1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
69 #include <net/tcp_memcontrol.h>
72 #include <asm/uaccess.h>
74 #include <trace/events/vmscan.h>
76 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
77 EXPORT_SYMBOL(memory_cgrp_subsys);
79 struct mem_cgroup *root_mem_cgroup __read_mostly;
81 #define MEM_CGROUP_RECLAIM_RETRIES 5
83 /* Whether the swap controller is active */
84 #ifdef CONFIG_MEMCG_SWAP
85 int do_swap_account __read_mostly;
87 #define do_swap_account 0
90 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 struct mem_cgroup *memcg;
297 struct cg_proto *cg_proto;
299 BUG_ON(!sk->sk_prot->proto_cgroup);
301 /* Socket cloning can throw us here with sk_cgrp already
302 * filled. It won't however, necessarily happen from
303 * process context. So the test for root memcg given
304 * the current task's memcg won't help us in this case.
306 * Respecting the original socket's memcg is a better
307 * decision in this case.
310 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
311 css_get(&sk->sk_cgrp->memcg->css);
316 memcg = mem_cgroup_from_task(current);
317 cg_proto = sk->sk_prot->proto_cgroup(memcg);
318 if (cg_proto && cg_proto->active &&
319 css_tryget_online(&memcg->css)) {
320 sk->sk_cgrp = cg_proto;
324 EXPORT_SYMBOL(sock_update_memcg);
326 void sock_release_memcg(struct sock *sk)
328 WARN_ON(!sk->sk_cgrp->memcg);
329 css_put(&sk->sk_cgrp->memcg->css);
332 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
334 if (!memcg || mem_cgroup_is_root(memcg))
337 return &memcg->tcp_mem;
339 EXPORT_SYMBOL(tcp_proto_cgroup);
342 * mem_cgroup_charge_skmem - charge socket memory
343 * @proto: proto to charge
344 * @nr_pages: number of pages to charge
346 * Charges @nr_pages to @proto. Returns %true if the charge fit within
347 * @proto's configured limit, %false if the charge had to be forced.
349 bool mem_cgroup_charge_skmem(struct cg_proto *proto, unsigned int nr_pages)
351 struct page_counter *counter;
353 if (page_counter_try_charge(&proto->memory_allocated,
354 nr_pages, &counter)) {
355 proto->memory_pressure = 0;
358 page_counter_charge(&proto->memory_allocated, nr_pages);
359 proto->memory_pressure = 1;
364 * mem_cgroup_uncharge_skmem - uncharge socket memory
365 * @proto - proto to uncharge
366 * @nr_pages - number of pages to uncharge
368 void mem_cgroup_uncharge_skmem(struct cg_proto *proto, unsigned int nr_pages)
370 page_counter_uncharge(&proto->memory_allocated, nr_pages);
375 #ifdef CONFIG_MEMCG_KMEM
377 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
378 * The main reason for not using cgroup id for this:
379 * this works better in sparse environments, where we have a lot of memcgs,
380 * but only a few kmem-limited. Or also, if we have, for instance, 200
381 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
382 * 200 entry array for that.
384 * The current size of the caches array is stored in memcg_nr_cache_ids. It
385 * will double each time we have to increase it.
387 static DEFINE_IDA(memcg_cache_ida);
388 int memcg_nr_cache_ids;
390 /* Protects memcg_nr_cache_ids */
391 static DECLARE_RWSEM(memcg_cache_ids_sem);
393 void memcg_get_cache_ids(void)
395 down_read(&memcg_cache_ids_sem);
398 void memcg_put_cache_ids(void)
400 up_read(&memcg_cache_ids_sem);
404 * MIN_SIZE is different than 1, because we would like to avoid going through
405 * the alloc/free process all the time. In a small machine, 4 kmem-limited
406 * cgroups is a reasonable guess. In the future, it could be a parameter or
407 * tunable, but that is strictly not necessary.
409 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
410 * this constant directly from cgroup, but it is understandable that this is
411 * better kept as an internal representation in cgroup.c. In any case, the
412 * cgrp_id space is not getting any smaller, and we don't have to necessarily
413 * increase ours as well if it increases.
415 #define MEMCG_CACHES_MIN_SIZE 4
416 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
419 * A lot of the calls to the cache allocation functions are expected to be
420 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
421 * conditional to this static branch, we'll have to allow modules that does
422 * kmem_cache_alloc and the such to see this symbol as well
424 struct static_key memcg_kmem_enabled_key;
425 EXPORT_SYMBOL(memcg_kmem_enabled_key);
427 #endif /* CONFIG_MEMCG_KMEM */
429 static struct mem_cgroup_per_zone *
430 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
432 int nid = zone_to_nid(zone);
433 int zid = zone_idx(zone);
435 return &memcg->nodeinfo[nid]->zoneinfo[zid];
439 * mem_cgroup_css_from_page - css of the memcg associated with a page
440 * @page: page of interest
442 * If memcg is bound to the default hierarchy, css of the memcg associated
443 * with @page is returned. The returned css remains associated with @page
444 * until it is released.
446 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
449 * XXX: The above description of behavior on the default hierarchy isn't
450 * strictly true yet as replace_page_cache_page() can modify the
451 * association before @page is released even on the default hierarchy;
452 * however, the current and planned usages don't mix the the two functions
453 * and replace_page_cache_page() will soon be updated to make the invariant
456 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
458 struct mem_cgroup *memcg;
462 memcg = page->mem_cgroup;
464 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
465 memcg = root_mem_cgroup;
472 * page_cgroup_ino - return inode number of the memcg a page is charged to
475 * Look up the closest online ancestor of the memory cgroup @page is charged to
476 * and return its inode number or 0 if @page is not charged to any cgroup. It
477 * is safe to call this function without holding a reference to @page.
479 * Note, this function is inherently racy, because there is nothing to prevent
480 * the cgroup inode from getting torn down and potentially reallocated a moment
481 * after page_cgroup_ino() returns, so it only should be used by callers that
482 * do not care (such as procfs interfaces).
484 ino_t page_cgroup_ino(struct page *page)
486 struct mem_cgroup *memcg;
487 unsigned long ino = 0;
490 memcg = READ_ONCE(page->mem_cgroup);
491 while (memcg && !(memcg->css.flags & CSS_ONLINE))
492 memcg = parent_mem_cgroup(memcg);
494 ino = cgroup_ino(memcg->css.cgroup);
499 static struct mem_cgroup_per_zone *
500 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
502 int nid = page_to_nid(page);
503 int zid = page_zonenum(page);
505 return &memcg->nodeinfo[nid]->zoneinfo[zid];
508 static struct mem_cgroup_tree_per_zone *
509 soft_limit_tree_node_zone(int nid, int zid)
511 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
514 static struct mem_cgroup_tree_per_zone *
515 soft_limit_tree_from_page(struct page *page)
517 int nid = page_to_nid(page);
518 int zid = page_zonenum(page);
520 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
523 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
524 struct mem_cgroup_tree_per_zone *mctz,
525 unsigned long new_usage_in_excess)
527 struct rb_node **p = &mctz->rb_root.rb_node;
528 struct rb_node *parent = NULL;
529 struct mem_cgroup_per_zone *mz_node;
534 mz->usage_in_excess = new_usage_in_excess;
535 if (!mz->usage_in_excess)
539 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
541 if (mz->usage_in_excess < mz_node->usage_in_excess)
544 * We can't avoid mem cgroups that are over their soft
545 * limit by the same amount
547 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
550 rb_link_node(&mz->tree_node, parent, p);
551 rb_insert_color(&mz->tree_node, &mctz->rb_root);
555 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
556 struct mem_cgroup_tree_per_zone *mctz)
560 rb_erase(&mz->tree_node, &mctz->rb_root);
564 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
565 struct mem_cgroup_tree_per_zone *mctz)
569 spin_lock_irqsave(&mctz->lock, flags);
570 __mem_cgroup_remove_exceeded(mz, mctz);
571 spin_unlock_irqrestore(&mctz->lock, flags);
574 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
576 unsigned long nr_pages = page_counter_read(&memcg->memory);
577 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
578 unsigned long excess = 0;
580 if (nr_pages > soft_limit)
581 excess = nr_pages - soft_limit;
586 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
588 unsigned long excess;
589 struct mem_cgroup_per_zone *mz;
590 struct mem_cgroup_tree_per_zone *mctz;
592 mctz = soft_limit_tree_from_page(page);
594 * Necessary to update all ancestors when hierarchy is used.
595 * because their event counter is not touched.
597 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
598 mz = mem_cgroup_page_zoneinfo(memcg, page);
599 excess = soft_limit_excess(memcg);
601 * We have to update the tree if mz is on RB-tree or
602 * mem is over its softlimit.
604 if (excess || mz->on_tree) {
607 spin_lock_irqsave(&mctz->lock, flags);
608 /* if on-tree, remove it */
610 __mem_cgroup_remove_exceeded(mz, mctz);
612 * Insert again. mz->usage_in_excess will be updated.
613 * If excess is 0, no tree ops.
615 __mem_cgroup_insert_exceeded(mz, mctz, excess);
616 spin_unlock_irqrestore(&mctz->lock, flags);
621 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
623 struct mem_cgroup_tree_per_zone *mctz;
624 struct mem_cgroup_per_zone *mz;
628 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
629 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
630 mctz = soft_limit_tree_node_zone(nid, zid);
631 mem_cgroup_remove_exceeded(mz, mctz);
636 static struct mem_cgroup_per_zone *
637 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
639 struct rb_node *rightmost = NULL;
640 struct mem_cgroup_per_zone *mz;
644 rightmost = rb_last(&mctz->rb_root);
646 goto done; /* Nothing to reclaim from */
648 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
650 * Remove the node now but someone else can add it back,
651 * we will to add it back at the end of reclaim to its correct
652 * position in the tree.
654 __mem_cgroup_remove_exceeded(mz, mctz);
655 if (!soft_limit_excess(mz->memcg) ||
656 !css_tryget_online(&mz->memcg->css))
662 static struct mem_cgroup_per_zone *
663 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
665 struct mem_cgroup_per_zone *mz;
667 spin_lock_irq(&mctz->lock);
668 mz = __mem_cgroup_largest_soft_limit_node(mctz);
669 spin_unlock_irq(&mctz->lock);
674 * Return page count for single (non recursive) @memcg.
676 * Implementation Note: reading percpu statistics for memcg.
678 * Both of vmstat[] and percpu_counter has threshold and do periodic
679 * synchronization to implement "quick" read. There are trade-off between
680 * reading cost and precision of value. Then, we may have a chance to implement
681 * a periodic synchronization of counter in memcg's counter.
683 * But this _read() function is used for user interface now. The user accounts
684 * memory usage by memory cgroup and he _always_ requires exact value because
685 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
686 * have to visit all online cpus and make sum. So, for now, unnecessary
687 * synchronization is not implemented. (just implemented for cpu hotplug)
689 * If there are kernel internal actions which can make use of some not-exact
690 * value, and reading all cpu value can be performance bottleneck in some
691 * common workload, threshold and synchronization as vmstat[] should be
695 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
700 /* Per-cpu values can be negative, use a signed accumulator */
701 for_each_possible_cpu(cpu)
702 val += per_cpu(memcg->stat->count[idx], cpu);
704 * Summing races with updates, so val may be negative. Avoid exposing
705 * transient negative values.
712 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
713 enum mem_cgroup_events_index idx)
715 unsigned long val = 0;
718 for_each_possible_cpu(cpu)
719 val += per_cpu(memcg->stat->events[idx], cpu);
723 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
728 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
729 * counted as CACHE even if it's on ANON LRU.
732 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
735 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
738 if (PageTransHuge(page))
739 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
742 /* pagein of a big page is an event. So, ignore page size */
744 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
746 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
747 nr_pages = -nr_pages; /* for event */
750 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
753 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
755 unsigned int lru_mask)
757 unsigned long nr = 0;
760 VM_BUG_ON((unsigned)nid >= nr_node_ids);
762 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
763 struct mem_cgroup_per_zone *mz;
767 if (!(BIT(lru) & lru_mask))
769 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
770 nr += mz->lru_size[lru];
776 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
777 unsigned int lru_mask)
779 unsigned long nr = 0;
782 for_each_node_state(nid, N_MEMORY)
783 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
787 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
788 enum mem_cgroup_events_target target)
790 unsigned long val, next;
792 val = __this_cpu_read(memcg->stat->nr_page_events);
793 next = __this_cpu_read(memcg->stat->targets[target]);
794 /* from time_after() in jiffies.h */
795 if ((long)next - (long)val < 0) {
797 case MEM_CGROUP_TARGET_THRESH:
798 next = val + THRESHOLDS_EVENTS_TARGET;
800 case MEM_CGROUP_TARGET_SOFTLIMIT:
801 next = val + SOFTLIMIT_EVENTS_TARGET;
803 case MEM_CGROUP_TARGET_NUMAINFO:
804 next = val + NUMAINFO_EVENTS_TARGET;
809 __this_cpu_write(memcg->stat->targets[target], next);
816 * Check events in order.
819 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
821 /* threshold event is triggered in finer grain than soft limit */
822 if (unlikely(mem_cgroup_event_ratelimit(memcg,
823 MEM_CGROUP_TARGET_THRESH))) {
825 bool do_numainfo __maybe_unused;
827 do_softlimit = mem_cgroup_event_ratelimit(memcg,
828 MEM_CGROUP_TARGET_SOFTLIMIT);
830 do_numainfo = mem_cgroup_event_ratelimit(memcg,
831 MEM_CGROUP_TARGET_NUMAINFO);
833 mem_cgroup_threshold(memcg);
834 if (unlikely(do_softlimit))
835 mem_cgroup_update_tree(memcg, page);
837 if (unlikely(do_numainfo))
838 atomic_inc(&memcg->numainfo_events);
843 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
846 * mm_update_next_owner() may clear mm->owner to NULL
847 * if it races with swapoff, page migration, etc.
848 * So this can be called with p == NULL.
853 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
855 EXPORT_SYMBOL(mem_cgroup_from_task);
857 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
859 struct mem_cgroup *memcg = NULL;
864 * Page cache insertions can happen withou an
865 * actual mm context, e.g. during disk probing
866 * on boot, loopback IO, acct() writes etc.
869 memcg = root_mem_cgroup;
871 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
872 if (unlikely(!memcg))
873 memcg = root_mem_cgroup;
875 } while (!css_tryget_online(&memcg->css));
881 * mem_cgroup_iter - iterate over memory cgroup hierarchy
882 * @root: hierarchy root
883 * @prev: previously returned memcg, NULL on first invocation
884 * @reclaim: cookie for shared reclaim walks, NULL for full walks
886 * Returns references to children of the hierarchy below @root, or
887 * @root itself, or %NULL after a full round-trip.
889 * Caller must pass the return value in @prev on subsequent
890 * invocations for reference counting, or use mem_cgroup_iter_break()
891 * to cancel a hierarchy walk before the round-trip is complete.
893 * Reclaimers can specify a zone and a priority level in @reclaim to
894 * divide up the memcgs in the hierarchy among all concurrent
895 * reclaimers operating on the same zone and priority.
