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;
298 /* Socket cloning can throw us here with sk_cgrp already
299 * filled. It won't however, necessarily happen from
300 * process context. So the test for root memcg given
301 * the current task's memcg won't help us in this case.
303 * Respecting the original socket's memcg is a better
304 * decision in this case.
307 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
308 css_get(&sk->sk_memcg->css);
313 memcg = mem_cgroup_from_task(current);
314 if (memcg != root_mem_cgroup &&
315 memcg->tcp_mem.active &&
316 css_tryget_online(&memcg->css))
317 sk->sk_memcg = memcg;
320 EXPORT_SYMBOL(sock_update_memcg);
322 void sock_release_memcg(struct sock *sk)
324 WARN_ON(!sk->sk_memcg);
325 css_put(&sk->sk_memcg->css);
329 * mem_cgroup_charge_skmem - charge socket memory
330 * @memcg: memcg to charge
331 * @nr_pages: number of pages to charge
333 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
334 * @memcg's configured limit, %false if the charge had to be forced.
336 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
338 struct page_counter *counter;
340 if (page_counter_try_charge(&memcg->tcp_mem.memory_allocated,
341 nr_pages, &counter)) {
342 memcg->tcp_mem.memory_pressure = 0;
345 page_counter_charge(&memcg->tcp_mem.memory_allocated, nr_pages);
346 memcg->tcp_mem.memory_pressure = 1;
351 * mem_cgroup_uncharge_skmem - uncharge socket memory
352 * @memcg - memcg to uncharge
353 * @nr_pages - number of pages to uncharge
355 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
357 page_counter_uncharge(&memcg->tcp_mem.memory_allocated, nr_pages);
362 #ifdef CONFIG_MEMCG_KMEM
364 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
365 * The main reason for not using cgroup id for this:
366 * this works better in sparse environments, where we have a lot of memcgs,
367 * but only a few kmem-limited. Or also, if we have, for instance, 200
368 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
369 * 200 entry array for that.
371 * The current size of the caches array is stored in memcg_nr_cache_ids. It
372 * will double each time we have to increase it.
374 static DEFINE_IDA(memcg_cache_ida);
375 int memcg_nr_cache_ids;
377 /* Protects memcg_nr_cache_ids */
378 static DECLARE_RWSEM(memcg_cache_ids_sem);
380 void memcg_get_cache_ids(void)
382 down_read(&memcg_cache_ids_sem);
385 void memcg_put_cache_ids(void)
387 up_read(&memcg_cache_ids_sem);
391 * MIN_SIZE is different than 1, because we would like to avoid going through
392 * the alloc/free process all the time. In a small machine, 4 kmem-limited
393 * cgroups is a reasonable guess. In the future, it could be a parameter or
394 * tunable, but that is strictly not necessary.
396 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
397 * this constant directly from cgroup, but it is understandable that this is
398 * better kept as an internal representation in cgroup.c. In any case, the
399 * cgrp_id space is not getting any smaller, and we don't have to necessarily
400 * increase ours as well if it increases.
402 #define MEMCG_CACHES_MIN_SIZE 4
403 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
406 * A lot of the calls to the cache allocation functions are expected to be
407 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
408 * conditional to this static branch, we'll have to allow modules that does
409 * kmem_cache_alloc and the such to see this symbol as well
411 struct static_key memcg_kmem_enabled_key;
412 EXPORT_SYMBOL(memcg_kmem_enabled_key);
414 #endif /* CONFIG_MEMCG_KMEM */
416 static struct mem_cgroup_per_zone *
417 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
419 int nid = zone_to_nid(zone);
420 int zid = zone_idx(zone);
422 return &memcg->nodeinfo[nid]->zoneinfo[zid];
426 * mem_cgroup_css_from_page - css of the memcg associated with a page
427 * @page: page of interest
429 * If memcg is bound to the default hierarchy, css of the memcg associated
430 * with @page is returned. The returned css remains associated with @page
431 * until it is released.
433 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
436 * XXX: The above description of behavior on the default hierarchy isn't
437 * strictly true yet as replace_page_cache_page() can modify the
438 * association before @page is released even on the default hierarchy;
439 * however, the current and planned usages don't mix the the two functions
440 * and replace_page_cache_page() will soon be updated to make the invariant
443 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
445 struct mem_cgroup *memcg;
449 memcg = page->mem_cgroup;
451 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
452 memcg = root_mem_cgroup;
459 * page_cgroup_ino - return inode number of the memcg a page is charged to
462 * Look up the closest online ancestor of the memory cgroup @page is charged to
463 * and return its inode number or 0 if @page is not charged to any cgroup. It
464 * is safe to call this function without holding a reference to @page.
466 * Note, this function is inherently racy, because there is nothing to prevent
467 * the cgroup inode from getting torn down and potentially reallocated a moment
468 * after page_cgroup_ino() returns, so it only should be used by callers that
469 * do not care (such as procfs interfaces).
471 ino_t page_cgroup_ino(struct page *page)
473 struct mem_cgroup *memcg;
474 unsigned long ino = 0;
477 memcg = READ_ONCE(page->mem_cgroup);
478 while (memcg && !(memcg->css.flags & CSS_ONLINE))
479 memcg = parent_mem_cgroup(memcg);
481 ino = cgroup_ino(memcg->css.cgroup);
486 static struct mem_cgroup_per_zone *
487 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
489 int nid = page_to_nid(page);
490 int zid = page_zonenum(page);
492 return &memcg->nodeinfo[nid]->zoneinfo[zid];
495 static struct mem_cgroup_tree_per_zone *
496 soft_limit_tree_node_zone(int nid, int zid)
498 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
501 static struct mem_cgroup_tree_per_zone *
502 soft_limit_tree_from_page(struct page *page)
504 int nid = page_to_nid(page);
505 int zid = page_zonenum(page);
507 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
510 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
511 struct mem_cgroup_tree_per_zone *mctz,
512 unsigned long new_usage_in_excess)
514 struct rb_node **p = &mctz->rb_root.rb_node;
515 struct rb_node *parent = NULL;
516 struct mem_cgroup_per_zone *mz_node;
521 mz->usage_in_excess = new_usage_in_excess;
522 if (!mz->usage_in_excess)
526 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
528 if (mz->usage_in_excess < mz_node->usage_in_excess)
531 * We can't avoid mem cgroups that are over their soft
532 * limit by the same amount
534 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
537 rb_link_node(&mz->tree_node, parent, p);
538 rb_insert_color(&mz->tree_node, &mctz->rb_root);
542 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
543 struct mem_cgroup_tree_per_zone *mctz)
547 rb_erase(&mz->tree_node, &mctz->rb_root);
551 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
552 struct mem_cgroup_tree_per_zone *mctz)
556 spin_lock_irqsave(&mctz->lock, flags);
557 __mem_cgroup_remove_exceeded(mz, mctz);
558 spin_unlock_irqrestore(&mctz->lock, flags);
561 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
563 unsigned long nr_pages = page_counter_read(&memcg->memory);
564 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
565 unsigned long excess = 0;
567 if (nr_pages > soft_limit)
568 excess = nr_pages - soft_limit;
573 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
575 unsigned long excess;
576 struct mem_cgroup_per_zone *mz;
577 struct mem_cgroup_tree_per_zone *mctz;
579 mctz = soft_limit_tree_from_page(page);
581 * Necessary to update all ancestors when hierarchy is used.
582 * because their event counter is not touched.
584 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
585 mz = mem_cgroup_page_zoneinfo(memcg, page);
586 excess = soft_limit_excess(memcg);
588 * We have to update the tree if mz is on RB-tree or
589 * mem is over its softlimit.
591 if (excess || mz->on_tree) {
594 spin_lock_irqsave(&mctz->lock, flags);
595 /* if on-tree, remove it */
597 __mem_cgroup_remove_exceeded(mz, mctz);
599 * Insert again. mz->usage_in_excess will be updated.
600 * If excess is 0, no tree ops.
602 __mem_cgroup_insert_exceeded(mz, mctz, excess);
603 spin_unlock_irqrestore(&mctz->lock, flags);
608 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
610 struct mem_cgroup_tree_per_zone *mctz;
611 struct mem_cgroup_per_zone *mz;
615 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
616 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
617 mctz = soft_limit_tree_node_zone(nid, zid);
618 mem_cgroup_remove_exceeded(mz, mctz);
623 static struct mem_cgroup_per_zone *
624 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
626 struct rb_node *rightmost = NULL;
627 struct mem_cgroup_per_zone *mz;
631 rightmost = rb_last(&mctz->rb_root);
633 goto done; /* Nothing to reclaim from */
635 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
637 * Remove the node now but someone else can add it back,
638 * we will to add it back at the end of reclaim to its correct
639 * position in the tree.
641 __mem_cgroup_remove_exceeded(mz, mctz);
642 if (!soft_limit_excess(mz->memcg) ||
643 !css_tryget_online(&mz->memcg->css))
649 static struct mem_cgroup_per_zone *
650 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
652 struct mem_cgroup_per_zone *mz;
654 spin_lock_irq(&mctz->lock);
655 mz = __mem_cgroup_largest_soft_limit_node(mctz);
656 spin_unlock_irq(&mctz->lock);
661 * Return page count for single (non recursive) @memcg.
663 * Implementation Note: reading percpu statistics for memcg.
665 * Both of vmstat[] and percpu_counter has threshold and do periodic
666 * synchronization to implement "quick" read. There are trade-off between
667 * reading cost and precision of value. Then, we may have a chance to implement
668 * a periodic synchronization of counter in memcg's counter.
670 * But this _read() function is used for user interface now. The user accounts
671 * memory usage by memory cgroup and he _always_ requires exact value because
672 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
673 * have to visit all online cpus and make sum. So, for now, unnecessary
674 * synchronization is not implemented. (just implemented for cpu hotplug)
676 * If there are kernel internal actions which can make use of some not-exact
677 * value, and reading all cpu value can be performance bottleneck in some
678 * common workload, threshold and synchronization as vmstat[] should be
682 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
687 /* Per-cpu values can be negative, use a signed accumulator */
688 for_each_possible_cpu(cpu)
689 val += per_cpu(memcg->stat->count[idx], cpu);
691 * Summing races with updates, so val may be negative. Avoid exposing
692 * transient negative values.
699 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
700 enum mem_cgroup_events_index idx)
702 unsigned long val = 0;
705 for_each_possible_cpu(cpu)
706 val += per_cpu(memcg->stat->events[idx], cpu);
710 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
715 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
716 * counted as CACHE even if it's on ANON LRU.
719 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
722 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
725 if (PageTransHuge(page))
726 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
729 /* pagein of a big page is an event. So, ignore page size */
731 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
733 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
734 nr_pages = -nr_pages; /* for event */
737 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
740 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
742 unsigned int lru_mask)
744 unsigned long nr = 0;
747 VM_BUG_ON((unsigned)nid >= nr_node_ids);
749 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
750 struct mem_cgroup_per_zone *mz;
754 if (!(BIT(lru) & lru_mask))
756 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
757 nr += mz->lru_size[lru];
763 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
764 unsigned int lru_mask)
766 unsigned long nr = 0;
769 for_each_node_state(nid, N_MEMORY)
770 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
774 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
775 enum mem_cgroup_events_target target)
777 unsigned long val, next;
779 val = __this_cpu_read(memcg->stat->nr_page_events);
780 next = __this_cpu_read(memcg->stat->targets[target]);
781 /* from time_after() in jiffies.h */
782 if ((long)next - (long)val < 0) {
784 case MEM_CGROUP_TARGET_THRESH:
785 next = val + THRESHOLDS_EVENTS_TARGET;
787 case MEM_CGROUP_TARGET_SOFTLIMIT:
788 next = val + SOFTLIMIT_EVENTS_TARGET;
790 case MEM_CGROUP_TARGET_NUMAINFO:
791 next = val + NUMAINFO_EVENTS_TARGET;
796 __this_cpu_write(memcg->stat->targets[target], next);
803 * Check events in order.
806 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
808 /* threshold event is triggered in finer grain than soft limit */
809 if (unlikely(mem_cgroup_event_ratelimit(memcg,
810 MEM_CGROUP_TARGET_THRESH))) {
812 bool do_numainfo __maybe_unused;
814 do_softlimit = mem_cgroup_event_ratelimit(memcg,
815 MEM_CGROUP_TARGET_SOFTLIMIT);
817 do_numainfo = mem_cgroup_event_ratelimit(memcg,
818 MEM_CGROUP_TARGET_NUMAINFO);
820 mem_cgroup_threshold(memcg);
821 if (unlikely(do_softlimit))
822 mem_cgroup_update_tree(memcg, page);
824 if (unlikely(do_numainfo))
825 atomic_inc(&memcg->numainfo_events);
830 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
833 * mm_update_next_owner() may clear mm->owner to NULL
834 * if it races with swapoff, page migration, etc.
835 * So this can be called with p == NULL.
840 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
842 EXPORT_SYMBOL(mem_cgroup_from_task);
844 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
846 struct mem_cgroup *memcg = NULL;
851 * Page cache insertions can happen withou an
852 * actual mm context, e.g. during disk probing
853 * on boot, loopback IO, acct() writes etc.
856 memcg = root_mem_cgroup;
858 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
859 if (unlikely(!memcg))
860 memcg = root_mem_cgroup;
862 } while (!css_tryget_online(&memcg->css));
868 * mem_cgroup_iter - iterate over memory cgroup hierarchy
869 * @root: hierarchy root
870 * @prev: previously returned memcg, NULL on first invocation
871 * @reclaim: cookie for shared reclaim walks, NULL for full walks
873 * Returns references to children of the hierarchy below @root, or
874 * @root itself, or %NULL after a full round-trip.
876 * Caller must pass the return value in @prev on subsequent
877 * invocations for reference counting, or use mem_cgroup_iter_break()
878 * to cancel a hierarchy walk before the round-trip is complete.
880 * Reclaimers can specify a zone and a priority level in @reclaim to
881 * divide up the memcgs in the hierarchy among all concurrent
882 * reclaimers operating on the same zone and priority.
884 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
885 struct mem_cgroup *prev,
886 struct mem_cgroup_reclaim_cookie *reclaim)
888 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
889 struct cgroup_subsys_state *css = NULL;
890 struct mem_cgroup *memcg = NULL;
891 struct mem_cgroup *pos = NULL;
893 if (mem_cgroup_disabled())
897 root = root_mem_cgroup;
899 if (prev && !reclaim)
902 if (!root->use_hierarchy && root != root_mem_cgroup) {
911 struct mem_cgroup_per_zone *mz;
913 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
914 iter = &mz->iter[reclaim->priority];
916 if (prev && reclaim->generation != iter->generation)
920 pos = READ_ONCE(iter->position);
921 if (!pos || css_tryget(&pos->css))
924 * css reference reached zero, so iter->position will
925 * be cleared by ->css_released. However, we should not
926 * rely on this happening soon, because ->css_released
927 * is called from a work queue, and by busy-waiting we
928 * might block it. So we clear iter->position right
931 (void)cmpxchg(&iter->position, pos, NULL);
939 css = css_next_descendant_pre(css, &root->css);
942 * Reclaimers share the hierarchy walk, and a
943 * new one might jump in right at the end of
944 * the hierarchy - make sure they see at least
945 * one group and restart from the beginning.
