2275 struct mem_cgroup *memcg = cw->memcg;
2276 struct kmem_cache *cachep = cw->cachep;
2278 memcg_create_kmem_cache(memcg, cachep);
2280 css_put(&memcg->css);
2285 * Enqueue the creation of a per-memcg kmem_cache.
2287 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2288 struct kmem_cache *cachep)
2290 struct memcg_kmem_cache_create_work *cw;
2292 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2296 css_get(&memcg->css);
2299 cw->cachep = cachep;
2300 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2302 schedule_work(&cw->work);
2305 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2306 struct kmem_cache *cachep)
2309 * We need to stop accounting when we kmalloc, because if the
2310 * corresponding kmalloc cache is not yet created, the first allocation
2311 * in __memcg_schedule_kmem_cache_create will recurse.
2313 * However, it is better to enclose the whole function. Depending on
2314 * the debugging options enabled, INIT_WORK(), for instance, can
2315 * trigger an allocation. This too, will make us recurse. Because at
2316 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2317 * the safest choice is to do it like this, wrapping the whole function.
2319 current->memcg_kmem_skip_account = 1;
2320 __memcg_schedule_kmem_cache_create(memcg, cachep);
2321 current->memcg_kmem_skip_account = 0;
2325 * Return the kmem_cache we're supposed to use for a slab allocation.
2326 * We try to use the current memcg's version of the cache.
2328 * If the cache does not exist yet, if we are the first user of it,
2329 * we either create it immediately, if possible, or create it asynchronously
2331 * In the latter case, we will let the current allocation go through with
2332 * the original cache.
2334 * Can't be called in interrupt context or from kernel threads.
2335 * This function needs to be called with rcu_read_lock() held.
2337 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2339 struct mem_cgroup *memcg;
2340 struct kmem_cache *memcg_cachep;
2343 VM_BUG_ON(!is_root_cache(cachep));
2345 if (current->memcg_kmem_skip_account)
2348 memcg = get_mem_cgroup_from_mm(current->mm);
2349 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2353 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2354 if (likely(memcg_cachep))
2355 return memcg_cachep;
2358 * If we are in a safe context (can wait, and not in interrupt
2359 * context), we could be be predictable and return right away.
2360 * This would guarantee that the allocation being performed
2361 * already belongs in the new cache.
2363 * However, there are some clashes that can arrive from locking.
2364 * For instance, because we acquire the slab_mutex while doing
2365 * memcg_create_kmem_cache, this means no further allocation
2366 * could happen with the slab_mutex held. So it's better to
2369 memcg_schedule_kmem_cache_create(memcg, cachep);
2371 css_put(&memcg->css);
2375 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2377 if (!is_root_cache(cachep))
2378 css_put(&cachep->memcg_params.memcg->css);
2381 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2382 struct mem_cgroup *memcg)
2384 unsigned int nr_pages = 1 << order;
2385 struct page_counter *counter;
2388 if (!memcg_kmem_is_active(memcg))
2391 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2395 ret = try_charge(memcg, gfp, nr_pages);
2397 page_counter_uncharge(&memcg->kmem, nr_pages);
2401 page->mem_cgroup = memcg;
2406 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2408 struct mem_cgroup *memcg;
2411 memcg = get_mem_cgroup_from_mm(current->mm);
2412 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2413 css_put(&memcg->css);
2417 void __memcg_kmem_uncharge(struct page *page, int order)
2419 struct mem_cgroup *memcg = page->mem_cgroup;
2420 unsigned int nr_pages = 1 << order;
2425 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2427 page_counter_uncharge(&memcg->kmem, nr_pages);
2428 page_counter_uncharge(&memcg->memory, nr_pages);
2429 if (do_swap_account)
2430 page_counter_uncharge(&memcg->memsw, nr_pages);
2432 page->mem_cgroup = NULL;
2433 css_put_many(&memcg->css, nr_pages);
2435 #endif /* CONFIG_MEMCG_KMEM */
2437 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2440 * Because tail pages are not marked as "used", set it. We're under
2441 * zone->lru_lock and migration entries setup in all page mappings.
2443 void mem_cgroup_split_huge_fixup(struct page *head)
2447 if (mem_cgroup_disabled())
2450 for (i = 1; i < HPAGE_PMD_NR; i++)
2451 head[i].mem_cgroup = head->mem_cgroup;
2453 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2456 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2458 #ifdef CONFIG_MEMCG_SWAP
2459 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2462 int val = (charge) ? 1 : -1;
2463 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2467 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2468 * @entry: swap entry to be moved
2469 * @from: mem_cgroup which the entry is moved from
2470 * @to: mem_cgroup which the entry is moved to
2472 * It succeeds only when the swap_cgroup's record for this entry is the same
2473 * as the mem_cgroup's id of @from.
2475 * Returns 0 on success, -EINVAL on failure.
2477 * The caller must have charged to @to, IOW, called page_counter_charge() about
2478 * both res and memsw, and called css_get().
2480 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2481 struct mem_cgroup *from, struct mem_cgroup *to)
2483 unsigned short old_id, new_id;
2485 old_id = mem_cgroup_id(from);
2486 new_id = mem_cgroup_id(to);
2488 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2489 mem_cgroup_swap_statistics(from, false);
2490 mem_cgroup_swap_statistics(to, true);
2496 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2497 struct mem_cgroup *from, struct mem_cgroup *to)
2503 static DEFINE_MUTEX(memcg_limit_mutex);
2505 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2506 unsigned long limit)
2508 unsigned long curusage;
2509 unsigned long oldusage;
2510 bool enlarge = false;
2515 * For keeping hierarchical_reclaim simple, how long we should retry
2516 * is depends on callers. We set our retry-count to be function
2517 * of # of children which we should visit in this loop.
2519 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2520 mem_cgroup_count_children(memcg);
2522 oldusage = page_counter_read(&memcg->memory);
2525 if (signal_pending(current)) {
2530 mutex_lock(&memcg_limit_mutex);
2531 if (limit > memcg->memsw.limit) {
2532 mutex_unlock(&memcg_limit_mutex);
2536 if (limit > memcg->memory.limit)
2538 ret = page_counter_limit(&memcg->memory, limit);
2539 mutex_unlock(&memcg_limit_mutex);
2544 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2546 curusage = page_counter_read(&memcg->memory);
2547 /* Usage is reduced ? */
2548 if (curusage >= oldusage)
2551 oldusage = curusage;
2552 } while (retry_count);
2554 if (!ret && enlarge)
2555 memcg_oom_recover(memcg);
2560 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2561 unsigned long limit)
2563 unsigned long curusage;
2564 unsigned long oldusage;
2565 bool enlarge = false;
2569 /* see mem_cgroup_resize_res_limit */
2570 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2571 mem_cgroup_count_children(memcg);
2573 oldusage = page_counter_read(&memcg->memsw);
2576 if (signal_pending(current)) {
2581 mutex_lock(&memcg_limit_mutex);
2582 if (limit < memcg->memory.limit) {
2583 mutex_unlock(&memcg_limit_mutex);
2587 if (limit > memcg->memsw.limit)
2589 ret = page_counter_limit(&memcg->memsw, limit);
2590 mutex_unlock(&memcg_limit_mutex);
2595 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2597 curusage = page_counter_read(&memcg->memsw);
2598 /* Usage is reduced ? */
2599 if (curusage >= oldusage)
2602 oldusage = curusage;
2603 } while (retry_count);
2605 if (!ret && enlarge)
2606 memcg_oom_recover(memcg);
2611 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2613 unsigned long *total_scanned)
2615 unsigned long nr_reclaimed = 0;
2616 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2617 unsigned long reclaimed;
2619 struct mem_cgroup_tree_per_zone *mctz;
2620 unsigned long excess;
2621 unsigned long nr_scanned;
2626 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2628 * This loop can run a while, specially if mem_cgroup's continuously
2629 * keep exceeding their soft limit and putting the system under
2636 mz = mem_cgroup_largest_soft_limit_node(mctz);
2641 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2642 gfp_mask, &nr_scanned);
2643 nr_reclaimed += reclaimed;
2644 *total_scanned += nr_scanned;
2645 spin_lock_irq(&mctz->lock);
2646 __mem_cgroup_remove_exceeded(mz, mctz);
2649 * If we failed to reclaim anything from this memory cgroup
2650 * it is time to move on to the next cgroup
2654 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2656 excess = soft_limit_excess(mz->memcg);
2658 * One school of thought says that we should not add
2659 * back the node to the tree if reclaim returns 0.
2660 * But our reclaim could return 0, simply because due
2661 * to priority we are exposing a smaller subset of
2662 * memory to reclaim from. Consider this as a longer
2665 /* If excess == 0, no tree ops */
2666 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2667 spin_unlock_irq(&mctz->lock);
2668 css_put(&mz->memcg->css);
2671 * Could not reclaim anything and there are no more
2672 * mem cgroups to try or we seem to be looping without
2673 * reclaiming anything.
2675 if (!nr_reclaimed &&
2677 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2679 } while (!nr_reclaimed);
2681 css_put(&next_mz->memcg->css);
2682 return nr_reclaimed;
2686 * Test whether @memcg has children, dead or alive. Note that this
2687 * function doesn't care whether @memcg has use_hierarchy enabled and
2688 * returns %true if there are child csses according to the cgroup
2689 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2691 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2696 * The lock does not prevent addition or deletion of children, but
2697 * it prevents a new child from being initialized based on this
2698 * parent in css_online(), so it's enough to decide whether
2699 * hierarchically inherited attributes can still be changed or not.
