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>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 struct mem_cgroup *root_mem_cgroup __read_mostly;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 static const char * const mem_cgroup_stat_names[] = {
111 static const char * const mem_cgroup_events_names[] = {
118 static const char * const mem_cgroup_lru_names[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_zone {
136 struct rb_root rb_root;
140 struct mem_cgroup_tree_per_node {
141 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
144 struct mem_cgroup_tree {
145 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
148 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
151 struct mem_cgroup_eventfd_list {
152 struct list_head list;
153 struct eventfd_ctx *eventfd;
157 * cgroup_event represents events which userspace want to receive.
159 struct mem_cgroup_event {
161 * memcg which the event belongs to.
163 struct mem_cgroup *memcg;
165 * eventfd to signal userspace about the event.
167 struct eventfd_ctx *eventfd;
169 * Each of these stored in a list by the cgroup.
171 struct list_head list;
173 * register_event() callback will be used to add new userspace
174 * waiter for changes related to this event. Use eventfd_signal()
175 * on eventfd to send notification to userspace.
177 int (*register_event)(struct mem_cgroup *memcg,
178 struct eventfd_ctx *eventfd, const char *args);
180 * unregister_event() callback will be called when userspace closes
181 * the eventfd or on cgroup removing. This callback must be set,
182 * if you want provide notification functionality.
184 void (*unregister_event)(struct mem_cgroup *memcg,
185 struct eventfd_ctx *eventfd);
187 * All fields below needed to unregister event when
188 * userspace closes eventfd.
191 wait_queue_head_t *wqh;
193 struct work_struct remove;
196 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
197 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
199 /* Stuffs for move charges at task migration. */
201 * Types of charges to be moved.
203 #define MOVE_ANON 0x1U
204 #define MOVE_FILE 0x2U
205 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
207 /* "mc" and its members are protected by cgroup_mutex */
208 static struct move_charge_struct {
209 spinlock_t lock; /* for from, to */
210 struct mm_struct *mm;
211 struct mem_cgroup *from;
212 struct mem_cgroup *to;
214 unsigned long precharge;
215 unsigned long moved_charge;
216 unsigned long moved_swap;
217 struct task_struct *moving_task; /* a task moving charges */
218 wait_queue_head_t waitq; /* a waitq for other context */
220 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
221 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
225 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
226 * limit reclaim to prevent infinite loops, if they ever occur.
228 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
229 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
232 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
233 MEM_CGROUP_CHARGE_TYPE_ANON,
234 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
235 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
239 /* for encoding cft->private value on file */
248 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
249 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
250 #define MEMFILE_ATTR(val) ((val) & 0xffff)
251 /* Used for OOM nofiier */
252 #define OOM_CONTROL (0)
254 /* Some nice accessors for the vmpressure. */
255 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
258 memcg = root_mem_cgroup;
259 return &memcg->vmpressure;
262 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
264 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
267 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
269 return (memcg == root_mem_cgroup);
274 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
275 * The main reason for not using cgroup id for this:
276 * this works better in sparse environments, where we have a lot of memcgs,
277 * but only a few kmem-limited. Or also, if we have, for instance, 200
278 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
279 * 200 entry array for that.
281 * The current size of the caches array is stored in memcg_nr_cache_ids. It
282 * will double each time we have to increase it.
284 static DEFINE_IDA(memcg_cache_ida);
285 int memcg_nr_cache_ids;
287 /* Protects memcg_nr_cache_ids */
288 static DECLARE_RWSEM(memcg_cache_ids_sem);
290 void memcg_get_cache_ids(void)
292 down_read(&memcg_cache_ids_sem);
295 void memcg_put_cache_ids(void)
297 up_read(&memcg_cache_ids_sem);
301 * MIN_SIZE is different than 1, because we would like to avoid going through
302 * the alloc/free process all the time. In a small machine, 4 kmem-limited
303 * cgroups is a reasonable guess. In the future, it could be a parameter or
304 * tunable, but that is strictly not necessary.
306 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
307 * this constant directly from cgroup, but it is understandable that this is
308 * better kept as an internal representation in cgroup.c. In any case, the
309 * cgrp_id space is not getting any smaller, and we don't have to necessarily
310 * increase ours as well if it increases.
312 #define MEMCG_CACHES_MIN_SIZE 4
313 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
316 * A lot of the calls to the cache allocation functions are expected to be
317 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
318 * conditional to this static branch, we'll have to allow modules that does
319 * kmem_cache_alloc and the such to see this symbol as well
321 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
322 EXPORT_SYMBOL(memcg_kmem_enabled_key);
324 #endif /* !CONFIG_SLOB */
327 * mem_cgroup_css_from_page - css of the memcg associated with a page
328 * @page: page of interest
330 * If memcg is bound to the default hierarchy, css of the memcg associated
331 * with @page is returned. The returned css remains associated with @page
332 * until it is released.
334 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
337 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
339 struct mem_cgroup *memcg;
341 memcg = page->mem_cgroup;
343 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
344 memcg = root_mem_cgroup;
350 * page_cgroup_ino - return inode number of the memcg a page is charged to
353 * Look up the closest online ancestor of the memory cgroup @page is charged to
354 * and return its inode number or 0 if @page is not charged to any cgroup. It
355 * is safe to call this function without holding a reference to @page.
357 * Note, this function is inherently racy, because there is nothing to prevent
358 * the cgroup inode from getting torn down and potentially reallocated a moment
359 * after page_cgroup_ino() returns, so it only should be used by callers that
360 * do not care (such as procfs interfaces).
362 ino_t page_cgroup_ino(struct page *page)
364 struct mem_cgroup *memcg;
365 unsigned long ino = 0;
368 memcg = READ_ONCE(page->mem_cgroup);
369 while (memcg && !(memcg->css.flags & CSS_ONLINE))
370 memcg = parent_mem_cgroup(memcg);
372 ino = cgroup_ino(memcg->css.cgroup);
377 static struct mem_cgroup_per_zone *
378 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
380 int nid = page_to_nid(page);
381 int zid = page_zonenum(page);
383 return &memcg->nodeinfo[nid]->zoneinfo[zid];
386 static struct mem_cgroup_tree_per_zone *
387 soft_limit_tree_node_zone(int nid, int zid)
389 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
392 static struct mem_cgroup_tree_per_zone *
393 soft_limit_tree_from_page(struct page *page)
395 int nid = page_to_nid(page);
396 int zid = page_zonenum(page);
398 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
401 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
402 struct mem_cgroup_tree_per_zone *mctz,
403 unsigned long new_usage_in_excess)
405 struct rb_node **p = &mctz->rb_root.rb_node;
406 struct rb_node *parent = NULL;
407 struct mem_cgroup_per_zone *mz_node;
412 mz->usage_in_excess = new_usage_in_excess;
413 if (!mz->usage_in_excess)
417 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
419 if (mz->usage_in_excess < mz_node->usage_in_excess)
422 * We can't avoid mem cgroups that are over their soft
423 * limit by the same amount
425 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
428 rb_link_node(&mz->tree_node, parent, p);
429 rb_insert_color(&mz->tree_node, &mctz->rb_root);
433 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
434 struct mem_cgroup_tree_per_zone *mctz)
438 rb_erase(&mz->tree_node, &mctz->rb_root);
442 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
443 struct mem_cgroup_tree_per_zone *mctz)
447 spin_lock_irqsave(&mctz->lock, flags);
448 __mem_cgroup_remove_exceeded(mz, mctz);
449 spin_unlock_irqrestore(&mctz->lock, flags);
452 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
454 unsigned long nr_pages = page_counter_read(&memcg->memory);
455 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
456 unsigned long excess = 0;
458 if (nr_pages > soft_limit)
459 excess = nr_pages - soft_limit;
464 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
466 unsigned long excess;
467 struct mem_cgroup_per_zone *mz;
468 struct mem_cgroup_tree_per_zone *mctz;
470 mctz = soft_limit_tree_from_page(page);
472 * Necessary to update all ancestors when hierarchy is used.
473 * because their event counter is not touched.
475 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
476 mz = mem_cgroup_page_zoneinfo(memcg, page);
477 excess = soft_limit_excess(memcg);
479 * We have to update the tree if mz is on RB-tree or
480 * mem is over its softlimit.
482 if (excess || mz->on_tree) {
485 spin_lock_irqsave(&mctz->lock, flags);
486 /* if on-tree, remove it */
488 __mem_cgroup_remove_exceeded(mz, mctz);
490 * Insert again. mz->usage_in_excess will be updated.
491 * If excess is 0, no tree ops.
493 __mem_cgroup_insert_exceeded(mz, mctz, excess);
494 spin_unlock_irqrestore(&mctz->lock, flags);
499 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
501 struct mem_cgroup_tree_per_zone *mctz;
502 struct mem_cgroup_per_zone *mz;
506 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
507 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
508 mctz = soft_limit_tree_node_zone(nid, zid);
509 mem_cgroup_remove_exceeded(mz, mctz);
514 static struct mem_cgroup_per_zone *
515 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
517 struct rb_node *rightmost = NULL;
518 struct mem_cgroup_per_zone *mz;
522 rightmost = rb_last(&mctz->rb_root);
524 goto done; /* Nothing to reclaim from */
526 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
528 * Remove the node now but someone else can add it back,
529 * we will to add it back at the end of reclaim to its correct
530 * position in the tree.
532 __mem_cgroup_remove_exceeded(mz, mctz);
533 if (!soft_limit_excess(mz->memcg) ||
534 !css_tryget_online(&mz->memcg->css))
540 static struct mem_cgroup_per_zone *
541 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
543 struct mem_cgroup_per_zone *mz;
545 spin_lock_irq(&mctz->lock);
546 mz = __mem_cgroup_largest_soft_limit_node(mctz);
547 spin_unlock_irq(&mctz->lock);
552 * Return page count for single (non recursive) @memcg.
554 * Implementation Note: reading percpu statistics for memcg.
556 * Both of vmstat[] and percpu_counter has threshold and do periodic
557 * synchronization to implement "quick" read. There are trade-off between
558 * reading cost and precision of value. Then, we may have a chance to implement
559 * a periodic synchronization of counter in memcg's counter.
561 * But this _read() function is used for user interface now. The user accounts
562 * memory usage by memory cgroup and he _always_ requires exact value because
563 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
564 * have to visit all online cpus and make sum. So, for now, unnecessary
565 * synchronization is not implemented. (just implemented for cpu hotplug)
567 * If there are kernel internal actions which can make use of some not-exact
568 * value, and reading all cpu value can be performance bottleneck in some
569 * common workload, threshold and synchronization as vmstat[] should be
573 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
578 /* Per-cpu values can be negative, use a signed accumulator */
579 for_each_possible_cpu(cpu)
580 val += per_cpu(memcg->stat->count[idx], cpu);
582 * Summing races with updates, so val may be negative. Avoid exposing
583 * transient negative values.
590 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
591 enum mem_cgroup_events_index idx)
593 unsigned long val = 0;
596 for_each_possible_cpu(cpu)
597 val += per_cpu(memcg->stat->events[idx], cpu);
601 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
603 bool compound, int nr_pages)
606 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
607 * counted as CACHE even if it's on ANON LRU.
610 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
613 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
617 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
618 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
622 /* pagein of a big page is an event. So, ignore page size */
624 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
626 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
627 nr_pages = -nr_pages; /* for event */
630 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
633 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
634 int nid, unsigned int lru_mask)
636 unsigned long nr = 0;
639 VM_BUG_ON((unsigned)nid >= nr_node_ids);
641 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
642 struct mem_cgroup_per_zone *mz;
646 if (!(BIT(lru) & lru_mask))
648 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
649 nr += mz->lru_size[lru];
655 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
656 unsigned int lru_mask)
658 unsigned long nr = 0;
661 for_each_node_state(nid, N_MEMORY)
662 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
666 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
667 enum mem_cgroup_events_target target)
669 unsigned long val, next;
671 val = __this_cpu_read(memcg->stat->nr_page_events);
672 next = __this_cpu_read(memcg->stat->targets[target]);
673 /* from time_after() in jiffies.h */
674 if ((long)next - (long)val < 0) {
676 case MEM_CGROUP_TARGET_THRESH:
677 next = val + THRESHOLDS_EVENTS_TARGET;
679 case MEM_CGROUP_TARGET_SOFTLIMIT:
680 next = val + SOFTLIMIT_EVENTS_TARGET;
682 case MEM_CGROUP_TARGET_NUMAINFO:
683 next = val + NUMAINFO_EVENTS_TARGET;
688 __this_cpu_write(memcg->stat->targets[target], next);
695 * Check events in order.
698 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
700 /* threshold event is triggered in finer grain than soft limit */
701 if (unlikely(mem_cgroup_event_ratelimit(memcg,
702 MEM_CGROUP_TARGET_THRESH))) {
704 bool do_numainfo __maybe_unused;
706 do_softlimit = mem_cgroup_event_ratelimit(memcg,
707 MEM_CGROUP_TARGET_SOFTLIMIT);
709 do_numainfo = mem_cgroup_event_ratelimit(memcg,
710 MEM_CGROUP_TARGET_NUMAINFO);
712 mem_cgroup_threshold(memcg);
713 if (unlikely(do_softlimit))
714 mem_cgroup_update_tree(memcg, page);
716 if (unlikely(do_numainfo))
717 atomic_inc(&memcg->numainfo_events);
722 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
725 * mm_update_next_owner() may clear mm->owner to NULL
726 * if it races with swapoff, page migration, etc.
727 * So this can be called with p == NULL.
732 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
734 EXPORT_SYMBOL(mem_cgroup_from_task);
736 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
738 struct mem_cgroup *memcg = NULL;
743 * Page cache insertions can happen withou an
744 * actual mm context, e.g. during disk probing
745 * on boot, loopback IO, acct() writes etc.
748 memcg = root_mem_cgroup;
750 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
751 if (unlikely(!memcg))
752 memcg = root_mem_cgroup;
754 } while (!css_tryget_online(&memcg->css));
760 * mem_cgroup_iter - iterate over memory cgroup hierarchy
761 * @root: hierarchy root
762 * @prev: previously returned memcg, NULL on first invocation
763 * @reclaim: cookie for shared reclaim walks, NULL for full walks
765 * Returns references to children of the hierarchy below @root, or
766 * @root itself, or %NULL after a full round-trip.
768 * Caller must pass the return value in @prev on subsequent
769 * invocations for reference counting, or use mem_cgroup_iter_break()
770 * to cancel a hierarchy walk before the round-trip is complete.
772 * Reclaimers can specify a zone and a priority level in @reclaim to
773 * divide up the memcgs in the hierarchy among all concurrent
774 * reclaimers operating on the same zone and priority.
776 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
777 struct mem_cgroup *prev,
778 struct mem_cgroup_reclaim_cookie *reclaim)
780 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
781 struct cgroup_subsys_state *css = NULL;
782 struct mem_cgroup *memcg = NULL;
783 struct mem_cgroup *pos = NULL;
785 if (mem_cgroup_disabled())
789 root = root_mem_cgroup;
791 if (prev && !reclaim)
794 if (!root->use_hierarchy && root != root_mem_cgroup) {
803 struct mem_cgroup_per_zone *mz;
805 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
806 iter = &mz->iter[reclaim->priority];
808 if (prev && reclaim->generation != iter->generation)
812 pos = READ_ONCE(iter->position);
813 if (!pos || css_tryget(&pos->css))
816 * css reference reached zero, so iter->position will
817 * be cleared by ->css_released. However, we should not
818 * rely on this happening soon, because ->css_released
819 * is called from a work queue, and by busy-waiting we
820 * might block it. So we clear iter->position right
823 (void)cmpxchg(&iter->position, pos, NULL);
831 css = css_next_descendant_pre(css, &root->css);
834 * Reclaimers share the hierarchy walk, and a
835 * new one might jump in right at the end of
836 * the hierarchy - make sure they see at least
837 * one group and restart from the beginning.
845 * Verify the css and acquire a reference. The root
846 * is provided by the caller, so we know it's alive
847 * and kicking, and don't take an extra reference.
849 memcg = mem_cgroup_from_css(css);
851 if (css == &root->css)
862 * The position could have already been updated by a competing
863 * thread, so check that the value hasn't changed since we read
864 * it to avoid reclaiming from the same cgroup twice.