953 * Verify the css and acquire a reference. The root
954 * is provided by the caller, so we know it's alive
955 * and kicking, and don't take an extra reference.
957 memcg = mem_cgroup_from_css(css);
959 if (css == &root->css)
962 if (css_tryget(css)) {
964 * Make sure the memcg is initialized:
965 * mem_cgroup_css_online() orders the the
966 * initialization against setting the flag.
968 if (smp_load_acquire(&memcg->initialized))
979 * The position could have already been updated by a competing
980 * thread, so check that the value hasn't changed since we read
981 * it to avoid reclaiming from the same cgroup twice.
983 (void)cmpxchg(&iter->position, pos, memcg);
991 reclaim->generation = iter->generation;
997 if (prev && prev != root)
1004 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1005 * @root: hierarchy root
1006 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1008 void mem_cgroup_iter_break(struct mem_cgroup *root,
1009 struct mem_cgroup *prev)
1012 root = root_mem_cgroup;
1013 if (prev && prev != root)
1014 css_put(&prev->css);
1017 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1019 struct mem_cgroup *memcg = dead_memcg;
1020 struct mem_cgroup_reclaim_iter *iter;
1021 struct mem_cgroup_per_zone *mz;
1025 while ((memcg = parent_mem_cgroup(memcg))) {
1026 for_each_node(nid) {
1027 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1028 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
1029 for (i = 0; i <= DEF_PRIORITY; i++) {
1030 iter = &mz->iter[i];
1031 cmpxchg(&iter->position,
1040 * Iteration constructs for visiting all cgroups (under a tree). If
1041 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1042 * be used for reference counting.
1044 #define for_each_mem_cgroup_tree(iter, root) \
1045 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1047 iter = mem_cgroup_iter(root, iter, NULL))
1049 #define for_each_mem_cgroup(iter) \
1050 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1052 iter = mem_cgroup_iter(NULL, iter, NULL))
1055 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1056 * @zone: zone of the wanted lruvec
1057 * @memcg: memcg of the wanted lruvec
1059 * Returns the lru list vector holding pages for the given @zone and
1060 * @mem. This can be the global zone lruvec, if the memory controller
1063 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1064 struct mem_cgroup *memcg)
1066 struct mem_cgroup_per_zone *mz;
1067 struct lruvec *lruvec;
1069 if (mem_cgroup_disabled()) {
1070 lruvec = &zone->lruvec;
1074 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1075 lruvec = &mz->lruvec;
1078 * Since a node can be onlined after the mem_cgroup was created,
1079 * we have to be prepared to initialize lruvec->zone here;
1080 * and if offlined then reonlined, we need to reinitialize it.
1082 if (unlikely(lruvec->zone != zone))
1083 lruvec->zone = zone;
1088 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1090 * @zone: zone of the page
1092 * This function is only safe when following the LRU page isolation
1093 * and putback protocol: the LRU lock must be held, and the page must
1094 * either be PageLRU() or the caller must have isolated/allocated it.
1096 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1098 struct mem_cgroup_per_zone *mz;
1099 struct mem_cgroup *memcg;
1100 struct lruvec *lruvec;
1102 if (mem_cgroup_disabled()) {
1103 lruvec = &zone->lruvec;
1107 memcg = page->mem_cgroup;
1109 * Swapcache readahead pages are added to the LRU - and
1110 * possibly migrated - before they are charged.
1113 memcg = root_mem_cgroup;
1115 mz = mem_cgroup_page_zoneinfo(memcg, page);
1116 lruvec = &mz->lruvec;
1119 * Since a node can be onlined after the mem_cgroup was created,
1120 * we have to be prepared to initialize lruvec->zone here;
1121 * and if offlined then reonlined, we need to reinitialize it.
1123 if (unlikely(lruvec->zone != zone))
1124 lruvec->zone = zone;
1129 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1130 * @lruvec: mem_cgroup per zone lru vector
1131 * @lru: index of lru list the page is sitting on
1132 * @nr_pages: positive when adding or negative when removing
1134 * This function must be called when a page is added to or removed from an
1137 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1140 struct mem_cgroup_per_zone *mz;
1141 unsigned long *lru_size;
1143 if (mem_cgroup_disabled())
1146 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1147 lru_size = mz->lru_size + lru;
1148 *lru_size += nr_pages;
1149 VM_BUG_ON((long)(*lru_size) < 0);
1152 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1154 struct mem_cgroup *task_memcg;
1155 struct task_struct *p;
1158 p = find_lock_task_mm(task);
1160 task_memcg = get_mem_cgroup_from_mm(p->mm);
1164 * All threads may have already detached their mm's, but the oom
1165 * killer still needs to detect if they have already been oom
1166 * killed to prevent needlessly killing additional tasks.
1169 task_memcg = mem_cgroup_from_task(task);
1170 css_get(&task_memcg->css);
1173 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1174 css_put(&task_memcg->css);
1178 #define mem_cgroup_from_counter(counter, member) \
1179 container_of(counter, struct mem_cgroup, member)
1182 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1183 * @memcg: the memory cgroup
1185 * Returns the maximum amount of memory @mem can be charged with, in
1188 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1190 unsigned long margin = 0;
1191 unsigned long count;
1192 unsigned long limit;
1194 count = page_counter_read(&memcg->memory);
1195 limit = READ_ONCE(memcg->memory.limit);
1197 margin = limit - count;
1199 if (do_swap_account) {
1200 count = page_counter_read(&memcg->memsw);
1201 limit = READ_ONCE(memcg->memsw.limit);
1203 margin = min(margin, limit - count);
1210 * A routine for checking "mem" is under move_account() or not.
1212 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1213 * moving cgroups. This is for waiting at high-memory pressure
1216 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1218 struct mem_cgroup *from;
1219 struct mem_cgroup *to;
1222 * Unlike task_move routines, we access mc.to, mc.from not under
1223 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1225 spin_lock(&mc.lock);
1231 ret = mem_cgroup_is_descendant(from, memcg) ||
1232 mem_cgroup_is_descendant(to, memcg);
1234 spin_unlock(&mc.lock);
1238 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1240 if (mc.moving_task && current != mc.moving_task) {
1241 if (mem_cgroup_under_move(memcg)) {
1243 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1244 /* moving charge context might have finished. */
1247 finish_wait(&mc.waitq, &wait);
1254 #define K(x) ((x) << (PAGE_SHIFT-10))
1256 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1257 * @memcg: The memory cgroup that went over limit
1258 * @p: Task that is going to be killed
1260 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1263 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1265 /* oom_info_lock ensures that parallel ooms do not interleave */
1266 static DEFINE_MUTEX(oom_info_lock);
1267 struct mem_cgroup *iter;
1270 mutex_lock(&oom_info_lock);
1274 pr_info("Task in ");
1275 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1276 pr_cont(" killed as a result of limit of ");
1278 pr_info("Memory limit reached of cgroup ");
1281 pr_cont_cgroup_path(memcg->css.cgroup);
1286 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1287 K((u64)page_counter_read(&memcg->memory)),
1288 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1289 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1290 K((u64)page_counter_read(&memcg->memsw)),
1291 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1292 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1293 K((u64)page_counter_read(&memcg->kmem)),
1294 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1296 for_each_mem_cgroup_tree(iter, memcg) {
1297 pr_info("Memory cgroup stats for ");
1298 pr_cont_cgroup_path(iter->css.cgroup);
1301 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1302 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1304 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1305 K(mem_cgroup_read_stat(iter, i)));
1308 for (i = 0; i < NR_LRU_LISTS; i++)
1309 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1310 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1314 mutex_unlock(&oom_info_lock);
1318 * This function returns the number of memcg under hierarchy tree. Returns
1319 * 1(self count) if no children.
1321 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1324 struct mem_cgroup *iter;
1326 for_each_mem_cgroup_tree(iter, memcg)
1332 * Return the memory (and swap, if configured) limit for a memcg.
1334 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1336 unsigned long limit;
1338 limit = memcg->memory.limit;
1339 if (mem_cgroup_swappiness(memcg)) {
1340 unsigned long memsw_limit;
1342 memsw_limit = memcg->memsw.limit;
1343 limit = min(limit + total_swap_pages, memsw_limit);
1348 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1351 struct oom_control oc = {
1354 .gfp_mask = gfp_mask,
1357 struct mem_cgroup *iter;
1358 unsigned long chosen_points = 0;
1359 unsigned long totalpages;
1360 unsigned int points = 0;
1361 struct task_struct *chosen = NULL;
1363 mutex_lock(&oom_lock);
1366 * If current has a pending SIGKILL or is exiting, then automatically
1367 * select it. The goal is to allow it to allocate so that it may
1368 * quickly exit and free its memory.
1370 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1371 mark_oom_victim(current);
1375 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1376 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1377 for_each_mem_cgroup_tree(iter, memcg) {
1378 struct css_task_iter it;
1379 struct task_struct *task;
1381 css_task_iter_start(&iter->css, &it);
1382 while ((task = css_task_iter_next(&it))) {
1383 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1384 case OOM_SCAN_SELECT:
1386 put_task_struct(chosen);
1388 chosen_points = ULONG_MAX;
1389 get_task_struct(chosen);
1391 case OOM_SCAN_CONTINUE:
1393 case OOM_SCAN_ABORT:
1394 css_task_iter_end(&it);
1395 mem_cgroup_iter_break(memcg, iter);
1397 put_task_struct(chosen);
1402 points = oom_badness(task, memcg, NULL, totalpages);
1403 if (!points || points < chosen_points)
1405 /* Prefer thread group leaders for display purposes */
1406 if (points == chosen_points &&
1407 thread_group_leader(chosen))
1411 put_task_struct(chosen);
1413 chosen_points = points;
1414 get_task_struct(chosen);
1416 css_task_iter_end(&it);
1420 points = chosen_points * 1000 / totalpages;
1421 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1422 "Memory cgroup out of memory");
1425 mutex_unlock(&oom_lock);
1428 #if MAX_NUMNODES > 1
1431 * test_mem_cgroup_node_reclaimable
1432 * @memcg: the target memcg
1433 * @nid: the node ID to be checked.
1434 * @noswap : specify true here if the user wants flle only information.
1436 * This function returns whether the specified memcg contains any
1437 * reclaimable pages on a node. Returns true if there are any reclaimable
1438 * pages in the node.
1440 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1441 int nid, bool noswap)
1443 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1445 if (noswap || !total_swap_pages)
1447 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1454 * Always updating the nodemask is not very good - even if we have an empty
1455 * list or the wrong list here, we can start from some node and traverse all
1456 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1459 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1463 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1464 * pagein/pageout changes since the last update.
1466 if (!atomic_read(&memcg->numainfo_events))
1468 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1471 /* make a nodemask where this memcg uses memory from */
1472 memcg->scan_nodes = node_states[N_MEMORY];
1474 for_each_node_mask(nid, node_states[N_MEMORY]) {
1476 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1477 node_clear(nid, memcg->scan_nodes);
1480 atomic_set(&memcg->numainfo_events, 0);
1481 atomic_set(&memcg->numainfo_updating, 0);
1485 * Selecting a node where we start reclaim from. Because what we need is just
1486 * reducing usage counter, start from anywhere is O,K. Considering
1487 * memory reclaim from current node, there are pros. and cons.
1489 * Freeing memory from current node means freeing memory from a node which
1490 * we'll use or we've used. So, it may make LRU bad. And if several threads
1491 * hit limits, it will see a contention on a node. But freeing from remote
1492 * node means more costs for memory reclaim because of memory latency.
1494 * Now, we use round-robin. Better algorithm is welcomed.
1496 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1500 mem_cgroup_may_update_nodemask(memcg);
1501 node = memcg->last_scanned_node;
1503 node = next_node(node, memcg->scan_nodes);
1504 if (node == MAX_NUMNODES)
1505 node = first_node(memcg->scan_nodes);
1507 * We call this when we hit limit, not when pages are added to LRU.
1508 * No LRU may hold pages because all pages are UNEVICTABLE or
1509 * memcg is too small and all pages are not on LRU. In that case,
1510 * we use curret node.
1512 if (unlikely(node == MAX_NUMNODES))
1513 node = numa_node_id();
1515 memcg->last_scanned_node = node;
1519 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1525 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1528 unsigned long *total_scanned)
1530 struct mem_cgroup *victim = NULL;
1533 unsigned long excess;
1534 unsigned long nr_scanned;
1535 struct mem_cgroup_reclaim_cookie reclaim = {
1540 excess = soft_limit_excess(root_memcg);
1543 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1548 * If we have not been able to reclaim
1549 * anything, it might because there are
1550 * no reclaimable pages under this hierarchy
1555 * We want to do more targeted reclaim.
1556 * excess >> 2 is not to excessive so as to
1557 * reclaim too much, nor too less that we keep
1558 * coming back to reclaim from this cgroup
1560 if (total >= (excess >> 2) ||
1561 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1566 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1568 *total_scanned += nr_scanned;
1569 if (!soft_limit_excess(root_memcg))
1572 mem_cgroup_iter_break(root_memcg, victim);
1576 #ifdef CONFIG_LOCKDEP
1577 static struct lockdep_map memcg_oom_lock_dep_map = {
1578 .name = "memcg_oom_lock",
1582 static DEFINE_SPINLOCK(memcg_oom_lock);
1585 * Check OOM-Killer is already running under our hierarchy.
1586 * If someone is running, return false.
1588 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1590 struct mem_cgroup *iter, *failed = NULL;
1592 spin_lock(&memcg_oom_lock);
1594 for_each_mem_cgroup_tree(iter, memcg) {
1595 if (iter->oom_lock) {
1597 * this subtree of our hierarchy is already locked
1598 * so we cannot give a lock.
1601 mem_cgroup_iter_break(memcg, iter);
1604 iter->oom_lock = true;
1609 * OK, we failed to lock the whole subtree so we have
1610 * to clean up what we set up to the failing subtree
1612 for_each_mem_cgroup_tree(iter, memcg) {
1613 if (iter == failed) {
1614 mem_cgroup_iter_break(memcg, iter);
1617 iter->oom_lock = false;
1620 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1622 spin_unlock(&memcg_oom_lock);
1627 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1629 struct mem_cgroup *iter;
1631 spin_lock(&memcg_oom_lock);
1632 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1633 for_each_mem_cgroup_tree(iter, memcg)
1634 iter->oom_lock = false;
1635 spin_unlock(&memcg_oom_lock);
1638 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1640 struct mem_cgroup *iter;
1642 spin_lock(&memcg_oom_lock);
1643 for_each_mem_cgroup_tree(iter, memcg)
1645 spin_unlock(&memcg_oom_lock);
1648 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1650 struct mem_cgroup *iter;
1653 * When a new child is created while the hierarchy is under oom,
1654 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1656 spin_lock(&memcg_oom_lock);
1657 for_each_mem_cgroup_tree(iter, memcg)
1658 if (iter->under_oom > 0)
1660 spin_unlock(&memcg_oom_lock);
1663 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1665 struct oom_wait_info {
1666 struct mem_cgroup *memcg;
1670 static int memcg_oom_wake_function(wait_queue_t *wait,
1671 unsigned mode, int sync, void *arg)
1673 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1674 struct mem_cgroup *oom_wait_memcg;
1675 struct oom_wait_info *oom_wait_info;
1677 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1678 oom_wait_memcg = oom_wait_info->memcg;
1680 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1681 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1683 return autoremove_wake_function(wait, mode, sync, arg);
1686 static void memcg_oom_recover(struct mem_cgroup *memcg)
1689 * For the following lockless ->under_oom test, the only required
1690 * guarantee is that it must see the state asserted by an OOM when
1691 * this function is called as a result of userland actions
1692 * triggered by the notification of the OOM. This is trivially
1693 * achieved by invoking mem_cgroup_mark_under_oom() before
1694 * triggering notification.