2701 lockdep_assert_held(&memcg_create_mutex);
2704 ret = css_next_child(NULL, &memcg->css);
2710 * Reclaims as many pages from the given memcg as possible and moves
2711 * the rest to the parent.
2713 * Caller is responsible for holding css reference for memcg.
2715 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2717 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2719 /* we call try-to-free pages for make this cgroup empty */
2720 lru_add_drain_all();
2721 /* try to free all pages in this cgroup */
2722 while (nr_retries && page_counter_read(&memcg->memory)) {
2725 if (signal_pending(current))
2728 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2732 /* maybe some writeback is necessary */
2733 congestion_wait(BLK_RW_ASYNC, HZ/10);
2741 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2742 char *buf, size_t nbytes,
2745 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2747 if (mem_cgroup_is_root(memcg))
2749 return mem_cgroup_force_empty(memcg) ?: nbytes;
2752 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2755 return mem_cgroup_from_css(css)->use_hierarchy;
2758 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2759 struct cftype *cft, u64 val)
2762 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2763 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2765 mutex_lock(&memcg_create_mutex);
2767 if (memcg->use_hierarchy == val)
2771 * If parent's use_hierarchy is set, we can't make any modifications
2772 * in the child subtrees. If it is unset, then the change can
2773 * occur, provided the current cgroup has no children.
2775 * For the root cgroup, parent_mem is NULL, we allow value to be
2776 * set if there are no children.
2778 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2779 (val == 1 || val == 0)) {
2780 if (!memcg_has_children(memcg))
2781 memcg->use_hierarchy = val;
2788 mutex_unlock(&memcg_create_mutex);
2793 static unsigned long tree_stat(struct mem_cgroup *memcg,
2794 enum mem_cgroup_stat_index idx)
2796 struct mem_cgroup *iter;
2797 unsigned long val = 0;
2799 for_each_mem_cgroup_tree(iter, memcg)
2800 val += mem_cgroup_read_stat(iter, idx);
2805 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2809 if (mem_cgroup_is_root(memcg)) {
2810 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2811 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2813 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2816 val = page_counter_read(&memcg->memory);
2818 val = page_counter_read(&memcg->memsw);
2820 return val << PAGE_SHIFT;
2831 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2834 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2835 struct page_counter *counter;
2837 switch (MEMFILE_TYPE(cft->private)) {
2839 counter = &memcg->memory;
2842 counter = &memcg->memsw;
2845 counter = &memcg->kmem;
2851 switch (MEMFILE_ATTR(cft->private)) {
2853 if (counter == &memcg->memory)
2854 return mem_cgroup_usage(memcg, false);
2855 if (counter == &memcg->memsw)
2856 return mem_cgroup_usage(memcg, true);
2857 return (u64)page_counter_read(counter) * PAGE_SIZE;
2859 return (u64)counter->limit * PAGE_SIZE;
2861 return (u64)counter->watermark * PAGE_SIZE;
2863 return counter->failcnt;
2864 case RES_SOFT_LIMIT:
2865 return (u64)memcg->soft_limit * PAGE_SIZE;
2871 #ifdef CONFIG_MEMCG_KMEM
2872 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2873 unsigned long nr_pages)
2878 BUG_ON(memcg->kmemcg_id >= 0);
2879 BUG_ON(memcg->kmem_acct_activated);
2880 BUG_ON(memcg->kmem_acct_active);
2883 * For simplicity, we won't allow this to be disabled. It also can't
2884 * be changed if the cgroup has children already, or if tasks had
2887 * If tasks join before we set the limit, a person looking at
2888 * kmem.usage_in_bytes will have no way to determine when it took
2889 * place, which makes the value quite meaningless.
2891 * After it first became limited, changes in the value of the limit are
2892 * of course permitted.
2894 mutex_lock(&memcg_create_mutex);
2895 if (cgroup_is_populated(memcg->css.cgroup) ||
2896 (memcg->use_hierarchy && memcg_has_children(memcg)))
2898 mutex_unlock(&memcg_create_mutex);
2902 memcg_id = memcg_alloc_cache_id();
2909 * We couldn't have accounted to this cgroup, because it hasn't got
2910 * activated yet, so this should succeed.
2912 err = page_counter_limit(&memcg->kmem, nr_pages);
2915 static_key_slow_inc(&memcg_kmem_enabled_key);
2917 * A memory cgroup is considered kmem-active as soon as it gets
2918 * kmemcg_id. Setting the id after enabling static branching will
2919 * guarantee no one starts accounting before all call sites are
2922 memcg->kmemcg_id = memcg_id;
2923 memcg->kmem_acct_activated = true;
2924 memcg->kmem_acct_active = true;
2929 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2930 unsigned long limit)
2934 mutex_lock(&memcg_limit_mutex);
2935 if (!memcg_kmem_is_active(memcg))
2936 ret = memcg_activate_kmem(memcg, limit);
2938 ret = page_counter_limit(&memcg->kmem, limit);
2939 mutex_unlock(&memcg_limit_mutex);
2943 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2946 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2951 mutex_lock(&memcg_limit_mutex);
2953 * If the parent cgroup is not kmem-active now, it cannot be activated
2954 * after this point, because it has at least one child already.
2956 if (memcg_kmem_is_active(parent))
2957 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
2958 mutex_unlock(&memcg_limit_mutex);
2962 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2963 unsigned long limit)
2967 #endif /* CONFIG_MEMCG_KMEM */
2970 * The user of this function is...
2973 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2974 char *buf, size_t nbytes, loff_t off)
2976 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2977 unsigned long nr_pages;
2980 buf = strstrip(buf);
2981 ret = page_counter_memparse(buf, "-1", &nr_pages);
2985 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2987 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2991 switch (MEMFILE_TYPE(of_cft(of)->private)) {
2993 ret = mem_cgroup_resize_limit(memcg, nr_pages);
2996 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
2999 ret = memcg_update_kmem_limit(memcg, nr_pages);
3003 case RES_SOFT_LIMIT:
3004 memcg->soft_limit = nr_pages;
3008 return ret ?: nbytes;
3011 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3012 size_t nbytes, loff_t off)
3014 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3015 struct page_counter *counter;
3017 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3019 counter = &memcg->memory;
3022 counter = &memcg->memsw;
3025 counter = &memcg->kmem;
3031 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3033 page_counter_reset_watermark(counter);
3036 counter->failcnt = 0;
3045 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3048 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3052 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3053 struct cftype *cft, u64 val)
3055 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3057 if (val & ~MOVE_MASK)
3061 * No kind of locking is needed in here, because ->can_attach() will
3062 * check this value once in the beginning of the process, and then carry
3063 * on with stale data. This means that changes to this value will only
3064 * affect task migrations starting after the change.
3066 memcg->move_charge_at_immigrate = val;
3070 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3071 struct cftype *cft, u64 val)
3078 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3082 unsigned int lru_mask;
3085 static const struct numa_stat stats[] = {
3086 { "total", LRU_ALL },
3087 { "file", LRU_ALL_FILE },
3088 { "anon", LRU_ALL_ANON },
3089 { "unevictable", BIT(LRU_UNEVICTABLE) },
3091 const struct numa_stat *stat;
3094 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3096 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3097 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3098 seq_printf(m, "%s=%lu", stat->name, nr);
3099 for_each_node_state(nid, N_MEMORY) {
3100 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3102 seq_printf(m, " N%d=%lu", nid, nr);
3107 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3108 struct mem_cgroup *iter;
3111 for_each_mem_cgroup_tree(iter, memcg)
3112 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3113 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3114 for_each_node_state(nid, N_MEMORY) {
3116 for_each_mem_cgroup_tree(iter, memcg)
3117 nr += mem_cgroup_node_nr_lru_pages(
3118 iter, nid, stat->lru_mask);
3119 seq_printf(m, " N%d=%lu", nid, nr);
3126 #endif /* CONFIG_NUMA */
3128 static int memcg_stat_show(struct seq_file *m, void *v)
3130 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3131 unsigned long memory, memsw;
3132 struct mem_cgroup *mi;
3135 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3136 MEM_CGROUP_STAT_NSTATS);
3137 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3138 MEM_CGROUP_EVENTS_NSTATS);
3139 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3141 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3142 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3144 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3145 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3148 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3149 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3150 mem_cgroup_read_events(memcg, i));
3152 for (i = 0; i < NR_LRU_LISTS; i++)
3153 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3154 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3156 /* Hierarchical information */
3157 memory = memsw = PAGE_COUNTER_MAX;
3158 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3159 memory = min(memory, mi->memory.limit);
3160 memsw = min(memsw, mi->memsw.limit);
3162 seq_printf(m, "hierarchical_memory_limit %llu\n",
3163 (u64)memory * PAGE_SIZE);
3164 if (do_swap_account)
3165 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3166 (u64)memsw * PAGE_SIZE);
3168 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3169 unsigned long long val = 0;
3171 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3173 for_each_mem_cgroup_tree(mi, memcg)
3174 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3175 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3178 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3179 unsigned long long val = 0;
3181 for_each_mem_cgroup_tree(mi, memcg)
3182 val += mem_cgroup_read_events(mi, i);
3183 seq_printf(m, "total_%s %llu\n",
3184 mem_cgroup_events_names[i], val);
3187 for (i = 0; i < NR_LRU_LISTS; i++) {
3188 unsigned long long val = 0;
3190 for_each_mem_cgroup_tree(mi, memcg)
3191 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3192 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3195 #ifdef CONFIG_DEBUG_VM
3198 struct mem_cgroup_per_zone *mz;
3199 struct zone_reclaim_stat *rstat;
3200 unsigned long recent_rotated[2] = {0, 0};
3201 unsigned long recent_scanned[2] = {0, 0};
3203 for_each_online_node(nid)
3204 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3205 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3206 rstat = &mz->lruvec.reclaim_stat;
3208 recent_rotated[0] += rstat->recent_rotated[0];
3209 recent_rotated[1] += rstat->recent_rotated[1];
3210 recent_scanned[0] += rstat->recent_scanned[0];
3211 recent_scanned[1] += rstat->recent_scanned[1];
3213 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3214 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3215 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3216 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3223 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3226 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3228 return mem_cgroup_swappiness(memcg);
3231 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3232 struct cftype *cft, u64 val)
3234 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3240 memcg->swappiness = val;
3242 vm_swappiness = val;
3247 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3249 struct mem_cgroup_threshold_ary *t;
3250 unsigned long usage;
3255 t = rcu_dereference(memcg->thresholds.primary);
3257 t = rcu_dereference(memcg->memsw_thresholds.primary);
3262 usage = mem_cgroup_usage(memcg, swap);
3265 * current_threshold points to threshold just below or equal to usage.