866 (void)cmpxchg(&iter->position, pos, memcg);
874 reclaim->generation = iter->generation;
880 if (prev && prev != root)
887 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
888 * @root: hierarchy root
889 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
891 void mem_cgroup_iter_break(struct mem_cgroup *root,
892 struct mem_cgroup *prev)
895 root = root_mem_cgroup;
896 if (prev && prev != root)
900 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
902 struct mem_cgroup *memcg = dead_memcg;
903 struct mem_cgroup_reclaim_iter *iter;
904 struct mem_cgroup_per_zone *mz;
908 while ((memcg = parent_mem_cgroup(memcg))) {
910 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
911 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
912 for (i = 0; i <= DEF_PRIORITY; i++) {
914 cmpxchg(&iter->position,
923 * Iteration constructs for visiting all cgroups (under a tree). If
924 * loops are exited prematurely (break), mem_cgroup_iter_break() must
925 * be used for reference counting.
927 #define for_each_mem_cgroup_tree(iter, root) \
928 for (iter = mem_cgroup_iter(root, NULL, NULL); \
930 iter = mem_cgroup_iter(root, iter, NULL))
932 #define for_each_mem_cgroup(iter) \
933 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
935 iter = mem_cgroup_iter(NULL, iter, NULL))
938 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
940 * @zone: zone of the page
942 * This function is only safe when following the LRU page isolation
943 * and putback protocol: the LRU lock must be held, and the page must
944 * either be PageLRU() or the caller must have isolated/allocated it.
946 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
948 struct mem_cgroup_per_zone *mz;
949 struct mem_cgroup *memcg;
950 struct lruvec *lruvec;
952 if (mem_cgroup_disabled()) {
953 lruvec = &zone->lruvec;
957 memcg = page->mem_cgroup;
959 * Swapcache readahead pages are added to the LRU - and
960 * possibly migrated - before they are charged.
963 memcg = root_mem_cgroup;
965 mz = mem_cgroup_page_zoneinfo(memcg, page);
966 lruvec = &mz->lruvec;
969 * Since a node can be onlined after the mem_cgroup was created,
970 * we have to be prepared to initialize lruvec->zone here;
971 * and if offlined then reonlined, we need to reinitialize it.
973 if (unlikely(lruvec->zone != zone))
979 * mem_cgroup_update_lru_size - account for adding or removing an lru page
980 * @lruvec: mem_cgroup per zone lru vector
981 * @lru: index of lru list the page is sitting on
982 * @nr_pages: positive when adding or negative when removing
984 * This function must be called under lru_lock, just before a page is added
985 * to or just after a page is removed from an lru list (that ordering being
986 * so as to allow it to check that lru_size 0 is consistent with list_empty).
988 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
991 struct mem_cgroup_per_zone *mz;
992 unsigned long *lru_size;
996 __update_lru_size(lruvec, lru, nr_pages);
998 if (mem_cgroup_disabled())
1001 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1002 lru_size = mz->lru_size + lru;
1003 empty = list_empty(lruvec->lists + lru);
1006 *lru_size += nr_pages;
1009 if (WARN_ONCE(size < 0 || empty != !size,
1010 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
1011 __func__, lruvec, lru, nr_pages, size, empty ? "" : "not ")) {
1017 *lru_size += nr_pages;
1020 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1022 struct mem_cgroup *task_memcg;
1023 struct task_struct *p;
1026 p = find_lock_task_mm(task);
1028 task_memcg = get_mem_cgroup_from_mm(p->mm);
1032 * All threads may have already detached their mm's, but the oom
1033 * killer still needs to detect if they have already been oom
1034 * killed to prevent needlessly killing additional tasks.
1037 task_memcg = mem_cgroup_from_task(task);
1038 css_get(&task_memcg->css);
1041 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1042 css_put(&task_memcg->css);
1047 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1048 * @memcg: the memory cgroup
1050 * Returns the maximum amount of memory @mem can be charged with, in
1053 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1055 unsigned long margin = 0;
1056 unsigned long count;
1057 unsigned long limit;
1059 count = page_counter_read(&memcg->memory);
1060 limit = READ_ONCE(memcg->memory.limit);
1062 margin = limit - count;
1064 if (do_memsw_account()) {
1065 count = page_counter_read(&memcg->memsw);
1066 limit = READ_ONCE(memcg->memsw.limit);
1068 margin = min(margin, limit - count);
1077 * A routine for checking "mem" is under move_account() or not.
1079 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1080 * moving cgroups. This is for waiting at high-memory pressure
1083 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1085 struct mem_cgroup *from;
1086 struct mem_cgroup *to;
1089 * Unlike task_move routines, we access mc.to, mc.from not under
1090 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1092 spin_lock(&mc.lock);
1098 ret = mem_cgroup_is_descendant(from, memcg) ||
1099 mem_cgroup_is_descendant(to, memcg);
1101 spin_unlock(&mc.lock);
1105 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1107 if (mc.moving_task && current != mc.moving_task) {
1108 if (mem_cgroup_under_move(memcg)) {
1110 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1111 /* moving charge context might have finished. */
1114 finish_wait(&mc.waitq, &wait);
1121 #define K(x) ((x) << (PAGE_SHIFT-10))
1123 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1124 * @memcg: The memory cgroup that went over limit
1125 * @p: Task that is going to be killed
1127 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1130 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1132 struct mem_cgroup *iter;
1138 pr_info("Task in ");
1139 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1140 pr_cont(" killed as a result of limit of ");
1142 pr_info("Memory limit reached of cgroup ");
1145 pr_cont_cgroup_path(memcg->css.cgroup);
1150 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1151 K((u64)page_counter_read(&memcg->memory)),
1152 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1153 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1154 K((u64)page_counter_read(&memcg->memsw)),
1155 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1156 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1157 K((u64)page_counter_read(&memcg->kmem)),
1158 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1160 for_each_mem_cgroup_tree(iter, memcg) {
1161 pr_info("Memory cgroup stats for ");
1162 pr_cont_cgroup_path(iter->css.cgroup);
1165 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1166 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1168 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1169 K(mem_cgroup_read_stat(iter, i)));
1172 for (i = 0; i < NR_LRU_LISTS; i++)
1173 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1174 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1181 * This function returns the number of memcg under hierarchy tree. Returns
1182 * 1(self count) if no children.
1184 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1187 struct mem_cgroup *iter;
1189 for_each_mem_cgroup_tree(iter, memcg)
1195 * Return the memory (and swap, if configured) limit for a memcg.
1197 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1199 unsigned long limit;
1201 limit = memcg->memory.limit;
1202 if (mem_cgroup_swappiness(memcg)) {
1203 unsigned long memsw_limit;
1204 unsigned long swap_limit;
1206 memsw_limit = memcg->memsw.limit;
1207 swap_limit = memcg->swap.limit;
1208 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1209 limit = min(limit + swap_limit, memsw_limit);
1214 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1217 struct oom_control oc = {
1221 .gfp_mask = gfp_mask,
1224 struct mem_cgroup *iter;
1225 unsigned long chosen_points = 0;
1226 unsigned long totalpages;
1227 unsigned int points = 0;
1228 struct task_struct *chosen = NULL;
1230 mutex_lock(&oom_lock);
1233 * If current has a pending SIGKILL or is exiting, then automatically
1234 * select it. The goal is to allow it to allocate so that it may
1235 * quickly exit and free its memory.
1237 if (task_will_free_mem(current)) {
1238 mark_oom_victim(current);
1239 wake_oom_reaper(current);
1243 check_panic_on_oom(&oc, CONSTRAINT_MEMCG);
1244 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1245 for_each_mem_cgroup_tree(iter, memcg) {
1246 struct css_task_iter it;
1247 struct task_struct *task;
1249 css_task_iter_start(&iter->css, &it);
1250 while ((task = css_task_iter_next(&it))) {
1251 switch (oom_scan_process_thread(&oc, task)) {
1252 case OOM_SCAN_SELECT:
1254 put_task_struct(chosen);
1256 chosen_points = ULONG_MAX;
1257 get_task_struct(chosen);
1259 case OOM_SCAN_CONTINUE:
1261 case OOM_SCAN_ABORT:
1262 css_task_iter_end(&it);
1263 mem_cgroup_iter_break(memcg, iter);
1265 put_task_struct(chosen);
1266 /* Set a dummy value to return "true". */
1267 chosen = (void *) 1;
1272 points = oom_badness(task, memcg, NULL, totalpages);
1273 if (!points || points < chosen_points)
1275 /* Prefer thread group leaders for display purposes */
1276 if (points == chosen_points &&
1277 thread_group_leader(chosen))
1281 put_task_struct(chosen);
1283 chosen_points = points;
1284 get_task_struct(chosen);
1286 css_task_iter_end(&it);
1290 points = chosen_points * 1000 / totalpages;
1291 oom_kill_process(&oc, chosen, points, totalpages,
1292 "Memory cgroup out of memory");
1295 mutex_unlock(&oom_lock);
1299 #if MAX_NUMNODES > 1
1302 * test_mem_cgroup_node_reclaimable
1303 * @memcg: the target memcg
1304 * @nid: the node ID to be checked.
1305 * @noswap : specify true here if the user wants flle only information.
1307 * This function returns whether the specified memcg contains any
1308 * reclaimable pages on a node. Returns true if there are any reclaimable
1309 * pages in the node.
1311 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1312 int nid, bool noswap)
1314 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1316 if (noswap || !total_swap_pages)
1318 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1325 * Always updating the nodemask is not very good - even if we have an empty
1326 * list or the wrong list here, we can start from some node and traverse all
1327 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1330 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1334 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1335 * pagein/pageout changes since the last update.
1337 if (!atomic_read(&memcg->numainfo_events))
1339 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1342 /* make a nodemask where this memcg uses memory from */
1343 memcg->scan_nodes = node_states[N_MEMORY];
1345 for_each_node_mask(nid, node_states[N_MEMORY]) {
1347 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1348 node_clear(nid, memcg->scan_nodes);
1351 atomic_set(&memcg->numainfo_events, 0);
1352 atomic_set(&memcg->numainfo_updating, 0);
1356 * Selecting a node where we start reclaim from. Because what we need is just
1357 * reducing usage counter, start from anywhere is O,K. Considering
1358 * memory reclaim from current node, there are pros. and cons.
1360 * Freeing memory from current node means freeing memory from a node which
1361 * we'll use or we've used. So, it may make LRU bad. And if several threads
1362 * hit limits, it will see a contention on a node. But freeing from remote
1363 * node means more costs for memory reclaim because of memory latency.
1365 * Now, we use round-robin. Better algorithm is welcomed.
1367 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1371 mem_cgroup_may_update_nodemask(memcg);
1372 node = memcg->last_scanned_node;
1374 node = next_node_in(node, memcg->scan_nodes);
1376 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1377 * last time it really checked all the LRUs due to rate limiting.
1378 * Fallback to the current node in that case for simplicity.
1380 if (unlikely(node == MAX_NUMNODES))
1381 node = numa_node_id();
1383 memcg->last_scanned_node = node;
1387 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1393 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1396 unsigned long *total_scanned)
1398 struct mem_cgroup *victim = NULL;
1401 unsigned long excess;
1402 unsigned long nr_scanned;
1403 struct mem_cgroup_reclaim_cookie reclaim = {
1408 excess = soft_limit_excess(root_memcg);
1411 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1416 * If we have not been able to reclaim
1417 * anything, it might because there are
1418 * no reclaimable pages under this hierarchy
1423 * We want to do more targeted reclaim.
1424 * excess >> 2 is not to excessive so as to
1425 * reclaim too much, nor too less that we keep
1426 * coming back to reclaim from this cgroup
1428 if (total >= (excess >> 2) ||
1429 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1434 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1436 *total_scanned += nr_scanned;
1437 if (!soft_limit_excess(root_memcg))
1440 mem_cgroup_iter_break(root_memcg, victim);
1444 #ifdef CONFIG_LOCKDEP
1445 static struct lockdep_map memcg_oom_lock_dep_map = {
1446 .name = "memcg_oom_lock",
1450 static DEFINE_SPINLOCK(memcg_oom_lock);
1453 * Check OOM-Killer is already running under our hierarchy.
1454 * If someone is running, return false.
1456 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1458 struct mem_cgroup *iter, *failed = NULL;
1460 spin_lock(&memcg_oom_lock);
1462 for_each_mem_cgroup_tree(iter, memcg) {
1463 if (iter->oom_lock) {
1465 * this subtree of our hierarchy is already locked
1466 * so we cannot give a lock.
1469 mem_cgroup_iter_break(memcg, iter);
1472 iter->oom_lock = true;
1477 * OK, we failed to lock the whole subtree so we have
1478 * to clean up what we set up to the failing subtree
1480 for_each_mem_cgroup_tree(iter, memcg) {
1481 if (iter == failed) {
1482 mem_cgroup_iter_break(memcg, iter);
1485 iter->oom_lock = false;
1488 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1490 spin_unlock(&memcg_oom_lock);
1495 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1497 struct mem_cgroup *iter;
1499 spin_lock(&memcg_oom_lock);
1500 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1501 for_each_mem_cgroup_tree(iter, memcg)
1502 iter->oom_lock = false;
1503 spin_unlock(&memcg_oom_lock);
1506 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1508 struct mem_cgroup *iter;
1510 spin_lock(&memcg_oom_lock);
1511 for_each_mem_cgroup_tree(iter, memcg)
1513 spin_unlock(&memcg_oom_lock);
1516 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1518 struct mem_cgroup *iter;
1521 * When a new child is created while the hierarchy is under oom,
1522 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1524 spin_lock(&memcg_oom_lock);
1525 for_each_mem_cgroup_tree(iter, memcg)
1526 if (iter->under_oom > 0)
1528 spin_unlock(&memcg_oom_lock);
1531 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1533 struct oom_wait_info {
1534 struct mem_cgroup *memcg;
1538 static int memcg_oom_wake_function(wait_queue_t *wait,
1539 unsigned mode, int sync, void *arg)
1541 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1542 struct mem_cgroup *oom_wait_memcg;
1543 struct oom_wait_info *oom_wait_info;
1545 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1546 oom_wait_memcg = oom_wait_info->memcg;
1548 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1549 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1551 return autoremove_wake_function(wait, mode, sync, arg);
1554 static void memcg_oom_recover(struct mem_cgroup *memcg)
1557 * For the following lockless ->under_oom test, the only required
1558 * guarantee is that it must see the state asserted by an OOM when
1559 * this function is called as a result of userland actions
1560 * triggered by the notification of the OOM. This is trivially
1561 * achieved by invoking mem_cgroup_mark_under_oom() before
1562 * triggering notification.
1564 if (memcg && memcg->under_oom)
1565 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1568 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1570 if (!current->memcg_may_oom)
1573 * We are in the middle of the charge context here, so we
1574 * don't want to block when potentially sitting on a callstack
1575 * that holds all kinds of filesystem and mm locks.
1577 * Also, the caller may handle a failed allocation gracefully
1578 * (like optional page cache readahead) and so an OOM killer
1579 * invocation might not even be necessary.
1581 * That's why we don't do anything here except remember the
1582 * OOM context and then deal with it at the end of the page
1583 * fault when the stack is unwound, the locks are released,
1584 * and when we know whether the fault was overall successful.
1586 css_get(&memcg->css);
1587 current->memcg_in_oom = memcg;
1588 current->memcg_oom_gfp_mask = mask;
1589 current->memcg_oom_order = order;
1593 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1594 * @handle: actually kill/wait or just clean up the OOM state
1596 * This has to be called at the end of a page fault if the memcg OOM
1597 * handler was enabled.
1599 * Memcg supports userspace OOM handling where failed allocations must
1600 * sleep on a waitqueue until the userspace task resolves the
1601 * situation. Sleeping directly in the charge context with all kinds
1602 * of locks held is not a good idea, instead we remember an OOM state
1603 * in the task and mem_cgroup_oom_synchronize() has to be called at
1604 * the end of the page fault to complete the OOM handling.
1606 * Returns %true if an ongoing memcg OOM situation was detected and
1607 * completed, %false otherwise.