1696 if (memcg && memcg->under_oom)
1697 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1700 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1702 if (!current->memcg_may_oom)
1705 * We are in the middle of the charge context here, so we
1706 * don't want to block when potentially sitting on a callstack
1707 * that holds all kinds of filesystem and mm locks.
1709 * Also, the caller may handle a failed allocation gracefully
1710 * (like optional page cache readahead) and so an OOM killer
1711 * invocation might not even be necessary.
1713 * That's why we don't do anything here except remember the
1714 * OOM context and then deal with it at the end of the page
1715 * fault when the stack is unwound, the locks are released,
1716 * and when we know whether the fault was overall successful.
1718 css_get(&memcg->css);
1719 current->memcg_in_oom = memcg;
1720 current->memcg_oom_gfp_mask = mask;
1721 current->memcg_oom_order = order;
1725 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1726 * @handle: actually kill/wait or just clean up the OOM state
1728 * This has to be called at the end of a page fault if the memcg OOM
1729 * handler was enabled.
1731 * Memcg supports userspace OOM handling where failed allocations must
1732 * sleep on a waitqueue until the userspace task resolves the
1733 * situation. Sleeping directly in the charge context with all kinds
1734 * of locks held is not a good idea, instead we remember an OOM state
1735 * in the task and mem_cgroup_oom_synchronize() has to be called at
1736 * the end of the page fault to complete the OOM handling.
1738 * Returns %true if an ongoing memcg OOM situation was detected and
1739 * completed, %false otherwise.
1741 bool mem_cgroup_oom_synchronize(bool handle)
1743 struct mem_cgroup *memcg = current->memcg_in_oom;
1744 struct oom_wait_info owait;
1747 /* OOM is global, do not handle */
1751 if (!handle || oom_killer_disabled)
1754 owait.memcg = memcg;
1755 owait.wait.flags = 0;
1756 owait.wait.func = memcg_oom_wake_function;
1757 owait.wait.private = current;
1758 INIT_LIST_HEAD(&owait.wait.task_list);
1760 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1761 mem_cgroup_mark_under_oom(memcg);
1763 locked = mem_cgroup_oom_trylock(memcg);
1766 mem_cgroup_oom_notify(memcg);
1768 if (locked && !memcg->oom_kill_disable) {
1769 mem_cgroup_unmark_under_oom(memcg);
1770 finish_wait(&memcg_oom_waitq, &owait.wait);
1771 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1772 current->memcg_oom_order);
1775 mem_cgroup_unmark_under_oom(memcg);
1776 finish_wait(&memcg_oom_waitq, &owait.wait);
1780 mem_cgroup_oom_unlock(memcg);
1782 * There is no guarantee that an OOM-lock contender
1783 * sees the wakeups triggered by the OOM kill
1784 * uncharges. Wake any sleepers explicitely.
1786 memcg_oom_recover(memcg);
1789 current->memcg_in_oom = NULL;
1790 css_put(&memcg->css);
1795 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1796 * @page: page that is going to change accounted state
1798 * This function must mark the beginning of an accounted page state
1799 * change to prevent double accounting when the page is concurrently
1800 * being moved to another memcg:
1802 * memcg = mem_cgroup_begin_page_stat(page);
1803 * if (TestClearPageState(page))
1804 * mem_cgroup_update_page_stat(memcg, state, -1);
1805 * mem_cgroup_end_page_stat(memcg);
1807 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1809 struct mem_cgroup *memcg;
1810 unsigned long flags;
1813 * The RCU lock is held throughout the transaction. The fast
1814 * path can get away without acquiring the memcg->move_lock
1815 * because page moving starts with an RCU grace period.
1817 * The RCU lock also protects the memcg from being freed when
1818 * the page state that is going to change is the only thing
1819 * preventing the page from being uncharged.
1820 * E.g. end-writeback clearing PageWriteback(), which allows
1821 * migration to go ahead and uncharge the page before the
1822 * account transaction might be complete.
1826 if (mem_cgroup_disabled())
1829 memcg = page->mem_cgroup;
1830 if (unlikely(!memcg))
1833 if (atomic_read(&memcg->moving_account) <= 0)
1836 spin_lock_irqsave(&memcg->move_lock, flags);
1837 if (memcg != page->mem_cgroup) {
1838 spin_unlock_irqrestore(&memcg->move_lock, flags);
1843 * When charge migration first begins, we can have locked and
1844 * unlocked page stat updates happening concurrently. Track
1845 * the task who has the lock for mem_cgroup_end_page_stat().
1847 memcg->move_lock_task = current;
1848 memcg->move_lock_flags = flags;
1852 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1855 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1856 * @memcg: the memcg that was accounted against
1858 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1860 if (memcg && memcg->move_lock_task == current) {
1861 unsigned long flags = memcg->move_lock_flags;
1863 memcg->move_lock_task = NULL;
1864 memcg->move_lock_flags = 0;
1866 spin_unlock_irqrestore(&memcg->move_lock, flags);
1871 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1874 * size of first charge trial. "32" comes from vmscan.c's magic value.
1875 * TODO: maybe necessary to use big numbers in big irons.
1877 #define CHARGE_BATCH 32U
1878 struct memcg_stock_pcp {
1879 struct mem_cgroup *cached; /* this never be root cgroup */
1880 unsigned int nr_pages;
1881 struct work_struct work;
1882 unsigned long flags;
1883 #define FLUSHING_CACHED_CHARGE 0
1885 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1886 static DEFINE_MUTEX(percpu_charge_mutex);
1889 * consume_stock: Try to consume stocked charge on this cpu.
1890 * @memcg: memcg to consume from.
1891 * @nr_pages: how many pages to charge.
1893 * The charges will only happen if @memcg matches the current cpu's memcg
1894 * stock, and at least @nr_pages are available in that stock. Failure to
1895 * service an allocation will refill the stock.
1897 * returns true if successful, false otherwise.
1899 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1901 struct memcg_stock_pcp *stock;
1904 if (nr_pages > CHARGE_BATCH)
1907 stock = &get_cpu_var(memcg_stock);
1908 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1909 stock->nr_pages -= nr_pages;
1912 put_cpu_var(memcg_stock);
1917 * Returns stocks cached in percpu and reset cached information.
1919 static void drain_stock(struct memcg_stock_pcp *stock)
1921 struct mem_cgroup *old = stock->cached;
1923 if (stock->nr_pages) {
1924 page_counter_uncharge(&old->memory, stock->nr_pages);
1925 if (do_swap_account)
1926 page_counter_uncharge(&old->memsw, stock->nr_pages);
1927 css_put_many(&old->css, stock->nr_pages);
1928 stock->nr_pages = 0;
1930 stock->cached = NULL;
1934 * This must be called under preempt disabled or must be called by
1935 * a thread which is pinned to local cpu.
1937 static void drain_local_stock(struct work_struct *dummy)
1939 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1941 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1945 * Cache charges(val) to local per_cpu area.
1946 * This will be consumed by consume_stock() function, later.
1948 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1950 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1952 if (stock->cached != memcg) { /* reset if necessary */
1954 stock->cached = memcg;
1956 stock->nr_pages += nr_pages;
1957 put_cpu_var(memcg_stock);
1961 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1962 * of the hierarchy under it.
1964 static void drain_all_stock(struct mem_cgroup *root_memcg)
1968 /* If someone's already draining, avoid adding running more workers. */
1969 if (!mutex_trylock(&percpu_charge_mutex))
1971 /* Notify other cpus that system-wide "drain" is running */
1974 for_each_online_cpu(cpu) {
1975 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1976 struct mem_cgroup *memcg;
1978 memcg = stock->cached;
1979 if (!memcg || !stock->nr_pages)
1981 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1983 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1985 drain_local_stock(&stock->work);
1987 schedule_work_on(cpu, &stock->work);
1992 mutex_unlock(&percpu_charge_mutex);
1995 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1996 unsigned long action,
1999 int cpu = (unsigned long)hcpu;
2000 struct memcg_stock_pcp *stock;
2002 if (action == CPU_ONLINE)
2005 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2008 stock = &per_cpu(memcg_stock, cpu);
2014 * Scheduled by try_charge() to be executed from the userland return path
2015 * and reclaims memory over the high limit.
2017 void mem_cgroup_handle_over_high(void)
2019 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2020 struct mem_cgroup *memcg, *pos;
2022 if (likely(!nr_pages))
2025 pos = memcg = get_mem_cgroup_from_mm(current->mm);
2028 if (page_counter_read(&pos->memory) <= pos->high)
2030 mem_cgroup_events(pos, MEMCG_HIGH, 1);
2031 try_to_free_mem_cgroup_pages(pos, nr_pages, GFP_KERNEL, true);
2032 } while ((pos = parent_mem_cgroup(pos)));
2034 css_put(&memcg->css);
2035 current->memcg_nr_pages_over_high = 0;
2038 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2039 unsigned int nr_pages)
2041 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2042 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2043 struct mem_cgroup *mem_over_limit;
2044 struct page_counter *counter;
2045 unsigned long nr_reclaimed;
2046 bool may_swap = true;
2047 bool drained = false;
2049 if (mem_cgroup_is_root(memcg))
2052 if (consume_stock(memcg, nr_pages))
2055 if (!do_swap_account ||
2056 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2057 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2059 if (do_swap_account)
2060 page_counter_uncharge(&memcg->memsw, batch);
2061 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2063 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2067 if (batch > nr_pages) {
2073 * Unlike in global OOM situations, memcg is not in a physical
2074 * memory shortage. Allow dying and OOM-killed tasks to
2075 * bypass the last charges so that they can exit quickly and
2076 * free their memory.
2078 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2079 fatal_signal_pending(current) ||
2080 current->flags & PF_EXITING))
2083 if (unlikely(task_in_memcg_oom(current)))
2086 if (!gfpflags_allow_blocking(gfp_mask))
2089 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2091 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2092 gfp_mask, may_swap);
2094 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2098 drain_all_stock(mem_over_limit);
2103 if (gfp_mask & __GFP_NORETRY)
2106 * Even though the limit is exceeded at this point, reclaim
2107 * may have been able to free some pages. Retry the charge
2108 * before killing the task.
2110 * Only for regular pages, though: huge pages are rather
2111 * unlikely to succeed so close to the limit, and we fall back
2112 * to regular pages anyway in case of failure.
2114 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2117 * At task move, charge accounts can be doubly counted. So, it's
2118 * better to wait until the end of task_move if something is going on.
2120 if (mem_cgroup_wait_acct_move(mem_over_limit))
2126 if (gfp_mask & __GFP_NOFAIL)
2129 if (fatal_signal_pending(current))
2132 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2134 mem_cgroup_oom(mem_over_limit, gfp_mask,
2135 get_order(nr_pages * PAGE_SIZE));
2137 if (!(gfp_mask & __GFP_NOFAIL))
2141 * The allocation either can't fail or will lead to more memory
2142 * being freed very soon. Allow memory usage go over the limit
2143 * temporarily by force charging it.
2145 page_counter_charge(&memcg->memory, nr_pages);
2146 if (do_swap_account)
2147 page_counter_charge(&memcg->memsw, nr_pages);
2148 css_get_many(&memcg->css, nr_pages);
2153 css_get_many(&memcg->css, batch);
2154 if (batch > nr_pages)
2155 refill_stock(memcg, batch - nr_pages);
2158 * If the hierarchy is above the normal consumption range, schedule
2159 * reclaim on returning to userland. We can perform reclaim here
2160 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2161 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2162 * not recorded as it most likely matches current's and won't
2163 * change in the meantime. As high limit is checked again before
2164 * reclaim, the cost of mismatch is negligible.
2167 if (page_counter_read(&memcg->memory) > memcg->high) {
2168 current->memcg_nr_pages_over_high += batch;
2169 set_notify_resume(current);
2172 } while ((memcg = parent_mem_cgroup(memcg)));
2177 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2179 if (mem_cgroup_is_root(memcg))
2182 page_counter_uncharge(&memcg->memory, nr_pages);
2183 if (do_swap_account)
2184 page_counter_uncharge(&memcg->memsw, nr_pages);
2186 css_put_many(&memcg->css, nr_pages);
2189 static void lock_page_lru(struct page *page, int *isolated)
2191 struct zone *zone = page_zone(page);
2193 spin_lock_irq(&zone->lru_lock);
2194 if (PageLRU(page)) {
2195 struct lruvec *lruvec;
2197 lruvec = mem_cgroup_page_lruvec(page, zone);
2199 del_page_from_lru_list(page, lruvec, page_lru(page));
2205 static void unlock_page_lru(struct page *page, int isolated)
2207 struct zone *zone = page_zone(page);
2210 struct lruvec *lruvec;
2212 lruvec = mem_cgroup_page_lruvec(page, zone);
2213 VM_BUG_ON_PAGE(PageLRU(page), page);
2215 add_page_to_lru_list(page, lruvec, page_lru(page));
2217 spin_unlock_irq(&zone->lru_lock);
2220 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2225 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2228 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2229 * may already be on some other mem_cgroup's LRU. Take care of it.
2232 lock_page_lru(page, &isolated);
2235 * Nobody should be changing or seriously looking at
2236 * page->mem_cgroup at this point:
2238 * - the page is uncharged
2240 * - the page is off-LRU
2242 * - an anonymous fault has exclusive page access, except for
2243 * a locked page table
2245 * - a page cache insertion, a swapin fault, or a migration
2246 * have the page locked
2248 page->mem_cgroup = memcg;
2251 unlock_page_lru(page, isolated);
2254 #ifdef CONFIG_MEMCG_KMEM
2255 static int memcg_alloc_cache_id(void)
2260 id = ida_simple_get(&memcg_cache_ida,
2261 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2265 if (id < memcg_nr_cache_ids)
2269 * There's no space for the new id in memcg_caches arrays,
2270 * so we have to grow them.