897 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
898 struct mem_cgroup *prev,
899 struct mem_cgroup_reclaim_cookie *reclaim)
901 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
902 struct cgroup_subsys_state *css = NULL;
903 struct mem_cgroup *memcg = NULL;
904 struct mem_cgroup *pos = NULL;
906 if (mem_cgroup_disabled())
910 root = root_mem_cgroup;
912 if (prev && !reclaim)
915 if (!root->use_hierarchy && root != root_mem_cgroup) {
924 struct mem_cgroup_per_zone *mz;
926 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
927 iter = &mz->iter[reclaim->priority];
929 if (prev && reclaim->generation != iter->generation)
933 pos = READ_ONCE(iter->position);
934 if (!pos || css_tryget(&pos->css))
937 * css reference reached zero, so iter->position will
938 * be cleared by ->css_released. However, we should not
939 * rely on this happening soon, because ->css_released
940 * is called from a work queue, and by busy-waiting we
941 * might block it. So we clear iter->position right
944 (void)cmpxchg(&iter->position, pos, NULL);
952 css = css_next_descendant_pre(css, &root->css);
955 * Reclaimers share the hierarchy walk, and a
956 * new one might jump in right at the end of
957 * the hierarchy - make sure they see at least
958 * one group and restart from the beginning.
966 * Verify the css and acquire a reference. The root
967 * is provided by the caller, so we know it's alive
968 * and kicking, and don't take an extra reference.
970 memcg = mem_cgroup_from_css(css);
972 if (css == &root->css)
975 if (css_tryget(css)) {
977 * Make sure the memcg is initialized:
978 * mem_cgroup_css_online() orders the the
979 * initialization against setting the flag.
981 if (smp_load_acquire(&memcg->initialized))
992 * The position could have already been updated by a competing
993 * thread, so check that the value hasn't changed since we read
994 * it to avoid reclaiming from the same cgroup twice.
996 (void)cmpxchg(&iter->position, pos, memcg);
1004 reclaim->generation = iter->generation;
1010 if (prev && prev != root)
1011 css_put(&prev->css);
1017 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1018 * @root: hierarchy root
1019 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1021 void mem_cgroup_iter_break(struct mem_cgroup *root,
1022 struct mem_cgroup *prev)
1025 root = root_mem_cgroup;
1026 if (prev && prev != root)
1027 css_put(&prev->css);
1030 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1032 struct mem_cgroup *memcg = dead_memcg;
1033 struct mem_cgroup_reclaim_iter *iter;
1034 struct mem_cgroup_per_zone *mz;
1038 while ((memcg = parent_mem_cgroup(memcg))) {
1039 for_each_node(nid) {
1040 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1041 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
1042 for (i = 0; i <= DEF_PRIORITY; i++) {
1043 iter = &mz->iter[i];
1044 cmpxchg(&iter->position,
1053 * Iteration constructs for visiting all cgroups (under a tree). If
1054 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1055 * be used for reference counting.
1057 #define for_each_mem_cgroup_tree(iter, root) \
1058 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1060 iter = mem_cgroup_iter(root, iter, NULL))
1062 #define for_each_mem_cgroup(iter) \
1063 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1065 iter = mem_cgroup_iter(NULL, iter, NULL))
1068 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1069 * @zone: zone of the wanted lruvec
1070 * @memcg: memcg of the wanted lruvec
1072 * Returns the lru list vector holding pages for the given @zone and
1073 * @mem. This can be the global zone lruvec, if the memory controller
1076 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1077 struct mem_cgroup *memcg)
1079 struct mem_cgroup_per_zone *mz;
1080 struct lruvec *lruvec;
1082 if (mem_cgroup_disabled()) {
1083 lruvec = &zone->lruvec;
1087 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1088 lruvec = &mz->lruvec;
1091 * Since a node can be onlined after the mem_cgroup was created,
1092 * we have to be prepared to initialize lruvec->zone here;
1093 * and if offlined then reonlined, we need to reinitialize it.
1095 if (unlikely(lruvec->zone != zone))
1096 lruvec->zone = zone;
1101 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1103 * @zone: zone of the page
1105 * This function is only safe when following the LRU page isolation
1106 * and putback protocol: the LRU lock must be held, and the page must
1107 * either be PageLRU() or the caller must have isolated/allocated it.
1109 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1111 struct mem_cgroup_per_zone *mz;
1112 struct mem_cgroup *memcg;
1113 struct lruvec *lruvec;
1115 if (mem_cgroup_disabled()) {
1116 lruvec = &zone->lruvec;
1120 memcg = page->mem_cgroup;
1122 * Swapcache readahead pages are added to the LRU - and
1123 * possibly migrated - before they are charged.
1126 memcg = root_mem_cgroup;
1128 mz = mem_cgroup_page_zoneinfo(memcg, page);
1129 lruvec = &mz->lruvec;
1132 * Since a node can be onlined after the mem_cgroup was created,
1133 * we have to be prepared to initialize lruvec->zone here;
1134 * and if offlined then reonlined, we need to reinitialize it.
1136 if (unlikely(lruvec->zone != zone))
1137 lruvec->zone = zone;
1142 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1143 * @lruvec: mem_cgroup per zone lru vector
1144 * @lru: index of lru list the page is sitting on
1145 * @nr_pages: positive when adding or negative when removing
1147 * This function must be called when a page is added to or removed from an
1150 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1153 struct mem_cgroup_per_zone *mz;
1154 unsigned long *lru_size;
1156 if (mem_cgroup_disabled())
1159 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1160 lru_size = mz->lru_size + lru;
1161 *lru_size += nr_pages;
1162 VM_BUG_ON((long)(*lru_size) < 0);
1165 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1167 struct mem_cgroup *task_memcg;
1168 struct task_struct *p;
1171 p = find_lock_task_mm(task);
1173 task_memcg = get_mem_cgroup_from_mm(p->mm);
1177 * All threads may have already detached their mm's, but the oom
1178 * killer still needs to detect if they have already been oom
1179 * killed to prevent needlessly killing additional tasks.
1182 task_memcg = mem_cgroup_from_task(task);
1183 css_get(&task_memcg->css);
1186 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1187 css_put(&task_memcg->css);
1191 #define mem_cgroup_from_counter(counter, member) \
1192 container_of(counter, struct mem_cgroup, member)
1195 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1196 * @memcg: the memory cgroup
1198 * Returns the maximum amount of memory @mem can be charged with, in
1201 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1203 unsigned long margin = 0;
1204 unsigned long count;
1205 unsigned long limit;
1207 count = page_counter_read(&memcg->memory);
1208 limit = READ_ONCE(memcg->memory.limit);
1210 margin = limit - count;
1212 if (do_swap_account) {
1213 count = page_counter_read(&memcg->memsw);
1214 limit = READ_ONCE(memcg->memsw.limit);
1216 margin = min(margin, limit - count);
1223 * A routine for checking "mem" is under move_account() or not.
1225 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1226 * moving cgroups. This is for waiting at high-memory pressure
1229 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1231 struct mem_cgroup *from;
1232 struct mem_cgroup *to;
1235 * Unlike task_move routines, we access mc.to, mc.from not under
1236 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1238 spin_lock(&mc.lock);
1244 ret = mem_cgroup_is_descendant(from, memcg) ||
1245 mem_cgroup_is_descendant(to, memcg);
1247 spin_unlock(&mc.lock);
1251 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1253 if (mc.moving_task && current != mc.moving_task) {
1254 if (mem_cgroup_under_move(memcg)) {
1256 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1257 /* moving charge context might have finished. */
1260 finish_wait(&mc.waitq, &wait);
1267 #define K(x) ((x) << (PAGE_SHIFT-10))
1269 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1270 * @memcg: The memory cgroup that went over limit
1271 * @p: Task that is going to be killed
1273 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1276 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1278 /* oom_info_lock ensures that parallel ooms do not interleave */
1279 static DEFINE_MUTEX(oom_info_lock);
1280 struct mem_cgroup *iter;
1283 mutex_lock(&oom_info_lock);
1287 pr_info("Task in ");
1288 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1289 pr_cont(" killed as a result of limit of ");
1291 pr_info("Memory limit reached of cgroup ");
1294 pr_cont_cgroup_path(memcg->css.cgroup);
1299 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1300 K((u64)page_counter_read(&memcg->memory)),
1301 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1302 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1303 K((u64)page_counter_read(&memcg->memsw)),
1304 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1305 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1306 K((u64)page_counter_read(&memcg->kmem)),
1307 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1309 for_each_mem_cgroup_tree(iter, memcg) {
1310 pr_info("Memory cgroup stats for ");
1311 pr_cont_cgroup_path(iter->css.cgroup);
1314 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1315 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1317 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1318 K(mem_cgroup_read_stat(iter, i)));
1321 for (i = 0; i < NR_LRU_LISTS; i++)
1322 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1323 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1327 mutex_unlock(&oom_info_lock);
1331 * This function returns the number of memcg under hierarchy tree. Returns
1332 * 1(self count) if no children.
1334 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1337 struct mem_cgroup *iter;
1339 for_each_mem_cgroup_tree(iter, memcg)
1345 * Return the memory (and swap, if configured) limit for a memcg.
1347 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1349 unsigned long limit;
1351 limit = memcg->memory.limit;
1352 if (mem_cgroup_swappiness(memcg)) {
1353 unsigned long memsw_limit;
1355 memsw_limit = memcg->memsw.limit;
1356 limit = min(limit + total_swap_pages, memsw_limit);
1361 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1364 struct oom_control oc = {
1367 .gfp_mask = gfp_mask,
1370 struct mem_cgroup *iter;
1371 unsigned long chosen_points = 0;
1372 unsigned long totalpages;
1373 unsigned int points = 0;
1374 struct task_struct *chosen = NULL;
1376 mutex_lock(&oom_lock);
1379 * If current has a pending SIGKILL or is exiting, then automatically
1380 * select it. The goal is to allow it to allocate so that it may
1381 * quickly exit and free its memory.
1383 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1384 mark_oom_victim(current);
1388 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1389 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1390 for_each_mem_cgroup_tree(iter, memcg) {
1391 struct css_task_iter it;
1392 struct task_struct *task;
1394 css_task_iter_start(&iter->css, &it);
1395 while ((task = css_task_iter_next(&it))) {
1396 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1397 case OOM_SCAN_SELECT:
1399 put_task_struct(chosen);
1401 chosen_points = ULONG_MAX;
1402 get_task_struct(chosen);
1404 case OOM_SCAN_CONTINUE:
1406 case OOM_SCAN_ABORT:
1407 css_task_iter_end(&it);
1408 mem_cgroup_iter_break(memcg, iter);
1410 put_task_struct(chosen);
1415 points = oom_badness(task, memcg, NULL, totalpages);
1416 if (!points || points < chosen_points)
1418 /* Prefer thread group leaders for display purposes */
1419 if (points == chosen_points &&
1420 thread_group_leader(chosen))
1424 put_task_struct(chosen);
1426 chosen_points = points;
1427 get_task_struct(chosen);
1429 css_task_iter_end(&it);
1433 points = chosen_points * 1000 / totalpages;
1434 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1435 "Memory cgroup out of memory");
1438 mutex_unlock(&oom_lock);
1441 #if MAX_NUMNODES > 1
1444 * test_mem_cgroup_node_reclaimable
1445 * @memcg: the target memcg
1446 * @nid: the node ID to be checked.
1447 * @noswap : specify true here if the user wants flle only information.
1449 * This function returns whether the specified memcg contains any
1450 * reclaimable pages on a node. Returns true if there are any reclaimable
1451 * pages in the node.
1453 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1454 int nid, bool noswap)
1456 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1458 if (noswap || !total_swap_pages)
1460 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1467 * Always updating the nodemask is not very good - even if we have an empty
1468 * list or the wrong list here, we can start from some node and traverse all
1469 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1472 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1476 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1477 * pagein/pageout changes since the last update.
1479 if (!atomic_read(&memcg->numainfo_events))
1481 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1484 /* make a nodemask where this memcg uses memory from */
1485 memcg->scan_nodes = node_states[N_MEMORY];
1487 for_each_node_mask(nid, node_states[N_MEMORY]) {
1489 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1490 node_clear(nid, memcg->scan_nodes);
1493 atomic_set(&memcg->numainfo_events, 0);
1494 atomic_set(&memcg->numainfo_updating, 0);
1498 * Selecting a node where we start reclaim from. Because what we need is just
1499 * reducing usage counter, start from anywhere is O,K. Considering
1500 * memory reclaim from current node, there are pros. and cons.
1502 * Freeing memory from current node means freeing memory from a node which
1503 * we'll use or we've used. So, it may make LRU bad. And if several threads
1504 * hit limits, it will see a contention on a node. But freeing from remote
1505 * node means more costs for memory reclaim because of memory latency.
1507 * Now, we use round-robin. Better algorithm is welcomed.
1509 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1513 mem_cgroup_may_update_nodemask(memcg);
1514 node = memcg->last_scanned_node;
1516 node = next_node(node, memcg->scan_nodes);
1517 if (node == MAX_NUMNODES)
1518 node = first_node(memcg->scan_nodes);
1520 * We call this when we hit limit, not when pages are added to LRU.
1521 * No LRU may hold pages because all pages are UNEVICTABLE or
1522 * memcg is too small and all pages are not on LRU. In that case,
1523 * we use curret node.
1525 if (unlikely(node == MAX_NUMNODES))
1526 node = numa_node_id();
1528 memcg->last_scanned_node = node;
1532 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1538 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1541 unsigned long *total_scanned)
1543 struct mem_cgroup *victim = NULL;
1546 unsigned long excess;
1547 unsigned long nr_scanned;
1548 struct mem_cgroup_reclaim_cookie reclaim = {
1553 excess = soft_limit_excess(root_memcg);
1556 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1561 * If we have not been able to reclaim
1562 * anything, it might because there are
1563 * no reclaimable pages under this hierarchy
1568 * We want to do more targeted reclaim.
1569 * excess >> 2 is not to excessive so as to
1570 * reclaim too much, nor too less that we keep
1571 * coming back to reclaim from this cgroup
1573 if (total >= (excess >> 2) ||
1574 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1579 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1581 *total_scanned += nr_scanned;
1582 if (!soft_limit_excess(root_memcg))
1585 mem_cgroup_iter_break(root_memcg, victim);
1589 #ifdef CONFIG_LOCKDEP
1590 static struct lockdep_map memcg_oom_lock_dep_map = {
1591 .name = "memcg_oom_lock",
1595 static DEFINE_SPINLOCK(memcg_oom_lock);
1598 * Check OOM-Killer is already running under our hierarchy.
1599 * If someone is running, return false.
1601 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1603 struct mem_cgroup *iter, *failed = NULL;
1605 spin_lock(&memcg_oom_lock);
1607 for_each_mem_cgroup_tree(iter, memcg) {
1608 if (iter->oom_lock) {
1610 * this subtree of our hierarchy is already locked
1611 * so we cannot give a lock.