1609 bool mem_cgroup_oom_synchronize(bool handle)
1611 struct mem_cgroup *memcg = current->memcg_in_oom;
1612 struct oom_wait_info owait;
1615 /* OOM is global, do not handle */
1619 if (!handle || oom_killer_disabled)
1622 owait.memcg = memcg;
1623 owait.wait.flags = 0;
1624 owait.wait.func = memcg_oom_wake_function;
1625 owait.wait.private = current;
1626 INIT_LIST_HEAD(&owait.wait.task_list);
1628 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1629 mem_cgroup_mark_under_oom(memcg);
1631 locked = mem_cgroup_oom_trylock(memcg);
1634 mem_cgroup_oom_notify(memcg);
1636 if (locked && !memcg->oom_kill_disable) {
1637 mem_cgroup_unmark_under_oom(memcg);
1638 finish_wait(&memcg_oom_waitq, &owait.wait);
1639 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1640 current->memcg_oom_order);
1643 mem_cgroup_unmark_under_oom(memcg);
1644 finish_wait(&memcg_oom_waitq, &owait.wait);
1648 mem_cgroup_oom_unlock(memcg);
1650 * There is no guarantee that an OOM-lock contender
1651 * sees the wakeups triggered by the OOM kill
1652 * uncharges. Wake any sleepers explicitely.
1654 memcg_oom_recover(memcg);
1657 current->memcg_in_oom = NULL;
1658 css_put(&memcg->css);
1663 * lock_page_memcg - lock a page->mem_cgroup binding
1666 * This function protects unlocked LRU pages from being moved to
1667 * another cgroup and stabilizes their page->mem_cgroup binding.
1669 void lock_page_memcg(struct page *page)
1671 struct mem_cgroup *memcg;
1672 unsigned long flags;
1675 * The RCU lock is held throughout the transaction. The fast
1676 * path can get away without acquiring the memcg->move_lock
1677 * because page moving starts with an RCU grace period.
1681 if (mem_cgroup_disabled())
1684 memcg = page->mem_cgroup;
1685 if (unlikely(!memcg))
1688 if (atomic_read(&memcg->moving_account) <= 0)
1691 spin_lock_irqsave(&memcg->move_lock, flags);
1692 if (memcg != page->mem_cgroup) {
1693 spin_unlock_irqrestore(&memcg->move_lock, flags);
1698 * When charge migration first begins, we can have locked and
1699 * unlocked page stat updates happening concurrently. Track
1700 * the task who has the lock for unlock_page_memcg().
1702 memcg->move_lock_task = current;
1703 memcg->move_lock_flags = flags;
1707 EXPORT_SYMBOL(lock_page_memcg);
1710 * unlock_page_memcg - unlock a page->mem_cgroup binding
1713 void unlock_page_memcg(struct page *page)
1715 struct mem_cgroup *memcg = page->mem_cgroup;
1717 if (memcg && memcg->move_lock_task == current) {
1718 unsigned long flags = memcg->move_lock_flags;
1720 memcg->move_lock_task = NULL;
1721 memcg->move_lock_flags = 0;
1723 spin_unlock_irqrestore(&memcg->move_lock, flags);
1728 EXPORT_SYMBOL(unlock_page_memcg);
1731 * size of first charge trial. "32" comes from vmscan.c's magic value.
1732 * TODO: maybe necessary to use big numbers in big irons.
1734 #define CHARGE_BATCH 32U
1735 struct memcg_stock_pcp {
1736 struct mem_cgroup *cached; /* this never be root cgroup */
1737 unsigned int nr_pages;
1738 struct work_struct work;
1739 unsigned long flags;
1740 #define FLUSHING_CACHED_CHARGE 0
1742 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1743 static DEFINE_MUTEX(percpu_charge_mutex);
1746 * consume_stock: Try to consume stocked charge on this cpu.
1747 * @memcg: memcg to consume from.
1748 * @nr_pages: how many pages to charge.
1750 * The charges will only happen if @memcg matches the current cpu's memcg
1751 * stock, and at least @nr_pages are available in that stock. Failure to
1752 * service an allocation will refill the stock.
1754 * returns true if successful, false otherwise.
1756 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1758 struct memcg_stock_pcp *stock;
1761 if (nr_pages > CHARGE_BATCH)
1764 stock = &get_cpu_var(memcg_stock);
1765 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1766 stock->nr_pages -= nr_pages;
1769 put_cpu_var(memcg_stock);
1774 * Returns stocks cached in percpu and reset cached information.
1776 static void drain_stock(struct memcg_stock_pcp *stock)
1778 struct mem_cgroup *old = stock->cached;
1780 if (stock->nr_pages) {
1781 page_counter_uncharge(&old->memory, stock->nr_pages);
1782 if (do_memsw_account())
1783 page_counter_uncharge(&old->memsw, stock->nr_pages);
1784 css_put_many(&old->css, stock->nr_pages);
1785 stock->nr_pages = 0;
1787 stock->cached = NULL;
1791 * This must be called under preempt disabled or must be called by
1792 * a thread which is pinned to local cpu.
1794 static void drain_local_stock(struct work_struct *dummy)
1796 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1798 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1802 * Cache charges(val) to local per_cpu area.
1803 * This will be consumed by consume_stock() function, later.
1805 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1807 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1809 if (stock->cached != memcg) { /* reset if necessary */
1811 stock->cached = memcg;
1813 stock->nr_pages += nr_pages;
1814 put_cpu_var(memcg_stock);
1818 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1819 * of the hierarchy under it.
1821 static void drain_all_stock(struct mem_cgroup *root_memcg)
1825 /* If someone's already draining, avoid adding running more workers. */
1826 if (!mutex_trylock(&percpu_charge_mutex))
1828 /* Notify other cpus that system-wide "drain" is running */
1831 for_each_online_cpu(cpu) {
1832 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1833 struct mem_cgroup *memcg;
1835 memcg = stock->cached;
1836 if (!memcg || !stock->nr_pages)
1838 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1840 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1842 drain_local_stock(&stock->work);
1844 schedule_work_on(cpu, &stock->work);
1849 mutex_unlock(&percpu_charge_mutex);
1852 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1853 unsigned long action,
1856 int cpu = (unsigned long)hcpu;
1857 struct memcg_stock_pcp *stock;
1859 if (action == CPU_ONLINE)
1862 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1865 stock = &per_cpu(memcg_stock, cpu);
1870 static void reclaim_high(struct mem_cgroup *memcg,
1871 unsigned int nr_pages,
1875 if (page_counter_read(&memcg->memory) <= memcg->high)
1877 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1878 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1879 } while ((memcg = parent_mem_cgroup(memcg)));
1882 static void high_work_func(struct work_struct *work)
1884 struct mem_cgroup *memcg;
1886 memcg = container_of(work, struct mem_cgroup, high_work);
1887 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1891 * Scheduled by try_charge() to be executed from the userland return path
1892 * and reclaims memory over the high limit.
1894 void mem_cgroup_handle_over_high(void)
1896 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1897 struct mem_cgroup *memcg;
1899 if (likely(!nr_pages))
1902 memcg = get_mem_cgroup_from_mm(current->mm);
1903 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1904 css_put(&memcg->css);
1905 current->memcg_nr_pages_over_high = 0;
1908 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1909 unsigned int nr_pages)
1911 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1912 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1913 struct mem_cgroup *mem_over_limit;
1914 struct page_counter *counter;
1915 unsigned long nr_reclaimed;
1916 bool may_swap = true;
1917 bool drained = false;
1919 if (mem_cgroup_is_root(memcg))
1922 if (consume_stock(memcg, nr_pages))
1925 if (!do_memsw_account() ||
1926 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1927 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1929 if (do_memsw_account())
1930 page_counter_uncharge(&memcg->memsw, batch);
1931 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1933 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1937 if (batch > nr_pages) {
1943 * Unlike in global OOM situations, memcg is not in a physical
1944 * memory shortage. Allow dying and OOM-killed tasks to
1945 * bypass the last charges so that they can exit quickly and
1946 * free their memory.
1948 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1949 fatal_signal_pending(current) ||
1950 current->flags & PF_EXITING))
1953 if (unlikely(task_in_memcg_oom(current)))
1956 if (!gfpflags_allow_blocking(gfp_mask))
1959 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1961 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1962 gfp_mask, may_swap);
1964 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1968 drain_all_stock(mem_over_limit);
1973 if (gfp_mask & __GFP_NORETRY)
1976 * Even though the limit is exceeded at this point, reclaim
1977 * may have been able to free some pages. Retry the charge
1978 * before killing the task.
1980 * Only for regular pages, though: huge pages are rather
1981 * unlikely to succeed so close to the limit, and we fall back
1982 * to regular pages anyway in case of failure.
1984 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
1987 * At task move, charge accounts can be doubly counted. So, it's
1988 * better to wait until the end of task_move if something is going on.
1990 if (mem_cgroup_wait_acct_move(mem_over_limit))
1996 if (gfp_mask & __GFP_NOFAIL)
1999 if (fatal_signal_pending(current))
2002 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2004 mem_cgroup_oom(mem_over_limit, gfp_mask,
2005 get_order(nr_pages * PAGE_SIZE));
2007 if (!(gfp_mask & __GFP_NOFAIL))
2011 * The allocation either can't fail or will lead to more memory
2012 * being freed very soon. Allow memory usage go over the limit
2013 * temporarily by force charging it.
2015 page_counter_charge(&memcg->memory, nr_pages);
2016 if (do_memsw_account())
2017 page_counter_charge(&memcg->memsw, nr_pages);
2018 css_get_many(&memcg->css, nr_pages);
2023 css_get_many(&memcg->css, batch);
2024 if (batch > nr_pages)
2025 refill_stock(memcg, batch - nr_pages);
2028 * If the hierarchy is above the normal consumption range, schedule
2029 * reclaim on returning to userland. We can perform reclaim here
2030 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2031 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2032 * not recorded as it most likely matches current's and won't
2033 * change in the meantime. As high limit is checked again before
2034 * reclaim, the cost of mismatch is negligible.
2037 if (page_counter_read(&memcg->memory) > memcg->high) {
2038 /* Don't bother a random interrupted task */
2039 if (in_interrupt()) {
2040 schedule_work(&memcg->high_work);
2043 current->memcg_nr_pages_over_high += batch;
2044 set_notify_resume(current);
2047 } while ((memcg = parent_mem_cgroup(memcg)));
2052 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2054 if (mem_cgroup_is_root(memcg))
2057 page_counter_uncharge(&memcg->memory, nr_pages);
2058 if (do_memsw_account())
2059 page_counter_uncharge(&memcg->memsw, nr_pages);
2061 css_put_many(&memcg->css, nr_pages);
2064 static void lock_page_lru(struct page *page, int *isolated)
2066 struct zone *zone = page_zone(page);
2068 spin_lock_irq(zone_lru_lock(zone));
2069 if (PageLRU(page)) {
2070 struct lruvec *lruvec;
2072 lruvec = mem_cgroup_page_lruvec(page, zone);
2074 del_page_from_lru_list(page, lruvec, page_lru(page));
2080 static void unlock_page_lru(struct page *page, int isolated)
2082 struct zone *zone = page_zone(page);
2085 struct lruvec *lruvec;
2087 lruvec = mem_cgroup_page_lruvec(page, zone);
2088 VM_BUG_ON_PAGE(PageLRU(page), page);
2090 add_page_to_lru_list(page, lruvec, page_lru(page));
2092 spin_unlock_irq(zone_lru_lock(zone));
2095 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2100 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2103 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2104 * may already be on some other mem_cgroup's LRU. Take care of it.
2107 lock_page_lru(page, &isolated);
2110 * Nobody should be changing or seriously looking at
2111 * page->mem_cgroup at this point:
2113 * - the page is uncharged
2115 * - the page is off-LRU
2117 * - an anonymous fault has exclusive page access, except for
2118 * a locked page table
2120 * - a page cache insertion, a swapin fault, or a migration
2121 * have the page locked
2123 page->mem_cgroup = memcg;
2126 unlock_page_lru(page, isolated);
2130 static int memcg_alloc_cache_id(void)
2135 id = ida_simple_get(&memcg_cache_ida,
2136 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2140 if (id < memcg_nr_cache_ids)
2144 * There's no space for the new id in memcg_caches arrays,
2145 * so we have to grow them.
2147 down_write(&memcg_cache_ids_sem);
2149 size = 2 * (id + 1);
2150 if (size < MEMCG_CACHES_MIN_SIZE)
2151 size = MEMCG_CACHES_MIN_SIZE;
2152 else if (size > MEMCG_CACHES_MAX_SIZE)
2153 size = MEMCG_CACHES_MAX_SIZE;
2155 err = memcg_update_all_caches(size);
2157 err = memcg_update_all_list_lrus(size);
2159 memcg_nr_cache_ids = size;
2161 up_write(&memcg_cache_ids_sem);
2164 ida_simple_remove(&memcg_cache_ida, id);
2170 static void memcg_free_cache_id(int id)
2172 ida_simple_remove(&memcg_cache_ida, id);
2175 struct memcg_kmem_cache_create_work {
2176 struct mem_cgroup *memcg;
2177 struct kmem_cache *cachep;
2178 struct work_struct work;
2181 static void memcg_kmem_cache_create_func(struct work_struct *w)
2183 struct memcg_kmem_cache_create_work *cw =
2184 container_of(w, struct memcg_kmem_cache_create_work, work);
2185 struct mem_cgroup *memcg = cw->memcg;
2186 struct kmem_cache *cachep = cw->cachep;
2188 memcg_create_kmem_cache(memcg, cachep);
2190 css_put(&memcg->css);
2195 * Enqueue the creation of a per-memcg kmem_cache.
2197 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2198 struct kmem_cache *cachep)
2200 struct memcg_kmem_cache_create_work *cw;
2202 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2206 css_get(&memcg->css);
2209 cw->cachep = cachep;
2210 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2212 schedule_work(&cw->work);
2215 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2216 struct kmem_cache *cachep)
2219 * We need to stop accounting when we kmalloc, because if the
2220 * corresponding kmalloc cache is not yet created, the first allocation
2221 * in __memcg_schedule_kmem_cache_create will recurse.
2223 * However, it is better to enclose the whole function. Depending on
2224 * the debugging options enabled, INIT_WORK(), for instance, can
2225 * trigger an allocation. This too, will make us recurse. Because at
2226 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2227 * the safest choice is to do it like this, wrapping the whole function.
2229 current->memcg_kmem_skip_account = 1;
2230 __memcg_schedule_kmem_cache_create(memcg, cachep);
2231 current->memcg_kmem_skip_account = 0;
2234 static inline bool memcg_kmem_bypass(void)
2236 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2242 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2243 * @cachep: the original global kmem cache
2245 * Return the kmem_cache we're supposed to use for a slab allocation.
2246 * We try to use the current memcg's version of the cache.
2248 * If the cache does not exist yet, if we are the first user of it, we
2249 * create it asynchronously in a workqueue and let the current allocation
2250 * go through with the original cache.
2252 * This function takes a reference to the cache it returns to assure it
2253 * won't get destroyed while we are working with it. Once the caller is
2254 * done with it, memcg_kmem_put_cache() must be called to release the
2257 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2259 struct mem_cgroup *memcg;
2260 struct kmem_cache *memcg_cachep;
2263 VM_BUG_ON(!is_root_cache(cachep));
2265 if (memcg_kmem_bypass())
2268 if (current->memcg_kmem_skip_account)
2271 memcg = get_mem_cgroup_from_mm(current->mm);
2272 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2276 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2277 if (likely(memcg_cachep))
2278 return memcg_cachep;
2281 * If we are in a safe context (can wait, and not in interrupt
2282 * context), we could be be predictable and return right away.
2283 * This would guarantee that the allocation being performed
2284 * already belongs in the new cache.
2286 * However, there are some clashes that can arrive from locking.