2272 down_write(&memcg_cache_ids_sem);
2274 size = 2 * (id + 1);
2275 if (size < MEMCG_CACHES_MIN_SIZE)
2276 size = MEMCG_CACHES_MIN_SIZE;
2277 else if (size > MEMCG_CACHES_MAX_SIZE)
2278 size = MEMCG_CACHES_MAX_SIZE;
2280 err = memcg_update_all_caches(size);
2282 err = memcg_update_all_list_lrus(size);
2284 memcg_nr_cache_ids = size;
2286 up_write(&memcg_cache_ids_sem);
2289 ida_simple_remove(&memcg_cache_ida, id);
2295 static void memcg_free_cache_id(int id)
2297 ida_simple_remove(&memcg_cache_ida, id);
2300 struct memcg_kmem_cache_create_work {
2301 struct mem_cgroup *memcg;
2302 struct kmem_cache *cachep;
2303 struct work_struct work;
2306 static void memcg_kmem_cache_create_func(struct work_struct *w)
2308 struct memcg_kmem_cache_create_work *cw =
2309 container_of(w, struct memcg_kmem_cache_create_work, work);
2310 struct mem_cgroup *memcg = cw->memcg;
2311 struct kmem_cache *cachep = cw->cachep;
2313 memcg_create_kmem_cache(memcg, cachep);
2315 css_put(&memcg->css);
2320 * Enqueue the creation of a per-memcg kmem_cache.
2322 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2323 struct kmem_cache *cachep)
2325 struct memcg_kmem_cache_create_work *cw;
2327 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2331 css_get(&memcg->css);
2334 cw->cachep = cachep;
2335 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2337 schedule_work(&cw->work);
2340 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2341 struct kmem_cache *cachep)
2344 * We need to stop accounting when we kmalloc, because if the
2345 * corresponding kmalloc cache is not yet created, the first allocation
2346 * in __memcg_schedule_kmem_cache_create will recurse.
2348 * However, it is better to enclose the whole function. Depending on
2349 * the debugging options enabled, INIT_WORK(), for instance, can
2350 * trigger an allocation. This too, will make us recurse. Because at
2351 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2352 * the safest choice is to do it like this, wrapping the whole function.
2354 current->memcg_kmem_skip_account = 1;
2355 __memcg_schedule_kmem_cache_create(memcg, cachep);
2356 current->memcg_kmem_skip_account = 0;
2360 * Return the kmem_cache we're supposed to use for a slab allocation.
2361 * We try to use the current memcg's version of the cache.
2363 * If the cache does not exist yet, if we are the first user of it,
2364 * we either create it immediately, if possible, or create it asynchronously
2366 * In the latter case, we will let the current allocation go through with
2367 * the original cache.
2369 * Can't be called in interrupt context or from kernel threads.
2370 * This function needs to be called with rcu_read_lock() held.
2372 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp)
2374 struct mem_cgroup *memcg;
2375 struct kmem_cache *memcg_cachep;
2378 VM_BUG_ON(!is_root_cache(cachep));
2380 if (cachep->flags & SLAB_ACCOUNT)
2381 gfp |= __GFP_ACCOUNT;
2383 if (!(gfp & __GFP_ACCOUNT))
2386 if (current->memcg_kmem_skip_account)
2389 memcg = get_mem_cgroup_from_mm(current->mm);
2390 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2394 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2395 if (likely(memcg_cachep))
2396 return memcg_cachep;
2399 * If we are in a safe context (can wait, and not in interrupt
2400 * context), we could be be predictable and return right away.
2401 * This would guarantee that the allocation being performed
2402 * already belongs in the new cache.
2404 * However, there are some clashes that can arrive from locking.
2405 * For instance, because we acquire the slab_mutex while doing
2406 * memcg_create_kmem_cache, this means no further allocation
2407 * could happen with the slab_mutex held. So it's better to
2410 memcg_schedule_kmem_cache_create(memcg, cachep);
2412 css_put(&memcg->css);
2416 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2418 if (!is_root_cache(cachep))
2419 css_put(&cachep->memcg_params.memcg->css);
2422 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2423 struct mem_cgroup *memcg)
2425 unsigned int nr_pages = 1 << order;
2426 struct page_counter *counter;
2429 if (!memcg_kmem_is_active(memcg))
2432 if (!page_counter_try_charge(&memcg->kmem, nr_pages, &counter))
2435 ret = try_charge(memcg, gfp, nr_pages);
2437 page_counter_uncharge(&memcg->kmem, nr_pages);
2441 page->mem_cgroup = memcg;
2446 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2448 struct mem_cgroup *memcg;
2451 memcg = get_mem_cgroup_from_mm(current->mm);
2452 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2453 css_put(&memcg->css);
2457 void __memcg_kmem_uncharge(struct page *page, int order)
2459 struct mem_cgroup *memcg = page->mem_cgroup;
2460 unsigned int nr_pages = 1 << order;
2465 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2467 page_counter_uncharge(&memcg->kmem, nr_pages);
2468 page_counter_uncharge(&memcg->memory, nr_pages);
2469 if (do_swap_account)
2470 page_counter_uncharge(&memcg->memsw, nr_pages);
2472 page->mem_cgroup = NULL;
2473 css_put_many(&memcg->css, nr_pages);
2475 #endif /* CONFIG_MEMCG_KMEM */
2477 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2480 * Because tail pages are not marked as "used", set it. We're under
2481 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2482 * charge/uncharge will be never happen and move_account() is done under
2483 * compound_lock(), so we don't have to take care of races.
2485 void mem_cgroup_split_huge_fixup(struct page *head)
2489 if (mem_cgroup_disabled())
2492 for (i = 1; i < HPAGE_PMD_NR; i++)
2493 head[i].mem_cgroup = head->mem_cgroup;
2495 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2498 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2500 #ifdef CONFIG_MEMCG_SWAP
2501 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2504 int val = (charge) ? 1 : -1;
2505 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2509 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2510 * @entry: swap entry to be moved
2511 * @from: mem_cgroup which the entry is moved from
2512 * @to: mem_cgroup which the entry is moved to
2514 * It succeeds only when the swap_cgroup's record for this entry is the same
2515 * as the mem_cgroup's id of @from.
2517 * Returns 0 on success, -EINVAL on failure.
2519 * The caller must have charged to @to, IOW, called page_counter_charge() about
2520 * both res and memsw, and called css_get().
2522 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2523 struct mem_cgroup *from, struct mem_cgroup *to)
2525 unsigned short old_id, new_id;
2527 old_id = mem_cgroup_id(from);
2528 new_id = mem_cgroup_id(to);
2530 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2531 mem_cgroup_swap_statistics(from, false);
2532 mem_cgroup_swap_statistics(to, true);
2538 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2539 struct mem_cgroup *from, struct mem_cgroup *to)
2545 static DEFINE_MUTEX(memcg_limit_mutex);
2547 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2548 unsigned long limit)
2550 unsigned long curusage;
2551 unsigned long oldusage;
2552 bool enlarge = false;
2557 * For keeping hierarchical_reclaim simple, how long we should retry
2558 * is depends on callers. We set our retry-count to be function
2559 * of # of children which we should visit in this loop.
2561 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2562 mem_cgroup_count_children(memcg);
2564 oldusage = page_counter_read(&memcg->memory);
2567 if (signal_pending(current)) {
2572 mutex_lock(&memcg_limit_mutex);
2573 if (limit > memcg->memsw.limit) {
2574 mutex_unlock(&memcg_limit_mutex);
2578 if (limit > memcg->memory.limit)
2580 ret = page_counter_limit(&memcg->memory, limit);
2581 mutex_unlock(&memcg_limit_mutex);
2586 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2588 curusage = page_counter_read(&memcg->memory);
2589 /* Usage is reduced ? */
2590 if (curusage >= oldusage)
2593 oldusage = curusage;
2594 } while (retry_count);
2596 if (!ret && enlarge)
2597 memcg_oom_recover(memcg);
2602 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2603 unsigned long limit)
2605 unsigned long curusage;
2606 unsigned long oldusage;
2607 bool enlarge = false;
2611 /* see mem_cgroup_resize_res_limit */
2612 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2613 mem_cgroup_count_children(memcg);
2615 oldusage = page_counter_read(&memcg->memsw);
2618 if (signal_pending(current)) {
2623 mutex_lock(&memcg_limit_mutex);
2624 if (limit < memcg->memory.limit) {
2625 mutex_unlock(&memcg_limit_mutex);
2629 if (limit > memcg->memsw.limit)
2631 ret = page_counter_limit(&memcg->memsw, limit);
2632 mutex_unlock(&memcg_limit_mutex);
2637 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2639 curusage = page_counter_read(&memcg->memsw);
2640 /* Usage is reduced ? */
2641 if (curusage >= oldusage)
2644 oldusage = curusage;
2645 } while (retry_count);
2647 if (!ret && enlarge)
2648 memcg_oom_recover(memcg);
2653 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2655 unsigned long *total_scanned)
2657 unsigned long nr_reclaimed = 0;
2658 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2659 unsigned long reclaimed;
2661 struct mem_cgroup_tree_per_zone *mctz;
2662 unsigned long excess;
2663 unsigned long nr_scanned;
2668 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2670 * This loop can run a while, specially if mem_cgroup's continuously
2671 * keep exceeding their soft limit and putting the system under
2678 mz = mem_cgroup_largest_soft_limit_node(mctz);
2683 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2684 gfp_mask, &nr_scanned);
2685 nr_reclaimed += reclaimed;
2686 *total_scanned += nr_scanned;
2687 spin_lock_irq(&mctz->lock);
2688 __mem_cgroup_remove_exceeded(mz, mctz);
2691 * If we failed to reclaim anything from this memory cgroup
2692 * it is time to move on to the next cgroup
2696 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2698 excess = soft_limit_excess(mz->memcg);
2700 * One school of thought says that we should not add
2701 * back the node to the tree if reclaim returns 0.
2702 * But our reclaim could return 0, simply because due
2703 * to priority we are exposing a smaller subset of
2704 * memory to reclaim from. Consider this as a longer
2707 /* If excess == 0, no tree ops */
2708 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2709 spin_unlock_irq(&mctz->lock);
2710 css_put(&mz->memcg->css);
2713 * Could not reclaim anything and there are no more
2714 * mem cgroups to try or we seem to be looping without
2715 * reclaiming anything.
2717 if (!nr_reclaimed &&
2719 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2721 } while (!nr_reclaimed);
2723 css_put(&next_mz->memcg->css);
2724 return nr_reclaimed;
2728 * Test whether @memcg has children, dead or alive. Note that this
2729 * function doesn't care whether @memcg has use_hierarchy enabled and
2730 * returns %true if there are child csses according to the cgroup
2731 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2733 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2738 * The lock does not prevent addition or deletion of children, but
2739 * it prevents a new child from being initialized based on this
2740 * parent in css_online(), so it's enough to decide whether
2741 * hierarchically inherited attributes can still be changed or not.
2743 lockdep_assert_held(&memcg_create_mutex);
2746 ret = css_next_child(NULL, &memcg->css);
2752 * Reclaims as many pages from the given memcg as possible and moves
2753 * the rest to the parent.
2755 * Caller is responsible for holding css reference for memcg.
2757 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2759 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2761 /* we call try-to-free pages for make this cgroup empty */
2762 lru_add_drain_all();
2763 /* try to free all pages in this cgroup */
2764 while (nr_retries && page_counter_read(&memcg->memory)) {
2767 if (signal_pending(current))
2770 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2774 /* maybe some writeback is necessary */
2775 congestion_wait(BLK_RW_ASYNC, HZ/10);
2783 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2784 char *buf, size_t nbytes,
2787 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2789 if (mem_cgroup_is_root(memcg))
2791 return mem_cgroup_force_empty(memcg) ?: nbytes;
2794 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2797 return mem_cgroup_from_css(css)->use_hierarchy;
2800 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2801 struct cftype *cft, u64 val)
2804 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2805 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2807 mutex_lock(&memcg_create_mutex);
2809 if (memcg->use_hierarchy == val)
2813 * If parent's use_hierarchy is set, we can't make any modifications
2814 * in the child subtrees. If it is unset, then the change can
2815 * occur, provided the current cgroup has no children.
2817 * For the root cgroup, parent_mem is NULL, we allow value to be
2818 * set if there are no children.
2820 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2821 (val == 1 || val == 0)) {
2822 if (!memcg_has_children(memcg))
2823 memcg->use_hierarchy = val;
2830 mutex_unlock(&memcg_create_mutex);
2835 static unsigned long tree_stat(struct mem_cgroup *memcg,
2836 enum mem_cgroup_stat_index idx)
2838 struct mem_cgroup *iter;
2839 unsigned long val = 0;
2841 for_each_mem_cgroup_tree(iter, memcg)
2842 val += mem_cgroup_read_stat(iter, idx);
2847 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2851 if (mem_cgroup_is_root(memcg)) {
2852 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2853 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2855 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2858 val = page_counter_read(&memcg->memory);
2860 val = page_counter_read(&memcg->memsw);
2873 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2876 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2877 struct page_counter *counter;
2879 switch (MEMFILE_TYPE(cft->private)) {
2881 counter = &memcg->memory;
2884 counter = &memcg->memsw;
2887 counter = &memcg->kmem;
2893 switch (MEMFILE_ATTR(cft->private)) {
2895 if (counter == &memcg->memory)
2896 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2897 if (counter == &memcg->memsw)
2898 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2899 return (u64)page_counter_read(counter) * PAGE_SIZE;
2901 return (u64)counter->limit * PAGE_SIZE;
2903 return (u64)counter->watermark * PAGE_SIZE;
2905 return counter->failcnt;
2906 case RES_SOFT_LIMIT:
2907 return (u64)memcg->soft_limit * PAGE_SIZE;
2913 #ifdef CONFIG_MEMCG_KMEM
2914 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2915 unsigned long nr_pages)
2920 BUG_ON(memcg->kmemcg_id >= 0);
2921 BUG_ON(memcg->kmem_acct_activated);
2922 BUG_ON(memcg->kmem_acct_active);
2925 * For simplicity, we won't allow this to be disabled. It also can't
2926 * be changed if the cgroup has children already, or if tasks had
2929 * If tasks join before we set the limit, a person looking at
2930 * kmem.usage_in_bytes will have no way to determine when it took
2931 * place, which makes the value quite meaningless.
2933 * After it first became limited, changes in the value of the limit are
2934 * of course permitted.
2936 mutex_lock(&memcg_create_mutex);
2937 if (cgroup_is_populated(memcg->css.cgroup) ||
2938 (memcg->use_hierarchy && memcg_has_children(memcg)))
2940 mutex_unlock(&memcg_create_mutex);
2944 memcg_id = memcg_alloc_cache_id();
2951 * We couldn't have accounted to this cgroup, because it hasn't got
2952 * activated yet, so this should succeed.