3266 * If it's not true, a threshold was crossed after last
3267 * call of __mem_cgroup_threshold().
3269 i = t->current_threshold;
3272 * Iterate backward over array of thresholds starting from
3273 * current_threshold and check if a threshold is crossed.
3274 * If none of thresholds below usage is crossed, we read
3275 * only one element of the array here.
3277 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3278 eventfd_signal(t->entries[i].eventfd, 1);
3280 /* i = current_threshold + 1 */
3284 * Iterate forward over array of thresholds starting from
3285 * current_threshold+1 and check if a threshold is crossed.
3286 * If none of thresholds above usage is crossed, we read
3287 * only one element of the array here.
3289 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3290 eventfd_signal(t->entries[i].eventfd, 1);
3292 /* Update current_threshold */
3293 t->current_threshold = i - 1;
3298 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3301 __mem_cgroup_threshold(memcg, false);
3302 if (do_swap_account)
3303 __mem_cgroup_threshold(memcg, true);
3305 memcg = parent_mem_cgroup(memcg);
3309 static int compare_thresholds(const void *a, const void *b)
3311 const struct mem_cgroup_threshold *_a = a;
3312 const struct mem_cgroup_threshold *_b = b;
3314 if (_a->threshold > _b->threshold)
3317 if (_a->threshold < _b->threshold)
3323 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3325 struct mem_cgroup_eventfd_list *ev;
3327 spin_lock(&memcg_oom_lock);
3329 list_for_each_entry(ev, &memcg->oom_notify, list)
3330 eventfd_signal(ev->eventfd, 1);
3332 spin_unlock(&memcg_oom_lock);
3336 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3338 struct mem_cgroup *iter;
3340 for_each_mem_cgroup_tree(iter, memcg)
3341 mem_cgroup_oom_notify_cb(iter);
3344 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3345 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3347 struct mem_cgroup_thresholds *thresholds;
3348 struct mem_cgroup_threshold_ary *new;
3349 unsigned long threshold;
3350 unsigned long usage;
3353 ret = page_counter_memparse(args, "-1", &threshold);
3356 threshold <<= PAGE_SHIFT;
3358 mutex_lock(&memcg->thresholds_lock);
3361 thresholds = &memcg->thresholds;
3362 usage = mem_cgroup_usage(memcg, false);
3363 } else if (type == _MEMSWAP) {
3364 thresholds = &memcg->memsw_thresholds;
3365 usage = mem_cgroup_usage(memcg, true);
3369 /* Check if a threshold crossed before adding a new one */
3370 if (thresholds->primary)
3371 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3373 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3375 /* Allocate memory for new array of thresholds */
3376 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3384 /* Copy thresholds (if any) to new array */
3385 if (thresholds->primary) {
3386 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3387 sizeof(struct mem_cgroup_threshold));
3390 /* Add new threshold */
3391 new->entries[size - 1].eventfd = eventfd;
3392 new->entries[size - 1].threshold = threshold;
3394 /* Sort thresholds. Registering of new threshold isn't time-critical */
3395 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3396 compare_thresholds, NULL);
3398 /* Find current threshold */
3399 new->current_threshold = -1;
3400 for (i = 0; i < size; i++) {
3401 if (new->entries[i].threshold <= usage) {
3403 * new->current_threshold will not be used until
3404 * rcu_assign_pointer(), so it's safe to increment
3407 ++new->current_threshold;
3412 /* Free old spare buffer and save old primary buffer as spare */
3413 kfree(thresholds->spare);
3414 thresholds->spare = thresholds->primary;
3416 rcu_assign_pointer(thresholds->primary, new);
3418 /* To be sure that nobody uses thresholds */
3422 mutex_unlock(&memcg->thresholds_lock);
3427 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3428 struct eventfd_ctx *eventfd, const char *args)
3430 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3433 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3434 struct eventfd_ctx *eventfd, const char *args)
3436 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3439 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3440 struct eventfd_ctx *eventfd, enum res_type type)
3442 struct mem_cgroup_thresholds *thresholds;
3443 struct mem_cgroup_threshold_ary *new;
3444 unsigned long usage;
3447 mutex_lock(&memcg->thresholds_lock);
3450 thresholds = &memcg->thresholds;
3451 usage = mem_cgroup_usage(memcg, false);
3452 } else if (type == _MEMSWAP) {
3453 thresholds = &memcg->memsw_thresholds;
3454 usage = mem_cgroup_usage(memcg, true);
3458 if (!thresholds->primary)
3461 /* Check if a threshold crossed before removing */
3462 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3464 /* Calculate new number of threshold */
3466 for (i = 0; i < thresholds->primary->size; i++) {
3467 if (thresholds->primary->entries[i].eventfd != eventfd)
3471 new = thresholds->spare;
3473 /* Set thresholds array to NULL if we don't have thresholds */
3482 /* Copy thresholds and find current threshold */
3483 new->current_threshold = -1;
3484 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3485 if (thresholds->primary->entries[i].eventfd == eventfd)
3488 new->entries[j] = thresholds->primary->entries[i];
3489 if (new->entries[j].threshold <= usage) {
3491 * new->current_threshold will not be used
3492 * until rcu_assign_pointer(), so it's safe to increment
3495 ++new->current_threshold;
3501 /* Swap primary and spare array */
3502 thresholds->spare = thresholds->primary;
3503 /* If all events are unregistered, free the spare array */
3505 kfree(thresholds->spare);
3506 thresholds->spare = NULL;
3509 rcu_assign_pointer(thresholds->primary, new);
3511 /* To be sure that nobody uses thresholds */
3514 mutex_unlock(&memcg->thresholds_lock);
3517 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3518 struct eventfd_ctx *eventfd)
3520 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3523 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3524 struct eventfd_ctx *eventfd)
3526 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3529 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3530 struct eventfd_ctx *eventfd, const char *args)
3532 struct mem_cgroup_eventfd_list *event;
3534 event = kmalloc(sizeof(*event), GFP_KERNEL);
3538 spin_lock(&memcg_oom_lock);
3540 event->eventfd = eventfd;
3541 list_add(&event->list, &memcg->oom_notify);
3543 /* already in OOM ? */
3544 if (memcg->under_oom)
3545 eventfd_signal(eventfd, 1);
3546 spin_unlock(&memcg_oom_lock);
3551 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3552 struct eventfd_ctx *eventfd)
3554 struct mem_cgroup_eventfd_list *ev, *tmp;
3556 spin_lock(&memcg_oom_lock);
3558 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3559 if (ev->eventfd == eventfd) {
3560 list_del(&ev->list);
3565 spin_unlock(&memcg_oom_lock);
3568 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3570 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3572 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3573 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3577 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3578 struct cftype *cft, u64 val)
3580 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3582 /* cannot set to root cgroup and only 0 and 1 are allowed */
3583 if (!css->parent || !((val == 0) || (val == 1)))
3586 memcg->oom_kill_disable = val;
3588 memcg_oom_recover(memcg);
3593 #ifdef CONFIG_MEMCG_KMEM
3594 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3598 ret = memcg_propagate_kmem(memcg);
3602 return mem_cgroup_sockets_init(memcg, ss);
3605 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3607 struct cgroup_subsys_state *css;
3608 struct mem_cgroup *parent, *child;
3611 if (!memcg->kmem_acct_active)
3615 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3616 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3617 * guarantees no cache will be created for this cgroup after we are
3618 * done (see memcg_create_kmem_cache()).