2287 * For instance, because we acquire the slab_mutex while doing
2288 * memcg_create_kmem_cache, this means no further allocation
2289 * could happen with the slab_mutex held. So it's better to
2292 memcg_schedule_kmem_cache_create(memcg, cachep);
2294 css_put(&memcg->css);
2299 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2300 * @cachep: the cache returned by memcg_kmem_get_cache
2302 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2304 if (!is_root_cache(cachep))
2305 css_put(&cachep->memcg_params.memcg->css);
2309 * memcg_kmem_charge: charge a kmem page
2310 * @page: page to charge
2311 * @gfp: reclaim mode
2312 * @order: allocation order
2313 * @memcg: memory cgroup to charge
2315 * Returns 0 on success, an error code on failure.
2317 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2318 struct mem_cgroup *memcg)
2320 unsigned int nr_pages = 1 << order;
2321 struct page_counter *counter;
2324 ret = try_charge(memcg, gfp, nr_pages);
2328 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2329 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2330 cancel_charge(memcg, nr_pages);
2334 page->mem_cgroup = memcg;
2340 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2341 * @page: page to charge
2342 * @gfp: reclaim mode
2343 * @order: allocation order
2345 * Returns 0 on success, an error code on failure.
2347 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2349 struct mem_cgroup *memcg;
2352 if (memcg_kmem_bypass())
2355 memcg = get_mem_cgroup_from_mm(current->mm);
2356 if (!mem_cgroup_is_root(memcg))
2357 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2358 css_put(&memcg->css);
2362 * memcg_kmem_uncharge: uncharge a kmem page
2363 * @page: page to uncharge
2364 * @order: allocation order
2366 void memcg_kmem_uncharge(struct page *page, int order)
2368 struct mem_cgroup *memcg = page->mem_cgroup;
2369 unsigned int nr_pages = 1 << order;
2374 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2376 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2377 page_counter_uncharge(&memcg->kmem, nr_pages);
2379 page_counter_uncharge(&memcg->memory, nr_pages);
2380 if (do_memsw_account())
2381 page_counter_uncharge(&memcg->memsw, nr_pages);
2383 page->mem_cgroup = NULL;
2384 css_put_many(&memcg->css, nr_pages);
2386 #endif /* !CONFIG_SLOB */
2388 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2391 * Because tail pages are not marked as "used", set it. We're under
2392 * zone_lru_lock and migration entries setup in all page mappings.
2394 void mem_cgroup_split_huge_fixup(struct page *head)
2398 if (mem_cgroup_disabled())
2401 for (i = 1; i < HPAGE_PMD_NR; i++)
2402 head[i].mem_cgroup = head->mem_cgroup;
2404 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2407 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2409 #ifdef CONFIG_MEMCG_SWAP
2410 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2413 int val = (charge) ? 1 : -1;
2414 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2418 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2419 * @entry: swap entry to be moved
2420 * @from: mem_cgroup which the entry is moved from
2421 * @to: mem_cgroup which the entry is moved to
2423 * It succeeds only when the swap_cgroup's record for this entry is the same
2424 * as the mem_cgroup's id of @from.
2426 * Returns 0 on success, -EINVAL on failure.
2428 * The caller must have charged to @to, IOW, called page_counter_charge() about
2429 * both res and memsw, and called css_get().
2431 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2432 struct mem_cgroup *from, struct mem_cgroup *to)
2434 unsigned short old_id, new_id;
2436 old_id = mem_cgroup_id(from);
2437 new_id = mem_cgroup_id(to);
2439 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2440 mem_cgroup_swap_statistics(from, false);
2441 mem_cgroup_swap_statistics(to, true);
2447 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2448 struct mem_cgroup *from, struct mem_cgroup *to)
2454 static DEFINE_MUTEX(memcg_limit_mutex);
2456 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2457 unsigned long limit)
2459 unsigned long curusage;
2460 unsigned long oldusage;
2461 bool enlarge = false;
2466 * For keeping hierarchical_reclaim simple, how long we should retry
2467 * is depends on callers. We set our retry-count to be function
2468 * of # of children which we should visit in this loop.
2470 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2471 mem_cgroup_count_children(memcg);
2473 oldusage = page_counter_read(&memcg->memory);
2476 if (signal_pending(current)) {
2481 mutex_lock(&memcg_limit_mutex);
2482 if (limit > memcg->memsw.limit) {
2483 mutex_unlock(&memcg_limit_mutex);
2487 if (limit > memcg->memory.limit)
2489 ret = page_counter_limit(&memcg->memory, limit);
2490 mutex_unlock(&memcg_limit_mutex);
2495 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2497 curusage = page_counter_read(&memcg->memory);
2498 /* Usage is reduced ? */
2499 if (curusage >= oldusage)
2502 oldusage = curusage;
2503 } while (retry_count);
2505 if (!ret && enlarge)
2506 memcg_oom_recover(memcg);
2511 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2512 unsigned long limit)
2514 unsigned long curusage;
2515 unsigned long oldusage;
2516 bool enlarge = false;
2520 /* see mem_cgroup_resize_res_limit */
2521 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2522 mem_cgroup_count_children(memcg);
2524 oldusage = page_counter_read(&memcg->memsw);
2527 if (signal_pending(current)) {
2532 mutex_lock(&memcg_limit_mutex);
2533 if (limit < memcg->memory.limit) {
2534 mutex_unlock(&memcg_limit_mutex);
2538 if (limit > memcg->memsw.limit)
2540 ret = page_counter_limit(&memcg->memsw, limit);
2541 mutex_unlock(&memcg_limit_mutex);
2546 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2548 curusage = page_counter_read(&memcg->memsw);
2549 /* Usage is reduced ? */
2550 if (curusage >= oldusage)
2553 oldusage = curusage;
2554 } while (retry_count);
2556 if (!ret && enlarge)
2557 memcg_oom_recover(memcg);
2562 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2564 unsigned long *total_scanned)
2566 unsigned long nr_reclaimed = 0;
2567 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2568 unsigned long reclaimed;
2570 struct mem_cgroup_tree_per_zone *mctz;
2571 unsigned long excess;
2572 unsigned long nr_scanned;
2577 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2579 * This loop can run a while, specially if mem_cgroup's continuously
2580 * keep exceeding their soft limit and putting the system under
2587 mz = mem_cgroup_largest_soft_limit_node(mctz);
2592 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2593 gfp_mask, &nr_scanned);
2594 nr_reclaimed += reclaimed;
2595 *total_scanned += nr_scanned;
2596 spin_lock_irq(&mctz->lock);
2597 __mem_cgroup_remove_exceeded(mz, mctz);
2600 * If we failed to reclaim anything from this memory cgroup
2601 * it is time to move on to the next cgroup
2605 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2607 excess = soft_limit_excess(mz->memcg);
2609 * One school of thought says that we should not add
2610 * back the node to the tree if reclaim returns 0.
2611 * But our reclaim could return 0, simply because due
2612 * to priority we are exposing a smaller subset of
2613 * memory to reclaim from. Consider this as a longer
2616 /* If excess == 0, no tree ops */
2617 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2618 spin_unlock_irq(&mctz->lock);
2619 css_put(&mz->memcg->css);
2622 * Could not reclaim anything and there are no more
2623 * mem cgroups to try or we seem to be looping without
2624 * reclaiming anything.
2626 if (!nr_reclaimed &&
2628 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2630 } while (!nr_reclaimed);
2632 css_put(&next_mz->memcg->css);
2633 return nr_reclaimed;
2637 * Test whether @memcg has children, dead or alive. Note that this
2638 * function doesn't care whether @memcg has use_hierarchy enabled and
2639 * returns %true if there are child csses according to the cgroup
2640 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2642 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2647 ret = css_next_child(NULL, &memcg->css);
2653 * Reclaims as many pages from the given memcg as possible.
2655 * Caller is responsible for holding css reference for memcg.
2657 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2659 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2661 /* we call try-to-free pages for make this cgroup empty */
2662 lru_add_drain_all();
2663 /* try to free all pages in this cgroup */
2664 while (nr_retries && page_counter_read(&memcg->memory)) {
2667 if (signal_pending(current))
2670 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2674 /* maybe some writeback is necessary */
2675 congestion_wait(BLK_RW_ASYNC, HZ/10);
2683 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2684 char *buf, size_t nbytes,
2687 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2689 if (mem_cgroup_is_root(memcg))
2691 return mem_cgroup_force_empty(memcg) ?: nbytes;
2694 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2697 return mem_cgroup_from_css(css)->use_hierarchy;
2700 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2701 struct cftype *cft, u64 val)
2704 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2705 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2707 if (memcg->use_hierarchy == val)
2711 * If parent's use_hierarchy is set, we can't make any modifications
2712 * in the child subtrees. If it is unset, then the change can
2713 * occur, provided the current cgroup has no children.
2715 * For the root cgroup, parent_mem is NULL, we allow value to be
2716 * set if there are no children.
2718 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2719 (val == 1 || val == 0)) {
2720 if (!memcg_has_children(memcg))
2721 memcg->use_hierarchy = val;
2730 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2732 struct mem_cgroup *iter;
2735 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2737 for_each_mem_cgroup_tree(iter, memcg) {
2738 for (i = 0; i < MEMCG_NR_STAT; i++)
2739 stat[i] += mem_cgroup_read_stat(iter, i);
2743 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2745 struct mem_cgroup *iter;
2748 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2750 for_each_mem_cgroup_tree(iter, memcg) {
2751 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2752 events[i] += mem_cgroup_read_events(iter, i);
2756 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2758 unsigned long val = 0;
2760 if (mem_cgroup_is_root(memcg)) {
2761 struct mem_cgroup *iter;
2763 for_each_mem_cgroup_tree(iter, memcg) {
2764 val += mem_cgroup_read_stat(iter,
2765 MEM_CGROUP_STAT_CACHE);
2766 val += mem_cgroup_read_stat(iter,
2767 MEM_CGROUP_STAT_RSS);
2769 val += mem_cgroup_read_stat(iter,
2770 MEM_CGROUP_STAT_SWAP);
2774 val = page_counter_read(&memcg->memory);
2776 val = page_counter_read(&memcg->memsw);
2789 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2792 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2793 struct page_counter *counter;
2795 switch (MEMFILE_TYPE(cft->private)) {
2797 counter = &memcg->memory;
2800 counter = &memcg->memsw;
2803 counter = &memcg->kmem;
2806 counter = &memcg->tcpmem;
2812 switch (MEMFILE_ATTR(cft->private)) {
2814 if (counter == &memcg->memory)
2815 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2816 if (counter == &memcg->memsw)
2817 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2818 return (u64)page_counter_read(counter) * PAGE_SIZE;
2820 return (u64)counter->limit * PAGE_SIZE;
2822 return (u64)counter->watermark * PAGE_SIZE;
2824 return counter->failcnt;
2825 case RES_SOFT_LIMIT:
2826 return (u64)memcg->soft_limit * PAGE_SIZE;
2833 static int memcg_online_kmem(struct mem_cgroup *memcg)
2837 if (cgroup_memory_nokmem)
2840 BUG_ON(memcg->kmemcg_id >= 0);
2841 BUG_ON(memcg->kmem_state);
2843 memcg_id = memcg_alloc_cache_id();
2847 static_branch_inc(&memcg_kmem_enabled_key);
2849 * A memory cgroup is considered kmem-online as soon as it gets
2850 * kmemcg_id. Setting the id after enabling static branching will
2851 * guarantee no one starts accounting before all call sites are
2854 memcg->kmemcg_id = memcg_id;
2855 memcg->kmem_state = KMEM_ONLINE;
2860 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2862 struct cgroup_subsys_state *css;
2863 struct mem_cgroup *parent, *child;
2866 if (memcg->kmem_state != KMEM_ONLINE)
2869 * Clear the online state before clearing memcg_caches array
2870 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2871 * guarantees that no cache will be created for this cgroup
2872 * after we are done (see memcg_create_kmem_cache()).
2874 memcg->kmem_state = KMEM_ALLOCATED;
2876 memcg_deactivate_kmem_caches(memcg);
2878 kmemcg_id = memcg->kmemcg_id;
2879 BUG_ON(kmemcg_id < 0);
2881 parent = parent_mem_cgroup(memcg);
2883 parent = root_mem_cgroup;
2886 * Change kmemcg_id of this cgroup and all its descendants to the
2887 * parent's id, and then move all entries from this cgroup's list_lrus
2888 * to ones of the parent. After we have finished, all list_lrus
2889 * corresponding to this cgroup are guaranteed to remain empty. The
2890 * ordering is imposed by list_lru_node->lock taken by
2891 * memcg_drain_all_list_lrus().
2893 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2894 css_for_each_descendant_pre(css, &memcg->css) {
2895 child = mem_cgroup_from_css(css);
2896 BUG_ON(child->kmemcg_id != kmemcg_id);
2897 child->kmemcg_id = parent->kmemcg_id;
2898 if (!memcg->use_hierarchy)
2903 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2905 memcg_free_cache_id(kmemcg_id);
2908 static void memcg_free_kmem(struct mem_cgroup *memcg)
2910 /* css_alloc() failed, offlining didn't happen */
2911 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2912 memcg_offline_kmem(memcg);
2914 if (memcg->kmem_state == KMEM_ALLOCATED) {
2915 memcg_destroy_kmem_caches(memcg);
2916 static_branch_dec(&memcg_kmem_enabled_key);
2917 WARN_ON(page_counter_read(&memcg->kmem));
2921 static int memcg_online_kmem(struct mem_cgroup *memcg)
2925 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2928 static void memcg_free_kmem(struct mem_cgroup *memcg)
2931 #endif /* !CONFIG_SLOB */
2933 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2934 unsigned long limit)
2938 mutex_lock(&memcg_limit_mutex);
2939 ret = page_counter_limit(&memcg->kmem, limit);
2940 mutex_unlock(&memcg_limit_mutex);
2944 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2948 mutex_lock(&memcg_limit_mutex);
2950 ret = page_counter_limit(&memcg->tcpmem, limit);
2954 if (!memcg->tcpmem_active) {
2956 * The active flag needs to be written after the static_key
2957 * update. This is what guarantees that the socket activation
2958 * function is the last one to run. See sock_update_memcg() for
2959 * details, and note that we don't mark any socket as belonging
2960 * to this memcg until that flag is up.
2962 * We need to do this, because static_keys will span multiple
2963 * sites, but we can't control their order. If we mark a socket
2964 * as accounted, but the accounting functions are not patched in
2965 * yet, we'll lose accounting.
2967 * We never race with the readers in sock_update_memcg(),
2968 * because when this value change, the code to process it is not
2971 static_branch_inc(&memcg_sockets_enabled_key);
2972 memcg->tcpmem_active = true;
2975 mutex_unlock(&memcg_limit_mutex);
2980 * The user of this function is...
2983 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2984 char *buf, size_t nbytes, loff_t off)
2986 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2987 unsigned long nr_pages;
2990 buf = strstrip(buf);
2991 ret = page_counter_memparse(buf, "-1", &nr_pages);
2995 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2997 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3001 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3003 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3006 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3009 ret = memcg_update_kmem_limit(memcg, nr_pages);
3012 ret = memcg_update_tcp_limit(memcg, nr_pages);
3016 case RES_SOFT_LIMIT:
3017 memcg->soft_limit = nr_pages;
3021 return ret ?: nbytes;
3024 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3025 size_t nbytes, loff_t off)
3027 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3028 struct page_counter *counter;
3030 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3032 counter = &memcg->memory;
3035 counter = &memcg->memsw;
3038 counter = &memcg->kmem;
3041 counter = &memcg->tcpmem;
3047 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3049 page_counter_reset_watermark(counter);
3052 counter->failcnt = 0;
3061 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3064 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3068 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3069 struct cftype *cft, u64 val)
3071 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3073 if (val & ~MOVE_MASK)
3077 * No kind of locking is needed in here, because ->can_attach() will
3078 * check this value once in the beginning of the process, and then carry
3079 * on with stale data. This means that changes to this value will only
3080 * affect task migrations starting after the change.