2954 err = page_counter_limit(&memcg->kmem, nr_pages);
2957 static_key_slow_inc(&memcg_kmem_enabled_key);
2959 * A memory cgroup is considered kmem-active as soon as it gets
2960 * kmemcg_id. Setting the id after enabling static branching will
2961 * guarantee no one starts accounting before all call sites are
2964 memcg->kmemcg_id = memcg_id;
2965 memcg->kmem_acct_activated = true;
2966 memcg->kmem_acct_active = true;
2971 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2972 unsigned long limit)
2976 mutex_lock(&memcg_limit_mutex);
2977 if (!memcg_kmem_is_active(memcg))
2978 ret = memcg_activate_kmem(memcg, limit);
2980 ret = page_counter_limit(&memcg->kmem, limit);
2981 mutex_unlock(&memcg_limit_mutex);
2985 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2988 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2993 mutex_lock(&memcg_limit_mutex);
2995 * If the parent cgroup is not kmem-active now, it cannot be activated
2996 * after this point, because it has at least one child already.
2998 if (memcg_kmem_is_active(parent))
2999 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3000 mutex_unlock(&memcg_limit_mutex);
3004 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3005 unsigned long limit)
3009 #endif /* CONFIG_MEMCG_KMEM */
3012 * The user of this function is...
3015 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3016 char *buf, size_t nbytes, loff_t off)
3018 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3019 unsigned long nr_pages;
3022 buf = strstrip(buf);
3023 ret = page_counter_memparse(buf, "-1", &nr_pages);
3027 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3029 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3033 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3035 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3038 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3041 ret = memcg_update_kmem_limit(memcg, nr_pages);
3045 case RES_SOFT_LIMIT:
3046 memcg->soft_limit = nr_pages;
3050 return ret ?: nbytes;
3053 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3054 size_t nbytes, loff_t off)
3056 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3057 struct page_counter *counter;
3059 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3061 counter = &memcg->memory;
3064 counter = &memcg->memsw;
3067 counter = &memcg->kmem;
3073 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3075 page_counter_reset_watermark(counter);
3078 counter->failcnt = 0;
3087 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3090 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3094 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3095 struct cftype *cft, u64 val)
3097 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3099 if (val & ~MOVE_MASK)
3103 * No kind of locking is needed in here, because ->can_attach() will
3104 * check this value once in the beginning of the process, and then carry
3105 * on with stale data. This means that changes to this value will only
3106 * affect task migrations starting after the change.
3108 memcg->move_charge_at_immigrate = val;
3112 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3113 struct cftype *cft, u64 val)
3120 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3124 unsigned int lru_mask;
3127 static const struct numa_stat stats[] = {
3128 { "total", LRU_ALL },
3129 { "file", LRU_ALL_FILE },
3130 { "anon", LRU_ALL_ANON },
3131 { "unevictable", BIT(LRU_UNEVICTABLE) },
3133 const struct numa_stat *stat;
3136 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3138 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3139 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3140 seq_printf(m, "%s=%lu", stat->name, nr);
3141 for_each_node_state(nid, N_MEMORY) {
3142 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3144 seq_printf(m, " N%d=%lu", nid, nr);
3149 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3150 struct mem_cgroup *iter;
3153 for_each_mem_cgroup_tree(iter, memcg)
3154 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3155 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3156 for_each_node_state(nid, N_MEMORY) {
3158 for_each_mem_cgroup_tree(iter, memcg)
3159 nr += mem_cgroup_node_nr_lru_pages(
3160 iter, nid, stat->lru_mask);
3161 seq_printf(m, " N%d=%lu", nid, nr);
3168 #endif /* CONFIG_NUMA */
3170 static int memcg_stat_show(struct seq_file *m, void *v)
3172 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3173 unsigned long memory, memsw;
3174 struct mem_cgroup *mi;
3177 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3178 MEM_CGROUP_STAT_NSTATS);
3179 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3180 MEM_CGROUP_EVENTS_NSTATS);
3181 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3183 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3184 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3186 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3187 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3190 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3191 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3192 mem_cgroup_read_events(memcg, i));
3194 for (i = 0; i < NR_LRU_LISTS; i++)
3195 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3196 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3198 /* Hierarchical information */
3199 memory = memsw = PAGE_COUNTER_MAX;
3200 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3201 memory = min(memory, mi->memory.limit);
3202 memsw = min(memsw, mi->memsw.limit);
3204 seq_printf(m, "hierarchical_memory_limit %llu\n",
3205 (u64)memory * PAGE_SIZE);
3206 if (do_swap_account)
3207 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3208 (u64)memsw * PAGE_SIZE);
3210 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3211 unsigned long long val = 0;
3213 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3215 for_each_mem_cgroup_tree(mi, memcg)
3216 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3217 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3220 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3221 unsigned long long val = 0;
3223 for_each_mem_cgroup_tree(mi, memcg)
3224 val += mem_cgroup_read_events(mi, i);
3225 seq_printf(m, "total_%s %llu\n",
3226 mem_cgroup_events_names[i], val);
3229 for (i = 0; i < NR_LRU_LISTS; i++) {
3230 unsigned long long val = 0;
3232 for_each_mem_cgroup_tree(mi, memcg)
3233 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3234 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3237 #ifdef CONFIG_DEBUG_VM
3240 struct mem_cgroup_per_zone *mz;
3241 struct zone_reclaim_stat *rstat;
3242 unsigned long recent_rotated[2] = {0, 0};
3243 unsigned long recent_scanned[2] = {0, 0};
3245 for_each_online_node(nid)
3246 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3247 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3248 rstat = &mz->lruvec.reclaim_stat;
3250 recent_rotated[0] += rstat->recent_rotated[0];
3251 recent_rotated[1] += rstat->recent_rotated[1];
3252 recent_scanned[0] += rstat->recent_scanned[0];
3253 recent_scanned[1] += rstat->recent_scanned[1];
3255 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3256 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3257 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3258 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3265 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3268 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3270 return mem_cgroup_swappiness(memcg);
3273 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3274 struct cftype *cft, u64 val)
3276 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3282 memcg->swappiness = val;
3284 vm_swappiness = val;
3289 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3291 struct mem_cgroup_threshold_ary *t;
3292 unsigned long usage;
3297 t = rcu_dereference(memcg->thresholds.primary);
3299 t = rcu_dereference(memcg->memsw_thresholds.primary);
3304 usage = mem_cgroup_usage(memcg, swap);
3307 * current_threshold points to threshold just below or equal to usage.
3308 * If it's not true, a threshold was crossed after last
3309 * call of __mem_cgroup_threshold().
3311 i = t->current_threshold;
3314 * Iterate backward over array of thresholds starting from
3315 * current_threshold and check if a threshold is crossed.
3316 * If none of thresholds below usage is crossed, we read
3317 * only one element of the array here.
3319 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3320 eventfd_signal(t->entries[i].eventfd, 1);
3322 /* i = current_threshold + 1 */
3326 * Iterate forward over array of thresholds starting from
3327 * current_threshold+1 and check if a threshold is crossed.
3328 * If none of thresholds above usage is crossed, we read
3329 * only one element of the array here.
3331 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3332 eventfd_signal(t->entries[i].eventfd, 1);
3334 /* Update current_threshold */
3335 t->current_threshold = i - 1;
3340 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3343 __mem_cgroup_threshold(memcg, false);
3344 if (do_swap_account)
3345 __mem_cgroup_threshold(memcg, true);
3347 memcg = parent_mem_cgroup(memcg);
3351 static int compare_thresholds(const void *a, const void *b)
3353 const struct mem_cgroup_threshold *_a = a;
3354 const struct mem_cgroup_threshold *_b = b;
3356 if (_a->threshold > _b->threshold)
3359 if (_a->threshold < _b->threshold)
3365 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3367 struct mem_cgroup_eventfd_list *ev;
3369 spin_lock(&memcg_oom_lock);
3371 list_for_each_entry(ev, &memcg->oom_notify, list)
3372 eventfd_signal(ev->eventfd, 1);
3374 spin_unlock(&memcg_oom_lock);
3378 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3380 struct mem_cgroup *iter;
3382 for_each_mem_cgroup_tree(iter, memcg)
3383 mem_cgroup_oom_notify_cb(iter);
3386 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3387 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3389 struct mem_cgroup_thresholds *thresholds;
3390 struct mem_cgroup_threshold_ary *new;
3391 unsigned long threshold;
3392 unsigned long usage;
3395 ret = page_counter_memparse(args, "-1", &threshold);
3399 mutex_lock(&memcg->thresholds_lock);
3402 thresholds = &memcg->thresholds;
3403 usage = mem_cgroup_usage(memcg, false);
3404 } else if (type == _MEMSWAP) {
3405 thresholds = &memcg->memsw_thresholds;
3406 usage = mem_cgroup_usage(memcg, true);
3410 /* Check if a threshold crossed before adding a new one */
3411 if (thresholds->primary)
3412 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3414 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3416 /* Allocate memory for new array of thresholds */
3417 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3425 /* Copy thresholds (if any) to new array */
3426 if (thresholds->primary) {
3427 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3428 sizeof(struct mem_cgroup_threshold));
3431 /* Add new threshold */
3432 new->entries[size - 1].eventfd = eventfd;
3433 new->entries[size - 1].threshold = threshold;
3435 /* Sort thresholds. Registering of new threshold isn't time-critical */
3436 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3437 compare_thresholds, NULL);
3439 /* Find current threshold */
3440 new->current_threshold = -1;
3441 for (i = 0; i < size; i++) {
3442 if (new->entries[i].threshold <= usage) {
3444 * new->current_threshold will not be used until
3445 * rcu_assign_pointer(), so it's safe to increment
3448 ++new->current_threshold;
3453 /* Free old spare buffer and save old primary buffer as spare */
3454 kfree(thresholds->spare);
3455 thresholds->spare = thresholds->primary;
3457 rcu_assign_pointer(thresholds->primary, new);
3459 /* To be sure that nobody uses thresholds */
3463 mutex_unlock(&memcg->thresholds_lock);
3468 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3469 struct eventfd_ctx *eventfd, const char *args)
3471 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3474 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3475 struct eventfd_ctx *eventfd, const char *args)
3477 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3480 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3481 struct eventfd_ctx *eventfd, enum res_type type)
3483 struct mem_cgroup_thresholds *thresholds;
3484 struct mem_cgroup_threshold_ary *new;
3485 unsigned long usage;
3488 mutex_lock(&memcg->thresholds_lock);
3491 thresholds = &memcg->thresholds;
3492 usage = mem_cgroup_usage(memcg, false);
3493 } else if (type == _MEMSWAP) {
3494 thresholds = &memcg->memsw_thresholds;
3495 usage = mem_cgroup_usage(memcg, true);
3499 if (!thresholds->primary)
3502 /* Check if a threshold crossed before removing */
3503 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3505 /* Calculate new number of threshold */
3507 for (i = 0; i < thresholds->primary->size; i++) {
3508 if (thresholds->primary->entries[i].eventfd != eventfd)
3512 new = thresholds->spare;
3514 /* Set thresholds array to NULL if we don't have thresholds */
3523 /* Copy thresholds and find current threshold */
3524 new->current_threshold = -1;
3525 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3526 if (thresholds->primary->entries[i].eventfd == eventfd)
3529 new->entries[j] = thresholds->primary->entries[i];
3530 if (new->entries[j].threshold <= usage) {
3532 * new->current_threshold will not be used
3533 * until rcu_assign_pointer(), so it's safe to increment
3536 ++new->current_threshold;
3542 /* Swap primary and spare array */
3543 thresholds->spare = thresholds->primary;
3544 /* If all events are unregistered, free the spare array */
3546 kfree(thresholds->spare);
3547 thresholds->spare = NULL;
3550 rcu_assign_pointer(thresholds->primary, new);
3552 /* To be sure that nobody uses thresholds */
3555 mutex_unlock(&memcg->thresholds_lock);
3558 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3559 struct eventfd_ctx *eventfd)
3561 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3564 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3565 struct eventfd_ctx *eventfd)
3567 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3570 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3571 struct eventfd_ctx *eventfd, const char *args)
3573 struct mem_cgroup_eventfd_list *event;
3575 event = kmalloc(sizeof(*event), GFP_KERNEL);
3579 spin_lock(&memcg_oom_lock);
3581 event->eventfd = eventfd;
3582 list_add(&event->list, &memcg->oom_notify);
3584 /* already in OOM ? */
3585 if (memcg->under_oom)
3586 eventfd_signal(eventfd, 1);
3587 spin_unlock(&memcg_oom_lock);
3592 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3593 struct eventfd_ctx *eventfd)
3595 struct mem_cgroup_eventfd_list *ev, *tmp;
3597 spin_lock(&memcg_oom_lock);
3599 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3600 if (ev->eventfd == eventfd) {
3601 list_del(&ev->list);
3606 spin_unlock(&memcg_oom_lock);
3609 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3611 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3613 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3614 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3618 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3619 struct cftype *cft, u64 val)
3621 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3623 /* cannot set to root cgroup and only 0 and 1 are allowed */
3624 if (!css->parent || !((val == 0) || (val == 1)))
3627 memcg->oom_kill_disable = val;
3629 memcg_oom_recover(memcg);
3634 #ifdef CONFIG_MEMCG_KMEM
3635 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3639 ret = memcg_propagate_kmem(memcg);
3643 return tcp_init_cgroup(memcg, ss);
3646 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3648 struct cgroup_subsys_state *css;
3649 struct mem_cgroup *parent, *child;
3652 if (!memcg->kmem_acct_active)
3656 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3657 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3658 * guarantees no cache will be created for this cgroup after we are
3659 * done (see memcg_create_kmem_cache()).
3661 memcg->kmem_acct_active = false;
3663 memcg_deactivate_kmem_caches(memcg);
3665 kmemcg_id = memcg->kmemcg_id;
3666 BUG_ON(kmemcg_id < 0);
3668 parent = parent_mem_cgroup(memcg);
3670 parent = root_mem_cgroup;
3673 * Change kmemcg_id of this cgroup and all its descendants to the
3674 * parent's id, and then move all entries from this cgroup's list_lrus
3675 * to ones of the parent. After we have finished, all list_lrus
3676 * corresponding to this cgroup are guaranteed to remain empty. The
3677 * ordering is imposed by list_lru_node->lock taken by
3678 * memcg_drain_all_list_lrus().
3680 css_for_each_descendant_pre(css, &memcg->css) {
3681 child = mem_cgroup_from_css(css);
3682 BUG_ON(child->kmemcg_id != kmemcg_id);
3683 child->kmemcg_id = parent->kmemcg_id;
3684 if (!memcg->use_hierarchy)
3687 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3689 memcg_free_cache_id(kmemcg_id);
3692 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3694 if (memcg->kmem_acct_activated) {
3695 memcg_destroy_kmem_caches(memcg);
3696 static_key_slow_dec(&memcg_kmem_enabled_key);
3697 WARN_ON(page_counter_read(&memcg->kmem));
3699 tcp_destroy_cgroup(memcg);
3702 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3707 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3711 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3716 #ifdef CONFIG_CGROUP_WRITEBACK
3718 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3720 return &memcg->cgwb_list;
3723 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3725 return wb_domain_init(&memcg->cgwb_domain, gfp);
3728 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3730 wb_domain_exit(&memcg->cgwb_domain);
3733 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3735 wb_domain_size_changed(&memcg->cgwb_domain);
3738 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3740 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3742 if (!memcg->css.parent)
3745 return &memcg->cgwb_domain;
3749 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3750 * @wb: bdi_writeback in question
3751 * @pfilepages: out parameter for number of file pages
3752 * @pheadroom: out parameter for number of allocatable pages according to memcg
3753 * @pdirty: out parameter for number of dirty pages
3754 * @pwriteback: out parameter for number of pages under writeback
3756 * Determine the numbers of file, headroom, dirty, and writeback pages in
3757 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3758 * is a bit more involved.