3620 memcg->kmem_acct_active = false;
3622 memcg_deactivate_kmem_caches(memcg);
3624 kmemcg_id = memcg->kmemcg_id;
3625 BUG_ON(kmemcg_id < 0);
3627 parent = parent_mem_cgroup(memcg);
3629 parent = root_mem_cgroup;
3632 * Change kmemcg_id of this cgroup and all its descendants to the
3633 * parent's id, and then move all entries from this cgroup's list_lrus
3634 * to ones of the parent. After we have finished, all list_lrus
3635 * corresponding to this cgroup are guaranteed to remain empty. The
3636 * ordering is imposed by list_lru_node->lock taken by
3637 * memcg_drain_all_list_lrus().
3639 css_for_each_descendant_pre(css, &memcg->css) {
3640 child = mem_cgroup_from_css(css);
3641 BUG_ON(child->kmemcg_id != kmemcg_id);
3642 child->kmemcg_id = parent->kmemcg_id;
3643 if (!memcg->use_hierarchy)
3646 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3648 memcg_free_cache_id(kmemcg_id);
3651 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3653 if (memcg->kmem_acct_activated) {
3654 memcg_destroy_kmem_caches(memcg);
3655 static_key_slow_dec(&memcg_kmem_enabled_key);
3656 WARN_ON(page_counter_read(&memcg->kmem));
3658 mem_cgroup_sockets_destroy(memcg);
3661 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3666 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3670 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3675 #ifdef CONFIG_CGROUP_WRITEBACK
3677 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3679 return &memcg->cgwb_list;
3682 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3684 return wb_domain_init(&memcg->cgwb_domain, gfp);
3687 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3689 wb_domain_exit(&memcg->cgwb_domain);
3692 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3694 wb_domain_size_changed(&memcg->cgwb_domain);
3697 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3699 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3701 if (!memcg->css.parent)
3704 return &memcg->cgwb_domain;
3708 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3709 * @wb: bdi_writeback in question
3710 * @pfilepages: out parameter for number of file pages
3711 * @pheadroom: out parameter for number of allocatable pages according to memcg
3712 * @pdirty: out parameter for number of dirty pages
3713 * @pwriteback: out parameter for number of pages under writeback
3715 * Determine the numbers of file, headroom, dirty, and writeback pages in
3716 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3717 * is a bit more involved.
3719 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3720 * headroom is calculated as the lowest headroom of itself and the
3721 * ancestors. Note that this doesn't consider the actual amount of
3722 * available memory in the system. The caller should further cap
3723 * *@pheadroom accordingly.
3725 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3726 unsigned long *pheadroom, unsigned long *pdirty,
3727 unsigned long *pwriteback)
3729 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3730 struct mem_cgroup *parent;
3732 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3734 /* this should eventually include NR_UNSTABLE_NFS */
3735 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3736 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3737 (1 << LRU_ACTIVE_FILE));
3738 *pheadroom = PAGE_COUNTER_MAX;
3740 while ((parent = parent_mem_cgroup(memcg))) {
3741 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3742 unsigned long used = page_counter_read(&memcg->memory);
3744 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3749 #else /* CONFIG_CGROUP_WRITEBACK */
3751 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3756 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3760 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3764 #endif /* CONFIG_CGROUP_WRITEBACK */
3767 * DO NOT USE IN NEW FILES.
3769 * "cgroup.event_control" implementation.
3771 * This is way over-engineered. It tries to support fully configurable
3772 * events for each user. Such level of flexibility is completely
3773 * unnecessary especially in the light of the planned unified hierarchy.
3775 * Please deprecate this and replace with something simpler if at all
3780 * Unregister event and free resources.
3782 * Gets called from workqueue.
3784 static void memcg_event_remove(struct work_struct *work)
3786 struct mem_cgroup_event *event =
3787 container_of(work, struct mem_cgroup_event, remove);
3788 struct mem_cgroup *memcg = event->memcg;
3790 remove_wait_queue(event->wqh, &event->wait);
3792 event->unregister_event(memcg, event->eventfd);
3794 /* Notify userspace the event is going away. */
3795 eventfd_signal(event->eventfd, 1);
3797 eventfd_ctx_put(event->eventfd);
3799 css_put(&memcg->css);
3803 * Gets called on POLLHUP on eventfd when user closes it.
3805 * Called with wqh->lock held and interrupts disabled.
3807 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3808 int sync, void *key)
3810 struct mem_cgroup_event *event =
3811 container_of(wait, struct mem_cgroup_event, wait);
3812 struct mem_cgroup *memcg = event->memcg;
3813 unsigned long flags = (unsigned long)key;
3815 if (flags & POLLHUP) {
3817 * If the event has been detached at cgroup removal, we
3818 * can simply return knowing the other side will cleanup
3821 * We can't race against event freeing since the other
3822 * side will require wqh->lock via remove_wait_queue(),
3825 spin_lock(&memcg->event_list_lock);
3826 if (!list_empty(&event->list)) {
3827 list_del_init(&event->list);
3829 * We are in atomic context, but cgroup_event_remove()
3830 * may sleep, so we have to call it in workqueue.
3832 schedule_work(&event->remove);
3834 spin_unlock(&memcg->event_list_lock);
3840 static void memcg_event_ptable_queue_proc(struct file *file,
3841 wait_queue_head_t *wqh, poll_table *pt)
3843 struct mem_cgroup_event *event =
3844 container_of(pt, struct mem_cgroup_event, pt);
3847 add_wait_queue(wqh, &event->wait);
3851 * DO NOT USE IN NEW FILES.
3853 * Parse input and register new cgroup event handler.
3855 * Input must be in format '<event_fd> <control_fd> <args>'.
3856 * Interpretation of args is defined by control file implementation.
3858 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3859 char *buf, size_t nbytes, loff_t off)
3861 struct cgroup_subsys_state *css = of_css(of);
3862 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3863 struct mem_cgroup_event *event;
3864 struct cgroup_subsys_state *cfile_css;
3865 unsigned int efd, cfd;
3872 buf = strstrip(buf);
3874 efd = simple_strtoul(buf, &endp, 10);
3879 cfd = simple_strtoul(buf, &endp, 10);
3880 if ((*endp != ' ') && (*endp != '\0'))
3884 event = kzalloc(sizeof(*event), GFP_KERNEL);
3888 event->memcg = memcg;
3889 INIT_LIST_HEAD(&event->list);
3890 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3891 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3892 INIT_WORK(&event->remove, memcg_event_remove);
3900 event->eventfd = eventfd_ctx_fileget(efile.file);
3901 if (IS_ERR(event->eventfd)) {
3902 ret = PTR_ERR(event->eventfd);
3909 goto out_put_eventfd;
3912 /* the process need read permission on control file */
3913 /* AV: shouldn't we check that it's been opened for read instead? */
3914 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3919 * Determine the event callbacks and set them in @event. This used
3920 * to be done via struct cftype but cgroup core no longer knows
3921 * about these events. The following is crude but the whole thing
3922 * is for compatibility anyway.
3924 * DO NOT ADD NEW FILES.
3926 name = cfile.file->f_path.dentry->d_name.name;
3928 if (!strcmp(name, "memory.usage_in_bytes")) {
3929 event->register_event = mem_cgroup_usage_register_event;
3930 event->unregister_event = mem_cgroup_usage_unregister_event;
3931 } else if (!strcmp(name, "memory.oom_control")) {
3932 event->register_event = mem_cgroup_oom_register_event;
3933 event->unregister_event = mem_cgroup_oom_unregister_event;
3934 } else if (!strcmp(name, "memory.pressure_level")) {
3935 event->register_event = vmpressure_register_event;
3936 event->unregister_event = vmpressure_unregister_event;
3937 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3938 event->register_event = memsw_cgroup_usage_register_event;
3939 event->unregister_event = memsw_cgroup_usage_unregister_event;
3946 * Verify @cfile should belong to @css. Also, remaining events are
3947 * automatically removed on cgroup destruction but the removal is
3948 * asynchronous, so take an extra ref on @css.