3082 memcg->move_charge_at_immigrate = val;
3086 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3087 struct cftype *cft, u64 val)
3094 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3098 unsigned int lru_mask;
3101 static const struct numa_stat stats[] = {
3102 { "total", LRU_ALL },
3103 { "file", LRU_ALL_FILE },
3104 { "anon", LRU_ALL_ANON },
3105 { "unevictable", BIT(LRU_UNEVICTABLE) },
3107 const struct numa_stat *stat;
3110 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3112 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3113 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3114 seq_printf(m, "%s=%lu", stat->name, nr);
3115 for_each_node_state(nid, N_MEMORY) {
3116 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3118 seq_printf(m, " N%d=%lu", nid, nr);
3123 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3124 struct mem_cgroup *iter;
3127 for_each_mem_cgroup_tree(iter, memcg)
3128 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3129 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3130 for_each_node_state(nid, N_MEMORY) {
3132 for_each_mem_cgroup_tree(iter, memcg)
3133 nr += mem_cgroup_node_nr_lru_pages(
3134 iter, nid, stat->lru_mask);
3135 seq_printf(m, " N%d=%lu", nid, nr);
3142 #endif /* CONFIG_NUMA */
3144 static int memcg_stat_show(struct seq_file *m, void *v)
3146 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3147 unsigned long memory, memsw;
3148 struct mem_cgroup *mi;
3151 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3152 MEM_CGROUP_STAT_NSTATS);
3153 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3154 MEM_CGROUP_EVENTS_NSTATS);
3155 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3157 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3158 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3160 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3161 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3164 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3165 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3166 mem_cgroup_read_events(memcg, i));
3168 for (i = 0; i < NR_LRU_LISTS; i++)
3169 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3170 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3172 /* Hierarchical information */
3173 memory = memsw = PAGE_COUNTER_MAX;
3174 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3175 memory = min(memory, mi->memory.limit);
3176 memsw = min(memsw, mi->memsw.limit);
3178 seq_printf(m, "hierarchical_memory_limit %llu\n",
3179 (u64)memory * PAGE_SIZE);
3180 if (do_memsw_account())
3181 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3182 (u64)memsw * PAGE_SIZE);
3184 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3185 unsigned long long val = 0;
3187 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3189 for_each_mem_cgroup_tree(mi, memcg)
3190 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3191 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3194 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3195 unsigned long long val = 0;
3197 for_each_mem_cgroup_tree(mi, memcg)
3198 val += mem_cgroup_read_events(mi, i);
3199 seq_printf(m, "total_%s %llu\n",
3200 mem_cgroup_events_names[i], val);
3203 for (i = 0; i < NR_LRU_LISTS; i++) {
3204 unsigned long long val = 0;
3206 for_each_mem_cgroup_tree(mi, memcg)
3207 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3208 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3211 #ifdef CONFIG_DEBUG_VM
3214 struct mem_cgroup_per_zone *mz;
3215 struct zone_reclaim_stat *rstat;
3216 unsigned long recent_rotated[2] = {0, 0};
3217 unsigned long recent_scanned[2] = {0, 0};
3219 for_each_online_node(nid)
3220 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3221 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3222 rstat = &mz->lruvec.reclaim_stat;
3224 recent_rotated[0] += rstat->recent_rotated[0];
3225 recent_rotated[1] += rstat->recent_rotated[1];
3226 recent_scanned[0] += rstat->recent_scanned[0];
3227 recent_scanned[1] += rstat->recent_scanned[1];
3229 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3230 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3231 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3232 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3239 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3242 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3244 return mem_cgroup_swappiness(memcg);
3247 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3248 struct cftype *cft, u64 val)
3250 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3256 memcg->swappiness = val;
3258 vm_swappiness = val;
3263 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3265 struct mem_cgroup_threshold_ary *t;
3266 unsigned long usage;
3271 t = rcu_dereference(memcg->thresholds.primary);
3273 t = rcu_dereference(memcg->memsw_thresholds.primary);
3278 usage = mem_cgroup_usage(memcg, swap);
3281 * current_threshold points to threshold just below or equal to usage.
3282 * If it's not true, a threshold was crossed after last
3283 * call of __mem_cgroup_threshold().
3285 i = t->current_threshold;
3288 * Iterate backward over array of thresholds starting from
3289 * current_threshold and check if a threshold is crossed.
3290 * If none of thresholds below usage is crossed, we read
3291 * only one element of the array here.
3293 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3294 eventfd_signal(t->entries[i].eventfd, 1);
3296 /* i = current_threshold + 1 */
3300 * Iterate forward over array of thresholds starting from
3301 * current_threshold+1 and check if a threshold is crossed.
3302 * If none of thresholds above usage is crossed, we read
3303 * only one element of the array here.
3305 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3306 eventfd_signal(t->entries[i].eventfd, 1);
3308 /* Update current_threshold */
3309 t->current_threshold = i - 1;
3314 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3317 __mem_cgroup_threshold(memcg, false);
3318 if (do_memsw_account())
3319 __mem_cgroup_threshold(memcg, true);
3321 memcg = parent_mem_cgroup(memcg);
3325 static int compare_thresholds(const void *a, const void *b)
3327 const struct mem_cgroup_threshold *_a = a;
3328 const struct mem_cgroup_threshold *_b = b;
3330 if (_a->threshold > _b->threshold)
3333 if (_a->threshold < _b->threshold)
3339 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3341 struct mem_cgroup_eventfd_list *ev;
3343 spin_lock(&memcg_oom_lock);
3345 list_for_each_entry(ev, &memcg->oom_notify, list)
3346 eventfd_signal(ev->eventfd, 1);
3348 spin_unlock(&memcg_oom_lock);
3352 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3354 struct mem_cgroup *iter;
3356 for_each_mem_cgroup_tree(iter, memcg)
3357 mem_cgroup_oom_notify_cb(iter);
3360 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3361 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3363 struct mem_cgroup_thresholds *thresholds;
3364 struct mem_cgroup_threshold_ary *new;
3365 unsigned long threshold;
3366 unsigned long usage;
3369 ret = page_counter_memparse(args, "-1", &threshold);
3373 mutex_lock(&memcg->thresholds_lock);
3376 thresholds = &memcg->thresholds;
3377 usage = mem_cgroup_usage(memcg, false);
3378 } else if (type == _MEMSWAP) {
3379 thresholds = &memcg->memsw_thresholds;
3380 usage = mem_cgroup_usage(memcg, true);
3384 /* Check if a threshold crossed before adding a new one */
3385 if (thresholds->primary)
3386 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3388 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3390 /* Allocate memory for new array of thresholds */
3391 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3399 /* Copy thresholds (if any) to new array */
3400 if (thresholds->primary) {
3401 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3402 sizeof(struct mem_cgroup_threshold));
3405 /* Add new threshold */
3406 new->entries[size - 1].eventfd = eventfd;
3407 new->entries[size - 1].threshold = threshold;
3409 /* Sort thresholds. Registering of new threshold isn't time-critical */
3410 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3411 compare_thresholds, NULL);
3413 /* Find current threshold */
3414 new->current_threshold = -1;
3415 for (i = 0; i < size; i++) {
3416 if (new->entries[i].threshold <= usage) {
3418 * new->current_threshold will not be used until
3419 * rcu_assign_pointer(), so it's safe to increment
3422 ++new->current_threshold;
3427 /* Free old spare buffer and save old primary buffer as spare */
3428 kfree(thresholds->spare);
3429 thresholds->spare = thresholds->primary;
3431 rcu_assign_pointer(thresholds->primary, new);
3433 /* To be sure that nobody uses thresholds */
3437 mutex_unlock(&memcg->thresholds_lock);
3442 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3443 struct eventfd_ctx *eventfd, const char *args)
3445 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3448 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3449 struct eventfd_ctx *eventfd, const char *args)
3451 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3454 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3455 struct eventfd_ctx *eventfd, enum res_type type)
3457 struct mem_cgroup_thresholds *thresholds;
3458 struct mem_cgroup_threshold_ary *new;
3459 unsigned long usage;
3462 mutex_lock(&memcg->thresholds_lock);
3465 thresholds = &memcg->thresholds;
3466 usage = mem_cgroup_usage(memcg, false);
3467 } else if (type == _MEMSWAP) {
3468 thresholds = &memcg->memsw_thresholds;
3469 usage = mem_cgroup_usage(memcg, true);
3473 if (!thresholds->primary)
3476 /* Check if a threshold crossed before removing */
3477 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3479 /* Calculate new number of threshold */
3481 for (i = 0; i < thresholds->primary->size; i++) {
3482 if (thresholds->primary->entries[i].eventfd != eventfd)
3486 new = thresholds->spare;
3488 /* Set thresholds array to NULL if we don't have thresholds */
3497 /* Copy thresholds and find current threshold */
3498 new->current_threshold = -1;
3499 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3500 if (thresholds->primary->entries[i].eventfd == eventfd)
3503 new->entries[j] = thresholds->primary->entries[i];
3504 if (new->entries[j].threshold <= usage) {
3506 * new->current_threshold will not be used
3507 * until rcu_assign_pointer(), so it's safe to increment
3510 ++new->current_threshold;
3516 /* Swap primary and spare array */
3517 thresholds->spare = thresholds->primary;
3519 rcu_assign_pointer(thresholds->primary, new);
3521 /* To be sure that nobody uses thresholds */
3524 /* If all events are unregistered, free the spare array */
3526 kfree(thresholds->spare);
3527 thresholds->spare = NULL;
3530 mutex_unlock(&memcg->thresholds_lock);
3533 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3534 struct eventfd_ctx *eventfd)
3536 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3539 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3540 struct eventfd_ctx *eventfd)
3542 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3545 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3546 struct eventfd_ctx *eventfd, const char *args)
3548 struct mem_cgroup_eventfd_list *event;
3550 event = kmalloc(sizeof(*event), GFP_KERNEL);
3554 spin_lock(&memcg_oom_lock);
3556 event->eventfd = eventfd;
3557 list_add(&event->list, &memcg->oom_notify);
3559 /* already in OOM ? */
3560 if (memcg->under_oom)
3561 eventfd_signal(eventfd, 1);
3562 spin_unlock(&memcg_oom_lock);
3567 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3568 struct eventfd_ctx *eventfd)
3570 struct mem_cgroup_eventfd_list *ev, *tmp;
3572 spin_lock(&memcg_oom_lock);
3574 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3575 if (ev->eventfd == eventfd) {
3576 list_del(&ev->list);
3581 spin_unlock(&memcg_oom_lock);
3584 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3586 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3588 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3589 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3593 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3594 struct cftype *cft, u64 val)
3596 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3598 /* cannot set to root cgroup and only 0 and 1 are allowed */
3599 if (!css->parent || !((val == 0) || (val == 1)))
3602 memcg->oom_kill_disable = val;
3604 memcg_oom_recover(memcg);
3609 #ifdef CONFIG_CGROUP_WRITEBACK
3611 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3613 return &memcg->cgwb_list;
3616 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3618 return wb_domain_init(&memcg->cgwb_domain, gfp);
3621 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3623 wb_domain_exit(&memcg->cgwb_domain);
3626 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3628 wb_domain_size_changed(&memcg->cgwb_domain);
3631 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3633 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3635 if (!memcg->css.parent)
3638 return &memcg->cgwb_domain;
3642 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3643 * @wb: bdi_writeback in question
3644 * @pfilepages: out parameter for number of file pages
3645 * @pheadroom: out parameter for number of allocatable pages according to memcg
3646 * @pdirty: out parameter for number of dirty pages
3647 * @pwriteback: out parameter for number of pages under writeback
3649 * Determine the numbers of file, headroom, dirty, and writeback pages in
3650 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3651 * is a bit more involved.
3653 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3654 * headroom is calculated as the lowest headroom of itself and the
3655 * ancestors. Note that this doesn't consider the actual amount of
3656 * available memory in the system. The caller should further cap
3657 * *@pheadroom accordingly.
3659 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3660 unsigned long *pheadroom, unsigned long *pdirty,
3661 unsigned long *pwriteback)
3663 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3664 struct mem_cgroup *parent;
3666 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3668 /* this should eventually include NR_UNSTABLE_NFS */
3669 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3670 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3671 (1 << LRU_ACTIVE_FILE));
3672 *pheadroom = PAGE_COUNTER_MAX;
3674 while ((parent = parent_mem_cgroup(memcg))) {
3675 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3676 unsigned long used = page_counter_read(&memcg->memory);
3678 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3683 #else /* CONFIG_CGROUP_WRITEBACK */
3685 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3690 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3694 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3698 #endif /* CONFIG_CGROUP_WRITEBACK */
3701 * DO NOT USE IN NEW FILES.
3703 * "cgroup.event_control" implementation.
3705 * This is way over-engineered. It tries to support fully configurable
3706 * events for each user. Such level of flexibility is completely
3707 * unnecessary especially in the light of the planned unified hierarchy.
3709 * Please deprecate this and replace with something simpler if at all
3714 * Unregister event and free resources.
3716 * Gets called from workqueue.
3718 static void memcg_event_remove(struct work_struct *work)
3720 struct mem_cgroup_event *event =
3721 container_of(work, struct mem_cgroup_event, remove);
3722 struct mem_cgroup *memcg = event->memcg;
3724 remove_wait_queue(event->wqh, &event->wait);
3726 event->unregister_event(memcg, event->eventfd);
3728 /* Notify userspace the event is going away. */
3729 eventfd_signal(event->eventfd, 1);
3731 eventfd_ctx_put(event->eventfd);
3733 css_put(&memcg->css);
3737 * Gets called on POLLHUP on eventfd when user closes it.
3739 * Called with wqh->lock held and interrupts disabled.
3741 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3742 int sync, void *key)
3744 struct mem_cgroup_event *event =
3745 container_of(wait, struct mem_cgroup_event, wait);
3746 struct mem_cgroup *memcg = event->memcg;
3747 unsigned long flags = (unsigned long)key;
3749 if (flags & POLLHUP) {
3751 * If the event has been detached at cgroup removal, we
3752 * can simply return knowing the other side will cleanup
3755 * We can't race against event freeing since the other
3756 * side will require wqh->lock via remove_wait_queue(),
3759 spin_lock(&memcg->event_list_lock);
3760 if (!list_empty(&event->list)) {
3761 list_del_init(&event->list);
3763 * We are in atomic context, but cgroup_event_remove()
3764 * may sleep, so we have to call it in workqueue.
3766 schedule_work(&event->remove);
3768 spin_unlock(&memcg->event_list_lock);
3774 static void memcg_event_ptable_queue_proc(struct file *file,
3775 wait_queue_head_t *wqh, poll_table *pt)
3777 struct mem_cgroup_event *event =
3778 container_of(pt, struct mem_cgroup_event, pt);
3781 add_wait_queue(wqh, &event->wait);
3785 * DO NOT USE IN NEW FILES.
3787 * Parse input and register new cgroup event handler.
3789 * Input must be in format '<event_fd> <control_fd> <args>'.
3790 * Interpretation of args is defined by control file implementation.
3792 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3793 char *buf, size_t nbytes, loff_t off)
3795 struct cgroup_subsys_state *css = of_css(of);
3796 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3797 struct mem_cgroup_event *event;
3798 struct cgroup_subsys_state *cfile_css;
3799 unsigned int efd, cfd;
3806 buf = strstrip(buf);
3808 efd = simple_strtoul(buf, &endp, 10);
3813 cfd = simple_strtoul(buf, &endp, 10);
3814 if ((*endp != ' ') && (*endp != '\0'))
3818 event = kzalloc(sizeof(*event), GFP_KERNEL);
3822 event->memcg = memcg;
3823 INIT_LIST_HEAD(&event->list);
3824 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3825 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3826 INIT_WORK(&event->remove, memcg_event_remove);
3834 event->eventfd = eventfd_ctx_fileget(efile.file);
3835 if (IS_ERR(event->eventfd)) {
3836 ret = PTR_ERR(event->eventfd);
3843 goto out_put_eventfd;
3846 /* the process need read permission on control file */
3847 /* AV: shouldn't we check that it's been opened for read instead? */
3848 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3853 * Determine the event callbacks and set them in @event. This used
3854 * to be done via struct cftype but cgroup core no longer knows
3855 * about these events. The following is crude but the whole thing
3856 * is for compatibility anyway.
3858 * DO NOT ADD NEW FILES.