1614 mem_cgroup_iter_break(memcg, iter);
1617 iter->oom_lock = true;
1622 * OK, we failed to lock the whole subtree so we have
1623 * to clean up what we set up to the failing subtree
1625 for_each_mem_cgroup_tree(iter, memcg) {
1626 if (iter == failed) {
1627 mem_cgroup_iter_break(memcg, iter);
1630 iter->oom_lock = false;
1633 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1635 spin_unlock(&memcg_oom_lock);
1640 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1642 struct mem_cgroup *iter;
1644 spin_lock(&memcg_oom_lock);
1645 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1646 for_each_mem_cgroup_tree(iter, memcg)
1647 iter->oom_lock = false;
1648 spin_unlock(&memcg_oom_lock);
1651 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1653 struct mem_cgroup *iter;
1655 spin_lock(&memcg_oom_lock);
1656 for_each_mem_cgroup_tree(iter, memcg)
1658 spin_unlock(&memcg_oom_lock);
1661 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1663 struct mem_cgroup *iter;
1666 * When a new child is created while the hierarchy is under oom,
1667 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1669 spin_lock(&memcg_oom_lock);
1670 for_each_mem_cgroup_tree(iter, memcg)
1671 if (iter->under_oom > 0)
1673 spin_unlock(&memcg_oom_lock);
1676 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1678 struct oom_wait_info {
1679 struct mem_cgroup *memcg;
1683 static int memcg_oom_wake_function(wait_queue_t *wait,
1684 unsigned mode, int sync, void *arg)
1686 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1687 struct mem_cgroup *oom_wait_memcg;
1688 struct oom_wait_info *oom_wait_info;
1690 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1691 oom_wait_memcg = oom_wait_info->memcg;
1693 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1694 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1696 return autoremove_wake_function(wait, mode, sync, arg);
1699 static void memcg_oom_recover(struct mem_cgroup *memcg)
1702 * For the following lockless ->under_oom test, the only required
1703 * guarantee is that it must see the state asserted by an OOM when
1704 * this function is called as a result of userland actions
1705 * triggered by the notification of the OOM. This is trivially
1706 * achieved by invoking mem_cgroup_mark_under_oom() before
1707 * triggering notification.
1709 if (memcg && memcg->under_oom)
1710 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1713 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1715 if (!current->memcg_may_oom)
1718 * We are in the middle of the charge context here, so we
1719 * don't want to block when potentially sitting on a callstack
1720 * that holds all kinds of filesystem and mm locks.
1722 * Also, the caller may handle a failed allocation gracefully
1723 * (like optional page cache readahead) and so an OOM killer
1724 * invocation might not even be necessary.
1726 * That's why we don't do anything here except remember the
1727 * OOM context and then deal with it at the end of the page
1728 * fault when the stack is unwound, the locks are released,
1729 * and when we know whether the fault was overall successful.
1731 css_get(&memcg->css);
1732 current->memcg_in_oom = memcg;
1733 current->memcg_oom_gfp_mask = mask;
1734 current->memcg_oom_order = order;
1738 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1739 * @handle: actually kill/wait or just clean up the OOM state
1741 * This has to be called at the end of a page fault if the memcg OOM
1742 * handler was enabled.
1744 * Memcg supports userspace OOM handling where failed allocations must
1745 * sleep on a waitqueue until the userspace task resolves the
1746 * situation. Sleeping directly in the charge context with all kinds
1747 * of locks held is not a good idea, instead we remember an OOM state
1748 * in the task and mem_cgroup_oom_synchronize() has to be called at
1749 * the end of the page fault to complete the OOM handling.
1751 * Returns %true if an ongoing memcg OOM situation was detected and
1752 * completed, %false otherwise.
1754 bool mem_cgroup_oom_synchronize(bool handle)
1756 struct mem_cgroup *memcg = current->memcg_in_oom;
1757 struct oom_wait_info owait;
1760 /* OOM is global, do not handle */
1764 if (!handle || oom_killer_disabled)
1767 owait.memcg = memcg;
1768 owait.wait.flags = 0;
1769 owait.wait.func = memcg_oom_wake_function;
1770 owait.wait.private = current;
1771 INIT_LIST_HEAD(&owait.wait.task_list);
1773 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1774 mem_cgroup_mark_under_oom(memcg);
1776 locked = mem_cgroup_oom_trylock(memcg);
1779 mem_cgroup_oom_notify(memcg);
1781 if (locked && !memcg->oom_kill_disable) {
1782 mem_cgroup_unmark_under_oom(memcg);
1783 finish_wait(&memcg_oom_waitq, &owait.wait);
1784 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1785 current->memcg_oom_order);
1788 mem_cgroup_unmark_under_oom(memcg);
1789 finish_wait(&memcg_oom_waitq, &owait.wait);
1793 mem_cgroup_oom_unlock(memcg);
1795 * There is no guarantee that an OOM-lock contender
1796 * sees the wakeups triggered by the OOM kill
1797 * uncharges. Wake any sleepers explicitely.
1799 memcg_oom_recover(memcg);
1802 current->memcg_in_oom = NULL;
1803 css_put(&memcg->css);
1808 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1809 * @page: page that is going to change accounted state
1811 * This function must mark the beginning of an accounted page state
1812 * change to prevent double accounting when the page is concurrently
1813 * being moved to another memcg:
1815 * memcg = mem_cgroup_begin_page_stat(page);
1816 * if (TestClearPageState(page))
1817 * mem_cgroup_update_page_stat(memcg, state, -1);
1818 * mem_cgroup_end_page_stat(memcg);
1820 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1822 struct mem_cgroup *memcg;
1823 unsigned long flags;
1826 * The RCU lock is held throughout the transaction. The fast
1827 * path can get away without acquiring the memcg->move_lock
1828 * because page moving starts with an RCU grace period.
1830 * The RCU lock also protects the memcg from being freed when
1831 * the page state that is going to change is the only thing
1832 * preventing the page from being uncharged.
1833 * E.g. end-writeback clearing PageWriteback(), which allows
1834 * migration to go ahead and uncharge the page before the
1835 * account transaction might be complete.
1839 if (mem_cgroup_disabled())
1842 memcg = page->mem_cgroup;
1843 if (unlikely(!memcg))
1846 if (atomic_read(&memcg->moving_account) <= 0)
1849 spin_lock_irqsave(&memcg->move_lock, flags);
1850 if (memcg != page->mem_cgroup) {
1851 spin_unlock_irqrestore(&memcg->move_lock, flags);
1856 * When charge migration first begins, we can have locked and
1857 * unlocked page stat updates happening concurrently. Track
1858 * the task who has the lock for mem_cgroup_end_page_stat().
1860 memcg->move_lock_task = current;
1861 memcg->move_lock_flags = flags;
1865 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1868 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1869 * @memcg: the memcg that was accounted against
1871 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1873 if (memcg && memcg->move_lock_task == current) {
1874 unsigned long flags = memcg->move_lock_flags;
1876 memcg->move_lock_task = NULL;
1877 memcg->move_lock_flags = 0;
1879 spin_unlock_irqrestore(&memcg->move_lock, flags);
1884 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1887 * size of first charge trial. "32" comes from vmscan.c's magic value.
1888 * TODO: maybe necessary to use big numbers in big irons.
1890 #define CHARGE_BATCH 32U
1891 struct memcg_stock_pcp {
1892 struct mem_cgroup *cached; /* this never be root cgroup */
1893 unsigned int nr_pages;
1894 struct work_struct work;
1895 unsigned long flags;
1896 #define FLUSHING_CACHED_CHARGE 0
1898 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1899 static DEFINE_MUTEX(percpu_charge_mutex);
1902 * consume_stock: Try to consume stocked charge on this cpu.
1903 * @memcg: memcg to consume from.
1904 * @nr_pages: how many pages to charge.
1906 * The charges will only happen if @memcg matches the current cpu's memcg
1907 * stock, and at least @nr_pages are available in that stock. Failure to
1908 * service an allocation will refill the stock.
1910 * returns true if successful, false otherwise.
1912 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1914 struct memcg_stock_pcp *stock;
1917 if (nr_pages > CHARGE_BATCH)
1920 stock = &get_cpu_var(memcg_stock);
1921 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1922 stock->nr_pages -= nr_pages;
1925 put_cpu_var(memcg_stock);
1930 * Returns stocks cached in percpu and reset cached information.
1932 static void drain_stock(struct memcg_stock_pcp *stock)
1934 struct mem_cgroup *old = stock->cached;
1936 if (stock->nr_pages) {
1937 page_counter_uncharge(&old->memory, stock->nr_pages);
1938 if (do_swap_account)
1939 page_counter_uncharge(&old->memsw, stock->nr_pages);
1940 css_put_many(&old->css, stock->nr_pages);
1941 stock->nr_pages = 0;
1943 stock->cached = NULL;
1947 * This must be called under preempt disabled or must be called by
1948 * a thread which is pinned to local cpu.
1950 static void drain_local_stock(struct work_struct *dummy)
1952 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1954 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1958 * Cache charges(val) to local per_cpu area.
1959 * This will be consumed by consume_stock() function, later.
1961 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1963 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1965 if (stock->cached != memcg) { /* reset if necessary */
1967 stock->cached = memcg;
1969 stock->nr_pages += nr_pages;
1970 put_cpu_var(memcg_stock);
1974 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1975 * of the hierarchy under it.
1977 static void drain_all_stock(struct mem_cgroup *root_memcg)
1981 /* If someone's already draining, avoid adding running more workers. */
1982 if (!mutex_trylock(&percpu_charge_mutex))
1984 /* Notify other cpus that system-wide "drain" is running */
1987 for_each_online_cpu(cpu) {
1988 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1989 struct mem_cgroup *memcg;
1991 memcg = stock->cached;
1992 if (!memcg || !stock->nr_pages)
1994 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1996 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1998 drain_local_stock(&stock->work);
2000 schedule_work_on(cpu, &stock->work);
2005 mutex_unlock(&percpu_charge_mutex);
2008 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2009 unsigned long action,
2012 int cpu = (unsigned long)hcpu;
2013 struct memcg_stock_pcp *stock;
2015 if (action == CPU_ONLINE)
2018 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2021 stock = &per_cpu(memcg_stock, cpu);
2027 * Scheduled by try_charge() to be executed from the userland return path
2028 * and reclaims memory over the high limit.
2030 void mem_cgroup_handle_over_high(void)
2032 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2033 struct mem_cgroup *memcg, *pos;
2035 if (likely(!nr_pages))
2038 pos = memcg = get_mem_cgroup_from_mm(current->mm);
2041 if (page_counter_read(&pos->memory) <= pos->high)
2043 mem_cgroup_events(pos, MEMCG_HIGH, 1);
2044 try_to_free_mem_cgroup_pages(pos, nr_pages, GFP_KERNEL, true);
2045 } while ((pos = parent_mem_cgroup(pos)));
2047 css_put(&memcg->css);
2048 current->memcg_nr_pages_over_high = 0;
2051 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2052 unsigned int nr_pages)
2054 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2055 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2056 struct mem_cgroup *mem_over_limit;
2057 struct page_counter *counter;
2058 unsigned long nr_reclaimed;
2059 bool may_swap = true;
2060 bool drained = false;
2062 if (mem_cgroup_is_root(memcg))
2065 if (consume_stock(memcg, nr_pages))
2068 if (!do_swap_account ||
2069 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2070 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2072 if (do_swap_account)
2073 page_counter_uncharge(&memcg->memsw, batch);
2074 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2076 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2080 if (batch > nr_pages) {
2086 * Unlike in global OOM situations, memcg is not in a physical
2087 * memory shortage. Allow dying and OOM-killed tasks to
2088 * bypass the last charges so that they can exit quickly and
2089 * free their memory.
2091 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2092 fatal_signal_pending(current) ||
2093 current->flags & PF_EXITING))
2096 if (unlikely(task_in_memcg_oom(current)))
2099 if (!gfpflags_allow_blocking(gfp_mask))
2102 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2104 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2105 gfp_mask, may_swap);
2107 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2111 drain_all_stock(mem_over_limit);
2116 if (gfp_mask & __GFP_NORETRY)
2119 * Even though the limit is exceeded at this point, reclaim
2120 * may have been able to free some pages. Retry the charge
2121 * before killing the task.
2123 * Only for regular pages, though: huge pages are rather
2124 * unlikely to succeed so close to the limit, and we fall back
2125 * to regular pages anyway in case of failure.
2127 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2130 * At task move, charge accounts can be doubly counted. So, it's
2131 * better to wait until the end of task_move if something is going on.
2133 if (mem_cgroup_wait_acct_move(mem_over_limit))
2139 if (gfp_mask & __GFP_NOFAIL)
2142 if (fatal_signal_pending(current))
2145 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2147 mem_cgroup_oom(mem_over_limit, gfp_mask,
2148 get_order(nr_pages * PAGE_SIZE));
2150 if (!(gfp_mask & __GFP_NOFAIL))
2154 * The allocation either can't fail or will lead to more memory
2155 * being freed very soon. Allow memory usage go over the limit
2156 * temporarily by force charging it.
2158 page_counter_charge(&memcg->memory, nr_pages);
2159 if (do_swap_account)
2160 page_counter_charge(&memcg->memsw, nr_pages);
2161 css_get_many(&memcg->css, nr_pages);
2166 css_get_many(&memcg->css, batch);
2167 if (batch > nr_pages)
2168 refill_stock(memcg, batch - nr_pages);
2171 * If the hierarchy is above the normal consumption range, schedule
2172 * reclaim on returning to userland. We can perform reclaim here
2173 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2174 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2175 * not recorded as it most likely matches current's and won't
2176 * change in the meantime. As high limit is checked again before
2177 * reclaim, the cost of mismatch is negligible.
2180 if (page_counter_read(&memcg->memory) > memcg->high) {
2181 current->memcg_nr_pages_over_high += batch;
2182 set_notify_resume(current);
2185 } while ((memcg = parent_mem_cgroup(memcg)));
2190 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2192 if (mem_cgroup_is_root(memcg))
2195 page_counter_uncharge(&memcg->memory, nr_pages);
2196 if (do_swap_account)
2197 page_counter_uncharge(&memcg->memsw, nr_pages);
2199 css_put_many(&memcg->css, nr_pages);
2202 static void lock_page_lru(struct page *page, int *isolated)
2204 struct zone *zone = page_zone(page);
2206 spin_lock_irq(&zone->lru_lock);
2207 if (PageLRU(page)) {
2208 struct lruvec *lruvec;
2210 lruvec = mem_cgroup_page_lruvec(page, zone);
2212 del_page_from_lru_list(page, lruvec, page_lru(page));
2218 static void unlock_page_lru(struct page *page, int isolated)
2220 struct zone *zone = page_zone(page);
2223 struct lruvec *lruvec;
2225 lruvec = mem_cgroup_page_lruvec(page, zone);
2226 VM_BUG_ON_PAGE(PageLRU(page), page);
2228 add_page_to_lru_list(page, lruvec, page_lru(page));
2230 spin_unlock_irq(&zone->lru_lock);
2233 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2238 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2241 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2242 * may already be on some other mem_cgroup's LRU. Take care of it.