3950 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3951 &memory_cgrp_subsys);
3953 if (IS_ERR(cfile_css))
3955 if (cfile_css != css) {
3960 ret = event->register_event(memcg, event->eventfd, buf);
3964 efile.file->f_op->poll(efile.file, &event->pt);
3966 spin_lock(&memcg->event_list_lock);
3967 list_add(&event->list, &memcg->event_list);
3968 spin_unlock(&memcg->event_list_lock);
3980 eventfd_ctx_put(event->eventfd);
3989 static struct cftype mem_cgroup_legacy_files[] = {
3991 .name = "usage_in_bytes",
3992 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3993 .read_u64 = mem_cgroup_read_u64,
3996 .name = "max_usage_in_bytes",
3997 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3998 .write = mem_cgroup_reset,
3999 .read_u64 = mem_cgroup_read_u64,
4002 .name = "limit_in_bytes",
4003 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4004 .write = mem_cgroup_write,
4005 .read_u64 = mem_cgroup_read_u64,
4008 .name = "soft_limit_in_bytes",
4009 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4010 .write = mem_cgroup_write,
4011 .read_u64 = mem_cgroup_read_u64,
4015 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4016 .write = mem_cgroup_reset,
4017 .read_u64 = mem_cgroup_read_u64,
4021 .seq_show = memcg_stat_show,
4024 .name = "force_empty",
4025 .write = mem_cgroup_force_empty_write,
4028 .name = "use_hierarchy",
4029 .write_u64 = mem_cgroup_hierarchy_write,
4030 .read_u64 = mem_cgroup_hierarchy_read,
4033 .name = "cgroup.event_control", /* XXX: for compat */
4034 .write = memcg_write_event_control,
4035 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4038 .name = "swappiness",
4039 .read_u64 = mem_cgroup_swappiness_read,
4040 .write_u64 = mem_cgroup_swappiness_write,
4043 .name = "move_charge_at_immigrate",
4044 .read_u64 = mem_cgroup_move_charge_read,
4045 .write_u64 = mem_cgroup_move_charge_write,
4048 .name = "oom_control",
4049 .seq_show = mem_cgroup_oom_control_read,
4050 .write_u64 = mem_cgroup_oom_control_write,
4051 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4054 .name = "pressure_level",
4058 .name = "numa_stat",
4059 .seq_show = memcg_numa_stat_show,
4062 #ifdef CONFIG_MEMCG_KMEM
4064 .name = "kmem.limit_in_bytes",
4065 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4066 .write = mem_cgroup_write,
4067 .read_u64 = mem_cgroup_read_u64,
4070 .name = "kmem.usage_in_bytes",
4071 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4072 .read_u64 = mem_cgroup_read_u64,
4075 .name = "kmem.failcnt",
4076 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4077 .write = mem_cgroup_reset,
4078 .read_u64 = mem_cgroup_read_u64,
4081 .name = "kmem.max_usage_in_bytes",
4082 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4083 .write = mem_cgroup_reset,
4084 .read_u64 = mem_cgroup_read_u64,
4086 #ifdef CONFIG_SLABINFO
4088 .name = "kmem.slabinfo",
4089 .seq_start = slab_start,
4090 .seq_next = slab_next,
4091 .seq_stop = slab_stop,
4092 .seq_show = memcg_slab_show,
4096 { }, /* terminate */
4099 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4101 struct mem_cgroup_per_node *pn;
4102 struct mem_cgroup_per_zone *mz;
4103 int zone, tmp = node;
4105 * This routine is called against possible nodes.
4106 * But it's BUG to call kmalloc() against offline node.
4108 * TODO: this routine can waste much memory for nodes which will
4109 * never be onlined. It's better to use memory hotplug callback
4112 if (!node_state(node, N_NORMAL_MEMORY))
4114 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4118 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4119 mz = &pn->zoneinfo[zone];
4120 lruvec_init(&mz->lruvec);
4121 mz->usage_in_excess = 0;
4122 mz->on_tree = false;
4125 memcg->nodeinfo[node] = pn;
4129 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4131 kfree(memcg->nodeinfo[node]);
4134 static struct mem_cgroup *mem_cgroup_alloc(void)
4136 struct mem_cgroup *memcg;
4139 size = sizeof(struct mem_cgroup);
4140 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4142 memcg = kzalloc(size, GFP_KERNEL);
4146 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4150 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4156 free_percpu(memcg->stat);
4163 * At destroying mem_cgroup, references from swap_cgroup can remain.
4164 * (scanning all at force_empty is too costly...)
4166 * Instead of clearing all references at force_empty, we remember
4167 * the number of reference from swap_cgroup and free mem_cgroup when
4168 * it goes down to 0.
4170 * Removal of cgroup itself succeeds regardless of refs from swap.
4173 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4177 mem_cgroup_remove_from_trees(memcg);
4180 free_mem_cgroup_per_zone_info(memcg, node);
4182 free_percpu(memcg->stat);
4183 memcg_wb_domain_exit(memcg);
4188 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4190 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4192 if (!memcg->memory.parent)
4194 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4196 EXPORT_SYMBOL(parent_mem_cgroup);
4198 static struct cgroup_subsys_state * __ref
4199 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4201 struct mem_cgroup *memcg;
4202 long error = -ENOMEM;
4205 memcg = mem_cgroup_alloc();
4207 return ERR_PTR(error);
4210 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4214 if (parent_css == NULL) {
4215 root_mem_cgroup = memcg;
4216 mem_cgroup_root_css = &memcg->css;
4217 page_counter_init(&memcg->memory, NULL);
4218 memcg->high = PAGE_COUNTER_MAX;
4219 memcg->soft_limit = PAGE_COUNTER_MAX;
4220 page_counter_init(&memcg->memsw, NULL);
4221 page_counter_init(&memcg->kmem, NULL);
4224 memcg->last_scanned_node = MAX_NUMNODES;
4225 INIT_LIST_HEAD(&memcg->oom_notify);
4226 memcg->move_charge_at_immigrate = 0;
4227 mutex_init(&memcg->thresholds_lock);
4228 spin_lock_init(&memcg->move_lock);
4229 vmpressure_init(&memcg->vmpressure);
4230 INIT_LIST_HEAD(&memcg->event_list);
4231 spin_lock_init(&memcg->event_list_lock);
4232 #ifdef CONFIG_MEMCG_KMEM
4233 memcg->kmemcg_id = -1;
4235 #ifdef CONFIG_CGROUP_WRITEBACK
4236 INIT_LIST_HEAD(&memcg->cgwb_list);
4241 __mem_cgroup_free(memcg);
4242 return ERR_PTR(error);
4246 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4248 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4249 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4252 if (css->id > MEM_CGROUP_ID_MAX)
4258 mutex_lock(&memcg_create_mutex);
4260 memcg->use_hierarchy = parent->use_hierarchy;
4261 memcg->oom_kill_disable = parent->oom_kill_disable;
4262 memcg->swappiness = mem_cgroup_swappiness(parent);
4264 if (parent->use_hierarchy) {
4265 page_counter_init(&memcg->memory, &parent->memory);
4266 memcg->high = PAGE_COUNTER_MAX;
4267 memcg->soft_limit = PAGE_COUNTER_MAX;
4268 page_counter_init(&memcg->memsw, &parent->memsw);
4269 page_counter_init(&memcg->kmem, &parent->kmem);
4272 * No need to take a reference to the parent because cgroup
4273 * core guarantees its existence.
4276 page_counter_init(&memcg->memory, NULL);
4277 memcg->high = PAGE_COUNTER_MAX;
4278 memcg->soft_limit = PAGE_COUNTER_MAX;
4279 page_counter_init(&memcg->memsw, NULL);
4280 page_counter_init(&memcg->kmem, NULL);
4282 * Deeper hierachy with use_hierarchy == false doesn't make
4283 * much sense so let cgroup subsystem know about this
4284 * unfortunate state in our controller.
4286 if (parent != root_mem_cgroup)
4287 memory_cgrp_subsys.broken_hierarchy = true;
4289 mutex_unlock(&memcg_create_mutex);
4291 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4296 * Make sure the memcg is initialized: mem_cgroup_iter()
4297 * orders reading memcg->initialized against its callers
4298 * reading the memcg members.
4300 smp_store_release(&memcg->initialized, 1);
4305 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4307 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4308 struct mem_cgroup_event *event, *tmp;
4311 * Unregister events and notify userspace.
4312 * Notify userspace about cgroup removing only after rmdir of cgroup
4313 * directory to avoid race between userspace and kernelspace.
4315 spin_lock(&memcg->event_list_lock);
4316 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4317 list_del_init(&event->list);
4318 schedule_work(&event->remove);
4320 spin_unlock(&memcg->event_list_lock);
4322 vmpressure_cleanup(&memcg->vmpressure);
4324 memcg_deactivate_kmem(memcg);
4326 wb_memcg_offline(memcg);
4329 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4331 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4333 memcg_destroy_kmem(memcg);
4334 __mem_cgroup_free(memcg);
4338 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4339 * @css: the target css
4341 * Reset the states of the mem_cgroup associated with @css. This is
4342 * invoked when the userland requests disabling on the default hierarchy
4343 * but the memcg is pinned through dependency. The memcg should stop
4344 * applying policies and should revert to the vanilla state as it may be
4345 * made visible again.
4347 * The current implementation only resets the essential configurations.
4348 * This needs to be expanded to cover all the visible parts.
4350 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4352 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4354 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4355 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4356 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4358 memcg->high = PAGE_COUNTER_MAX;
4359 memcg->soft_limit = PAGE_COUNTER_MAX;
4360 memcg_wb_domain_size_changed(memcg);
4364 /* Handlers for move charge at task migration. */
4365 static int mem_cgroup_do_precharge(unsigned long count)
4369 /* Try a single bulk charge without reclaim first, kswapd may wake */
4370 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4372 mc.precharge += count;
4376 /* Try charges one by one with reclaim */
4378 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4388 * get_mctgt_type - get target type of moving charge
4389 * @vma: the vma the pte to be checked belongs
4390 * @addr: the address corresponding to the pte to be checked
4391 * @ptent: the pte to be checked
4392 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4395 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4396 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4397 * move charge. if @target is not NULL, the page is stored in target->page
4398 * with extra refcnt got(Callers should handle it).
4399 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4400 * target for charge migration. if @target is not NULL, the entry is stored
4403 * Called with pte lock held.