3760 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3761 * headroom is calculated as the lowest headroom of itself and the
3762 * ancestors. Note that this doesn't consider the actual amount of
3763 * available memory in the system. The caller should further cap
3764 * *@pheadroom accordingly.
3766 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3767 unsigned long *pheadroom, unsigned long *pdirty,
3768 unsigned long *pwriteback)
3770 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3771 struct mem_cgroup *parent;
3773 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3775 /* this should eventually include NR_UNSTABLE_NFS */
3776 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3777 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3778 (1 << LRU_ACTIVE_FILE));
3779 *pheadroom = PAGE_COUNTER_MAX;
3781 while ((parent = parent_mem_cgroup(memcg))) {
3782 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3783 unsigned long used = page_counter_read(&memcg->memory);
3785 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3790 #else /* CONFIG_CGROUP_WRITEBACK */
3792 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3797 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3801 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3805 #endif /* CONFIG_CGROUP_WRITEBACK */
3808 * DO NOT USE IN NEW FILES.
3810 * "cgroup.event_control" implementation.
3812 * This is way over-engineered. It tries to support fully configurable
3813 * events for each user. Such level of flexibility is completely
3814 * unnecessary especially in the light of the planned unified hierarchy.
3816 * Please deprecate this and replace with something simpler if at all
3821 * Unregister event and free resources.
3823 * Gets called from workqueue.
3825 static void memcg_event_remove(struct work_struct *work)
3827 struct mem_cgroup_event *event =
3828 container_of(work, struct mem_cgroup_event, remove);
3829 struct mem_cgroup *memcg = event->memcg;
3831 remove_wait_queue(event->wqh, &event->wait);
3833 event->unregister_event(memcg, event->eventfd);
3835 /* Notify userspace the event is going away. */
3836 eventfd_signal(event->eventfd, 1);
3838 eventfd_ctx_put(event->eventfd);
3840 css_put(&memcg->css);
3844 * Gets called on POLLHUP on eventfd when user closes it.
3846 * Called with wqh->lock held and interrupts disabled.
3848 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3849 int sync, void *key)
3851 struct mem_cgroup_event *event =
3852 container_of(wait, struct mem_cgroup_event, wait);
3853 struct mem_cgroup *memcg = event->memcg;
3854 unsigned long flags = (unsigned long)key;
3856 if (flags & POLLHUP) {
3858 * If the event has been detached at cgroup removal, we
3859 * can simply return knowing the other side will cleanup
3862 * We can't race against event freeing since the other
3863 * side will require wqh->lock via remove_wait_queue(),
3866 spin_lock(&memcg->event_list_lock);
3867 if (!list_empty(&event->list)) {
3868 list_del_init(&event->list);
3870 * We are in atomic context, but cgroup_event_remove()
3871 * may sleep, so we have to call it in workqueue.
3873 schedule_work(&event->remove);
3875 spin_unlock(&memcg->event_list_lock);
3881 static void memcg_event_ptable_queue_proc(struct file *file,
3882 wait_queue_head_t *wqh, poll_table *pt)
3884 struct mem_cgroup_event *event =
3885 container_of(pt, struct mem_cgroup_event, pt);
3888 add_wait_queue(wqh, &event->wait);
3892 * DO NOT USE IN NEW FILES.
3894 * Parse input and register new cgroup event handler.
3896 * Input must be in format '<event_fd> <control_fd> <args>'.
3897 * Interpretation of args is defined by control file implementation.
3899 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3900 char *buf, size_t nbytes, loff_t off)
3902 struct cgroup_subsys_state *css = of_css(of);
3903 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3904 struct mem_cgroup_event *event;
3905 struct cgroup_subsys_state *cfile_css;
3906 unsigned int efd, cfd;
3913 buf = strstrip(buf);
3915 efd = simple_strtoul(buf, &endp, 10);
3920 cfd = simple_strtoul(buf, &endp, 10);
3921 if ((*endp != ' ') && (*endp != '\0'))
3925 event = kzalloc(sizeof(*event), GFP_KERNEL);
3929 event->memcg = memcg;
3930 INIT_LIST_HEAD(&event->list);
3931 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3932 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3933 INIT_WORK(&event->remove, memcg_event_remove);
3941 event->eventfd = eventfd_ctx_fileget(efile.file);
3942 if (IS_ERR(event->eventfd)) {
3943 ret = PTR_ERR(event->eventfd);
3950 goto out_put_eventfd;
3953 /* the process need read permission on control file */
3954 /* AV: shouldn't we check that it's been opened for read instead? */
3955 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3960 * Determine the event callbacks and set them in @event. This used
3961 * to be done via struct cftype but cgroup core no longer knows
3962 * about these events. The following is crude but the whole thing
3963 * is for compatibility anyway.
3965 * DO NOT ADD NEW FILES.
3967 name = cfile.file->f_path.dentry->d_name.name;
3969 if (!strcmp(name, "memory.usage_in_bytes")) {
3970 event->register_event = mem_cgroup_usage_register_event;
3971 event->unregister_event = mem_cgroup_usage_unregister_event;
3972 } else if (!strcmp(name, "memory.oom_control")) {
3973 event->register_event = mem_cgroup_oom_register_event;
3974 event->unregister_event = mem_cgroup_oom_unregister_event;
3975 } else if (!strcmp(name, "memory.pressure_level")) {
3976 event->register_event = vmpressure_register_event;
3977 event->unregister_event = vmpressure_unregister_event;
3978 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3979 event->register_event = memsw_cgroup_usage_register_event;
3980 event->unregister_event = memsw_cgroup_usage_unregister_event;
3987 * Verify @cfile should belong to @css. Also, remaining events are
3988 * automatically removed on cgroup destruction but the removal is
3989 * asynchronous, so take an extra ref on @css.
3991 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3992 &memory_cgrp_subsys);
3994 if (IS_ERR(cfile_css))
3996 if (cfile_css != css) {
4001 ret = event->register_event(memcg, event->eventfd, buf);
4005 efile.file->f_op->poll(efile.file, &event->pt);
4007 spin_lock(&memcg->event_list_lock);
4008 list_add(&event->list, &memcg->event_list);
4009 spin_unlock(&memcg->event_list_lock);
4021 eventfd_ctx_put(event->eventfd);
4030 static struct cftype mem_cgroup_legacy_files[] = {
4032 .name = "usage_in_bytes",
4033 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4034 .read_u64 = mem_cgroup_read_u64,
4037 .name = "max_usage_in_bytes",
4038 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4039 .write = mem_cgroup_reset,
4040 .read_u64 = mem_cgroup_read_u64,
4043 .name = "limit_in_bytes",
4044 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4045 .write = mem_cgroup_write,
4046 .read_u64 = mem_cgroup_read_u64,
4049 .name = "soft_limit_in_bytes",
4050 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4051 .write = mem_cgroup_write,
4052 .read_u64 = mem_cgroup_read_u64,
4056 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4057 .write = mem_cgroup_reset,
4058 .read_u64 = mem_cgroup_read_u64,
4062 .seq_show = memcg_stat_show,
4065 .name = "force_empty",
4066 .write = mem_cgroup_force_empty_write,
4069 .name = "use_hierarchy",
4070 .write_u64 = mem_cgroup_hierarchy_write,
4071 .read_u64 = mem_cgroup_hierarchy_read,
4074 .name = "cgroup.event_control", /* XXX: for compat */
4075 .write = memcg_write_event_control,
4076 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4079 .name = "swappiness",
4080 .read_u64 = mem_cgroup_swappiness_read,
4081 .write_u64 = mem_cgroup_swappiness_write,
4084 .name = "move_charge_at_immigrate",
4085 .read_u64 = mem_cgroup_move_charge_read,
4086 .write_u64 = mem_cgroup_move_charge_write,
4089 .name = "oom_control",
4090 .seq_show = mem_cgroup_oom_control_read,
4091 .write_u64 = mem_cgroup_oom_control_write,
4092 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4095 .name = "pressure_level",
4099 .name = "numa_stat",
4100 .seq_show = memcg_numa_stat_show,
4103 #ifdef CONFIG_MEMCG_KMEM
4105 .name = "kmem.limit_in_bytes",
4106 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4107 .write = mem_cgroup_write,
4108 .read_u64 = mem_cgroup_read_u64,
4111 .name = "kmem.usage_in_bytes",
4112 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4113 .read_u64 = mem_cgroup_read_u64,
4116 .name = "kmem.failcnt",
4117 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4118 .write = mem_cgroup_reset,
4119 .read_u64 = mem_cgroup_read_u64,
4122 .name = "kmem.max_usage_in_bytes",
4123 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4124 .write = mem_cgroup_reset,
4125 .read_u64 = mem_cgroup_read_u64,
4127 #ifdef CONFIG_SLABINFO
4129 .name = "kmem.slabinfo",
4130 .seq_start = slab_start,
4131 .seq_next = slab_next,
4132 .seq_stop = slab_stop,
4133 .seq_show = memcg_slab_show,
4137 { }, /* terminate */
4140 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4142 struct mem_cgroup_per_node *pn;
4143 struct mem_cgroup_per_zone *mz;
4144 int zone, tmp = node;
4146 * This routine is called against possible nodes.
4147 * But it's BUG to call kmalloc() against offline node.
4149 * TODO: this routine can waste much memory for nodes which will
4150 * never be onlined. It's better to use memory hotplug callback
4153 if (!node_state(node, N_NORMAL_MEMORY))
4155 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4159 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4160 mz = &pn->zoneinfo[zone];
4161 lruvec_init(&mz->lruvec);
4162 mz->usage_in_excess = 0;
4163 mz->on_tree = false;
4166 memcg->nodeinfo[node] = pn;
4170 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4172 kfree(memcg->nodeinfo[node]);
4175 static struct mem_cgroup *mem_cgroup_alloc(void)
4177 struct mem_cgroup *memcg;
4180 size = sizeof(struct mem_cgroup);
4181 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4183 memcg = kzalloc(size, GFP_KERNEL);
4187 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4191 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4197 free_percpu(memcg->stat);
4204 * At destroying mem_cgroup, references from swap_cgroup can remain.
4205 * (scanning all at force_empty is too costly...)
4207 * Instead of clearing all references at force_empty, we remember
4208 * the number of reference from swap_cgroup and free mem_cgroup when
4209 * it goes down to 0.
4211 * Removal of cgroup itself succeeds regardless of refs from swap.
4214 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4218 mem_cgroup_remove_from_trees(memcg);
4221 free_mem_cgroup_per_zone_info(memcg, node);
4223 free_percpu(memcg->stat);
4224 memcg_wb_domain_exit(memcg);
4229 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4231 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4233 if (!memcg->memory.parent)
4235 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4237 EXPORT_SYMBOL(parent_mem_cgroup);
4239 static struct cgroup_subsys_state * __ref
4240 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4242 struct mem_cgroup *memcg;
4243 long error = -ENOMEM;
4246 memcg = mem_cgroup_alloc();
4248 return ERR_PTR(error);
4251 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4255 if (parent_css == NULL) {
4256 root_mem_cgroup = memcg;
4257 page_counter_init(&memcg->memory, NULL);
4258 memcg->high = PAGE_COUNTER_MAX;
4259 memcg->soft_limit = PAGE_COUNTER_MAX;
4260 page_counter_init(&memcg->memsw, NULL);
4261 page_counter_init(&memcg->kmem, NULL);
4264 memcg->last_scanned_node = MAX_NUMNODES;
4265 INIT_LIST_HEAD(&memcg->oom_notify);
4266 memcg->move_charge_at_immigrate = 0;
4267 mutex_init(&memcg->thresholds_lock);
4268 spin_lock_init(&memcg->move_lock);
4269 vmpressure_init(&memcg->vmpressure);
4270 INIT_LIST_HEAD(&memcg->event_list);
4271 spin_lock_init(&memcg->event_list_lock);
4272 #ifdef CONFIG_MEMCG_KMEM
4273 memcg->kmemcg_id = -1;
4275 #ifdef CONFIG_CGROUP_WRITEBACK
4276 INIT_LIST_HEAD(&memcg->cgwb_list);
4281 __mem_cgroup_free(memcg);
4282 return ERR_PTR(error);
4286 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4288 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4289 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4292 if (css->id > MEM_CGROUP_ID_MAX)
4298 mutex_lock(&memcg_create_mutex);
4300 memcg->use_hierarchy = parent->use_hierarchy;
4301 memcg->oom_kill_disable = parent->oom_kill_disable;
4302 memcg->swappiness = mem_cgroup_swappiness(parent);
4304 if (parent->use_hierarchy) {
4305 page_counter_init(&memcg->memory, &parent->memory);
4306 memcg->high = PAGE_COUNTER_MAX;
4307 memcg->soft_limit = PAGE_COUNTER_MAX;
4308 page_counter_init(&memcg->memsw, &parent->memsw);
4309 page_counter_init(&memcg->kmem, &parent->kmem);
4312 * No need to take a reference to the parent because cgroup
4313 * core guarantees its existence.
4316 page_counter_init(&memcg->memory, NULL);
4317 memcg->high = PAGE_COUNTER_MAX;
4318 memcg->soft_limit = PAGE_COUNTER_MAX;
4319 page_counter_init(&memcg->memsw, NULL);
4320 page_counter_init(&memcg->kmem, NULL);
4322 * Deeper hierachy with use_hierarchy == false doesn't make
4323 * much sense so let cgroup subsystem know about this
4324 * unfortunate state in our controller.
4326 if (parent != root_mem_cgroup)
4327 memory_cgrp_subsys.broken_hierarchy = true;
4329 mutex_unlock(&memcg_create_mutex);
4331 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4336 * Make sure the memcg is initialized: mem_cgroup_iter()
4337 * orders reading memcg->initialized against its callers
4338 * reading the memcg members.
4340 smp_store_release(&memcg->initialized, 1);
4345 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4347 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4348 struct mem_cgroup_event *event, *tmp;
4351 * Unregister events and notify userspace.
4352 * Notify userspace about cgroup removing only after rmdir of cgroup
4353 * directory to avoid race between userspace and kernelspace.
4355 spin_lock(&memcg->event_list_lock);
4356 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4357 list_del_init(&event->list);
4358 schedule_work(&event->remove);
4360 spin_unlock(&memcg->event_list_lock);
4362 vmpressure_cleanup(&memcg->vmpressure);
4364 memcg_deactivate_kmem(memcg);
4366 wb_memcg_offline(memcg);
4369 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4371 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4373 invalidate_reclaim_iterators(memcg);
4376 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4378 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4380 memcg_destroy_kmem(memcg);
4381 __mem_cgroup_free(memcg);
4385 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4386 * @css: the target css
4388 * Reset the states of the mem_cgroup associated with @css. This is
4389 * invoked when the userland requests disabling on the default hierarchy
4390 * but the memcg is pinned through dependency. The memcg should stop
4391 * applying policies and should revert to the vanilla state as it may be
4392 * made visible again.