2245 lock_page_lru(page, &isolated);
2248 * Nobody should be changing or seriously looking at
2249 * page->mem_cgroup at this point:
2251 * - the page is uncharged
2253 * - the page is off-LRU
2255 * - an anonymous fault has exclusive page access, except for
2256 * a locked page table
2258 * - a page cache insertion, a swapin fault, or a migration
2259 * have the page locked
2261 page->mem_cgroup = memcg;
2264 unlock_page_lru(page, isolated);
2267 #ifdef CONFIG_MEMCG_KMEM
2268 static int memcg_alloc_cache_id(void)
2273 id = ida_simple_get(&memcg_cache_ida,
2274 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2278 if (id < memcg_nr_cache_ids)
2282 * There's no space for the new id in memcg_caches arrays,
2283 * so we have to grow them.
2285 down_write(&memcg_cache_ids_sem);
2287 size = 2 * (id + 1);
2288 if (size < MEMCG_CACHES_MIN_SIZE)
2289 size = MEMCG_CACHES_MIN_SIZE;
2290 else if (size > MEMCG_CACHES_MAX_SIZE)
2291 size = MEMCG_CACHES_MAX_SIZE;
2293 err = memcg_update_all_caches(size);
2295 err = memcg_update_all_list_lrus(size);
2297 memcg_nr_cache_ids = size;
2299 up_write(&memcg_cache_ids_sem);
2302 ida_simple_remove(&memcg_cache_ida, id);
2308 static void memcg_free_cache_id(int id)
2310 ida_simple_remove(&memcg_cache_ida, id);
2313 struct memcg_kmem_cache_create_work {
2314 struct mem_cgroup *memcg;
2315 struct kmem_cache *cachep;
2316 struct work_struct work;
2319 static void memcg_kmem_cache_create_func(struct work_struct *w)
2321 struct memcg_kmem_cache_create_work *cw =
2322 container_of(w, struct memcg_kmem_cache_create_work, work);
2323 struct mem_cgroup *memcg = cw->memcg;
2324 struct kmem_cache *cachep = cw->cachep;
2326 memcg_create_kmem_cache(memcg, cachep);
2328 css_put(&memcg->css);
2333 * Enqueue the creation of a per-memcg kmem_cache.
2335 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2336 struct kmem_cache *cachep)
2338 struct memcg_kmem_cache_create_work *cw;
2340 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2344 css_get(&memcg->css);
2347 cw->cachep = cachep;
2348 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2350 schedule_work(&cw->work);
2353 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2354 struct kmem_cache *cachep)
2357 * We need to stop accounting when we kmalloc, because if the
2358 * corresponding kmalloc cache is not yet created, the first allocation
2359 * in __memcg_schedule_kmem_cache_create will recurse.
2361 * However, it is better to enclose the whole function. Depending on
2362 * the debugging options enabled, INIT_WORK(), for instance, can
2363 * trigger an allocation. This too, will make us recurse. Because at
2364 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2365 * the safest choice is to do it like this, wrapping the whole function.
2367 current->memcg_kmem_skip_account = 1;
2368 __memcg_schedule_kmem_cache_create(memcg, cachep);
2369 current->memcg_kmem_skip_account = 0;
2373 * Return the kmem_cache we're supposed to use for a slab allocation.
2374 * We try to use the current memcg's version of the cache.
2376 * If the cache does not exist yet, if we are the first user of it,
2377 * we either create it immediately, if possible, or create it asynchronously
2379 * In the latter case, we will let the current allocation go through with
2380 * the original cache.
2382 * Can't be called in interrupt context or from kernel threads.
2383 * This function needs to be called with rcu_read_lock() held.
2385 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp)
2387 struct mem_cgroup *memcg;
2388 struct kmem_cache *memcg_cachep;
2391 VM_BUG_ON(!is_root_cache(cachep));
2393 if (cachep->flags & SLAB_ACCOUNT)
2394 gfp |= __GFP_ACCOUNT;
2396 if (!(gfp & __GFP_ACCOUNT))
2399 if (current->memcg_kmem_skip_account)
2402 memcg = get_mem_cgroup_from_mm(current->mm);
2403 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2407 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2408 if (likely(memcg_cachep))
2409 return memcg_cachep;
2412 * If we are in a safe context (can wait, and not in interrupt
2413 * context), we could be be predictable and return right away.
2414 * This would guarantee that the allocation being performed
2415 * already belongs in the new cache.
2417 * However, there are some clashes that can arrive from locking.
2418 * For instance, because we acquire the slab_mutex while doing
2419 * memcg_create_kmem_cache, this means no further allocation
2420 * could happen with the slab_mutex held. So it's better to
2423 memcg_schedule_kmem_cache_create(memcg, cachep);
2425 css_put(&memcg->css);
2429 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2431 if (!is_root_cache(cachep))
2432 css_put(&cachep->memcg_params.memcg->css);
2435 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2436 struct mem_cgroup *memcg)
2438 unsigned int nr_pages = 1 << order;
2439 struct page_counter *counter;
2442 if (!memcg_kmem_is_active(memcg))
2445 if (!page_counter_try_charge(&memcg->kmem, nr_pages, &counter))
2448 ret = try_charge(memcg, gfp, nr_pages);
2450 page_counter_uncharge(&memcg->kmem, nr_pages);
2454 page->mem_cgroup = memcg;
2459 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2461 struct mem_cgroup *memcg;
2464 memcg = get_mem_cgroup_from_mm(current->mm);
2465 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2466 css_put(&memcg->css);
2470 void __memcg_kmem_uncharge(struct page *page, int order)
2472 struct mem_cgroup *memcg = page->mem_cgroup;
2473 unsigned int nr_pages = 1 << order;
2478 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2480 page_counter_uncharge(&memcg->kmem, nr_pages);
2481 page_counter_uncharge(&memcg->memory, nr_pages);
2482 if (do_swap_account)
2483 page_counter_uncharge(&memcg->memsw, nr_pages);
2485 page->mem_cgroup = NULL;
2486 css_put_many(&memcg->css, nr_pages);
2488 #endif /* CONFIG_MEMCG_KMEM */
2490 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2493 * Because tail pages are not marked as "used", set it. We're under
2494 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2495 * charge/uncharge will be never happen and move_account() is done under
2496 * compound_lock(), so we don't have to take care of races.
2498 void mem_cgroup_split_huge_fixup(struct page *head)
2502 if (mem_cgroup_disabled())
2505 for (i = 1; i < HPAGE_PMD_NR; i++)
2506 head[i].mem_cgroup = head->mem_cgroup;
2508 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2511 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2513 #ifdef CONFIG_MEMCG_SWAP
2514 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2517 int val = (charge) ? 1 : -1;
2518 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2522 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2523 * @entry: swap entry to be moved
2524 * @from: mem_cgroup which the entry is moved from
2525 * @to: mem_cgroup which the entry is moved to
2527 * It succeeds only when the swap_cgroup's record for this entry is the same
2528 * as the mem_cgroup's id of @from.
2530 * Returns 0 on success, -EINVAL on failure.
2532 * The caller must have charged to @to, IOW, called page_counter_charge() about
2533 * both res and memsw, and called css_get().
2535 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2536 struct mem_cgroup *from, struct mem_cgroup *to)
2538 unsigned short old_id, new_id;
2540 old_id = mem_cgroup_id(from);
2541 new_id = mem_cgroup_id(to);
2543 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2544 mem_cgroup_swap_statistics(from, false);
2545 mem_cgroup_swap_statistics(to, true);
2551 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2552 struct mem_cgroup *from, struct mem_cgroup *to)
2558 static DEFINE_MUTEX(memcg_limit_mutex);
2560 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2561 unsigned long limit)
2563 unsigned long curusage;
2564 unsigned long oldusage;
2565 bool enlarge = false;
2570 * For keeping hierarchical_reclaim simple, how long we should retry
2571 * is depends on callers. We set our retry-count to be function
2572 * of # of children which we should visit in this loop.
2574 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2575 mem_cgroup_count_children(memcg);
2577 oldusage = page_counter_read(&memcg->memory);
2580 if (signal_pending(current)) {
2585 mutex_lock(&memcg_limit_mutex);
2586 if (limit > memcg->memsw.limit) {
2587 mutex_unlock(&memcg_limit_mutex);
2591 if (limit > memcg->memory.limit)
2593 ret = page_counter_limit(&memcg->memory, limit);
2594 mutex_unlock(&memcg_limit_mutex);
2599 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2601 curusage = page_counter_read(&memcg->memory);
2602 /* Usage is reduced ? */
2603 if (curusage >= oldusage)
2606 oldusage = curusage;
2607 } while (retry_count);
2609 if (!ret && enlarge)
2610 memcg_oom_recover(memcg);
2615 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2616 unsigned long limit)
2618 unsigned long curusage;
2619 unsigned long oldusage;
2620 bool enlarge = false;
2624 /* see mem_cgroup_resize_res_limit */
2625 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2626 mem_cgroup_count_children(memcg);
2628 oldusage = page_counter_read(&memcg->memsw);
2631 if (signal_pending(current)) {
2636 mutex_lock(&memcg_limit_mutex);
2637 if (limit < memcg->memory.limit) {
2638 mutex_unlock(&memcg_limit_mutex);
2642 if (limit > memcg->memsw.limit)
2644 ret = page_counter_limit(&memcg->memsw, limit);
2645 mutex_unlock(&memcg_limit_mutex);
2650 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2652 curusage = page_counter_read(&memcg->memsw);
2653 /* Usage is reduced ? */
2654 if (curusage >= oldusage)
2657 oldusage = curusage;
2658 } while (retry_count);
2660 if (!ret && enlarge)
2661 memcg_oom_recover(memcg);
2666 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2668 unsigned long *total_scanned)
2670 unsigned long nr_reclaimed = 0;
2671 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2672 unsigned long reclaimed;
2674 struct mem_cgroup_tree_per_zone *mctz;
2675 unsigned long excess;
2676 unsigned long nr_scanned;
2681 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2683 * This loop can run a while, specially if mem_cgroup's continuously
2684 * keep exceeding their soft limit and putting the system under
2691 mz = mem_cgroup_largest_soft_limit_node(mctz);
2696 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2697 gfp_mask, &nr_scanned);
2698 nr_reclaimed += reclaimed;
2699 *total_scanned += nr_scanned;
2700 spin_lock_irq(&mctz->lock);
2701 __mem_cgroup_remove_exceeded(mz, mctz);
2704 * If we failed to reclaim anything from this memory cgroup
2705 * it is time to move on to the next cgroup
2709 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2711 excess = soft_limit_excess(mz->memcg);
2713 * One school of thought says that we should not add
2714 * back the node to the tree if reclaim returns 0.
2715 * But our reclaim could return 0, simply because due
2716 * to priority we are exposing a smaller subset of
2717 * memory to reclaim from. Consider this as a longer
2720 /* If excess == 0, no tree ops */
2721 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2722 spin_unlock_irq(&mctz->lock);
2723 css_put(&mz->memcg->css);
2726 * Could not reclaim anything and there are no more
2727 * mem cgroups to try or we seem to be looping without
2728 * reclaiming anything.
2730 if (!nr_reclaimed &&
2732 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2734 } while (!nr_reclaimed);
2736 css_put(&next_mz->memcg->css);
2737 return nr_reclaimed;
2741 * Test whether @memcg has children, dead or alive. Note that this
2742 * function doesn't care whether @memcg has use_hierarchy enabled and
2743 * returns %true if there are child csses according to the cgroup
2744 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2746 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2751 * The lock does not prevent addition or deletion of children, but
2752 * it prevents a new child from being initialized based on this
2753 * parent in css_online(), so it's enough to decide whether
2754 * hierarchically inherited attributes can still be changed or not.
2756 lockdep_assert_held(&memcg_create_mutex);
2759 ret = css_next_child(NULL, &memcg->css);
2765 * Reclaims as many pages from the given memcg as possible and moves
2766 * the rest to the parent.
2768 * Caller is responsible for holding css reference for memcg.
2770 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2772 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2774 /* we call try-to-free pages for make this cgroup empty */
2775 lru_add_drain_all();
2776 /* try to free all pages in this cgroup */
2777 while (nr_retries && page_counter_read(&memcg->memory)) {
2780 if (signal_pending(current))
2783 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2787 /* maybe some writeback is necessary */
2788 congestion_wait(BLK_RW_ASYNC, HZ/10);
2796 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2797 char *buf, size_t nbytes,
2800 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2802 if (mem_cgroup_is_root(memcg))
2804 return mem_cgroup_force_empty(memcg) ?: nbytes;
2807 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2810 return mem_cgroup_from_css(css)->use_hierarchy;
2813 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2814 struct cftype *cft, u64 val)
2817 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2818 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2820 mutex_lock(&memcg_create_mutex);
2822 if (memcg->use_hierarchy == val)
2826 * If parent's use_hierarchy is set, we can't make any modifications
2827 * in the child subtrees. If it is unset, then the change can
2828 * occur, provided the current cgroup has no children.
2830 * For the root cgroup, parent_mem is NULL, we allow value to be
2831 * set if there are no children.
2833 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2834 (val == 1 || val == 0)) {
2835 if (!memcg_has_children(memcg))
2836 memcg->use_hierarchy = val;
2843 mutex_unlock(&memcg_create_mutex);
2848 static unsigned long tree_stat(struct mem_cgroup *memcg,
2849 enum mem_cgroup_stat_index idx)
2851 struct mem_cgroup *iter;
2852 unsigned long val = 0;
2854 for_each_mem_cgroup_tree(iter, memcg)
2855 val += mem_cgroup_read_stat(iter, idx);
2860 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2864 if (mem_cgroup_is_root(memcg)) {
2865 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2866 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2868 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2871 val = page_counter_read(&memcg->memory);
2873 val = page_counter_read(&memcg->memsw);
2886 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2889 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2890 struct page_counter *counter;
2892 switch (MEMFILE_TYPE(cft->private)) {
2894 counter = &memcg->memory;
2897 counter = &memcg->memsw;
2900 counter = &memcg->kmem;
2906 switch (MEMFILE_ATTR(cft->private)) {
2908 if (counter == &memcg->memory)
2909 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2910 if (counter == &memcg->memsw)
2911 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2912 return (u64)page_counter_read(counter) * PAGE_SIZE;
2914 return (u64)counter->limit * PAGE_SIZE;
2916 return (u64)counter->watermark * PAGE_SIZE;
2918 return counter->failcnt;
2919 case RES_SOFT_LIMIT:
2920 return (u64)memcg->soft_limit * PAGE_SIZE;
2926 #ifdef CONFIG_MEMCG_KMEM
2927 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2928 unsigned long nr_pages)
2933 BUG_ON(memcg->kmemcg_id >= 0);
2934 BUG_ON(memcg->kmem_acct_activated);
2935 BUG_ON(memcg->kmem_acct_active);
2938 * For simplicity, we won't allow this to be disabled. It also can't
2939 * be changed if the cgroup has children already, or if tasks had
2942 * If tasks join before we set the limit, a person looking at
2943 * kmem.usage_in_bytes will have no way to determine when it took
2944 * place, which makes the value quite meaningless.
2946 * After it first became limited, changes in the value of the limit are
2947 * of course permitted.
2949 mutex_lock(&memcg_create_mutex);
2950 if (cgroup_is_populated(memcg->css.cgroup) ||
2951 (memcg->use_hierarchy && memcg_has_children(memcg)))
2953 mutex_unlock(&memcg_create_mutex);
2957 memcg_id = memcg_alloc_cache_id();
2964 * We couldn't have accounted to this cgroup, because it hasn't got
2965 * activated yet, so this should succeed.