4410 enum mc_target_type {
4416 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4417 unsigned long addr, pte_t ptent)
4419 struct page *page = vm_normal_page(vma, addr, ptent);
4421 if (!page || !page_mapped(page))
4423 if (PageAnon(page)) {
4424 if (!(mc.flags & MOVE_ANON))
4427 if (!(mc.flags & MOVE_FILE))
4430 if (!get_page_unless_zero(page))
4437 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4438 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4440 struct page *page = NULL;
4441 swp_entry_t ent = pte_to_swp_entry(ptent);
4443 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4446 * Because lookup_swap_cache() updates some statistics counter,
4447 * we call find_get_page() with swapper_space directly.
4449 page = find_get_page(swap_address_space(ent), ent.val);
4450 if (do_swap_account)
4451 entry->val = ent.val;
4456 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4457 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4463 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4464 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4466 struct page *page = NULL;
4467 struct address_space *mapping;
4470 if (!vma->vm_file) /* anonymous vma */
4472 if (!(mc.flags & MOVE_FILE))
4475 mapping = vma->vm_file->f_mapping;
4476 pgoff = linear_page_index(vma, addr);
4478 /* page is moved even if it's not RSS of this task(page-faulted). */
4480 /* shmem/tmpfs may report page out on swap: account for that too. */
4481 if (shmem_mapping(mapping)) {
4482 page = find_get_entry(mapping, pgoff);
4483 if (radix_tree_exceptional_entry(page)) {
4484 swp_entry_t swp = radix_to_swp_entry(page);
4485 if (do_swap_account)
4487 page = find_get_page(swap_address_space(swp), swp.val);
4490 page = find_get_page(mapping, pgoff);
4492 page = find_get_page(mapping, pgoff);
4498 * mem_cgroup_move_account - move account of the page
4500 * @nr_pages: number of regular pages (>1 for huge pages)
4501 * @from: mem_cgroup which the page is moved from.
4502 * @to: mem_cgroup which the page is moved to. @from != @to.
4504 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4506 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4509 static int mem_cgroup_move_account(struct page *page,
4511 struct mem_cgroup *from,
4512 struct mem_cgroup *to)
4514 unsigned long flags;
4515 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4519 VM_BUG_ON(from == to);
4520 VM_BUG_ON_PAGE(PageLRU(page), page);
4521 VM_BUG_ON(compound && !PageTransHuge(page));
4524 * Prevent mem_cgroup_replace_page() from looking at
4525 * page->mem_cgroup of its source page while we change it.
4528 if (!trylock_page(page))
4532 if (page->mem_cgroup != from)
4535 anon = PageAnon(page);
4537 spin_lock_irqsave(&from->move_lock, flags);
4539 if (!anon && page_mapped(page)) {
4540 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4542 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4547 * move_lock grabbed above and caller set from->moving_account, so
4548 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4549 * So mapping should be stable for dirty pages.
4551 if (!anon && PageDirty(page)) {
4552 struct address_space *mapping = page_mapping(page);
4554 if (mapping_cap_account_dirty(mapping)) {
4555 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4557 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4562 if (PageWriteback(page)) {
4563 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4565 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4570 * It is safe to change page->mem_cgroup here because the page
4571 * is referenced, charged, and isolated - we can't race with
4572 * uncharging, charging, migration, or LRU putback.
4575 /* caller should have done css_get */
4576 page->mem_cgroup = to;
4577 spin_unlock_irqrestore(&from->move_lock, flags);
4581 local_irq_disable();
4582 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4583 memcg_check_events(to, page);
4584 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4585 memcg_check_events(from, page);
4593 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4594 unsigned long addr, pte_t ptent, union mc_target *target)
4596 struct page *page = NULL;
4597 enum mc_target_type ret = MC_TARGET_NONE;
4598 swp_entry_t ent = { .val = 0 };
4600 if (pte_present(ptent))
4601 page = mc_handle_present_pte(vma, addr, ptent);
4602 else if (is_swap_pte(ptent))
4603 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4604 else if (pte_none(ptent))
4605 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4607 if (!page && !ent.val)
4611 * Do only loose check w/o serialization.
4612 * mem_cgroup_move_account() checks the page is valid or
4613 * not under LRU exclusion.
4615 if (page->mem_cgroup == mc.from) {
4616 ret = MC_TARGET_PAGE;
4618 target->page = page;
4620 if (!ret || !target)
4623 /* There is a swap entry and a page doesn't exist or isn't charged */
4624 if (ent.val && !ret &&
4625 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4626 ret = MC_TARGET_SWAP;
4633 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4635 * We don't consider swapping or file mapped pages because THP does not
4636 * support them for now.
4637 * Caller should make sure that pmd_trans_huge(pmd) is true.
4639 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4640 unsigned long addr, pmd_t pmd, union mc_target *target)
4642 struct page *page = NULL;
4643 enum mc_target_type ret = MC_TARGET_NONE;
4645 page = pmd_page(pmd);
4646 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4647 if (!(mc.flags & MOVE_ANON))
4649 if (page->mem_cgroup == mc.from) {
4650 ret = MC_TARGET_PAGE;
4653 target->page = page;
4659 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4660 unsigned long addr, pmd_t pmd, union mc_target *target)
4662 return MC_TARGET_NONE;
4666 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4667 unsigned long addr, unsigned long end,
4668 struct mm_walk *walk)
4670 struct vm_area_struct *vma = walk->vma;
4674 if (pmd_trans_huge_lock(pmd, vma, &ptl)) {
4675 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4676 mc.precharge += HPAGE_PMD_NR;
4681 if (pmd_trans_unstable(pmd))
4683 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4684 for (; addr != end; pte++, addr += PAGE_SIZE)
4685 if (get_mctgt_type(vma, addr, *pte, NULL))
4686 mc.precharge++; /* increment precharge temporarily */
4687 pte_unmap_unlock(pte - 1, ptl);
4693 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4695 unsigned long precharge;
4697 struct mm_walk mem_cgroup_count_precharge_walk = {
4698 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4701 down_read(&mm->mmap_sem);
4702 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4703 up_read(&mm->mmap_sem);
4705 precharge = mc.precharge;
4711 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4713 unsigned long precharge = mem_cgroup_count_precharge(mm);
4715 VM_BUG_ON(mc.moving_task);
4716 mc.moving_task = current;
4717 return mem_cgroup_do_precharge(precharge);
4720 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4721 static void __mem_cgroup_clear_mc(void)
4723 struct mem_cgroup *from = mc.from;
4724 struct mem_cgroup *to = mc.to;
4726 /* we must uncharge all the leftover precharges from mc.to */
4728 cancel_charge(mc.to, mc.precharge);
4732 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4733 * we must uncharge here.
4735 if (mc.moved_charge) {
4736 cancel_charge(mc.from, mc.moved_charge);
4737 mc.moved_charge = 0;
4739 /* we must fixup refcnts and charges */
4740 if (mc.moved_swap) {
4741 /* uncharge swap account from the old cgroup */
4742 if (!mem_cgroup_is_root(mc.from))
4743 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4746 * we charged both to->memory and to->memsw, so we
4747 * should uncharge to->memory.
4749 if (!mem_cgroup_is_root(mc.to))
4750 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4752 css_put_many(&mc.from->css, mc.moved_swap);
4754 /* we've already done css_get(mc.to) */
4757 memcg_oom_recover(from);
4758 memcg_oom_recover(to);
4759 wake_up_all(&mc.waitq);
4762 static void mem_cgroup_clear_mc(void)
4765 * we must clear moving_task before waking up waiters at the end of
4768 mc.moving_task = NULL;
4769 __mem_cgroup_clear_mc();
4770 spin_lock(&mc.lock);
4773 spin_unlock(&mc.lock);
4776 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
4777 struct cgroup_taskset *tset)
4779 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4780 struct mem_cgroup *from;
4781 struct task_struct *leader, *p;
4782 struct mm_struct *mm;
4783 unsigned long move_flags;
4787 * We are now commited to this value whatever it is. Changes in this
4788 * tunable will only affect upcoming migrations, not the current one.
4789 * So we need to save it, and keep it going.
4791 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4796 * Multi-process migrations only happen on the default hierarchy
4797 * where charge immigration is not used. Perform charge
4798 * immigration if @tset contains a leader and whine if there are
4802 cgroup_taskset_for_each_leader(leader, tset) {
4809 from = mem_cgroup_from_task(p);
4811 VM_BUG_ON(from == memcg);
4813 mm = get_task_mm(p);
4816 /* We move charges only when we move a owner of the mm */
4817 if (mm->owner == p) {
4820 VM_BUG_ON(mc.precharge);
4821 VM_BUG_ON(mc.moved_charge);
4822 VM_BUG_ON(mc.moved_swap);
4824 spin_lock(&mc.lock);
4827 mc.flags = move_flags;
4828 spin_unlock(&mc.lock);
4829 /* We set mc.moving_task later */
4831 ret = mem_cgroup_precharge_mc(mm);
4833 mem_cgroup_clear_mc();
4839 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
4840 struct cgroup_taskset *tset)
4843 mem_cgroup_clear_mc();
4846 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4847 unsigned long addr, unsigned long end,
4848 struct mm_walk *walk)
4851 struct vm_area_struct *vma = walk->vma;
4854 enum mc_target_type target_type;
4855 union mc_target target;
4858 if (pmd_trans_huge_lock(pmd, vma, &ptl)) {
4859 if (mc.precharge < HPAGE_PMD_NR) {
4863 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4864 if (target_type == MC_TARGET_PAGE) {
4866 if (!isolate_lru_page(page)) {
4867 if (!mem_cgroup_move_account(page, true,
4869 mc.precharge -= HPAGE_PMD_NR;
4870 mc.moved_charge += HPAGE_PMD_NR;
4872 putback_lru_page(page);
4880 if (pmd_trans_unstable(pmd))
4883 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4884 for (; addr != end; addr += PAGE_SIZE) {
4885 pte_t ptent = *(pte++);
4891 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4892 case MC_TARGET_PAGE:
4894 if (isolate_lru_page(page))
4896 if (!mem_cgroup_move_account(page, false,
4899 /* we uncharge from mc.from later. */
4902 putback_lru_page(page);
4903 put: /* get_mctgt_type() gets the page */
4906 case MC_TARGET_SWAP:
4908 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4910 /* we fixup refcnts and charges later. */
4918 pte_unmap_unlock(pte - 1, ptl);
4923 * We have consumed all precharges we got in can_attach().