3860 name = cfile.file->f_path.dentry->d_name.name;
3862 if (!strcmp(name, "memory.usage_in_bytes")) {
3863 event->register_event = mem_cgroup_usage_register_event;
3864 event->unregister_event = mem_cgroup_usage_unregister_event;
3865 } else if (!strcmp(name, "memory.oom_control")) {
3866 event->register_event = mem_cgroup_oom_register_event;
3867 event->unregister_event = mem_cgroup_oom_unregister_event;
3868 } else if (!strcmp(name, "memory.pressure_level")) {
3869 event->register_event = vmpressure_register_event;
3870 event->unregister_event = vmpressure_unregister_event;
3871 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3872 event->register_event = memsw_cgroup_usage_register_event;
3873 event->unregister_event = memsw_cgroup_usage_unregister_event;
3880 * Verify @cfile should belong to @css. Also, remaining events are
3881 * automatically removed on cgroup destruction but the removal is
3882 * asynchronous, so take an extra ref on @css.
3884 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3885 &memory_cgrp_subsys);
3887 if (IS_ERR(cfile_css))
3889 if (cfile_css != css) {
3894 ret = event->register_event(memcg, event->eventfd, buf);
3898 efile.file->f_op->poll(efile.file, &event->pt);
3900 spin_lock(&memcg->event_list_lock);
3901 list_add(&event->list, &memcg->event_list);
3902 spin_unlock(&memcg->event_list_lock);
3914 eventfd_ctx_put(event->eventfd);
3923 static struct cftype mem_cgroup_legacy_files[] = {
3925 .name = "usage_in_bytes",
3926 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3927 .read_u64 = mem_cgroup_read_u64,
3930 .name = "max_usage_in_bytes",
3931 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3932 .write = mem_cgroup_reset,
3933 .read_u64 = mem_cgroup_read_u64,
3936 .name = "limit_in_bytes",
3937 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3938 .write = mem_cgroup_write,
3939 .read_u64 = mem_cgroup_read_u64,
3942 .name = "soft_limit_in_bytes",
3943 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3944 .write = mem_cgroup_write,
3945 .read_u64 = mem_cgroup_read_u64,
3949 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3950 .write = mem_cgroup_reset,
3951 .read_u64 = mem_cgroup_read_u64,
3955 .seq_show = memcg_stat_show,
3958 .name = "force_empty",
3959 .write = mem_cgroup_force_empty_write,
3962 .name = "use_hierarchy",
3963 .write_u64 = mem_cgroup_hierarchy_write,
3964 .read_u64 = mem_cgroup_hierarchy_read,
3967 .name = "cgroup.event_control", /* XXX: for compat */
3968 .write = memcg_write_event_control,
3969 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3972 .name = "swappiness",
3973 .read_u64 = mem_cgroup_swappiness_read,
3974 .write_u64 = mem_cgroup_swappiness_write,
3977 .name = "move_charge_at_immigrate",
3978 .read_u64 = mem_cgroup_move_charge_read,
3979 .write_u64 = mem_cgroup_move_charge_write,
3982 .name = "oom_control",
3983 .seq_show = mem_cgroup_oom_control_read,
3984 .write_u64 = mem_cgroup_oom_control_write,
3985 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3988 .name = "pressure_level",
3992 .name = "numa_stat",
3993 .seq_show = memcg_numa_stat_show,
3997 .name = "kmem.limit_in_bytes",
3998 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
3999 .write = mem_cgroup_write,
4000 .read_u64 = mem_cgroup_read_u64,
4003 .name = "kmem.usage_in_bytes",
4004 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4005 .read_u64 = mem_cgroup_read_u64,
4008 .name = "kmem.failcnt",
4009 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4010 .write = mem_cgroup_reset,
4011 .read_u64 = mem_cgroup_read_u64,
4014 .name = "kmem.max_usage_in_bytes",
4015 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4016 .write = mem_cgroup_reset,
4017 .read_u64 = mem_cgroup_read_u64,
4019 #ifdef CONFIG_SLABINFO
4021 .name = "kmem.slabinfo",
4022 .seq_start = slab_start,
4023 .seq_next = slab_next,
4024 .seq_stop = slab_stop,
4025 .seq_show = memcg_slab_show,
4029 .name = "kmem.tcp.limit_in_bytes",
4030 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4031 .write = mem_cgroup_write,
4032 .read_u64 = mem_cgroup_read_u64,
4035 .name = "kmem.tcp.usage_in_bytes",
4036 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4037 .read_u64 = mem_cgroup_read_u64,
4040 .name = "kmem.tcp.failcnt",
4041 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4042 .write = mem_cgroup_reset,
4043 .read_u64 = mem_cgroup_read_u64,
4046 .name = "kmem.tcp.max_usage_in_bytes",
4047 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4048 .write = mem_cgroup_reset,
4049 .read_u64 = mem_cgroup_read_u64,
4051 { }, /* terminate */
4055 * Private memory cgroup IDR
4057 * Swap-out records and page cache shadow entries need to store memcg
4058 * references in constrained space, so we maintain an ID space that is
4059 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4060 * memory-controlled cgroups to 64k.
4062 * However, there usually are many references to the oflline CSS after
4063 * the cgroup has been destroyed, such as page cache or reclaimable
4064 * slab objects, that don't need to hang on to the ID. We want to keep
4065 * those dead CSS from occupying IDs, or we might quickly exhaust the
4066 * relatively small ID space and prevent the creation of new cgroups
4067 * even when there are much fewer than 64k cgroups - possibly none.
4069 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4070 * be freed and recycled when it's no longer needed, which is usually
4071 * when the CSS is offlined.
4073 * The only exception to that are records of swapped out tmpfs/shmem
4074 * pages that need to be attributed to live ancestors on swapin. But
4075 * those references are manageable from userspace.
4078 static DEFINE_IDR(mem_cgroup_idr);
4080 static void mem_cgroup_id_get(struct mem_cgroup *memcg)
4082 atomic_inc(&memcg->id.ref);
4085 static void mem_cgroup_id_put(struct mem_cgroup *memcg)
4087 if (atomic_dec_and_test(&memcg->id.ref)) {
4088 idr_remove(&mem_cgroup_idr, memcg->id.id);
4091 /* Memcg ID pins CSS */
4092 css_put(&memcg->css);
4097 * mem_cgroup_from_id - look up a memcg from a memcg id
4098 * @id: the memcg id to look up
4100 * Caller must hold rcu_read_lock().
4102 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4104 WARN_ON_ONCE(!rcu_read_lock_held());
4105 return idr_find(&mem_cgroup_idr, id);
4108 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4110 struct mem_cgroup_per_node *pn;
4111 struct mem_cgroup_per_zone *mz;
4112 int zone, tmp = node;
4114 * This routine is called against possible nodes.
4115 * But it's BUG to call kmalloc() against offline node.
4117 * TODO: this routine can waste much memory for nodes which will
4118 * never be onlined. It's better to use memory hotplug callback
4121 if (!node_state(node, N_NORMAL_MEMORY))
4123 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4127 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4128 mz = &pn->zoneinfo[zone];
4129 lruvec_init(&mz->lruvec);
4130 mz->usage_in_excess = 0;
4131 mz->on_tree = false;
4134 memcg->nodeinfo[node] = pn;
4138 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4140 kfree(memcg->nodeinfo[node]);
4143 static void mem_cgroup_free(struct mem_cgroup *memcg)
4147 memcg_wb_domain_exit(memcg);
4149 free_mem_cgroup_per_zone_info(memcg, node);
4150 free_percpu(memcg->stat);
4154 static struct mem_cgroup *mem_cgroup_alloc(void)
4156 struct mem_cgroup *memcg;
4160 size = sizeof(struct mem_cgroup);
4161 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4163 memcg = kzalloc(size, GFP_KERNEL);
4167 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4168 1, MEM_CGROUP_ID_MAX,
4170 if (memcg->id.id < 0)
4173 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4178 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4181 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4184 INIT_WORK(&memcg->high_work, high_work_func);
4185 memcg->last_scanned_node = MAX_NUMNODES;
4186 INIT_LIST_HEAD(&memcg->oom_notify);
4187 mutex_init(&memcg->thresholds_lock);
4188 spin_lock_init(&memcg->move_lock);
4189 vmpressure_init(&memcg->vmpressure);
4190 INIT_LIST_HEAD(&memcg->event_list);
4191 spin_lock_init(&memcg->event_list_lock);
4192 memcg->socket_pressure = jiffies;
4194 memcg->kmemcg_id = -1;
4196 #ifdef CONFIG_CGROUP_WRITEBACK
4197 INIT_LIST_HEAD(&memcg->cgwb_list);
4199 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4202 if (memcg->id.id > 0)
4203 idr_remove(&mem_cgroup_idr, memcg->id.id);
4204 mem_cgroup_free(memcg);
4208 static struct cgroup_subsys_state * __ref
4209 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4211 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4212 struct mem_cgroup *memcg;
4213 long error = -ENOMEM;
4215 memcg = mem_cgroup_alloc();
4217 return ERR_PTR(error);
4219 memcg->high = PAGE_COUNTER_MAX;
4220 memcg->soft_limit = PAGE_COUNTER_MAX;
4222 memcg->swappiness = mem_cgroup_swappiness(parent);
4223 memcg->oom_kill_disable = parent->oom_kill_disable;
4225 if (parent && parent->use_hierarchy) {
4226 memcg->use_hierarchy = true;
4227 page_counter_init(&memcg->memory, &parent->memory);
4228 page_counter_init(&memcg->swap, &parent->swap);
4229 page_counter_init(&memcg->memsw, &parent->memsw);
4230 page_counter_init(&memcg->kmem, &parent->kmem);
4231 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4233 page_counter_init(&memcg->memory, NULL);
4234 page_counter_init(&memcg->swap, NULL);
4235 page_counter_init(&memcg->memsw, NULL);
4236 page_counter_init(&memcg->kmem, NULL);
4237 page_counter_init(&memcg->tcpmem, NULL);
4239 * Deeper hierachy with use_hierarchy == false doesn't make
4240 * much sense so let cgroup subsystem know about this
4241 * unfortunate state in our controller.
4243 if (parent != root_mem_cgroup)
4244 memory_cgrp_subsys.broken_hierarchy = true;
4247 /* The following stuff does not apply to the root */
4249 root_mem_cgroup = memcg;
4253 error = memcg_online_kmem(memcg);
4257 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4258 static_branch_inc(&memcg_sockets_enabled_key);
4262 mem_cgroup_free(memcg);
4263 return ERR_PTR(-ENOMEM);
4266 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4268 /* Online state pins memcg ID, memcg ID pins CSS */
4269 mem_cgroup_id_get(mem_cgroup_from_css(css));
4274 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4276 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4277 struct mem_cgroup_event *event, *tmp;
4280 * Unregister events and notify userspace.
4281 * Notify userspace about cgroup removing only after rmdir of cgroup
4282 * directory to avoid race between userspace and kernelspace.
4284 spin_lock(&memcg->event_list_lock);
4285 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4286 list_del_init(&event->list);
4287 schedule_work(&event->remove);
4289 spin_unlock(&memcg->event_list_lock);
4291 memcg_offline_kmem(memcg);
4292 wb_memcg_offline(memcg);
4294 mem_cgroup_id_put(memcg);
4297 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4299 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4301 invalidate_reclaim_iterators(memcg);
4304 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4306 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4308 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4309 static_branch_dec(&memcg_sockets_enabled_key);
4311 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4312 static_branch_dec(&memcg_sockets_enabled_key);
4314 vmpressure_cleanup(&memcg->vmpressure);
4315 cancel_work_sync(&memcg->high_work);
4316 mem_cgroup_remove_from_trees(memcg);
4317 memcg_free_kmem(memcg);
4318 mem_cgroup_free(memcg);
4322 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4323 * @css: the target css
4325 * Reset the states of the mem_cgroup associated with @css. This is
4326 * invoked when the userland requests disabling on the default hierarchy
4327 * but the memcg is pinned through dependency. The memcg should stop
4328 * applying policies and should revert to the vanilla state as it may be
4329 * made visible again.
4331 * The current implementation only resets the essential configurations.
4332 * This needs to be expanded to cover all the visible parts.
4334 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4336 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4338 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4339 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4340 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4341 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4342 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4344 memcg->high = PAGE_COUNTER_MAX;
4345 memcg->soft_limit = PAGE_COUNTER_MAX;
4346 memcg_wb_domain_size_changed(memcg);
4350 /* Handlers for move charge at task migration. */
4351 static int mem_cgroup_do_precharge(unsigned long count)
4355 /* Try a single bulk charge without reclaim first, kswapd may wake */
4356 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4358 mc.precharge += count;
4362 /* Try charges one by one with reclaim */
4364 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4378 enum mc_target_type {
4384 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4385 unsigned long addr, pte_t ptent)
4387 struct page *page = vm_normal_page(vma, addr, ptent);
4389 if (!page || !page_mapped(page))
4391 if (PageAnon(page)) {
4392 if (!(mc.flags & MOVE_ANON))
4395 if (!(mc.flags & MOVE_FILE))
4398 if (!get_page_unless_zero(page))
4405 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4406 pte_t ptent, swp_entry_t *entry)
4408 struct page *page = NULL;
4409 swp_entry_t ent = pte_to_swp_entry(ptent);
4411 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4414 * Because lookup_swap_cache() updates some statistics counter,
4415 * we call find_get_page() with swapper_space directly.
4417 page = find_get_page(swap_address_space(ent), ent.val);
4418 if (do_memsw_account())
4419 entry->val = ent.val;
4424 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4425 pte_t ptent, swp_entry_t *entry)
4431 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4432 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4434 struct page *page = NULL;
4435 struct address_space *mapping;
4438 if (!vma->vm_file) /* anonymous vma */
4440 if (!(mc.flags & MOVE_FILE))
4443 mapping = vma->vm_file->f_mapping;
4444 pgoff = linear_page_index(vma, addr);
4446 /* page is moved even if it's not RSS of this task(page-faulted). */
4448 /* shmem/tmpfs may report page out on swap: account for that too. */
4449 if (shmem_mapping(mapping)) {
4450 page = find_get_entry(mapping, pgoff);
4451 if (radix_tree_exceptional_entry(page)) {
4452 swp_entry_t swp = radix_to_swp_entry(page);
4453 if (do_memsw_account())
4455 page = find_get_page(swap_address_space(swp), swp.val);
4458 page = find_get_page(mapping, pgoff);
4460 page = find_get_page(mapping, pgoff);
4466 * mem_cgroup_move_account - move account of the page
4468 * @compound: charge the page as compound or small page
4469 * @from: mem_cgroup which the page is moved from.
4470 * @to: mem_cgroup which the page is moved to. @from != @to.
4472 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4474 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4477 static int mem_cgroup_move_account(struct page *page,
4479 struct mem_cgroup *from,
4480 struct mem_cgroup *to)
4482 unsigned long flags;
4483 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4487 VM_BUG_ON(from == to);
4488 VM_BUG_ON_PAGE(PageLRU(page), page);
4489 VM_BUG_ON(compound && !PageTransHuge(page));
4492 * Prevent mem_cgroup_migrate() from looking at
4493 * page->mem_cgroup of its source page while we change it.
4496 if (!trylock_page(page))
4500 if (page->mem_cgroup != from)
4503 anon = PageAnon(page);
4505 spin_lock_irqsave(&from->move_lock, flags);
4507 if (!anon && page_mapped(page)) {
4508 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4510 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4515 * move_lock grabbed above and caller set from->moving_account, so
4516 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4517 * So mapping should be stable for dirty pages.
4519 if (!anon && PageDirty(page)) {
4520 struct address_space *mapping = page_mapping(page);
4522 if (mapping_cap_account_dirty(mapping)) {
4523 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4525 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4530 if (PageWriteback(page)) {
4531 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4533 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4538 * It is safe to change page->mem_cgroup here because the page
4539 * is referenced, charged, and isolated - we can't race with
4540 * uncharging, charging, migration, or LRU putback.
4543 /* caller should have done css_get */
4544 page->mem_cgroup = to;
4545 spin_unlock_irqrestore(&from->move_lock, flags);
4549 local_irq_disable();
4550 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4551 memcg_check_events(to, page);
4552 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4553 memcg_check_events(from, page);
4562 * get_mctgt_type - get target type of moving charge
4563 * @vma: the vma the pte to be checked belongs
4564 * @addr: the address corresponding to the pte to be checked
4565 * @ptent: the pte to be checked
4566 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4569 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4570 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4571 * move charge. if @target is not NULL, the page is stored in target->page
4572 * with extra refcnt got(Callers should handle it).