2967 err = page_counter_limit(&memcg->kmem, nr_pages);
2970 static_key_slow_inc(&memcg_kmem_enabled_key);
2972 * A memory cgroup is considered kmem-active as soon as it gets
2973 * kmemcg_id. Setting the id after enabling static branching will
2974 * guarantee no one starts accounting before all call sites are
2977 memcg->kmemcg_id = memcg_id;
2978 memcg->kmem_acct_activated = true;
2979 memcg->kmem_acct_active = true;
2984 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2985 unsigned long limit)
2989 mutex_lock(&memcg_limit_mutex);
2990 if (!memcg_kmem_is_active(memcg))
2991 ret = memcg_activate_kmem(memcg, limit);
2993 ret = page_counter_limit(&memcg->kmem, limit);
2994 mutex_unlock(&memcg_limit_mutex);
2998 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3001 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3006 mutex_lock(&memcg_limit_mutex);
3008 * If the parent cgroup is not kmem-active now, it cannot be activated
3009 * after this point, because it has at least one child already.
3011 if (memcg_kmem_is_active(parent))
3012 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3013 mutex_unlock(&memcg_limit_mutex);
3017 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3018 unsigned long limit)
3022 #endif /* CONFIG_MEMCG_KMEM */
3025 * The user of this function is...
3028 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3029 char *buf, size_t nbytes, loff_t off)
3031 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3032 unsigned long nr_pages;
3035 buf = strstrip(buf);
3036 ret = page_counter_memparse(buf, "-1", &nr_pages);
3040 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3042 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3046 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3048 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3051 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3054 ret = memcg_update_kmem_limit(memcg, nr_pages);
3058 case RES_SOFT_LIMIT:
3059 memcg->soft_limit = nr_pages;
3063 return ret ?: nbytes;
3066 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3067 size_t nbytes, loff_t off)
3069 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3070 struct page_counter *counter;
3072 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3074 counter = &memcg->memory;
3077 counter = &memcg->memsw;
3080 counter = &memcg->kmem;
3086 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3088 page_counter_reset_watermark(counter);
3091 counter->failcnt = 0;
3100 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3103 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3107 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3108 struct cftype *cft, u64 val)
3110 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3112 if (val & ~MOVE_MASK)
3116 * No kind of locking is needed in here, because ->can_attach() will
3117 * check this value once in the beginning of the process, and then carry
3118 * on with stale data. This means that changes to this value will only
3119 * affect task migrations starting after the change.
3121 memcg->move_charge_at_immigrate = val;
3125 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3126 struct cftype *cft, u64 val)
3133 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3137 unsigned int lru_mask;
3140 static const struct numa_stat stats[] = {
3141 { "total", LRU_ALL },
3142 { "file", LRU_ALL_FILE },
3143 { "anon", LRU_ALL_ANON },
3144 { "unevictable", BIT(LRU_UNEVICTABLE) },
3146 const struct numa_stat *stat;
3149 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3151 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3152 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3153 seq_printf(m, "%s=%lu", stat->name, nr);
3154 for_each_node_state(nid, N_MEMORY) {
3155 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3157 seq_printf(m, " N%d=%lu", nid, nr);
3162 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3163 struct mem_cgroup *iter;
3166 for_each_mem_cgroup_tree(iter, memcg)
3167 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3168 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3169 for_each_node_state(nid, N_MEMORY) {
3171 for_each_mem_cgroup_tree(iter, memcg)
3172 nr += mem_cgroup_node_nr_lru_pages(
3173 iter, nid, stat->lru_mask);
3174 seq_printf(m, " N%d=%lu", nid, nr);
3181 #endif /* CONFIG_NUMA */
3183 static int memcg_stat_show(struct seq_file *m, void *v)
3185 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3186 unsigned long memory, memsw;
3187 struct mem_cgroup *mi;
3190 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3191 MEM_CGROUP_STAT_NSTATS);
3192 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3193 MEM_CGROUP_EVENTS_NSTATS);
3194 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3196 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3197 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3199 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3200 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3203 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3204 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3205 mem_cgroup_read_events(memcg, i));
3207 for (i = 0; i < NR_LRU_LISTS; i++)
3208 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3209 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3211 /* Hierarchical information */
3212 memory = memsw = PAGE_COUNTER_MAX;
3213 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3214 memory = min(memory, mi->memory.limit);
3215 memsw = min(memsw, mi->memsw.limit);
3217 seq_printf(m, "hierarchical_memory_limit %llu\n",
3218 (u64)memory * PAGE_SIZE);
3219 if (do_swap_account)
3220 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3221 (u64)memsw * PAGE_SIZE);
3223 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3224 unsigned long long val = 0;
3226 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3228 for_each_mem_cgroup_tree(mi, memcg)
3229 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3230 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3233 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3234 unsigned long long val = 0;
3236 for_each_mem_cgroup_tree(mi, memcg)
3237 val += mem_cgroup_read_events(mi, i);
3238 seq_printf(m, "total_%s %llu\n",
3239 mem_cgroup_events_names[i], val);
3242 for (i = 0; i < NR_LRU_LISTS; i++) {
3243 unsigned long long val = 0;
3245 for_each_mem_cgroup_tree(mi, memcg)
3246 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3247 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3250 #ifdef CONFIG_DEBUG_VM
3253 struct mem_cgroup_per_zone *mz;
3254 struct zone_reclaim_stat *rstat;
3255 unsigned long recent_rotated[2] = {0, 0};
3256 unsigned long recent_scanned[2] = {0, 0};
3258 for_each_online_node(nid)
3259 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3260 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3261 rstat = &mz->lruvec.reclaim_stat;
3263 recent_rotated[0] += rstat->recent_rotated[0];
3264 recent_rotated[1] += rstat->recent_rotated[1];
3265 recent_scanned[0] += rstat->recent_scanned[0];
3266 recent_scanned[1] += rstat->recent_scanned[1];
3268 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3269 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3270 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3271 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3278 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3281 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3283 return mem_cgroup_swappiness(memcg);
3286 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3287 struct cftype *cft, u64 val)
3289 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3295 memcg->swappiness = val;
3297 vm_swappiness = val;
3302 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3304 struct mem_cgroup_threshold_ary *t;
3305 unsigned long usage;
3310 t = rcu_dereference(memcg->thresholds.primary);
3312 t = rcu_dereference(memcg->memsw_thresholds.primary);
3317 usage = mem_cgroup_usage(memcg, swap);
3320 * current_threshold points to threshold just below or equal to usage.
3321 * If it's not true, a threshold was crossed after last
3322 * call of __mem_cgroup_threshold().
3324 i = t->current_threshold;
3327 * Iterate backward over array of thresholds starting from
3328 * current_threshold and check if a threshold is crossed.
3329 * If none of thresholds below usage is crossed, we read
3330 * only one element of the array here.
3332 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3333 eventfd_signal(t->entries[i].eventfd, 1);
3335 /* i = current_threshold + 1 */
3339 * Iterate forward over array of thresholds starting from
3340 * current_threshold+1 and check if a threshold is crossed.
3341 * If none of thresholds above usage is crossed, we read
3342 * only one element of the array here.
3344 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3345 eventfd_signal(t->entries[i].eventfd, 1);
3347 /* Update current_threshold */
3348 t->current_threshold = i - 1;
3353 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3356 __mem_cgroup_threshold(memcg, false);
3357 if (do_swap_account)
3358 __mem_cgroup_threshold(memcg, true);
3360 memcg = parent_mem_cgroup(memcg);
3364 static int compare_thresholds(const void *a, const void *b)
3366 const struct mem_cgroup_threshold *_a = a;
3367 const struct mem_cgroup_threshold *_b = b;
3369 if (_a->threshold > _b->threshold)
3372 if (_a->threshold < _b->threshold)
3378 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3380 struct mem_cgroup_eventfd_list *ev;
3382 spin_lock(&memcg_oom_lock);
3384 list_for_each_entry(ev, &memcg->oom_notify, list)
3385 eventfd_signal(ev->eventfd, 1);
3387 spin_unlock(&memcg_oom_lock);
3391 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3393 struct mem_cgroup *iter;
3395 for_each_mem_cgroup_tree(iter, memcg)
3396 mem_cgroup_oom_notify_cb(iter);
3399 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3400 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3402 struct mem_cgroup_thresholds *thresholds;
3403 struct mem_cgroup_threshold_ary *new;
3404 unsigned long threshold;
3405 unsigned long usage;
3408 ret = page_counter_memparse(args, "-1", &threshold);
3412 mutex_lock(&memcg->thresholds_lock);
3415 thresholds = &memcg->thresholds;
3416 usage = mem_cgroup_usage(memcg, false);
3417 } else if (type == _MEMSWAP) {
3418 thresholds = &memcg->memsw_thresholds;
3419 usage = mem_cgroup_usage(memcg, true);
3423 /* Check if a threshold crossed before adding a new one */
3424 if (thresholds->primary)
3425 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3427 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3429 /* Allocate memory for new array of thresholds */
3430 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3438 /* Copy thresholds (if any) to new array */
3439 if (thresholds->primary) {
3440 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3441 sizeof(struct mem_cgroup_threshold));
3444 /* Add new threshold */
3445 new->entries[size - 1].eventfd = eventfd;
3446 new->entries[size - 1].threshold = threshold;
3448 /* Sort thresholds. Registering of new threshold isn't time-critical */
3449 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3450 compare_thresholds, NULL);
3452 /* Find current threshold */
3453 new->current_threshold = -1;
3454 for (i = 0; i < size; i++) {
3455 if (new->entries[i].threshold <= usage) {
3457 * new->current_threshold will not be used until
3458 * rcu_assign_pointer(), so it's safe to increment
3461 ++new->current_threshold;
3466 /* Free old spare buffer and save old primary buffer as spare */
3467 kfree(thresholds->spare);
3468 thresholds->spare = thresholds->primary;
3470 rcu_assign_pointer(thresholds->primary, new);
3472 /* To be sure that nobody uses thresholds */
3476 mutex_unlock(&memcg->thresholds_lock);
3481 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3482 struct eventfd_ctx *eventfd, const char *args)
3484 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3487 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3488 struct eventfd_ctx *eventfd, const char *args)
3490 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3493 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3494 struct eventfd_ctx *eventfd, enum res_type type)
3496 struct mem_cgroup_thresholds *thresholds;
3497 struct mem_cgroup_threshold_ary *new;
3498 unsigned long usage;
3501 mutex_lock(&memcg->thresholds_lock);
3504 thresholds = &memcg->thresholds;
3505 usage = mem_cgroup_usage(memcg, false);
3506 } else if (type == _MEMSWAP) {
3507 thresholds = &memcg->memsw_thresholds;
3508 usage = mem_cgroup_usage(memcg, true);
3512 if (!thresholds->primary)
3515 /* Check if a threshold crossed before removing */
3516 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3518 /* Calculate new number of threshold */
3520 for (i = 0; i < thresholds->primary->size; i++) {
3521 if (thresholds->primary->entries[i].eventfd != eventfd)
3525 new = thresholds->spare;
3527 /* Set thresholds array to NULL if we don't have thresholds */
3536 /* Copy thresholds and find current threshold */
3537 new->current_threshold = -1;
3538 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3539 if (thresholds->primary->entries[i].eventfd == eventfd)
3542 new->entries[j] = thresholds->primary->entries[i];
3543 if (new->entries[j].threshold <= usage) {
3545 * new->current_threshold will not be used
3546 * until rcu_assign_pointer(), so it's safe to increment
3549 ++new->current_threshold;
3555 /* Swap primary and spare array */
3556 thresholds->spare = thresholds->primary;
3557 /* If all events are unregistered, free the spare array */
3559 kfree(thresholds->spare);
3560 thresholds->spare = NULL;
3563 rcu_assign_pointer(thresholds->primary, new);
3565 /* To be sure that nobody uses thresholds */
3568 mutex_unlock(&memcg->thresholds_lock);
3571 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3572 struct eventfd_ctx *eventfd)
3574 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3577 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3578 struct eventfd_ctx *eventfd)
3580 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3583 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3584 struct eventfd_ctx *eventfd, const char *args)
3586 struct mem_cgroup_eventfd_list *event;
3588 event = kmalloc(sizeof(*event), GFP_KERNEL);
3592 spin_lock(&memcg_oom_lock);
3594 event->eventfd = eventfd;
3595 list_add(&event->list, &memcg->oom_notify);
3597 /* already in OOM ? */
3598 if (memcg->under_oom)
3599 eventfd_signal(eventfd, 1);
3600 spin_unlock(&memcg_oom_lock);
3605 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3606 struct eventfd_ctx *eventfd)
3608 struct mem_cgroup_eventfd_list *ev, *tmp;
3610 spin_lock(&memcg_oom_lock);
3612 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3613 if (ev->eventfd == eventfd) {
3614 list_del(&ev->list);
3619 spin_unlock(&memcg_oom_lock);
3622 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3624 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3626 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3627 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3631 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3632 struct cftype *cft, u64 val)
3634 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3636 /* cannot set to root cgroup and only 0 and 1 are allowed */
3637 if (!css->parent || !((val == 0) || (val == 1)))
3640 memcg->oom_kill_disable = val;
3642 memcg_oom_recover(memcg);
3647 #ifdef CONFIG_MEMCG_KMEM
3648 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3652 ret = memcg_propagate_kmem(memcg);
3656 return mem_cgroup_sockets_init(memcg, ss);
3659 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3661 struct cgroup_subsys_state *css;
3662 struct mem_cgroup *parent, *child;
3665 if (!memcg->kmem_acct_active)
3669 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3670 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3671 * guarantees no cache will be created for this cgroup after we are
3672 * done (see memcg_create_kmem_cache()).
3674 memcg->kmem_acct_active = false;
3676 memcg_deactivate_kmem_caches(memcg);
3678 kmemcg_id = memcg->kmemcg_id;
3679 BUG_ON(kmemcg_id < 0);
3681 parent = parent_mem_cgroup(memcg);
3683 parent = root_mem_cgroup;
3686 * Change kmemcg_id of this cgroup and all its descendants to the
3687 * parent's id, and then move all entries from this cgroup's list_lrus
3688 * to ones of the parent. After we have finished, all list_lrus
3689 * corresponding to this cgroup are guaranteed to remain empty. The
3690 * ordering is imposed by list_lru_node->lock taken by
3691 * memcg_drain_all_list_lrus().