4924 * We try charge one by one, but don't do any additional
4925 * charges to mc.to if we have failed in charge once in attach()
4928 ret = mem_cgroup_do_precharge(1);
4936 static void mem_cgroup_move_charge(struct mm_struct *mm)
4938 struct mm_walk mem_cgroup_move_charge_walk = {
4939 .pmd_entry = mem_cgroup_move_charge_pte_range,
4943 lru_add_drain_all();
4945 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4946 * move_lock while we're moving its pages to another memcg.
4947 * Then wait for already started RCU-only updates to finish.
4949 atomic_inc(&mc.from->moving_account);
4952 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4954 * Someone who are holding the mmap_sem might be waiting in
4955 * waitq. So we cancel all extra charges, wake up all waiters,
4956 * and retry. Because we cancel precharges, we might not be able
4957 * to move enough charges, but moving charge is a best-effort
4958 * feature anyway, so it wouldn't be a big problem.
4960 __mem_cgroup_clear_mc();
4965 * When we have consumed all precharges and failed in doing
4966 * additional charge, the page walk just aborts.
4968 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
4969 up_read(&mm->mmap_sem);
4970 atomic_dec(&mc.from->moving_account);
4973 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
4974 struct cgroup_taskset *tset)
4976 struct task_struct *p = cgroup_taskset_first(tset);
4977 struct mm_struct *mm = get_task_mm(p);
4981 mem_cgroup_move_charge(mm);
4985 mem_cgroup_clear_mc();
4987 #else /* !CONFIG_MMU */
4988 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
4989 struct cgroup_taskset *tset)
4993 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
4994 struct cgroup_taskset *tset)
4997 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
4998 struct cgroup_taskset *tset)
5004 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5005 * to verify whether we're attached to the default hierarchy on each mount
5008 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5011 * use_hierarchy is forced on the default hierarchy. cgroup core
5012 * guarantees that @root doesn't have any children, so turning it
5013 * on for the root memcg is enough.
5015 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5016 root_mem_cgroup->use_hierarchy = true;
5018 root_mem_cgroup->use_hierarchy = false;
5021 static u64 memory_current_read(struct cgroup_subsys_state *css,
5024 return page_counter_read(&mem_cgroup_from_css(css)->memory);
5027 static int memory_low_show(struct seq_file *m, void *v)
5029 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5030 unsigned long low = READ_ONCE(memcg->low);
5032 if (low == PAGE_COUNTER_MAX)
5033 seq_puts(m, "max\n");
5035 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5040 static ssize_t memory_low_write(struct kernfs_open_file *of,
5041 char *buf, size_t nbytes, loff_t off)
5043 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5047 buf = strstrip(buf);
5048 err = page_counter_memparse(buf, "max", &low);
5057 static int memory_high_show(struct seq_file *m, void *v)
5059 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5060 unsigned long high = READ_ONCE(memcg->high);
5062 if (high == PAGE_COUNTER_MAX)
5063 seq_puts(m, "max\n");
5065 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5070 static ssize_t memory_high_write(struct kernfs_open_file *of,
5071 char *buf, size_t nbytes, loff_t off)
5073 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5077 buf = strstrip(buf);
5078 err = page_counter_memparse(buf, "max", &high);
5084 memcg_wb_domain_size_changed(memcg);
5088 static int memory_max_show(struct seq_file *m, void *v)
5090 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5091 unsigned long max = READ_ONCE(memcg->memory.limit);
5093 if (max == PAGE_COUNTER_MAX)
5094 seq_puts(m, "max\n");
5096 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5101 static ssize_t memory_max_write(struct kernfs_open_file *of,
5102 char *buf, size_t nbytes, loff_t off)
5104 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5108 buf = strstrip(buf);
5109 err = page_counter_memparse(buf, "max", &max);
5113 err = mem_cgroup_resize_limit(memcg, max);
5117 memcg_wb_domain_size_changed(memcg);
5121 static int memory_events_show(struct seq_file *m, void *v)
5123 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5125 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5126 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5127 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5128 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5133 static struct cftype memory_files[] = {
5136 .flags = CFTYPE_NOT_ON_ROOT,
5137 .read_u64 = memory_current_read,
5141 .flags = CFTYPE_NOT_ON_ROOT,
5142 .seq_show = memory_low_show,
5143 .write = memory_low_write,
5147 .flags = CFTYPE_NOT_ON_ROOT,
5148 .seq_show = memory_high_show,
5149 .write = memory_high_write,
5153 .flags = CFTYPE_NOT_ON_ROOT,
5154 .seq_show = memory_max_show,
5155 .write = memory_max_write,
5159 .flags = CFTYPE_NOT_ON_ROOT,
5160 .file_offset = offsetof(struct mem_cgroup, events_file),
5161 .seq_show = memory_events_show,
5166 struct cgroup_subsys memory_cgrp_subsys = {
5167 .css_alloc = mem_cgroup_css_alloc,
5168 .css_online = mem_cgroup_css_online,
5169 .css_offline = mem_cgroup_css_offline,
5170 .css_free = mem_cgroup_css_free,
5171 .css_reset = mem_cgroup_css_reset,
5172 .can_attach = mem_cgroup_can_attach,
5173 .cancel_attach = mem_cgroup_cancel_attach,
5174 .attach = mem_cgroup_move_task,
5175 .bind = mem_cgroup_bind,
5176 .dfl_cftypes = memory_files,
5177 .legacy_cftypes = mem_cgroup_legacy_files,
5182 * mem_cgroup_low - check if memory consumption is below the normal range
5183 * @root: the highest ancestor to consider
5184 * @memcg: the memory cgroup to check
5186 * Returns %true if memory consumption of @memcg, and that of all
5187 * configurable ancestors up to @root, is below the normal range.
5189 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5191 if (mem_cgroup_disabled())
5195 * The toplevel group doesn't have a configurable range, so
5196 * it's never low when looked at directly, and it is not
5197 * considered an ancestor when assessing the hierarchy.
5200 if (memcg == root_mem_cgroup)
5203 if (page_counter_read(&memcg->memory) >= memcg->low)
5206 while (memcg != root) {
5207 memcg = parent_mem_cgroup(memcg);
5209 if (memcg == root_mem_cgroup)
5212 if (page_counter_read(&memcg->memory) >= memcg->low)
5219 * mem_cgroup_try_charge - try charging a page
5220 * @page: page to charge
5221 * @mm: mm context of the victim
5222 * @gfp_mask: reclaim mode
5223 * @memcgp: charged memcg return
5225 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5226 * pages according to @gfp_mask if necessary.
5228 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5229 * Otherwise, an error code is returned.
5231 * After page->mapping has been set up, the caller must finalize the
5232 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5233 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5235 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5236 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5239 struct mem_cgroup *memcg = NULL;
5240 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5243 if (mem_cgroup_disabled())
5246 if (PageSwapCache(page)) {
5248 * Every swap fault against a single page tries to charge the
5249 * page, bail as early as possible. shmem_unuse() encounters
5250 * already charged pages, too. The USED bit is protected by
5251 * the page lock, which serializes swap cache removal, which
5252 * in turn serializes uncharging.
5254 VM_BUG_ON_PAGE(!PageLocked(page), page);
5255 if (page->mem_cgroup)
5258 if (do_swap_account) {
5259 swp_entry_t ent = { .val = page_private(page), };
5260 unsigned short id = lookup_swap_cgroup_id(ent);
5263 memcg = mem_cgroup_from_id(id);
5264 if (memcg && !css_tryget_online(&memcg->css))
5271 memcg = get_mem_cgroup_from_mm(mm);
5273 ret = try_charge(memcg, gfp_mask, nr_pages);
5275 css_put(&memcg->css);
5282 * mem_cgroup_commit_charge - commit a page charge
5283 * @page: page to charge
5284 * @memcg: memcg to charge the page to
5285 * @lrucare: page might be on LRU already
5287 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5288 * after page->mapping has been set up. This must happen atomically
5289 * as part of the page instantiation, i.e. under the page table lock
5290 * for anonymous pages, under the page lock for page and swap cache.
5292 * In addition, the page must not be on the LRU during the commit, to
5293 * prevent racing with task migration. If it might be, use @lrucare.