4394 * The current implementation only resets the essential configurations.
4395 * This needs to be expanded to cover all the visible parts.
4397 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4399 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4401 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4402 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4403 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4405 memcg->high = PAGE_COUNTER_MAX;
4406 memcg->soft_limit = PAGE_COUNTER_MAX;
4407 memcg_wb_domain_size_changed(memcg);
4411 /* Handlers for move charge at task migration. */
4412 static int mem_cgroup_do_precharge(unsigned long count)
4416 /* Try a single bulk charge without reclaim first, kswapd may wake */
4417 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4419 mc.precharge += count;
4423 /* Try charges one by one with reclaim */
4425 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4435 * get_mctgt_type - get target type of moving charge
4436 * @vma: the vma the pte to be checked belongs
4437 * @addr: the address corresponding to the pte to be checked
4438 * @ptent: the pte to be checked
4439 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4442 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4443 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4444 * move charge. if @target is not NULL, the page is stored in target->page
4445 * with extra refcnt got(Callers should handle it).
4446 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4447 * target for charge migration. if @target is not NULL, the entry is stored
4450 * Called with pte lock held.
4457 enum mc_target_type {
4463 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4464 unsigned long addr, pte_t ptent)
4466 struct page *page = vm_normal_page(vma, addr, ptent);
4468 if (!page || !page_mapped(page))
4470 if (PageAnon(page)) {
4471 if (!(mc.flags & MOVE_ANON))
4474 if (!(mc.flags & MOVE_FILE))
4477 if (!get_page_unless_zero(page))
4484 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4485 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4487 struct page *page = NULL;
4488 swp_entry_t ent = pte_to_swp_entry(ptent);
4490 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4493 * Because lookup_swap_cache() updates some statistics counter,
4494 * we call find_get_page() with swapper_space directly.
4496 page = find_get_page(swap_address_space(ent), ent.val);
4497 if (do_swap_account)
4498 entry->val = ent.val;
4503 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4504 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4510 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4511 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4513 struct page *page = NULL;
4514 struct address_space *mapping;
4517 if (!vma->vm_file) /* anonymous vma */
4519 if (!(mc.flags & MOVE_FILE))
4522 mapping = vma->vm_file->f_mapping;
4523 pgoff = linear_page_index(vma, addr);
4525 /* page is moved even if it's not RSS of this task(page-faulted). */
4527 /* shmem/tmpfs may report page out on swap: account for that too. */
4528 if (shmem_mapping(mapping)) {
4529 page = find_get_entry(mapping, pgoff);
4530 if (radix_tree_exceptional_entry(page)) {
4531 swp_entry_t swp = radix_to_swp_entry(page);
4532 if (do_swap_account)
4534 page = find_get_page(swap_address_space(swp), swp.val);
4537 page = find_get_page(mapping, pgoff);
4539 page = find_get_page(mapping, pgoff);
4545 * mem_cgroup_move_account - move account of the page
4547 * @nr_pages: number of regular pages (>1 for huge pages)
4548 * @from: mem_cgroup which the page is moved from.
4549 * @to: mem_cgroup which the page is moved to. @from != @to.
4551 * The caller must confirm following.
4552 * - page is not on LRU (isolate_page() is useful.)
4553 * - compound_lock is held when nr_pages > 1
4555 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4558 static int mem_cgroup_move_account(struct page *page,
4559 unsigned int nr_pages,
4560 struct mem_cgroup *from,
4561 struct mem_cgroup *to)
4563 unsigned long flags;
4567 VM_BUG_ON(from == to);
4568 VM_BUG_ON_PAGE(PageLRU(page), page);
4570 * The page is isolated from LRU. So, collapse function
4571 * will not handle this page. But page splitting can happen.
4572 * Do this check under compound_page_lock(). The caller should
4576 if (nr_pages > 1 && !PageTransHuge(page))
4580 * Prevent mem_cgroup_replace_page() from looking at
4581 * page->mem_cgroup of its source page while we change it.
4583 if (!trylock_page(page))
4587 if (page->mem_cgroup != from)
4590 anon = PageAnon(page);
4592 spin_lock_irqsave(&from->move_lock, flags);
4594 if (!anon && page_mapped(page)) {
4595 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4597 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4602 * move_lock grabbed above and caller set from->moving_account, so
4603 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4604 * So mapping should be stable for dirty pages.
4606 if (!anon && PageDirty(page)) {
4607 struct address_space *mapping = page_mapping(page);
4609 if (mapping_cap_account_dirty(mapping)) {
4610 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4612 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4617 if (PageWriteback(page)) {
4618 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4620 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4625 * It is safe to change page->mem_cgroup here because the page
4626 * is referenced, charged, and isolated - we can't race with
4627 * uncharging, charging, migration, or LRU putback.
4630 /* caller should have done css_get */
4631 page->mem_cgroup = to;
4632 spin_unlock_irqrestore(&from->move_lock, flags);
4636 local_irq_disable();
4637 mem_cgroup_charge_statistics(to, page, nr_pages);
4638 memcg_check_events(to, page);
4639 mem_cgroup_charge_statistics(from, page, -nr_pages);
4640 memcg_check_events(from, page);
4648 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4649 unsigned long addr, pte_t ptent, union mc_target *target)
4651 struct page *page = NULL;
4652 enum mc_target_type ret = MC_TARGET_NONE;
4653 swp_entry_t ent = { .val = 0 };
4655 if (pte_present(ptent))
4656 page = mc_handle_present_pte(vma, addr, ptent);
4657 else if (is_swap_pte(ptent))
4658 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4659 else if (pte_none(ptent))
4660 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4662 if (!page && !ent.val)
4666 * Do only loose check w/o serialization.
4667 * mem_cgroup_move_account() checks the page is valid or
4668 * not under LRU exclusion.
4670 if (page->mem_cgroup == mc.from) {
4671 ret = MC_TARGET_PAGE;
4673 target->page = page;
4675 if (!ret || !target)
4678 /* There is a swap entry and a page doesn't exist or isn't charged */
4679 if (ent.val && !ret &&
4680 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4681 ret = MC_TARGET_SWAP;
4688 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4690 * We don't consider swapping or file mapped pages because THP does not
4691 * support them for now.
4692 * Caller should make sure that pmd_trans_huge(pmd) is true.
4694 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4695 unsigned long addr, pmd_t pmd, union mc_target *target)
4697 struct page *page = NULL;
4698 enum mc_target_type ret = MC_TARGET_NONE;
4700 page = pmd_page(pmd);
4701 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4702 if (!(mc.flags & MOVE_ANON))
4704 if (page->mem_cgroup == mc.from) {
4705 ret = MC_TARGET_PAGE;
4708 target->page = page;
4714 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4715 unsigned long addr, pmd_t pmd, union mc_target *target)
4717 return MC_TARGET_NONE;
4721 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4722 unsigned long addr, unsigned long end,
4723 struct mm_walk *walk)
4725 struct vm_area_struct *vma = walk->vma;
4729 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4730 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4731 mc.precharge += HPAGE_PMD_NR;
4736 if (pmd_trans_unstable(pmd))
4738 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4739 for (; addr != end; pte++, addr += PAGE_SIZE)
4740 if (get_mctgt_type(vma, addr, *pte, NULL))
4741 mc.precharge++; /* increment precharge temporarily */
4742 pte_unmap_unlock(pte - 1, ptl);
4748 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4750 unsigned long precharge;
4752 struct mm_walk mem_cgroup_count_precharge_walk = {
4753 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4756 down_read(&mm->mmap_sem);
4757 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4758 up_read(&mm->mmap_sem);
4760 precharge = mc.precharge;
4766 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4768 unsigned long precharge = mem_cgroup_count_precharge(mm);
4770 VM_BUG_ON(mc.moving_task);
4771 mc.moving_task = current;
4772 return mem_cgroup_do_precharge(precharge);
4775 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4776 static void __mem_cgroup_clear_mc(void)
4778 struct mem_cgroup *from = mc.from;
4779 struct mem_cgroup *to = mc.to;
4781 /* we must uncharge all the leftover precharges from mc.to */
4783 cancel_charge(mc.to, mc.precharge);
4787 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4788 * we must uncharge here.
4790 if (mc.moved_charge) {
4791 cancel_charge(mc.from, mc.moved_charge);
4792 mc.moved_charge = 0;
4794 /* we must fixup refcnts and charges */
4795 if (mc.moved_swap) {
4796 /* uncharge swap account from the old cgroup */
4797 if (!mem_cgroup_is_root(mc.from))
4798 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4801 * we charged both to->memory and to->memsw, so we
4802 * should uncharge to->memory.
4804 if (!mem_cgroup_is_root(mc.to))
4805 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4807 css_put_many(&mc.from->css, mc.moved_swap);
4809 /* we've already done css_get(mc.to) */
4812 memcg_oom_recover(from);
4813 memcg_oom_recover(to);
4814 wake_up_all(&mc.waitq);
4817 static void mem_cgroup_clear_mc(void)
4820 * we must clear moving_task before waking up waiters at the end of
4823 mc.moving_task = NULL;
4824 __mem_cgroup_clear_mc();
4825 spin_lock(&mc.lock);
4828 spin_unlock(&mc.lock);
4831 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4833 struct cgroup_subsys_state *css;
4834 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4835 struct mem_cgroup *from;
4836 struct task_struct *leader, *p;
4837 struct mm_struct *mm;
4838 unsigned long move_flags;
4841 /* charge immigration isn't supported on the default hierarchy */
4842 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4846 * Multi-process migrations only happen on the default hierarchy
4847 * where charge immigration is not used. Perform charge
4848 * immigration if @tset contains a leader and whine if there are
4852 cgroup_taskset_for_each_leader(leader, css, tset) {
4855 memcg = mem_cgroup_from_css(css);
4861 * We are now commited to this value whatever it is. Changes in this
4862 * tunable will only affect upcoming migrations, not the current one.
4863 * So we need to save it, and keep it going.
4865 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4869 from = mem_cgroup_from_task(p);
4871 VM_BUG_ON(from == memcg);
4873 mm = get_task_mm(p);
4876 /* We move charges only when we move a owner of the mm */
4877 if (mm->owner == p) {
4880 VM_BUG_ON(mc.precharge);
4881 VM_BUG_ON(mc.moved_charge);
4882 VM_BUG_ON(mc.moved_swap);
4884 spin_lock(&mc.lock);
4887 mc.flags = move_flags;
4888 spin_unlock(&mc.lock);
4889 /* We set mc.moving_task later */
4891 ret = mem_cgroup_precharge_mc(mm);
4893 mem_cgroup_clear_mc();
4899 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4902 mem_cgroup_clear_mc();
4905 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4906 unsigned long addr, unsigned long end,
4907 struct mm_walk *walk)
4910 struct vm_area_struct *vma = walk->vma;
4913 enum mc_target_type target_type;
4914 union mc_target target;
4918 * We don't take compound_lock() here but no race with splitting thp
4920 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4921 * under splitting, which means there's no concurrent thp split,
4922 * - if another thread runs into split_huge_page() just after we
4923 * entered this if-block, the thread must wait for page table lock
4924 * to be unlocked in __split_huge_page_splitting(), where the main
4925 * part of thp split is not executed yet.
4927 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4928 if (mc.precharge < HPAGE_PMD_NR) {
4932 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4933 if (target_type == MC_TARGET_PAGE) {
4935 if (!isolate_lru_page(page)) {
4936 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
4938 mc.precharge -= HPAGE_PMD_NR;
4939 mc.moved_charge += HPAGE_PMD_NR;
4941 putback_lru_page(page);
4949 if (pmd_trans_unstable(pmd))
4952 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4953 for (; addr != end; addr += PAGE_SIZE) {
4954 pte_t ptent = *(pte++);
4960 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4961 case MC_TARGET_PAGE:
4963 if (isolate_lru_page(page))
4965 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
4967 /* we uncharge from mc.from later. */
4970 putback_lru_page(page);
4971 put: /* get_mctgt_type() gets the page */
4974 case MC_TARGET_SWAP:
4976 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4978 /* we fixup refcnts and charges later. */
4986 pte_unmap_unlock(pte - 1, ptl);
4991 * We have consumed all precharges we got in can_attach().
4992 * We try charge one by one, but don't do any additional
4993 * charges to mc.to if we have failed in charge once in attach()
4996 ret = mem_cgroup_do_precharge(1);
5004 static void mem_cgroup_move_charge(struct mm_struct *mm)
5006 struct mm_walk mem_cgroup_move_charge_walk = {
5007 .pmd_entry = mem_cgroup_move_charge_pte_range,
5011 lru_add_drain_all();
5013 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5014 * move_lock while we're moving its pages to another memcg.
5015 * Then wait for already started RCU-only updates to finish.
5017 atomic_inc(&mc.from->moving_account);
5020 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5022 * Someone who are holding the mmap_sem might be waiting in
5023 * waitq. So we cancel all extra charges, wake up all waiters,
5024 * and retry. Because we cancel precharges, we might not be able
5025 * to move enough charges, but moving charge is a best-effort
5026 * feature anyway, so it wouldn't be a big problem.
5028 __mem_cgroup_clear_mc();
5033 * When we have consumed all precharges and failed in doing
5034 * additional charge, the page walk just aborts.
5036 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5037 up_read(&mm->mmap_sem);
5038 atomic_dec(&mc.from->moving_account);
5041 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5043 struct cgroup_subsys_state *css;
5044 struct task_struct *p = cgroup_taskset_first(tset, &css);
5045 struct mm_struct *mm = get_task_mm(p);
5049 mem_cgroup_move_charge(mm);
5053 mem_cgroup_clear_mc();
5055 #else /* !CONFIG_MMU */
5056 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5060 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5063 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5069 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5070 * to verify whether we're attached to the default hierarchy on each mount
5073 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5076 * use_hierarchy is forced on the default hierarchy. cgroup core
5077 * guarantees that @root doesn't have any children, so turning it
5078 * on for the root memcg is enough.