3693 css_for_each_descendant_pre(css, &memcg->css) {
3694 child = mem_cgroup_from_css(css);
3695 BUG_ON(child->kmemcg_id != kmemcg_id);
3696 child->kmemcg_id = parent->kmemcg_id;
3697 if (!memcg->use_hierarchy)
3700 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3702 memcg_free_cache_id(kmemcg_id);
3705 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3707 if (memcg->kmem_acct_activated) {
3708 memcg_destroy_kmem_caches(memcg);
3709 static_key_slow_dec(&memcg_kmem_enabled_key);
3710 WARN_ON(page_counter_read(&memcg->kmem));
3712 mem_cgroup_sockets_destroy(memcg);
3715 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3720 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3724 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3729 #ifdef CONFIG_CGROUP_WRITEBACK
3731 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3733 return &memcg->cgwb_list;
3736 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3738 return wb_domain_init(&memcg->cgwb_domain, gfp);
3741 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3743 wb_domain_exit(&memcg->cgwb_domain);
3746 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3748 wb_domain_size_changed(&memcg->cgwb_domain);
3751 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3753 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3755 if (!memcg->css.parent)
3758 return &memcg->cgwb_domain;
3762 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3763 * @wb: bdi_writeback in question
3764 * @pfilepages: out parameter for number of file pages
3765 * @pheadroom: out parameter for number of allocatable pages according to memcg
3766 * @pdirty: out parameter for number of dirty pages
3767 * @pwriteback: out parameter for number of pages under writeback
3769 * Determine the numbers of file, headroom, dirty, and writeback pages in
3770 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3771 * is a bit more involved.
3773 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3774 * headroom is calculated as the lowest headroom of itself and the
3775 * ancestors. Note that this doesn't consider the actual amount of
3776 * available memory in the system. The caller should further cap
3777 * *@pheadroom accordingly.
3779 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3780 unsigned long *pheadroom, unsigned long *pdirty,
3781 unsigned long *pwriteback)
3783 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3784 struct mem_cgroup *parent;
3786 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3788 /* this should eventually include NR_UNSTABLE_NFS */
3789 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3790 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3791 (1 << LRU_ACTIVE_FILE));
3792 *pheadroom = PAGE_COUNTER_MAX;
3794 while ((parent = parent_mem_cgroup(memcg))) {
3795 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3796 unsigned long used = page_counter_read(&memcg->memory);
3798 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3803 #else /* CONFIG_CGROUP_WRITEBACK */
3805 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3810 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3814 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3818 #endif /* CONFIG_CGROUP_WRITEBACK */
3821 * DO NOT USE IN NEW FILES.
3823 * "cgroup.event_control" implementation.
3825 * This is way over-engineered. It tries to support fully configurable
3826 * events for each user. Such level of flexibility is completely
3827 * unnecessary especially in the light of the planned unified hierarchy.
3829 * Please deprecate this and replace with something simpler if at all
3834 * Unregister event and free resources.
3836 * Gets called from workqueue.
3838 static void memcg_event_remove(struct work_struct *work)
3840 struct mem_cgroup_event *event =
3841 container_of(work, struct mem_cgroup_event, remove);
3842 struct mem_cgroup *memcg = event->memcg;
3844 remove_wait_queue(event->wqh, &event->wait);
3846 event->unregister_event(memcg, event->eventfd);
3848 /* Notify userspace the event is going away. */
3849 eventfd_signal(event->eventfd, 1);
3851 eventfd_ctx_put(event->eventfd);
3853 css_put(&memcg->css);
3857 * Gets called on POLLHUP on eventfd when user closes it.
3859 * Called with wqh->lock held and interrupts disabled.
3861 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3862 int sync, void *key)
3864 struct mem_cgroup_event *event =
3865 container_of(wait, struct mem_cgroup_event, wait);
3866 struct mem_cgroup *memcg = event->memcg;
3867 unsigned long flags = (unsigned long)key;
3869 if (flags & POLLHUP) {
3871 * If the event has been detached at cgroup removal, we
3872 * can simply return knowing the other side will cleanup
3875 * We can't race against event freeing since the other
3876 * side will require wqh->lock via remove_wait_queue(),
3879 spin_lock(&memcg->event_list_lock);
3880 if (!list_empty(&event->list)) {
3881 list_del_init(&event->list);
3883 * We are in atomic context, but cgroup_event_remove()
3884 * may sleep, so we have to call it in workqueue.
3886 schedule_work(&event->remove);
3888 spin_unlock(&memcg->event_list_lock);
3894 static void memcg_event_ptable_queue_proc(struct file *file,
3895 wait_queue_head_t *wqh, poll_table *pt)
3897 struct mem_cgroup_event *event =
3898 container_of(pt, struct mem_cgroup_event, pt);
3901 add_wait_queue(wqh, &event->wait);
3905 * DO NOT USE IN NEW FILES.
3907 * Parse input and register new cgroup event handler.
3909 * Input must be in format '<event_fd> <control_fd> <args>'.
3910 * Interpretation of args is defined by control file implementation.
3912 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3913 char *buf, size_t nbytes, loff_t off)
3915 struct cgroup_subsys_state *css = of_css(of);
3916 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3917 struct mem_cgroup_event *event;
3918 struct cgroup_subsys_state *cfile_css;
3919 unsigned int efd, cfd;
3926 buf = strstrip(buf);
3928 efd = simple_strtoul(buf, &endp, 10);
3933 cfd = simple_strtoul(buf, &endp, 10);
3934 if ((*endp != ' ') && (*endp != '\0'))
3938 event = kzalloc(sizeof(*event), GFP_KERNEL);
3942 event->memcg = memcg;
3943 INIT_LIST_HEAD(&event->list);
3944 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3945 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3946 INIT_WORK(&event->remove, memcg_event_remove);
3954 event->eventfd = eventfd_ctx_fileget(efile.file);
3955 if (IS_ERR(event->eventfd)) {
3956 ret = PTR_ERR(event->eventfd);
3963 goto out_put_eventfd;
3966 /* the process need read permission on control file */
3967 /* AV: shouldn't we check that it's been opened for read instead? */
3968 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3973 * Determine the event callbacks and set them in @event. This used
3974 * to be done via struct cftype but cgroup core no longer knows
3975 * about these events. The following is crude but the whole thing
3976 * is for compatibility anyway.
3978 * DO NOT ADD NEW FILES.
3980 name = cfile.file->f_path.dentry->d_name.name;
3982 if (!strcmp(name, "memory.usage_in_bytes")) {
3983 event->register_event = mem_cgroup_usage_register_event;
3984 event->unregister_event = mem_cgroup_usage_unregister_event;
3985 } else if (!strcmp(name, "memory.oom_control")) {
3986 event->register_event = mem_cgroup_oom_register_event;
3987 event->unregister_event = mem_cgroup_oom_unregister_event;
3988 } else if (!strcmp(name, "memory.pressure_level")) {
3989 event->register_event = vmpressure_register_event;
3990 event->unregister_event = vmpressure_unregister_event;
3991 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3992 event->register_event = memsw_cgroup_usage_register_event;
3993 event->unregister_event = memsw_cgroup_usage_unregister_event;
4000 * Verify @cfile should belong to @css. Also, remaining events are
4001 * automatically removed on cgroup destruction but the removal is
4002 * asynchronous, so take an extra ref on @css.
4004 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4005 &memory_cgrp_subsys);
4007 if (IS_ERR(cfile_css))
4009 if (cfile_css != css) {
4014 ret = event->register_event(memcg, event->eventfd, buf);
4018 efile.file->f_op->poll(efile.file, &event->pt);
4020 spin_lock(&memcg->event_list_lock);
4021 list_add(&event->list, &memcg->event_list);
4022 spin_unlock(&memcg->event_list_lock);
4034 eventfd_ctx_put(event->eventfd);
4043 static struct cftype mem_cgroup_legacy_files[] = {
4045 .name = "usage_in_bytes",
4046 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4047 .read_u64 = mem_cgroup_read_u64,
4050 .name = "max_usage_in_bytes",
4051 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4052 .write = mem_cgroup_reset,
4053 .read_u64 = mem_cgroup_read_u64,
4056 .name = "limit_in_bytes",
4057 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4058 .write = mem_cgroup_write,
4059 .read_u64 = mem_cgroup_read_u64,
4062 .name = "soft_limit_in_bytes",
4063 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4064 .write = mem_cgroup_write,
4065 .read_u64 = mem_cgroup_read_u64,
4069 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4070 .write = mem_cgroup_reset,
4071 .read_u64 = mem_cgroup_read_u64,
4075 .seq_show = memcg_stat_show,
4078 .name = "force_empty",
4079 .write = mem_cgroup_force_empty_write,
4082 .name = "use_hierarchy",
4083 .write_u64 = mem_cgroup_hierarchy_write,
4084 .read_u64 = mem_cgroup_hierarchy_read,
4087 .name = "cgroup.event_control", /* XXX: for compat */
4088 .write = memcg_write_event_control,
4089 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4092 .name = "swappiness",
4093 .read_u64 = mem_cgroup_swappiness_read,
4094 .write_u64 = mem_cgroup_swappiness_write,
4097 .name = "move_charge_at_immigrate",
4098 .read_u64 = mem_cgroup_move_charge_read,
4099 .write_u64 = mem_cgroup_move_charge_write,
4102 .name = "oom_control",
4103 .seq_show = mem_cgroup_oom_control_read,
4104 .write_u64 = mem_cgroup_oom_control_write,
4105 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4108 .name = "pressure_level",
4112 .name = "numa_stat",
4113 .seq_show = memcg_numa_stat_show,
4116 #ifdef CONFIG_MEMCG_KMEM
4118 .name = "kmem.limit_in_bytes",
4119 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4120 .write = mem_cgroup_write,
4121 .read_u64 = mem_cgroup_read_u64,
4124 .name = "kmem.usage_in_bytes",
4125 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4126 .read_u64 = mem_cgroup_read_u64,
4129 .name = "kmem.failcnt",
4130 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4131 .write = mem_cgroup_reset,
4132 .read_u64 = mem_cgroup_read_u64,
4135 .name = "kmem.max_usage_in_bytes",
4136 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4137 .write = mem_cgroup_reset,
4138 .read_u64 = mem_cgroup_read_u64,
4140 #ifdef CONFIG_SLABINFO
4142 .name = "kmem.slabinfo",
4143 .seq_start = slab_start,
4144 .seq_next = slab_next,
4145 .seq_stop = slab_stop,
4146 .seq_show = memcg_slab_show,
4150 { }, /* terminate */
4153 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4155 struct mem_cgroup_per_node *pn;
4156 struct mem_cgroup_per_zone *mz;
4157 int zone, tmp = node;
4159 * This routine is called against possible nodes.
4160 * But it's BUG to call kmalloc() against offline node.
4162 * TODO: this routine can waste much memory for nodes which will
4163 * never be onlined. It's better to use memory hotplug callback
4166 if (!node_state(node, N_NORMAL_MEMORY))
4168 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4172 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4173 mz = &pn->zoneinfo[zone];
4174 lruvec_init(&mz->lruvec);
4175 mz->usage_in_excess = 0;
4176 mz->on_tree = false;
4179 memcg->nodeinfo[node] = pn;
4183 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4185 kfree(memcg->nodeinfo[node]);
4188 static struct mem_cgroup *mem_cgroup_alloc(void)
4190 struct mem_cgroup *memcg;
4193 size = sizeof(struct mem_cgroup);
4194 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4196 memcg = kzalloc(size, GFP_KERNEL);
4200 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4204 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4210 free_percpu(memcg->stat);
4217 * At destroying mem_cgroup, references from swap_cgroup can remain.
4218 * (scanning all at force_empty is too costly...)
4220 * Instead of clearing all references at force_empty, we remember
4221 * the number of reference from swap_cgroup and free mem_cgroup when
4222 * it goes down to 0.
4224 * Removal of cgroup itself succeeds regardless of refs from swap.
4227 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4231 mem_cgroup_remove_from_trees(memcg);
4234 free_mem_cgroup_per_zone_info(memcg, node);
4236 free_percpu(memcg->stat);
4237 memcg_wb_domain_exit(memcg);
4242 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4244 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4246 if (!memcg->memory.parent)
4248 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4250 EXPORT_SYMBOL(parent_mem_cgroup);
4252 static struct cgroup_subsys_state * __ref
4253 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4255 struct mem_cgroup *memcg;
4256 long error = -ENOMEM;
4259 memcg = mem_cgroup_alloc();
4261 return ERR_PTR(error);
4264 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4268 if (parent_css == NULL) {
4269 root_mem_cgroup = memcg;
4270 page_counter_init(&memcg->memory, NULL);
4271 memcg->high = PAGE_COUNTER_MAX;
4272 memcg->soft_limit = PAGE_COUNTER_MAX;
4273 page_counter_init(&memcg->memsw, NULL);
4274 page_counter_init(&memcg->kmem, NULL);
4277 memcg->last_scanned_node = MAX_NUMNODES;
4278 INIT_LIST_HEAD(&memcg->oom_notify);
4279 memcg->move_charge_at_immigrate = 0;
4280 mutex_init(&memcg->thresholds_lock);
4281 spin_lock_init(&memcg->move_lock);
4282 vmpressure_init(&memcg->vmpressure);
4283 INIT_LIST_HEAD(&memcg->event_list);
4284 spin_lock_init(&memcg->event_list_lock);
4285 #ifdef CONFIG_MEMCG_KMEM
4286 memcg->kmemcg_id = -1;
4288 #ifdef CONFIG_CGROUP_WRITEBACK
4289 INIT_LIST_HEAD(&memcg->cgwb_list);
4294 __mem_cgroup_free(memcg);
4295 return ERR_PTR(error);
4299 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4301 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4302 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4305 if (css->id > MEM_CGROUP_ID_MAX)
4311 mutex_lock(&memcg_create_mutex);
4313 memcg->use_hierarchy = parent->use_hierarchy;
4314 memcg->oom_kill_disable = parent->oom_kill_disable;
4315 memcg->swappiness = mem_cgroup_swappiness(parent);
4317 if (parent->use_hierarchy) {
4318 page_counter_init(&memcg->memory, &parent->memory);
4319 memcg->high = PAGE_COUNTER_MAX;
4320 memcg->soft_limit = PAGE_COUNTER_MAX;
4321 page_counter_init(&memcg->memsw, &parent->memsw);
4322 page_counter_init(&memcg->kmem, &parent->kmem);
4325 * No need to take a reference to the parent because cgroup
4326 * core guarantees its existence.
4329 page_counter_init(&memcg->memory, NULL);
4330 memcg->high = PAGE_COUNTER_MAX;
4331 memcg->soft_limit = PAGE_COUNTER_MAX;
4332 page_counter_init(&memcg->memsw, NULL);
4333 page_counter_init(&memcg->kmem, NULL);
4335 * Deeper hierachy with use_hierarchy == false doesn't make
4336 * much sense so let cgroup subsystem know about this
4337 * unfortunate state in our controller.
4339 if (parent != root_mem_cgroup)
4340 memory_cgrp_subsys.broken_hierarchy = true;
4342 mutex_unlock(&memcg_create_mutex);
4344 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4349 * Make sure the memcg is initialized: mem_cgroup_iter()
4350 * orders reading memcg->initialized against its callers
4351 * reading the memcg members.
4353 smp_store_release(&memcg->initialized, 1);
4358 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4360 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4361 struct mem_cgroup_event *event, *tmp;
4364 * Unregister events and notify userspace.
4365 * Notify userspace about cgroup removing only after rmdir of cgroup
4366 * directory to avoid race between userspace and kernelspace.