4573 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4574 * target for charge migration. if @target is not NULL, the entry is stored
4577 * Called with pte lock held.
4580 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4581 unsigned long addr, pte_t ptent, union mc_target *target)
4583 struct page *page = NULL;
4584 enum mc_target_type ret = MC_TARGET_NONE;
4585 swp_entry_t ent = { .val = 0 };
4587 if (pte_present(ptent))
4588 page = mc_handle_present_pte(vma, addr, ptent);
4589 else if (is_swap_pte(ptent))
4590 page = mc_handle_swap_pte(vma, ptent, &ent);
4591 else if (pte_none(ptent))
4592 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4594 if (!page && !ent.val)
4598 * Do only loose check w/o serialization.
4599 * mem_cgroup_move_account() checks the page is valid or
4600 * not under LRU exclusion.
4602 if (page->mem_cgroup == mc.from) {
4603 ret = MC_TARGET_PAGE;
4605 target->page = page;
4607 if (!ret || !target)
4610 /* There is a swap entry and a page doesn't exist or isn't charged */
4611 if (ent.val && !ret &&
4612 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4613 ret = MC_TARGET_SWAP;
4620 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4622 * We don't consider swapping or file mapped pages because THP does not
4623 * support them for now.
4624 * Caller should make sure that pmd_trans_huge(pmd) is true.
4626 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4627 unsigned long addr, pmd_t pmd, union mc_target *target)
4629 struct page *page = NULL;
4630 enum mc_target_type ret = MC_TARGET_NONE;
4632 page = pmd_page(pmd);
4633 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4634 if (!(mc.flags & MOVE_ANON))
4636 if (page->mem_cgroup == mc.from) {
4637 ret = MC_TARGET_PAGE;
4640 target->page = page;
4646 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4647 unsigned long addr, pmd_t pmd, union mc_target *target)
4649 return MC_TARGET_NONE;
4653 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4654 unsigned long addr, unsigned long end,
4655 struct mm_walk *walk)
4657 struct vm_area_struct *vma = walk->vma;
4661 ptl = pmd_trans_huge_lock(pmd, vma);
4663 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4664 mc.precharge += HPAGE_PMD_NR;
4669 if (pmd_trans_unstable(pmd))
4671 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4672 for (; addr != end; pte++, addr += PAGE_SIZE)
4673 if (get_mctgt_type(vma, addr, *pte, NULL))
4674 mc.precharge++; /* increment precharge temporarily */
4675 pte_unmap_unlock(pte - 1, ptl);
4681 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4683 unsigned long precharge;
4685 struct mm_walk mem_cgroup_count_precharge_walk = {
4686 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4689 down_read(&mm->mmap_sem);
4690 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4691 up_read(&mm->mmap_sem);
4693 precharge = mc.precharge;
4699 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4701 unsigned long precharge = mem_cgroup_count_precharge(mm);
4703 VM_BUG_ON(mc.moving_task);
4704 mc.moving_task = current;
4705 return mem_cgroup_do_precharge(precharge);
4708 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4709 static void __mem_cgroup_clear_mc(void)
4711 struct mem_cgroup *from = mc.from;
4712 struct mem_cgroup *to = mc.to;
4714 /* we must uncharge all the leftover precharges from mc.to */
4716 cancel_charge(mc.to, mc.precharge);
4720 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4721 * we must uncharge here.
4723 if (mc.moved_charge) {
4724 cancel_charge(mc.from, mc.moved_charge);
4725 mc.moved_charge = 0;
4727 /* we must fixup refcnts and charges */
4728 if (mc.moved_swap) {
4729 /* uncharge swap account from the old cgroup */
4730 if (!mem_cgroup_is_root(mc.from))
4731 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4734 * we charged both to->memory and to->memsw, so we
4735 * should uncharge to->memory.
4737 if (!mem_cgroup_is_root(mc.to))
4738 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4740 css_put_many(&mc.from->css, mc.moved_swap);
4742 /* we've already done css_get(mc.to) */
4745 memcg_oom_recover(from);
4746 memcg_oom_recover(to);
4747 wake_up_all(&mc.waitq);
4750 static void mem_cgroup_clear_mc(void)
4752 struct mm_struct *mm = mc.mm;
4755 * we must clear moving_task before waking up waiters at the end of
4758 mc.moving_task = NULL;
4759 __mem_cgroup_clear_mc();
4760 spin_lock(&mc.lock);
4764 spin_unlock(&mc.lock);
4769 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4771 struct cgroup_subsys_state *css;
4772 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4773 struct mem_cgroup *from;
4774 struct task_struct *leader, *p;
4775 struct mm_struct *mm;
4776 unsigned long move_flags;
4779 /* charge immigration isn't supported on the default hierarchy */
4780 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4784 * Multi-process migrations only happen on the default hierarchy
4785 * where charge immigration is not used. Perform charge
4786 * immigration if @tset contains a leader and whine if there are
4790 cgroup_taskset_for_each_leader(leader, css, tset) {
4793 memcg = mem_cgroup_from_css(css);
4799 * We are now commited to this value whatever it is. Changes in this
4800 * tunable will only affect upcoming migrations, not the current one.
4801 * So we need to save it, and keep it going.
4803 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4807 from = mem_cgroup_from_task(p);
4809 VM_BUG_ON(from == memcg);
4811 mm = get_task_mm(p);
4814 /* We move charges only when we move a owner of the mm */
4815 if (mm->owner == p) {
4818 VM_BUG_ON(mc.precharge);
4819 VM_BUG_ON(mc.moved_charge);
4820 VM_BUG_ON(mc.moved_swap);
4822 spin_lock(&mc.lock);
4826 mc.flags = move_flags;
4827 spin_unlock(&mc.lock);
4828 /* We set mc.moving_task later */
4830 ret = mem_cgroup_precharge_mc(mm);
4832 mem_cgroup_clear_mc();
4839 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4842 mem_cgroup_clear_mc();
4845 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4846 unsigned long addr, unsigned long end,
4847 struct mm_walk *walk)
4850 struct vm_area_struct *vma = walk->vma;
4853 enum mc_target_type target_type;
4854 union mc_target target;
4857 ptl = pmd_trans_huge_lock(pmd, vma);
4859 if (mc.precharge < HPAGE_PMD_NR) {
4863 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4864 if (target_type == MC_TARGET_PAGE) {
4866 if (!isolate_lru_page(page)) {
4867 if (!mem_cgroup_move_account(page, true,
4869 mc.precharge -= HPAGE_PMD_NR;
4870 mc.moved_charge += HPAGE_PMD_NR;
4872 putback_lru_page(page);
4880 if (pmd_trans_unstable(pmd))
4883 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4884 for (; addr != end; addr += PAGE_SIZE) {
4885 pte_t ptent = *(pte++);
4891 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4892 case MC_TARGET_PAGE:
4895 * We can have a part of the split pmd here. Moving it
4896 * can be done but it would be too convoluted so simply
4897 * ignore such a partial THP and keep it in original
4898 * memcg. There should be somebody mapping the head.
4900 if (PageTransCompound(page))
4902 if (isolate_lru_page(page))
4904 if (!mem_cgroup_move_account(page, false,
4907 /* we uncharge from mc.from later. */
4910 putback_lru_page(page);
4911 put: /* get_mctgt_type() gets the page */
4914 case MC_TARGET_SWAP:
4916 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4918 /* we fixup refcnts and charges later. */
4926 pte_unmap_unlock(pte - 1, ptl);
4931 * We have consumed all precharges we got in can_attach().
4932 * We try charge one by one, but don't do any additional
4933 * charges to mc.to if we have failed in charge once in attach()
4936 ret = mem_cgroup_do_precharge(1);
4944 static void mem_cgroup_move_charge(void)
4946 struct mm_walk mem_cgroup_move_charge_walk = {
4947 .pmd_entry = mem_cgroup_move_charge_pte_range,
4951 lru_add_drain_all();
4953 * Signal lock_page_memcg() to take the memcg's move_lock
4954 * while we're moving its pages to another memcg. Then wait
4955 * for already started RCU-only updates to finish.
4957 atomic_inc(&mc.from->moving_account);
4960 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
4962 * Someone who are holding the mmap_sem might be waiting in
4963 * waitq. So we cancel all extra charges, wake up all waiters,
4964 * and retry. Because we cancel precharges, we might not be able
4965 * to move enough charges, but moving charge is a best-effort
4966 * feature anyway, so it wouldn't be a big problem.
4968 __mem_cgroup_clear_mc();
4973 * When we have consumed all precharges and failed in doing
4974 * additional charge, the page walk just aborts.
4976 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
4977 up_read(&mc.mm->mmap_sem);
4978 atomic_dec(&mc.from->moving_account);
4981 static void mem_cgroup_move_task(void)
4984 mem_cgroup_move_charge();
4985 mem_cgroup_clear_mc();
4988 #else /* !CONFIG_MMU */
4989 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4993 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4996 static void mem_cgroup_move_task(void)
5002 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5003 * to verify whether we're attached to the default hierarchy on each mount
5006 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5009 * use_hierarchy is forced on the default hierarchy. cgroup core
5010 * guarantees that @root doesn't have any children, so turning it
5011 * on for the root memcg is enough.
5013 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5014 root_mem_cgroup->use_hierarchy = true;
5016 root_mem_cgroup->use_hierarchy = false;
5019 static u64 memory_current_read(struct cgroup_subsys_state *css,
5022 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5024 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5027 static int memory_low_show(struct seq_file *m, void *v)
5029 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5030 unsigned long low = READ_ONCE(memcg->low);
5032 if (low == PAGE_COUNTER_MAX)
5033 seq_puts(m, "max\n");
5035 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5040 static ssize_t memory_low_write(struct kernfs_open_file *of,
5041 char *buf, size_t nbytes, loff_t off)
5043 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5047 buf = strstrip(buf);
5048 err = page_counter_memparse(buf, "max", &low);
5057 static int memory_high_show(struct seq_file *m, void *v)
5059 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5060 unsigned long high = READ_ONCE(memcg->high);
5062 if (high == PAGE_COUNTER_MAX)
5063 seq_puts(m, "max\n");
5065 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5070 static ssize_t memory_high_write(struct kernfs_open_file *of,
5071 char *buf, size_t nbytes, loff_t off)
5073 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5074 unsigned long nr_pages;
5078 buf = strstrip(buf);
5079 err = page_counter_memparse(buf, "max", &high);
5085 nr_pages = page_counter_read(&memcg->memory);
5086 if (nr_pages > high)
5087 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5090 memcg_wb_domain_size_changed(memcg);
5094 static int memory_max_show(struct seq_file *m, void *v)
5096 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5097 unsigned long max = READ_ONCE(memcg->memory.limit);
5099 if (max == PAGE_COUNTER_MAX)
5100 seq_puts(m, "max\n");
5102 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5107 static ssize_t memory_max_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));
5111 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5112 bool drained = false;
5116 buf = strstrip(buf);
5117 err = page_counter_memparse(buf, "max", &max);
5121 xchg(&memcg->memory.limit, max);
5124 unsigned long nr_pages = page_counter_read(&memcg->memory);
5126 if (nr_pages <= max)
5129 if (signal_pending(current)) {
5135 drain_all_stock(memcg);
5141 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5147 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5148 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5152 memcg_wb_domain_size_changed(memcg);
5156 static int memory_events_show(struct seq_file *m, void *v)
5158 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5160 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5161 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5162 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5163 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5168 static int memory_stat_show(struct seq_file *m, void *v)
5170 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5171 unsigned long stat[MEMCG_NR_STAT];
5172 unsigned long events[MEMCG_NR_EVENTS];
5176 * Provide statistics on the state of the memory subsystem as
5177 * well as cumulative event counters that show past behavior.
5179 * This list is ordered following a combination of these gradients:
5180 * 1) generic big picture -> specifics and details
5181 * 2) reflecting userspace activity -> reflecting kernel heuristics
5183 * Current memory state:
5186 tree_stat(memcg, stat);
5187 tree_events(memcg, events);
5189 seq_printf(m, "anon %llu\n",
5190 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5191 seq_printf(m, "file %llu\n",
5192 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5193 seq_printf(m, "kernel_stack %llu\n",
5194 (u64)stat[MEMCG_KERNEL_STACK] * PAGE_SIZE);
5195 seq_printf(m, "slab %llu\n",
5196 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5197 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5198 seq_printf(m, "sock %llu\n",
5199 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5201 seq_printf(m, "file_mapped %llu\n",
5202 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5203 seq_printf(m, "file_dirty %llu\n",
5204 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5205 seq_printf(m, "file_writeback %llu\n",
5206 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5208 for (i = 0; i < NR_LRU_LISTS; i++) {
5209 struct mem_cgroup *mi;
5210 unsigned long val = 0;
5212 for_each_mem_cgroup_tree(mi, memcg)
5213 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5214 seq_printf(m, "%s %llu\n",
5215 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5218 seq_printf(m, "slab_reclaimable %llu\n",
5219 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5220 seq_printf(m, "slab_unreclaimable %llu\n",
5221 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5223 /* Accumulated memory events */
5225 seq_printf(m, "pgfault %lu\n",
5226 events[MEM_CGROUP_EVENTS_PGFAULT]);
5227 seq_printf(m, "pgmajfault %lu\n",
5228 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5233 static struct cftype memory_files[] = {
5236 .flags = CFTYPE_NOT_ON_ROOT,
5237 .read_u64 = memory_current_read,
5241 .flags = CFTYPE_NOT_ON_ROOT,
5242 .seq_show = memory_low_show,
5243 .write = memory_low_write,
5247 .flags = CFTYPE_NOT_ON_ROOT,
5248 .seq_show = memory_high_show,
5249 .write = memory_high_write,
5253 .flags = CFTYPE_NOT_ON_ROOT,
5254 .seq_show = memory_max_show,
5255 .write = memory_max_write,
5259 .flags = CFTYPE_NOT_ON_ROOT,
5260 .file_offset = offsetof(struct mem_cgroup, events_file),
5261 .seq_show = memory_events_show,
5265 .flags = CFTYPE_NOT_ON_ROOT,
5266 .seq_show = memory_stat_show,
5271 struct cgroup_subsys memory_cgrp_subsys = {
5272 .css_alloc = mem_cgroup_css_alloc,
5273 .css_online = mem_cgroup_css_online,
5274 .css_offline = mem_cgroup_css_offline,
5275 .css_released = mem_cgroup_css_released,
5276 .css_free = mem_cgroup_css_free,
5277 .css_reset = mem_cgroup_css_reset,
5278 .can_attach = mem_cgroup_can_attach,
5279 .cancel_attach = mem_cgroup_cancel_attach,
5280 .post_attach = mem_cgroup_move_task,
5281 .bind = mem_cgroup_bind,
5282 .dfl_cftypes = memory_files,
5283 .legacy_cftypes = mem_cgroup_legacy_files,
5288 * mem_cgroup_low - check if memory consumption is below the normal range
5289 * @root: the highest ancestor to consider
5290 * @memcg: the memory cgroup to check
5292 * Returns %true if memory consumption of @memcg, and that of all
5293 * configurable ancestors up to @root, is below the normal range.
5295 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5297 if (mem_cgroup_disabled())
5301 * The toplevel group doesn't have a configurable range, so
5302 * it's never low when looked at directly, and it is not
5303 * considered an ancestor when assessing the hierarchy.
5306 if (memcg == root_mem_cgroup)
5309 if (page_counter_read(&memcg->memory) >= memcg->low)
5312 while (memcg != root) {
5313 memcg = parent_mem_cgroup(memcg);
5315 if (memcg == root_mem_cgroup)
5318 if (page_counter_read(&memcg->memory) >= memcg->low)
5325 * mem_cgroup_try_charge - try charging a page
5326 * @page: page to charge
5327 * @mm: mm context of the victim
5328 * @gfp_mask: reclaim mode
5329 * @memcgp: charged memcg return
5330 * @compound: charge the page as compound or small page
5332 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5333 * pages according to @gfp_mask if necessary.