5295 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5297 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5298 bool lrucare, bool compound)
5300 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5302 VM_BUG_ON_PAGE(!page->mapping, page);
5303 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5305 if (mem_cgroup_disabled())
5308 * Swap faults will attempt to charge the same page multiple
5309 * times. But reuse_swap_page() might have removed the page
5310 * from swapcache already, so we can't check PageSwapCache().
5315 commit_charge(page, memcg, lrucare);
5317 local_irq_disable();
5318 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5319 memcg_check_events(memcg, page);
5322 if (do_swap_account && PageSwapCache(page)) {
5323 swp_entry_t entry = { .val = page_private(page) };
5325 * The swap entry might not get freed for a long time,
5326 * let's not wait for it. The page already received a
5327 * memory+swap charge, drop the swap entry duplicate.
5329 mem_cgroup_uncharge_swap(entry);
5334 * mem_cgroup_cancel_charge - cancel a page charge
5335 * @page: page to charge
5336 * @memcg: memcg to charge the page to
5338 * Cancel a charge transaction started by mem_cgroup_try_charge().
5340 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5343 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5345 if (mem_cgroup_disabled())
5348 * Swap faults will attempt to charge the same page multiple
5349 * times. But reuse_swap_page() might have removed the page
5350 * from swapcache already, so we can't check PageSwapCache().
5355 cancel_charge(memcg, nr_pages);
5358 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5359 unsigned long nr_anon, unsigned long nr_file,
5360 unsigned long nr_huge, struct page *dummy_page)
5362 unsigned long nr_pages = nr_anon + nr_file;
5363 unsigned long flags;
5365 if (!mem_cgroup_is_root(memcg)) {
5366 page_counter_uncharge(&memcg->memory, nr_pages);
5367 if (do_swap_account)
5368 page_counter_uncharge(&memcg->memsw, nr_pages);
5369 memcg_oom_recover(memcg);
5372 local_irq_save(flags);
5373 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5374 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5375 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5376 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5377 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5378 memcg_check_events(memcg, dummy_page);
5379 local_irq_restore(flags);
5381 if (!mem_cgroup_is_root(memcg))
5382 css_put_many(&memcg->css, nr_pages);
5385 static void uncharge_list(struct list_head *page_list)
5387 struct mem_cgroup *memcg = NULL;
5388 unsigned long nr_anon = 0;
5389 unsigned long nr_file = 0;
5390 unsigned long nr_huge = 0;
5391 unsigned long pgpgout = 0;
5392 struct list_head *next;
5395 next = page_list->next;
5397 unsigned int nr_pages = 1;
5399 page = list_entry(next, struct page, lru);
5400 next = page->lru.next;
5402 VM_BUG_ON_PAGE(PageLRU(page), page);
5403 VM_BUG_ON_PAGE(page_count(page), page);
5405 if (!page->mem_cgroup)
5409 * Nobody should be changing or seriously looking at
5410 * page->mem_cgroup at this point, we have fully
5411 * exclusive access to the page.
5414 if (memcg != page->mem_cgroup) {
5416 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5418 pgpgout = nr_anon = nr_file = nr_huge = 0;
5420 memcg = page->mem_cgroup;
5423 if (PageTransHuge(page)) {
5424 nr_pages <<= compound_order(page);
5425 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5426 nr_huge += nr_pages;
5430 nr_anon += nr_pages;
5432 nr_file += nr_pages;
5434 page->mem_cgroup = NULL;
5437 } while (next != page_list);
5440 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5445 * mem_cgroup_uncharge - uncharge a page
5446 * @page: page to uncharge
5448 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5449 * mem_cgroup_commit_charge().
5451 void mem_cgroup_uncharge(struct page *page)
5453 if (mem_cgroup_disabled())
5456 /* Don't touch page->lru of any random page, pre-check: */
5457 if (!page->mem_cgroup)
5460 INIT_LIST_HEAD(&page->lru);
5461 uncharge_list(&page->lru);
5465 * mem_cgroup_uncharge_list - uncharge a list of page
5466 * @page_list: list of pages to uncharge
5468 * Uncharge a list of pages previously charged with
5469 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5471 void mem_cgroup_uncharge_list(struct list_head *page_list)
5473 if (mem_cgroup_disabled())
5476 if (!list_empty(page_list))
5477 uncharge_list(page_list);
5481 * mem_cgroup_replace_page - migrate a charge to another page
5482 * @oldpage: currently charged page
5483 * @newpage: page to transfer the charge to
5484 * @lrucare: either or both pages might be on the LRU already
5486 * Migrate the charge from @oldpage to @newpage.
5488 * Both pages must be locked, @newpage->mapping must be set up.
5490 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5492 struct mem_cgroup *memcg;
5495 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5496 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5497 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5498 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5501 if (mem_cgroup_disabled())
5504 /* Page cache replacement: new page already charged? */
5505 if (newpage->mem_cgroup)
5508 /* Swapcache readahead pages can get replaced before being charged */
5509 memcg = oldpage->mem_cgroup;
5513 lock_page_lru(oldpage, &isolated);
5514 oldpage->mem_cgroup = NULL;
5515 unlock_page_lru(oldpage, isolated);
5517 commit_charge(newpage, memcg, true);
5521 * subsys_initcall() for memory controller.
5523 * Some parts like hotcpu_notifier() have to be initialized from this context
5524 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5525 * everything that doesn't depend on a specific mem_cgroup structure should
5526 * be initialized from here.
5528 static int __init mem_cgroup_init(void)
5532 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5534 for_each_possible_cpu(cpu)
5535 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5538 for_each_node(node) {
5539 struct mem_cgroup_tree_per_node *rtpn;
5542 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5543 node_online(node) ? node : NUMA_NO_NODE);
5545 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5546 struct mem_cgroup_tree_per_zone *rtpz;
5548 rtpz = &rtpn->rb_tree_per_zone[zone];
5549 rtpz->rb_root = RB_ROOT;
5550 spin_lock_init(&rtpz->lock);
5552 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5557 subsys_initcall(mem_cgroup_init);
5559 #ifdef CONFIG_MEMCG_SWAP
5561 * mem_cgroup_swapout - transfer a memsw charge to swap
5562 * @page: page whose memsw charge to transfer
5563 * @entry: swap entry to move the charge to
5565 * Transfer the memsw charge of @page to @entry.
5567 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5569 struct mem_cgroup *memcg;
5570 unsigned short oldid;
5572 VM_BUG_ON_PAGE(PageLRU(page), page);
5573 VM_BUG_ON_PAGE(page_count(page), page);
5575 if (!do_swap_account)
5578 memcg = page->mem_cgroup;
5580 /* Readahead page, never charged */
5584 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5585 VM_BUG_ON_PAGE(oldid, page);
5586 mem_cgroup_swap_statistics(memcg, true);
5588 page->mem_cgroup = NULL;
5590 if (!mem_cgroup_is_root(memcg))
5591 page_counter_uncharge(&memcg->memory, 1);
5594 * Interrupts should be disabled here because the caller holds the
5595 * mapping->tree_lock lock which is taken with interrupts-off. It is
5596 * important here to have the interrupts disabled because it is the
5597 * only synchronisation we have for udpating the per-CPU variables.
5599 VM_BUG_ON(!irqs_disabled());
5600 mem_cgroup_charge_statistics(memcg, page, false, -1);
5601 memcg_check_events(memcg, page);
5605 * mem_cgroup_uncharge_swap - uncharge a swap entry
5606 * @entry: swap entry to uncharge
5608 * Drop the memsw charge associated with @entry.
5610 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5612 struct mem_cgroup *memcg;
5615 if (!do_swap_account)
5618 id = swap_cgroup_record(entry, 0);
5620 memcg = mem_cgroup_from_id(id);
5622 if (!mem_cgroup_is_root(memcg))
5623 page_counter_uncharge(&memcg->memsw, 1);
5624 mem_cgroup_swap_statistics(memcg, false);
5625 css_put(&memcg->css);
5630 /* for remember boot option*/
5631 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5632 static int really_do_swap_account __initdata = 1;
5634 static int really_do_swap_account __initdata;
5637 static int __init enable_swap_account(char *s)
5639 if (!strcmp(s, "1"))
5640 really_do_swap_account = 1;
5641 else if (!strcmp(s, "0"))
5642 really_do_swap_account = 0;
5645 __setup("swapaccount=", enable_swap_account);
5647 static struct cftype memsw_cgroup_files[] = {
5649 .name = "memsw.usage_in_bytes",
5650 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5651 .read_u64 = mem_cgroup_read_u64,
5654 .name = "memsw.max_usage_in_bytes",
5655 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5656 .write = mem_cgroup_reset,
5657 .read_u64 = mem_cgroup_read_u64,
5660 .name = "memsw.limit_in_bytes",
5661 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5662 .write = mem_cgroup_write,
5663 .read_u64 = mem_cgroup_read_u64,
5666 .name = "memsw.failcnt",
5667 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5668 .write = mem_cgroup_reset,
5669 .read_u64 = mem_cgroup_read_u64,
5671 { }, /* terminate */
5674 static int __init mem_cgroup_swap_init(void)
5676 if (!mem_cgroup_disabled() && really_do_swap_account) {
5677 do_swap_account = 1;
5678 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5679 memsw_cgroup_files));
5683 subsys_initcall(mem_cgroup_swap_init);
5685 #endif /* CONFIG_MEMCG_SWAP */