4368 spin_lock(&memcg->event_list_lock);
4369 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4370 list_del_init(&event->list);
4371 schedule_work(&event->remove);
4373 spin_unlock(&memcg->event_list_lock);
4375 vmpressure_cleanup(&memcg->vmpressure);
4377 memcg_deactivate_kmem(memcg);
4379 wb_memcg_offline(memcg);
4382 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4384 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4386 invalidate_reclaim_iterators(memcg);
4389 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4391 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4393 memcg_destroy_kmem(memcg);
4394 __mem_cgroup_free(memcg);
4398 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4399 * @css: the target css
4401 * Reset the states of the mem_cgroup associated with @css. This is
4402 * invoked when the userland requests disabling on the default hierarchy
4403 * but the memcg is pinned through dependency. The memcg should stop
4404 * applying policies and should revert to the vanilla state as it may be
4405 * made visible again.
4407 * The current implementation only resets the essential configurations.
4408 * This needs to be expanded to cover all the visible parts.
4410 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4412 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4414 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4415 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4416 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4418 memcg->high = PAGE_COUNTER_MAX;
4419 memcg->soft_limit = PAGE_COUNTER_MAX;
4420 memcg_wb_domain_size_changed(memcg);
4424 /* Handlers for move charge at task migration. */
4425 static int mem_cgroup_do_precharge(unsigned long count)
4429 /* Try a single bulk charge without reclaim first, kswapd may wake */
4430 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4432 mc.precharge += count;
4436 /* Try charges one by one with reclaim */
4438 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4448 * get_mctgt_type - get target type of moving charge
4449 * @vma: the vma the pte to be checked belongs
4450 * @addr: the address corresponding to the pte to be checked
4451 * @ptent: the pte to be checked
4452 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4455 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4456 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4457 * move charge. if @target is not NULL, the page is stored in target->page
4458 * with extra refcnt got(Callers should handle it).
4459 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4460 * target for charge migration. if @target is not NULL, the entry is stored
4463 * Called with pte lock held.
4470 enum mc_target_type {
4476 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4477 unsigned long addr, pte_t ptent)
4479 struct page *page = vm_normal_page(vma, addr, ptent);
4481 if (!page || !page_mapped(page))
4483 if (PageAnon(page)) {
4484 if (!(mc.flags & MOVE_ANON))
4487 if (!(mc.flags & MOVE_FILE))
4490 if (!get_page_unless_zero(page))
4497 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4498 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4500 struct page *page = NULL;
4501 swp_entry_t ent = pte_to_swp_entry(ptent);
4503 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4506 * Because lookup_swap_cache() updates some statistics counter,
4507 * we call find_get_page() with swapper_space directly.
4509 page = find_get_page(swap_address_space(ent), ent.val);
4510 if (do_swap_account)
4511 entry->val = ent.val;
4516 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4517 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4523 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4524 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4526 struct page *page = NULL;
4527 struct address_space *mapping;
4530 if (!vma->vm_file) /* anonymous vma */
4532 if (!(mc.flags & MOVE_FILE))
4535 mapping = vma->vm_file->f_mapping;
4536 pgoff = linear_page_index(vma, addr);
4538 /* page is moved even if it's not RSS of this task(page-faulted). */
4540 /* shmem/tmpfs may report page out on swap: account for that too. */
4541 if (shmem_mapping(mapping)) {
4542 page = find_get_entry(mapping, pgoff);
4543 if (radix_tree_exceptional_entry(page)) {
4544 swp_entry_t swp = radix_to_swp_entry(page);
4545 if (do_swap_account)
4547 page = find_get_page(swap_address_space(swp), swp.val);
4550 page = find_get_page(mapping, pgoff);
4552 page = find_get_page(mapping, pgoff);
4558 * mem_cgroup_move_account - move account of the page
4560 * @nr_pages: number of regular pages (>1 for huge pages)
4561 * @from: mem_cgroup which the page is moved from.
4562 * @to: mem_cgroup which the page is moved to. @from != @to.
4564 * The caller must confirm following.
4565 * - page is not on LRU (isolate_page() is useful.)
4566 * - compound_lock is held when nr_pages > 1
4568 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4571 static int mem_cgroup_move_account(struct page *page,
4572 unsigned int nr_pages,
4573 struct mem_cgroup *from,
4574 struct mem_cgroup *to)
4576 unsigned long flags;
4580 VM_BUG_ON(from == to);
4581 VM_BUG_ON_PAGE(PageLRU(page), page);
4583 * The page is isolated from LRU. So, collapse function
4584 * will not handle this page. But page splitting can happen.
4585 * Do this check under compound_page_lock(). The caller should
4589 if (nr_pages > 1 && !PageTransHuge(page))
4593 * Prevent mem_cgroup_replace_page() from looking at
4594 * page->mem_cgroup of its source page while we change it.
4596 if (!trylock_page(page))
4600 if (page->mem_cgroup != from)
4603 anon = PageAnon(page);
4605 spin_lock_irqsave(&from->move_lock, flags);
4607 if (!anon && page_mapped(page)) {
4608 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4610 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4615 * move_lock grabbed above and caller set from->moving_account, so
4616 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4617 * So mapping should be stable for dirty pages.
4619 if (!anon && PageDirty(page)) {
4620 struct address_space *mapping = page_mapping(page);
4622 if (mapping_cap_account_dirty(mapping)) {
4623 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4625 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4630 if (PageWriteback(page)) {
4631 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4633 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4638 * It is safe to change page->mem_cgroup here because the page
4639 * is referenced, charged, and isolated - we can't race with
4640 * uncharging, charging, migration, or LRU putback.
4643 /* caller should have done css_get */
4644 page->mem_cgroup = to;
4645 spin_unlock_irqrestore(&from->move_lock, flags);
4649 local_irq_disable();
4650 mem_cgroup_charge_statistics(to, page, nr_pages);
4651 memcg_check_events(to, page);
4652 mem_cgroup_charge_statistics(from, page, -nr_pages);
4653 memcg_check_events(from, page);
4661 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4662 unsigned long addr, pte_t ptent, union mc_target *target)
4664 struct page *page = NULL;
4665 enum mc_target_type ret = MC_TARGET_NONE;
4666 swp_entry_t ent = { .val = 0 };
4668 if (pte_present(ptent))
4669 page = mc_handle_present_pte(vma, addr, ptent);
4670 else if (is_swap_pte(ptent))
4671 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4672 else if (pte_none(ptent))
4673 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4675 if (!page && !ent.val)
4679 * Do only loose check w/o serialization.
4680 * mem_cgroup_move_account() checks the page is valid or
4681 * not under LRU exclusion.
4683 if (page->mem_cgroup == mc.from) {
4684 ret = MC_TARGET_PAGE;
4686 target->page = page;
4688 if (!ret || !target)
4691 /* There is a swap entry and a page doesn't exist or isn't charged */
4692 if (ent.val && !ret &&
4693 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4694 ret = MC_TARGET_SWAP;
4701 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4703 * We don't consider swapping or file mapped pages because THP does not
4704 * support them for now.
4705 * Caller should make sure that pmd_trans_huge(pmd) is true.
4707 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4708 unsigned long addr, pmd_t pmd, union mc_target *target)
4710 struct page *page = NULL;
4711 enum mc_target_type ret = MC_TARGET_NONE;
4713 page = pmd_page(pmd);
4714 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4715 if (!(mc.flags & MOVE_ANON))
4717 if (page->mem_cgroup == mc.from) {
4718 ret = MC_TARGET_PAGE;
4721 target->page = page;
4727 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4728 unsigned long addr, pmd_t pmd, union mc_target *target)
4730 return MC_TARGET_NONE;
4734 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4735 unsigned long addr, unsigned long end,
4736 struct mm_walk *walk)
4738 struct vm_area_struct *vma = walk->vma;
4742 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4743 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4744 mc.precharge += HPAGE_PMD_NR;
4749 if (pmd_trans_unstable(pmd))
4751 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4752 for (; addr != end; pte++, addr += PAGE_SIZE)
4753 if (get_mctgt_type(vma, addr, *pte, NULL))
4754 mc.precharge++; /* increment precharge temporarily */
4755 pte_unmap_unlock(pte - 1, ptl);
4761 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4763 unsigned long precharge;
4765 struct mm_walk mem_cgroup_count_precharge_walk = {
4766 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4769 down_read(&mm->mmap_sem);
4770 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4771 up_read(&mm->mmap_sem);
4773 precharge = mc.precharge;
4779 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4781 unsigned long precharge = mem_cgroup_count_precharge(mm);
4783 VM_BUG_ON(mc.moving_task);
4784 mc.moving_task = current;
4785 return mem_cgroup_do_precharge(precharge);
4788 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4789 static void __mem_cgroup_clear_mc(void)
4791 struct mem_cgroup *from = mc.from;
4792 struct mem_cgroup *to = mc.to;
4794 /* we must uncharge all the leftover precharges from mc.to */
4796 cancel_charge(mc.to, mc.precharge);
4800 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4801 * we must uncharge here.
4803 if (mc.moved_charge) {
4804 cancel_charge(mc.from, mc.moved_charge);
4805 mc.moved_charge = 0;
4807 /* we must fixup refcnts and charges */
4808 if (mc.moved_swap) {
4809 /* uncharge swap account from the old cgroup */
4810 if (!mem_cgroup_is_root(mc.from))
4811 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4814 * we charged both to->memory and to->memsw, so we
4815 * should uncharge to->memory.
4817 if (!mem_cgroup_is_root(mc.to))
4818 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4820 css_put_many(&mc.from->css, mc.moved_swap);
4822 /* we've already done css_get(mc.to) */
4825 memcg_oom_recover(from);
4826 memcg_oom_recover(to);
4827 wake_up_all(&mc.waitq);
4830 static void mem_cgroup_clear_mc(void)
4833 * we must clear moving_task before waking up waiters at the end of
4836 mc.moving_task = NULL;
4837 __mem_cgroup_clear_mc();
4838 spin_lock(&mc.lock);
4841 spin_unlock(&mc.lock);
4844 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4846 struct cgroup_subsys_state *css;
4847 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4848 struct mem_cgroup *from;
4849 struct task_struct *leader, *p;
4850 struct mm_struct *mm;
4851 unsigned long move_flags;
4854 /* charge immigration isn't supported on the default hierarchy */
4855 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4859 * Multi-process migrations only happen on the default hierarchy
4860 * where charge immigration is not used. Perform charge
4861 * immigration if @tset contains a leader and whine if there are
4865 cgroup_taskset_for_each_leader(leader, css, tset) {
4868 memcg = mem_cgroup_from_css(css);
4874 * We are now commited to this value whatever it is. Changes in this
4875 * tunable will only affect upcoming migrations, not the current one.
4876 * So we need to save it, and keep it going.
4878 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4882 from = mem_cgroup_from_task(p);
4884 VM_BUG_ON(from == memcg);
4886 mm = get_task_mm(p);
4889 /* We move charges only when we move a owner of the mm */
4890 if (mm->owner == p) {
4893 VM_BUG_ON(mc.precharge);
4894 VM_BUG_ON(mc.moved_charge);
4895 VM_BUG_ON(mc.moved_swap);
4897 spin_lock(&mc.lock);
4900 mc.flags = move_flags;
4901 spin_unlock(&mc.lock);
4902 /* We set mc.moving_task later */
4904 ret = mem_cgroup_precharge_mc(mm);
4906 mem_cgroup_clear_mc();
4912 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4915 mem_cgroup_clear_mc();
4918 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4919 unsigned long addr, unsigned long end,
4920 struct mm_walk *walk)
4923 struct vm_area_struct *vma = walk->vma;
4926 enum mc_target_type target_type;
4927 union mc_target target;
4931 * We don't take compound_lock() here but no race with splitting thp
4933 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4934 * under splitting, which means there's no concurrent thp split,
4935 * - if another thread runs into split_huge_page() just after we
4936 * entered this if-block, the thread must wait for page table lock
4937 * to be unlocked in __split_huge_page_splitting(), where the main
4938 * part of thp split is not executed yet.
4940 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4941 if (mc.precharge < HPAGE_PMD_NR) {
4945 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4946 if (target_type == MC_TARGET_PAGE) {
4948 if (!isolate_lru_page(page)) {
4949 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
4951 mc.precharge -= HPAGE_PMD_NR;
4952 mc.moved_charge += HPAGE_PMD_NR;
4954 putback_lru_page(page);
4962 if (pmd_trans_unstable(pmd))
4965 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4966 for (; addr != end; addr += PAGE_SIZE) {
4967 pte_t ptent = *(pte++);
4973 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4974 case MC_TARGET_PAGE:
4976 if (isolate_lru_page(page))
4978 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
4980 /* we uncharge from mc.from later. */
4983 putback_lru_page(page);
4984 put: /* get_mctgt_type() gets the page */
4987 case MC_TARGET_SWAP:
4989 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4991 /* we fixup refcnts and charges later. */
4999 pte_unmap_unlock(pte - 1, ptl);
5004 * We have consumed all precharges we got in can_attach().
5005 * We try charge one by one, but don't do any additional
5006 * charges to mc.to if we have failed in charge once in attach()
5009 ret = mem_cgroup_do_precharge(1);
5017 static void mem_cgroup_move_charge(struct mm_struct *mm)
5019 struct mm_walk mem_cgroup_move_charge_walk = {
5020 .pmd_entry = mem_cgroup_move_charge_pte_range,
5024 lru_add_drain_all();
5026 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5027 * move_lock while we're moving its pages to another memcg.
5028 * Then wait for already started RCU-only updates to finish.
5030 atomic_inc(&mc.from->moving_account);
5033 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5035 * Someone who are holding the mmap_sem might be waiting in
5036 * waitq. So we cancel all extra charges, wake up all waiters,
5037 * and retry. Because we cancel precharges, we might not be able
5038 * to move enough charges, but moving charge is a best-effort
5039 * feature anyway, so it wouldn't be a big problem.
5041 __mem_cgroup_clear_mc();
5046 * When we have consumed all precharges and failed in doing
5047 * additional charge, the page walk just aborts.
5049 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5050 up_read(&mm->mmap_sem);
5051 atomic_dec(&mc.from->moving_account);
5054 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5056 struct cgroup_subsys_state *css;
5057 struct task_struct *p = cgroup_taskset_first(tset, &css);
5058 struct mm_struct *mm = get_task_mm(p);
5062 mem_cgroup_move_charge(mm);
5066 mem_cgroup_clear_mc();
5068 #else /* !CONFIG_MMU */
5069 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5073 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5076 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5082 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5083 * to verify whether we're attached to the default hierarchy on each mount
5086 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5089 * use_hierarchy is forced on the default hierarchy. cgroup core
5090 * guarantees that @root doesn't have any children, so turning it
5091 * on for the root memcg is enough.