5080 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5081 root_mem_cgroup->use_hierarchy = true;
5083 root_mem_cgroup->use_hierarchy = false;
5086 static u64 memory_current_read(struct cgroup_subsys_state *css,
5089 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5091 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5094 static int memory_low_show(struct seq_file *m, void *v)
5096 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5097 unsigned long low = READ_ONCE(memcg->low);
5099 if (low == PAGE_COUNTER_MAX)
5100 seq_puts(m, "max\n");
5102 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5107 static ssize_t memory_low_write(struct kernfs_open_file *of,
5108 char *buf, size_t nbytes, loff_t off)
5110 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5114 buf = strstrip(buf);
5115 err = page_counter_memparse(buf, "max", &low);
5124 static int memory_high_show(struct seq_file *m, void *v)
5126 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5127 unsigned long high = READ_ONCE(memcg->high);
5129 if (high == PAGE_COUNTER_MAX)
5130 seq_puts(m, "max\n");
5132 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5137 static ssize_t memory_high_write(struct kernfs_open_file *of,
5138 char *buf, size_t nbytes, loff_t off)
5140 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5144 buf = strstrip(buf);
5145 err = page_counter_memparse(buf, "max", &high);
5151 memcg_wb_domain_size_changed(memcg);
5155 static int memory_max_show(struct seq_file *m, void *v)
5157 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5158 unsigned long max = READ_ONCE(memcg->memory.limit);
5160 if (max == PAGE_COUNTER_MAX)
5161 seq_puts(m, "max\n");
5163 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5168 static ssize_t memory_max_write(struct kernfs_open_file *of,
5169 char *buf, size_t nbytes, loff_t off)
5171 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5175 buf = strstrip(buf);
5176 err = page_counter_memparse(buf, "max", &max);
5180 err = mem_cgroup_resize_limit(memcg, max);
5184 memcg_wb_domain_size_changed(memcg);
5188 static int memory_events_show(struct seq_file *m, void *v)
5190 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5192 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5193 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5194 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5195 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5200 static struct cftype memory_files[] = {
5203 .flags = CFTYPE_NOT_ON_ROOT,
5204 .read_u64 = memory_current_read,
5208 .flags = CFTYPE_NOT_ON_ROOT,
5209 .seq_show = memory_low_show,
5210 .write = memory_low_write,
5214 .flags = CFTYPE_NOT_ON_ROOT,
5215 .seq_show = memory_high_show,
5216 .write = memory_high_write,
5220 .flags = CFTYPE_NOT_ON_ROOT,
5221 .seq_show = memory_max_show,
5222 .write = memory_max_write,
5226 .flags = CFTYPE_NOT_ON_ROOT,
5227 .file_offset = offsetof(struct mem_cgroup, events_file),
5228 .seq_show = memory_events_show,
5233 struct cgroup_subsys memory_cgrp_subsys = {
5234 .css_alloc = mem_cgroup_css_alloc,
5235 .css_online = mem_cgroup_css_online,
5236 .css_offline = mem_cgroup_css_offline,
5237 .css_released = mem_cgroup_css_released,
5238 .css_free = mem_cgroup_css_free,
5239 .css_reset = mem_cgroup_css_reset,
5240 .can_attach = mem_cgroup_can_attach,
5241 .cancel_attach = mem_cgroup_cancel_attach,
5242 .attach = mem_cgroup_move_task,
5243 .bind = mem_cgroup_bind,
5244 .dfl_cftypes = memory_files,
5245 .legacy_cftypes = mem_cgroup_legacy_files,
5250 * mem_cgroup_low - check if memory consumption is below the normal range
5251 * @root: the highest ancestor to consider
5252 * @memcg: the memory cgroup to check
5254 * Returns %true if memory consumption of @memcg, and that of all
5255 * configurable ancestors up to @root, is below the normal range.
5257 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5259 if (mem_cgroup_disabled())
5263 * The toplevel group doesn't have a configurable range, so
5264 * it's never low when looked at directly, and it is not
5265 * considered an ancestor when assessing the hierarchy.
5268 if (memcg == root_mem_cgroup)
5271 if (page_counter_read(&memcg->memory) >= memcg->low)
5274 while (memcg != root) {
5275 memcg = parent_mem_cgroup(memcg);
5277 if (memcg == root_mem_cgroup)
5280 if (page_counter_read(&memcg->memory) >= memcg->low)
5287 * mem_cgroup_try_charge - try charging a page
5288 * @page: page to charge
5289 * @mm: mm context of the victim
5290 * @gfp_mask: reclaim mode
5291 * @memcgp: charged memcg return
5293 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5294 * pages according to @gfp_mask if necessary.
5296 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5297 * Otherwise, an error code is returned.
5299 * After page->mapping has been set up, the caller must finalize the
5300 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5301 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5303 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5304 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5306 struct mem_cgroup *memcg = NULL;
5307 unsigned int nr_pages = 1;
5310 if (mem_cgroup_disabled())
5313 if (PageSwapCache(page)) {
5315 * Every swap fault against a single page tries to charge the
5316 * page, bail as early as possible. shmem_unuse() encounters
5317 * already charged pages, too. The USED bit is protected by
5318 * the page lock, which serializes swap cache removal, which
5319 * in turn serializes uncharging.
5321 VM_BUG_ON_PAGE(!PageLocked(page), page);
5322 if (page->mem_cgroup)
5325 if (do_swap_account) {
5326 swp_entry_t ent = { .val = page_private(page), };
5327 unsigned short id = lookup_swap_cgroup_id(ent);
5330 memcg = mem_cgroup_from_id(id);
5331 if (memcg && !css_tryget_online(&memcg->css))
5337 if (PageTransHuge(page)) {
5338 nr_pages <<= compound_order(page);
5339 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5343 memcg = get_mem_cgroup_from_mm(mm);
5345 ret = try_charge(memcg, gfp_mask, nr_pages);
5347 css_put(&memcg->css);
5354 * mem_cgroup_commit_charge - commit a page charge
5355 * @page: page to charge
5356 * @memcg: memcg to charge the page to
5357 * @lrucare: page might be on LRU already
5359 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5360 * after page->mapping has been set up. This must happen atomically
5361 * as part of the page instantiation, i.e. under the page table lock
5362 * for anonymous pages, under the page lock for page and swap cache.
5364 * In addition, the page must not be on the LRU during the commit, to
5365 * prevent racing with task migration. If it might be, use @lrucare.
5367 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5369 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5372 unsigned int nr_pages = 1;
5374 VM_BUG_ON_PAGE(!page->mapping, page);
5375 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5377 if (mem_cgroup_disabled())
5380 * Swap faults will attempt to charge the same page multiple
5381 * times. But reuse_swap_page() might have removed the page
5382 * from swapcache already, so we can't check PageSwapCache().
5387 commit_charge(page, memcg, lrucare);
5389 if (PageTransHuge(page)) {
5390 nr_pages <<= compound_order(page);
5391 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5394 local_irq_disable();
5395 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5396 memcg_check_events(memcg, page);
5399 if (do_swap_account && PageSwapCache(page)) {
5400 swp_entry_t entry = { .val = page_private(page) };
5402 * The swap entry might not get freed for a long time,
5403 * let's not wait for it. The page already received a
5404 * memory+swap charge, drop the swap entry duplicate.
5406 mem_cgroup_uncharge_swap(entry);
5411 * mem_cgroup_cancel_charge - cancel a page charge
5412 * @page: page to charge
5413 * @memcg: memcg to charge the page to
5415 * Cancel a charge transaction started by mem_cgroup_try_charge().
5417 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5419 unsigned int nr_pages = 1;
5421 if (mem_cgroup_disabled())
5424 * Swap faults will attempt to charge the same page multiple
5425 * times. But reuse_swap_page() might have removed the page
5426 * from swapcache already, so we can't check PageSwapCache().
5431 if (PageTransHuge(page)) {
5432 nr_pages <<= compound_order(page);
5433 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5436 cancel_charge(memcg, nr_pages);
5439 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5440 unsigned long nr_anon, unsigned long nr_file,
5441 unsigned long nr_huge, struct page *dummy_page)
5443 unsigned long nr_pages = nr_anon + nr_file;
5444 unsigned long flags;
5446 if (!mem_cgroup_is_root(memcg)) {
5447 page_counter_uncharge(&memcg->memory, nr_pages);
5448 if (do_swap_account)
5449 page_counter_uncharge(&memcg->memsw, nr_pages);
5450 memcg_oom_recover(memcg);
5453 local_irq_save(flags);
5454 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5455 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5456 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5457 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5458 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5459 memcg_check_events(memcg, dummy_page);
5460 local_irq_restore(flags);
5462 if (!mem_cgroup_is_root(memcg))
5463 css_put_many(&memcg->css, nr_pages);
5466 static void uncharge_list(struct list_head *page_list)
5468 struct mem_cgroup *memcg = NULL;
5469 unsigned long nr_anon = 0;
5470 unsigned long nr_file = 0;
5471 unsigned long nr_huge = 0;
5472 unsigned long pgpgout = 0;
5473 struct list_head *next;
5476 next = page_list->next;
5478 unsigned int nr_pages = 1;
5480 page = list_entry(next, struct page, lru);
5481 next = page->lru.next;
5483 VM_BUG_ON_PAGE(PageLRU(page), page);
5484 VM_BUG_ON_PAGE(page_count(page), page);
5486 if (!page->mem_cgroup)
5490 * Nobody should be changing or seriously looking at
5491 * page->mem_cgroup at this point, we have fully
5492 * exclusive access to the page.
5495 if (memcg != page->mem_cgroup) {
5497 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5499 pgpgout = nr_anon = nr_file = nr_huge = 0;
5501 memcg = page->mem_cgroup;
5504 if (PageTransHuge(page)) {
5505 nr_pages <<= compound_order(page);
5506 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5507 nr_huge += nr_pages;
5511 nr_anon += nr_pages;
5513 nr_file += nr_pages;
5515 page->mem_cgroup = NULL;
5518 } while (next != page_list);
5521 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5526 * mem_cgroup_uncharge - uncharge a page
5527 * @page: page to uncharge
5529 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5530 * mem_cgroup_commit_charge().
5532 void mem_cgroup_uncharge(struct page *page)
5534 if (mem_cgroup_disabled())
5537 /* Don't touch page->lru of any random page, pre-check: */
5538 if (!page->mem_cgroup)
5541 INIT_LIST_HEAD(&page->lru);
5542 uncharge_list(&page->lru);
5546 * mem_cgroup_uncharge_list - uncharge a list of page
5547 * @page_list: list of pages to uncharge
5549 * Uncharge a list of pages previously charged with
5550 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5552 void mem_cgroup_uncharge_list(struct list_head *page_list)
5554 if (mem_cgroup_disabled())
5557 if (!list_empty(page_list))
5558 uncharge_list(page_list);
5562 * mem_cgroup_replace_page - migrate a charge to another page
5563 * @oldpage: currently charged page
5564 * @newpage: page to transfer the charge to
5566 * Migrate the charge from @oldpage to @newpage.
5568 * Both pages must be locked, @newpage->mapping must be set up.
5569 * Either or both pages might be on the LRU already.
5571 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5573 struct mem_cgroup *memcg;
5576 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5577 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5578 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5579 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5582 if (mem_cgroup_disabled())
5585 /* Page cache replacement: new page already charged? */
5586 if (newpage->mem_cgroup)
5589 /* Swapcache readahead pages can get replaced before being charged */
5590 memcg = oldpage->mem_cgroup;
5594 lock_page_lru(oldpage, &isolated);
5595 oldpage->mem_cgroup = NULL;
5596 unlock_page_lru(oldpage, isolated);
5598 commit_charge(newpage, memcg, true);
5602 * subsys_initcall() for memory controller.
5604 * Some parts like hotcpu_notifier() have to be initialized from this context
5605 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5606 * everything that doesn't depend on a specific mem_cgroup structure should
5607 * be initialized from here.
5609 static int __init mem_cgroup_init(void)
5613 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5615 for_each_possible_cpu(cpu)
5616 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5619 for_each_node(node) {
5620 struct mem_cgroup_tree_per_node *rtpn;
5623 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5624 node_online(node) ? node : NUMA_NO_NODE);
5626 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5627 struct mem_cgroup_tree_per_zone *rtpz;
5629 rtpz = &rtpn->rb_tree_per_zone[zone];
5630 rtpz->rb_root = RB_ROOT;
5631 spin_lock_init(&rtpz->lock);
5633 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5638 subsys_initcall(mem_cgroup_init);
5640 #ifdef CONFIG_MEMCG_SWAP
5642 * mem_cgroup_swapout - transfer a memsw charge to swap
5643 * @page: page whose memsw charge to transfer
5644 * @entry: swap entry to move the charge to
5646 * Transfer the memsw charge of @page to @entry.
5648 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5650 struct mem_cgroup *memcg;
5651 unsigned short oldid;
5653 VM_BUG_ON_PAGE(PageLRU(page), page);
5654 VM_BUG_ON_PAGE(page_count(page), page);
5656 if (!do_swap_account)
5659 memcg = page->mem_cgroup;
5661 /* Readahead page, never charged */
5665 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5666 VM_BUG_ON_PAGE(oldid, page);
5667 mem_cgroup_swap_statistics(memcg, true);
5669 page->mem_cgroup = NULL;
5671 if (!mem_cgroup_is_root(memcg))
5672 page_counter_uncharge(&memcg->memory, 1);
5675 * Interrupts should be disabled here because the caller holds the
5676 * mapping->tree_lock lock which is taken with interrupts-off. It is
5677 * important here to have the interrupts disabled because it is the
5678 * only synchronisation we have for udpating the per-CPU variables.
5680 VM_BUG_ON(!irqs_disabled());
5681 mem_cgroup_charge_statistics(memcg, page, -1);
5682 memcg_check_events(memcg, page);
5686 * mem_cgroup_uncharge_swap - uncharge a swap entry
5687 * @entry: swap entry to uncharge
5689 * Drop the memsw charge associated with @entry.
5691 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5693 struct mem_cgroup *memcg;
5696 if (!do_swap_account)
5699 id = swap_cgroup_record(entry, 0);
5701 memcg = mem_cgroup_from_id(id);
5703 if (!mem_cgroup_is_root(memcg))
5704 page_counter_uncharge(&memcg->memsw, 1);
5705 mem_cgroup_swap_statistics(memcg, false);
5706 css_put(&memcg->css);
5711 /* for remember boot option*/
5712 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5713 static int really_do_swap_account __initdata = 1;
5715 static int really_do_swap_account __initdata;
5718 static int __init enable_swap_account(char *s)
5720 if (!strcmp(s, "1"))
5721 really_do_swap_account = 1;
5722 else if (!strcmp(s, "0"))
5723 really_do_swap_account = 0;
5726 __setup("swapaccount=", enable_swap_account);
5728 static struct cftype memsw_cgroup_files[] = {
5730 .name = "memsw.usage_in_bytes",
5731 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5732 .read_u64 = mem_cgroup_read_u64,
5735 .name = "memsw.max_usage_in_bytes",
5736 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5737 .write = mem_cgroup_reset,
5738 .read_u64 = mem_cgroup_read_u64,
5741 .name = "memsw.limit_in_bytes",
5742 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5743 .write = mem_cgroup_write,
5744 .read_u64 = mem_cgroup_read_u64,
5747 .name = "memsw.failcnt",
5748 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5749 .write = mem_cgroup_reset,
5750 .read_u64 = mem_cgroup_read_u64,
5752 { }, /* terminate */
5755 static int __init mem_cgroup_swap_init(void)
5757 if (!mem_cgroup_disabled() && really_do_swap_account) {
5758 do_swap_account = 1;
5759 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5760 memsw_cgroup_files));
5764 subsys_initcall(mem_cgroup_swap_init);
5766 #endif /* CONFIG_MEMCG_SWAP */