5093 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5094 root_mem_cgroup->use_hierarchy = true;
5096 root_mem_cgroup->use_hierarchy = false;
5099 static u64 memory_current_read(struct cgroup_subsys_state *css,
5102 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5104 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5107 static int memory_low_show(struct seq_file *m, void *v)
5109 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5110 unsigned long low = READ_ONCE(memcg->low);
5112 if (low == PAGE_COUNTER_MAX)
5113 seq_puts(m, "max\n");
5115 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5120 static ssize_t memory_low_write(struct kernfs_open_file *of,
5121 char *buf, size_t nbytes, loff_t off)
5123 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5127 buf = strstrip(buf);
5128 err = page_counter_memparse(buf, "max", &low);
5137 static int memory_high_show(struct seq_file *m, void *v)
5139 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5140 unsigned long high = READ_ONCE(memcg->high);
5142 if (high == PAGE_COUNTER_MAX)
5143 seq_puts(m, "max\n");
5145 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5150 static ssize_t memory_high_write(struct kernfs_open_file *of,
5151 char *buf, size_t nbytes, loff_t off)
5153 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5157 buf = strstrip(buf);
5158 err = page_counter_memparse(buf, "max", &high);
5164 memcg_wb_domain_size_changed(memcg);
5168 static int memory_max_show(struct seq_file *m, void *v)
5170 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5171 unsigned long max = READ_ONCE(memcg->memory.limit);
5173 if (max == PAGE_COUNTER_MAX)
5174 seq_puts(m, "max\n");
5176 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5181 static ssize_t memory_max_write(struct kernfs_open_file *of,
5182 char *buf, size_t nbytes, loff_t off)
5184 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5188 buf = strstrip(buf);
5189 err = page_counter_memparse(buf, "max", &max);
5193 err = mem_cgroup_resize_limit(memcg, max);
5197 memcg_wb_domain_size_changed(memcg);
5201 static int memory_events_show(struct seq_file *m, void *v)
5203 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5205 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5206 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5207 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5208 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5213 static struct cftype memory_files[] = {
5216 .flags = CFTYPE_NOT_ON_ROOT,
5217 .read_u64 = memory_current_read,
5221 .flags = CFTYPE_NOT_ON_ROOT,
5222 .seq_show = memory_low_show,
5223 .write = memory_low_write,
5227 .flags = CFTYPE_NOT_ON_ROOT,
5228 .seq_show = memory_high_show,
5229 .write = memory_high_write,
5233 .flags = CFTYPE_NOT_ON_ROOT,
5234 .seq_show = memory_max_show,
5235 .write = memory_max_write,
5239 .flags = CFTYPE_NOT_ON_ROOT,
5240 .file_offset = offsetof(struct mem_cgroup, events_file),
5241 .seq_show = memory_events_show,
5246 struct cgroup_subsys memory_cgrp_subsys = {
5247 .css_alloc = mem_cgroup_css_alloc,
5248 .css_online = mem_cgroup_css_online,
5249 .css_offline = mem_cgroup_css_offline,
5250 .css_released = mem_cgroup_css_released,
5251 .css_free = mem_cgroup_css_free,
5252 .css_reset = mem_cgroup_css_reset,
5253 .can_attach = mem_cgroup_can_attach,
5254 .cancel_attach = mem_cgroup_cancel_attach,
5255 .attach = mem_cgroup_move_task,
5256 .bind = mem_cgroup_bind,
5257 .dfl_cftypes = memory_files,
5258 .legacy_cftypes = mem_cgroup_legacy_files,
5263 * mem_cgroup_low - check if memory consumption is below the normal range
5264 * @root: the highest ancestor to consider
5265 * @memcg: the memory cgroup to check
5267 * Returns %true if memory consumption of @memcg, and that of all
5268 * configurable ancestors up to @root, is below the normal range.
5270 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5272 if (mem_cgroup_disabled())
5276 * The toplevel group doesn't have a configurable range, so
5277 * it's never low when looked at directly, and it is not
5278 * considered an ancestor when assessing the hierarchy.
5281 if (memcg == root_mem_cgroup)
5284 if (page_counter_read(&memcg->memory) >= memcg->low)
5287 while (memcg != root) {
5288 memcg = parent_mem_cgroup(memcg);
5290 if (memcg == root_mem_cgroup)
5293 if (page_counter_read(&memcg->memory) >= memcg->low)
5300 * mem_cgroup_try_charge - try charging a page
5301 * @page: page to charge
5302 * @mm: mm context of the victim
5303 * @gfp_mask: reclaim mode
5304 * @memcgp: charged memcg return
5306 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5307 * pages according to @gfp_mask if necessary.
5309 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5310 * Otherwise, an error code is returned.
5312 * After page->mapping has been set up, the caller must finalize the
5313 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5314 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5316 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5317 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5319 struct mem_cgroup *memcg = NULL;
5320 unsigned int nr_pages = 1;
5323 if (mem_cgroup_disabled())
5326 if (PageSwapCache(page)) {
5328 * Every swap fault against a single page tries to charge the
5329 * page, bail as early as possible. shmem_unuse() encounters
5330 * already charged pages, too. The USED bit is protected by
5331 * the page lock, which serializes swap cache removal, which
5332 * in turn serializes uncharging.
5334 VM_BUG_ON_PAGE(!PageLocked(page), page);
5335 if (page->mem_cgroup)
5338 if (do_swap_account) {
5339 swp_entry_t ent = { .val = page_private(page), };
5340 unsigned short id = lookup_swap_cgroup_id(ent);
5343 memcg = mem_cgroup_from_id(id);
5344 if (memcg && !css_tryget_online(&memcg->css))
5350 if (PageTransHuge(page)) {
5351 nr_pages <<= compound_order(page);
5352 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5356 memcg = get_mem_cgroup_from_mm(mm);
5358 ret = try_charge(memcg, gfp_mask, nr_pages);
5360 css_put(&memcg->css);
5367 * mem_cgroup_commit_charge - commit a page charge
5368 * @page: page to charge
5369 * @memcg: memcg to charge the page to
5370 * @lrucare: page might be on LRU already
5372 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5373 * after page->mapping has been set up. This must happen atomically
5374 * as part of the page instantiation, i.e. under the page table lock
5375 * for anonymous pages, under the page lock for page and swap cache.
5377 * In addition, the page must not be on the LRU during the commit, to
5378 * prevent racing with task migration. If it might be, use @lrucare.
5380 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5382 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5385 unsigned int nr_pages = 1;
5387 VM_BUG_ON_PAGE(!page->mapping, page);
5388 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5390 if (mem_cgroup_disabled())
5393 * Swap faults will attempt to charge the same page multiple
5394 * times. But reuse_swap_page() might have removed the page
5395 * from swapcache already, so we can't check PageSwapCache().
5400 commit_charge(page, memcg, lrucare);
5402 if (PageTransHuge(page)) {
5403 nr_pages <<= compound_order(page);
5404 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5407 local_irq_disable();
5408 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5409 memcg_check_events(memcg, page);
5412 if (do_swap_account && PageSwapCache(page)) {
5413 swp_entry_t entry = { .val = page_private(page) };
5415 * The swap entry might not get freed for a long time,
5416 * let's not wait for it. The page already received a
5417 * memory+swap charge, drop the swap entry duplicate.
5419 mem_cgroup_uncharge_swap(entry);
5424 * mem_cgroup_cancel_charge - cancel a page charge
5425 * @page: page to charge
5426 * @memcg: memcg to charge the page to
5428 * Cancel a charge transaction started by mem_cgroup_try_charge().
5430 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5432 unsigned int nr_pages = 1;
5434 if (mem_cgroup_disabled())
5437 * Swap faults will attempt to charge the same page multiple
5438 * times. But reuse_swap_page() might have removed the page
5439 * from swapcache already, so we can't check PageSwapCache().
5444 if (PageTransHuge(page)) {
5445 nr_pages <<= compound_order(page);
5446 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5449 cancel_charge(memcg, nr_pages);
5452 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5453 unsigned long nr_anon, unsigned long nr_file,
5454 unsigned long nr_huge, struct page *dummy_page)
5456 unsigned long nr_pages = nr_anon + nr_file;
5457 unsigned long flags;
5459 if (!mem_cgroup_is_root(memcg)) {
5460 page_counter_uncharge(&memcg->memory, nr_pages);
5461 if (do_swap_account)
5462 page_counter_uncharge(&memcg->memsw, nr_pages);
5463 memcg_oom_recover(memcg);
5466 local_irq_save(flags);
5467 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5468 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5469 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5470 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5471 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5472 memcg_check_events(memcg, dummy_page);
5473 local_irq_restore(flags);
5475 if (!mem_cgroup_is_root(memcg))
5476 css_put_many(&memcg->css, nr_pages);
5479 static void uncharge_list(struct list_head *page_list)
5481 struct mem_cgroup *memcg = NULL;
5482 unsigned long nr_anon = 0;
5483 unsigned long nr_file = 0;
5484 unsigned long nr_huge = 0;
5485 unsigned long pgpgout = 0;
5486 struct list_head *next;
5489 next = page_list->next;
5491 unsigned int nr_pages = 1;
5493 page = list_entry(next, struct page, lru);
5494 next = page->lru.next;
5496 VM_BUG_ON_PAGE(PageLRU(page), page);
5497 VM_BUG_ON_PAGE(page_count(page), page);
5499 if (!page->mem_cgroup)
5503 * Nobody should be changing or seriously looking at
5504 * page->mem_cgroup at this point, we have fully
5505 * exclusive access to the page.
5508 if (memcg != page->mem_cgroup) {
5510 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5512 pgpgout = nr_anon = nr_file = nr_huge = 0;
5514 memcg = page->mem_cgroup;
5517 if (PageTransHuge(page)) {
5518 nr_pages <<= compound_order(page);
5519 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5520 nr_huge += nr_pages;
5524 nr_anon += nr_pages;
5526 nr_file += nr_pages;
5528 page->mem_cgroup = NULL;
5531 } while (next != page_list);
5534 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5539 * mem_cgroup_uncharge - uncharge a page
5540 * @page: page to uncharge
5542 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5543 * mem_cgroup_commit_charge().
5545 void mem_cgroup_uncharge(struct page *page)
5547 if (mem_cgroup_disabled())
5550 /* Don't touch page->lru of any random page, pre-check: */
5551 if (!page->mem_cgroup)
5554 INIT_LIST_HEAD(&page->lru);
5555 uncharge_list(&page->lru);
5559 * mem_cgroup_uncharge_list - uncharge a list of page
5560 * @page_list: list of pages to uncharge
5562 * Uncharge a list of pages previously charged with
5563 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5565 void mem_cgroup_uncharge_list(struct list_head *page_list)
5567 if (mem_cgroup_disabled())
5570 if (!list_empty(page_list))
5571 uncharge_list(page_list);
5575 * mem_cgroup_replace_page - migrate a charge to another page
5576 * @oldpage: currently charged page
5577 * @newpage: page to transfer the charge to
5579 * Migrate the charge from @oldpage to @newpage.
5581 * Both pages must be locked, @newpage->mapping must be set up.
5582 * Either or both pages might be on the LRU already.
5584 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5586 struct mem_cgroup *memcg;
5589 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5590 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5591 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5592 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5595 if (mem_cgroup_disabled())
5598 /* Page cache replacement: new page already charged? */
5599 if (newpage->mem_cgroup)
5602 /* Swapcache readahead pages can get replaced before being charged */
5603 memcg = oldpage->mem_cgroup;
5607 lock_page_lru(oldpage, &isolated);
5608 oldpage->mem_cgroup = NULL;
5609 unlock_page_lru(oldpage, isolated);
5611 commit_charge(newpage, memcg, true);
5615 * subsys_initcall() for memory controller.
5617 * Some parts like hotcpu_notifier() have to be initialized from this context
5618 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5619 * everything that doesn't depend on a specific mem_cgroup structure should
5620 * be initialized from here.
5622 static int __init mem_cgroup_init(void)
5626 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5628 for_each_possible_cpu(cpu)
5629 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5632 for_each_node(node) {
5633 struct mem_cgroup_tree_per_node *rtpn;
5636 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5637 node_online(node) ? node : NUMA_NO_NODE);
5639 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5640 struct mem_cgroup_tree_per_zone *rtpz;
5642 rtpz = &rtpn->rb_tree_per_zone[zone];
5643 rtpz->rb_root = RB_ROOT;
5644 spin_lock_init(&rtpz->lock);
5646 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5651 subsys_initcall(mem_cgroup_init);
5653 #ifdef CONFIG_MEMCG_SWAP
5655 * mem_cgroup_swapout - transfer a memsw charge to swap
5656 * @page: page whose memsw charge to transfer
5657 * @entry: swap entry to move the charge to
5659 * Transfer the memsw charge of @page to @entry.
5661 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5663 struct mem_cgroup *memcg;
5664 unsigned short oldid;
5666 VM_BUG_ON_PAGE(PageLRU(page), page);
5667 VM_BUG_ON_PAGE(page_count(page), page);
5669 if (!do_swap_account)
5672 memcg = page->mem_cgroup;
5674 /* Readahead page, never charged */
5678 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5679 VM_BUG_ON_PAGE(oldid, page);
5680 mem_cgroup_swap_statistics(memcg, true);
5682 page->mem_cgroup = NULL;
5684 if (!mem_cgroup_is_root(memcg))
5685 page_counter_uncharge(&memcg->memory, 1);
5688 * Interrupts should be disabled here because the caller holds the
5689 * mapping->tree_lock lock which is taken with interrupts-off. It is
5690 * important here to have the interrupts disabled because it is the
5691 * only synchronisation we have for udpating the per-CPU variables.
5693 VM_BUG_ON(!irqs_disabled());
5694 mem_cgroup_charge_statistics(memcg, page, -1);
5695 memcg_check_events(memcg, page);
5699 * mem_cgroup_uncharge_swap - uncharge a swap entry
5700 * @entry: swap entry to uncharge
5702 * Drop the memsw charge associated with @entry.
5704 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5706 struct mem_cgroup *memcg;
5709 if (!do_swap_account)
5712 id = swap_cgroup_record(entry, 0);
5714 memcg = mem_cgroup_from_id(id);
5716 if (!mem_cgroup_is_root(memcg))
5717 page_counter_uncharge(&memcg->memsw, 1);
5718 mem_cgroup_swap_statistics(memcg, false);
5719 css_put(&memcg->css);
5724 /* for remember boot option*/
5725 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5726 static int really_do_swap_account __initdata = 1;
5728 static int really_do_swap_account __initdata;
5731 static int __init enable_swap_account(char *s)
5733 if (!strcmp(s, "1"))
5734 really_do_swap_account = 1;
5735 else if (!strcmp(s, "0"))
5736 really_do_swap_account = 0;
5739 __setup("swapaccount=", enable_swap_account);
5741 static struct cftype memsw_cgroup_files[] = {
5743 .name = "memsw.usage_in_bytes",
5744 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5745 .read_u64 = mem_cgroup_read_u64,
5748 .name = "memsw.max_usage_in_bytes",
5749 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5750 .write = mem_cgroup_reset,
5751 .read_u64 = mem_cgroup_read_u64,
5754 .name = "memsw.limit_in_bytes",
5755 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5756 .write = mem_cgroup_write,
5757 .read_u64 = mem_cgroup_read_u64,
5760 .name = "memsw.failcnt",
5761 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5762 .write = mem_cgroup_reset,
5763 .read_u64 = mem_cgroup_read_u64,
5765 { }, /* terminate */
5768 static int __init mem_cgroup_swap_init(void)
5770 if (!mem_cgroup_disabled() && really_do_swap_account) {
5771 do_swap_account = 1;
5772 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5773 memsw_cgroup_files));
5777 subsys_initcall(mem_cgroup_swap_init);
5779 #endif /* CONFIG_MEMCG_SWAP */