5335 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5336 * Otherwise, an error code is returned.
5338 * After page->mapping has been set up, the caller must finalize the
5339 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5340 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5342 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5343 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5346 struct mem_cgroup *memcg = NULL;
5347 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5350 if (mem_cgroup_disabled())
5353 if (PageSwapCache(page)) {
5355 * Every swap fault against a single page tries to charge the
5356 * page, bail as early as possible. shmem_unuse() encounters
5357 * already charged pages, too. The USED bit is protected by
5358 * the page lock, which serializes swap cache removal, which
5359 * in turn serializes uncharging.
5361 VM_BUG_ON_PAGE(!PageLocked(page), page);
5362 if (page->mem_cgroup)
5365 if (do_swap_account) {
5366 swp_entry_t ent = { .val = page_private(page), };
5367 unsigned short id = lookup_swap_cgroup_id(ent);
5370 memcg = mem_cgroup_from_id(id);
5371 if (memcg && !css_tryget_online(&memcg->css))
5378 memcg = get_mem_cgroup_from_mm(mm);
5380 ret = try_charge(memcg, gfp_mask, nr_pages);
5382 css_put(&memcg->css);
5389 * mem_cgroup_commit_charge - commit a page charge
5390 * @page: page to charge
5391 * @memcg: memcg to charge the page to
5392 * @lrucare: page might be on LRU already
5393 * @compound: charge the page as compound or small page
5395 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5396 * after page->mapping has been set up. This must happen atomically
5397 * as part of the page instantiation, i.e. under the page table lock
5398 * for anonymous pages, under the page lock for page and swap cache.
5400 * In addition, the page must not be on the LRU during the commit, to
5401 * prevent racing with task migration. If it might be, use @lrucare.
5403 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5405 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5406 bool lrucare, bool compound)
5408 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5410 VM_BUG_ON_PAGE(!page->mapping, page);
5411 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5413 if (mem_cgroup_disabled())
5416 * Swap faults will attempt to charge the same page multiple
5417 * times. But reuse_swap_page() might have removed the page
5418 * from swapcache already, so we can't check PageSwapCache().
5423 commit_charge(page, memcg, lrucare);
5425 local_irq_disable();
5426 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5427 memcg_check_events(memcg, page);
5430 if (do_memsw_account() && PageSwapCache(page)) {
5431 swp_entry_t entry = { .val = page_private(page) };
5433 * The swap entry might not get freed for a long time,
5434 * let's not wait for it. The page already received a
5435 * memory+swap charge, drop the swap entry duplicate.
5437 mem_cgroup_uncharge_swap(entry);
5442 * mem_cgroup_cancel_charge - cancel a page charge
5443 * @page: page to charge
5444 * @memcg: memcg to charge the page to
5445 * @compound: charge the page as compound or small page
5447 * Cancel a charge transaction started by mem_cgroup_try_charge().
5449 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5452 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5454 if (mem_cgroup_disabled())
5457 * Swap faults will attempt to charge the same page multiple
5458 * times. But reuse_swap_page() might have removed the page
5459 * from swapcache already, so we can't check PageSwapCache().
5464 cancel_charge(memcg, nr_pages);
5467 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5468 unsigned long nr_anon, unsigned long nr_file,
5469 unsigned long nr_huge, unsigned long nr_kmem,
5470 struct page *dummy_page)
5472 unsigned long nr_pages = nr_anon + nr_file + nr_kmem;
5473 unsigned long flags;
5475 if (!mem_cgroup_is_root(memcg)) {
5476 page_counter_uncharge(&memcg->memory, nr_pages);
5477 if (do_memsw_account())
5478 page_counter_uncharge(&memcg->memsw, nr_pages);
5479 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && nr_kmem)
5480 page_counter_uncharge(&memcg->kmem, nr_kmem);
5481 memcg_oom_recover(memcg);
5484 local_irq_save(flags);
5485 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5486 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5487 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5488 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5489 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5490 memcg_check_events(memcg, dummy_page);
5491 local_irq_restore(flags);
5493 if (!mem_cgroup_is_root(memcg))
5494 css_put_many(&memcg->css, nr_pages);
5497 static void uncharge_list(struct list_head *page_list)
5499 struct mem_cgroup *memcg = NULL;
5500 unsigned long nr_anon = 0;
5501 unsigned long nr_file = 0;
5502 unsigned long nr_huge = 0;
5503 unsigned long nr_kmem = 0;
5504 unsigned long pgpgout = 0;
5505 struct list_head *next;
5509 * Note that the list can be a single page->lru; hence the
5510 * do-while loop instead of a simple list_for_each_entry().
5512 next = page_list->next;
5514 page = list_entry(next, struct page, lru);
5515 next = page->lru.next;
5517 VM_BUG_ON_PAGE(PageLRU(page), page);
5518 VM_BUG_ON_PAGE(page_count(page), page);
5520 if (!page->mem_cgroup)
5524 * Nobody should be changing or seriously looking at
5525 * page->mem_cgroup at this point, we have fully
5526 * exclusive access to the page.
5529 if (memcg != page->mem_cgroup) {
5531 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5532 nr_huge, nr_kmem, page);
5533 pgpgout = nr_anon = nr_file =
5534 nr_huge = nr_kmem = 0;
5536 memcg = page->mem_cgroup;
5539 if (!PageKmemcg(page)) {
5540 unsigned int nr_pages = 1;
5542 if (PageTransHuge(page)) {
5543 nr_pages <<= compound_order(page);
5544 nr_huge += nr_pages;
5547 nr_anon += nr_pages;
5549 nr_file += nr_pages;
5552 nr_kmem += 1 << compound_order(page);
5554 page->mem_cgroup = NULL;
5555 } while (next != page_list);
5558 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5559 nr_huge, nr_kmem, page);
5563 * mem_cgroup_uncharge - uncharge a page
5564 * @page: page to uncharge
5566 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5567 * mem_cgroup_commit_charge().
5569 void mem_cgroup_uncharge(struct page *page)
5571 if (mem_cgroup_disabled())
5574 /* Don't touch page->lru of any random page, pre-check: */
5575 if (!page->mem_cgroup)
5578 INIT_LIST_HEAD(&page->lru);
5579 uncharge_list(&page->lru);
5583 * mem_cgroup_uncharge_list - uncharge a list of page
5584 * @page_list: list of pages to uncharge
5586 * Uncharge a list of pages previously charged with
5587 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5589 void mem_cgroup_uncharge_list(struct list_head *page_list)
5591 if (mem_cgroup_disabled())
5594 if (!list_empty(page_list))
5595 uncharge_list(page_list);
5599 * mem_cgroup_migrate - charge a page's replacement
5600 * @oldpage: currently circulating page
5601 * @newpage: replacement page
5603 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5604 * be uncharged upon free.
5606 * Both pages must be locked, @newpage->mapping must be set up.
5608 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5610 struct mem_cgroup *memcg;
5611 unsigned int nr_pages;
5613 unsigned long flags;
5615 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5616 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5617 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5618 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5621 if (mem_cgroup_disabled())
5624 /* Page cache replacement: new page already charged? */
5625 if (newpage->mem_cgroup)
5628 /* Swapcache readahead pages can get replaced before being charged */
5629 memcg = oldpage->mem_cgroup;
5633 /* Force-charge the new page. The old one will be freed soon */
5634 compound = PageTransHuge(newpage);
5635 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5637 page_counter_charge(&memcg->memory, nr_pages);
5638 if (do_memsw_account())
5639 page_counter_charge(&memcg->memsw, nr_pages);
5640 css_get_many(&memcg->css, nr_pages);
5642 commit_charge(newpage, memcg, false);
5644 local_irq_save(flags);
5645 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5646 memcg_check_events(memcg, newpage);
5647 local_irq_restore(flags);
5650 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5651 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5653 void sock_update_memcg(struct sock *sk)
5655 struct mem_cgroup *memcg;
5657 /* Socket cloning can throw us here with sk_cgrp already
5658 * filled. It won't however, necessarily happen from
5659 * process context. So the test for root memcg given
5660 * the current task's memcg won't help us in this case.
5662 * Respecting the original socket's memcg is a better
5663 * decision in this case.
5666 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5667 css_get(&sk->sk_memcg->css);
5672 memcg = mem_cgroup_from_task(current);
5673 if (memcg == root_mem_cgroup)
5675 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5677 if (css_tryget_online(&memcg->css))
5678 sk->sk_memcg = memcg;
5682 EXPORT_SYMBOL(sock_update_memcg);
5684 void sock_release_memcg(struct sock *sk)
5686 WARN_ON(!sk->sk_memcg);
5687 css_put(&sk->sk_memcg->css);
5691 * mem_cgroup_charge_skmem - charge socket memory
5692 * @memcg: memcg to charge
5693 * @nr_pages: number of pages to charge
5695 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5696 * @memcg's configured limit, %false if the charge had to be forced.
5698 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5700 gfp_t gfp_mask = GFP_KERNEL;
5702 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5703 struct page_counter *fail;
5705 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5706 memcg->tcpmem_pressure = 0;
5709 page_counter_charge(&memcg->tcpmem, nr_pages);
5710 memcg->tcpmem_pressure = 1;
5714 /* Don't block in the packet receive path */
5716 gfp_mask = GFP_NOWAIT;
5718 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5720 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5723 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5728 * mem_cgroup_uncharge_skmem - uncharge socket memory
5729 * @memcg - memcg to uncharge
5730 * @nr_pages - number of pages to uncharge
5732 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5734 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5735 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5739 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5741 page_counter_uncharge(&memcg->memory, nr_pages);
5742 css_put_many(&memcg->css, nr_pages);
5745 static int __init cgroup_memory(char *s)
5749 while ((token = strsep(&s, ",")) != NULL) {
5752 if (!strcmp(token, "nosocket"))
5753 cgroup_memory_nosocket = true;
5754 if (!strcmp(token, "nokmem"))
5755 cgroup_memory_nokmem = true;
5759 __setup("cgroup.memory=", cgroup_memory);
5762 * subsys_initcall() for memory controller.
5764 * Some parts like hotcpu_notifier() have to be initialized from this context
5765 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5766 * everything that doesn't depend on a specific mem_cgroup structure should
5767 * be initialized from here.
5769 static int __init mem_cgroup_init(void)
5773 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5775 for_each_possible_cpu(cpu)
5776 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5779 for_each_node(node) {
5780 struct mem_cgroup_tree_per_node *rtpn;
5783 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5784 node_online(node) ? node : NUMA_NO_NODE);
5786 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5787 struct mem_cgroup_tree_per_zone *rtpz;
5789 rtpz = &rtpn->rb_tree_per_zone[zone];
5790 rtpz->rb_root = RB_ROOT;
5791 spin_lock_init(&rtpz->lock);
5793 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5798 subsys_initcall(mem_cgroup_init);
5800 #ifdef CONFIG_MEMCG_SWAP
5802 * mem_cgroup_swapout - transfer a memsw charge to swap
5803 * @page: page whose memsw charge to transfer
5804 * @entry: swap entry to move the charge to
5806 * Transfer the memsw charge of @page to @entry.
5808 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5810 struct mem_cgroup *memcg;
5811 unsigned short oldid;
5813 VM_BUG_ON_PAGE(PageLRU(page), page);
5814 VM_BUG_ON_PAGE(page_count(page), page);
5816 if (!do_memsw_account())
5819 memcg = page->mem_cgroup;
5821 /* Readahead page, never charged */
5825 mem_cgroup_id_get(memcg);
5826 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5827 VM_BUG_ON_PAGE(oldid, page);
5828 mem_cgroup_swap_statistics(memcg, true);
5830 page->mem_cgroup = NULL;
5832 if (!mem_cgroup_is_root(memcg))
5833 page_counter_uncharge(&memcg->memory, 1);
5836 * Interrupts should be disabled here because the caller holds the
5837 * mapping->tree_lock lock which is taken with interrupts-off. It is
5838 * important here to have the interrupts disabled because it is the
5839 * only synchronisation we have for udpating the per-CPU variables.
5841 VM_BUG_ON(!irqs_disabled());
5842 mem_cgroup_charge_statistics(memcg, page, false, -1);
5843 memcg_check_events(memcg, page);
5845 if (!mem_cgroup_is_root(memcg))
5846 css_put(&memcg->css);
5850 * mem_cgroup_try_charge_swap - try charging a swap entry
5851 * @page: page being added to swap
5852 * @entry: swap entry to charge
5854 * Try to charge @entry to the memcg that @page belongs to.
5856 * Returns 0 on success, -ENOMEM on failure.
5858 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5860 struct mem_cgroup *memcg;
5861 struct page_counter *counter;
5862 unsigned short oldid;
5864 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5867 memcg = page->mem_cgroup;
5869 /* Readahead page, never charged */
5873 if (!mem_cgroup_is_root(memcg) &&
5874 !page_counter_try_charge(&memcg->swap, 1, &counter))
5877 mem_cgroup_id_get(memcg);
5878 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5879 VM_BUG_ON_PAGE(oldid, page);
5880 mem_cgroup_swap_statistics(memcg, true);
5886 * mem_cgroup_uncharge_swap - uncharge a swap entry
5887 * @entry: swap entry to uncharge
5889 * Drop the swap charge associated with @entry.
5891 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5893 struct mem_cgroup *memcg;
5896 if (!do_swap_account)
5899 id = swap_cgroup_record(entry, 0);
5901 memcg = mem_cgroup_from_id(id);
5903 if (!mem_cgroup_is_root(memcg)) {
5904 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5905 page_counter_uncharge(&memcg->swap, 1);
5907 page_counter_uncharge(&memcg->memsw, 1);
5909 mem_cgroup_swap_statistics(memcg, false);
5910 mem_cgroup_id_put(memcg);
5915 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
5917 long nr_swap_pages = get_nr_swap_pages();
5919 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5920 return nr_swap_pages;
5921 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5922 nr_swap_pages = min_t(long, nr_swap_pages,
5923 READ_ONCE(memcg->swap.limit) -
5924 page_counter_read(&memcg->swap));
5925 return nr_swap_pages;
5928 bool mem_cgroup_swap_full(struct page *page)
5930 struct mem_cgroup *memcg;
5932 VM_BUG_ON_PAGE(!PageLocked(page), page);
5936 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5939 memcg = page->mem_cgroup;
5943 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5944 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
5950 /* for remember boot option*/
5951 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5952 static int really_do_swap_account __initdata = 1;
5954 static int really_do_swap_account __initdata;
5957 static int __init enable_swap_account(char *s)
5959 if (!strcmp(s, "1"))
5960 really_do_swap_account = 1;
5961 else if (!strcmp(s, "0"))
5962 really_do_swap_account = 0;
5965 __setup("swapaccount=", enable_swap_account);
5967 static u64 swap_current_read(struct cgroup_subsys_state *css,
5970 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5972 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
5975 static int swap_max_show(struct seq_file *m, void *v)
5977 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5978 unsigned long max = READ_ONCE(memcg->swap.limit);
5980 if (max == PAGE_COUNTER_MAX)
5981 seq_puts(m, "max\n");
5983 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5988 static ssize_t swap_max_write(struct kernfs_open_file *of,
5989 char *buf, size_t nbytes, loff_t off)
5991 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5995 buf = strstrip(buf);
5996 err = page_counter_memparse(buf, "max", &max);
6000 mutex_lock(&memcg_limit_mutex);
6001 err = page_counter_limit(&memcg->swap, max);
6002 mutex_unlock(&memcg_limit_mutex);
6009 static struct cftype swap_files[] = {
6011 .name = "swap.current",
6012 .flags = CFTYPE_NOT_ON_ROOT,
6013 .read_u64 = swap_current_read,
6017 .flags = CFTYPE_NOT_ON_ROOT,
6018 .seq_show = swap_max_show,
6019 .write = swap_max_write,
6024 static struct cftype memsw_cgroup_files[] = {
6026 .name = "memsw.usage_in_bytes",
6027 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6028 .read_u64 = mem_cgroup_read_u64,
6031 .name = "memsw.max_usage_in_bytes",
6032 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6033 .write = mem_cgroup_reset,
6034 .read_u64 = mem_cgroup_read_u64,
6037 .name = "memsw.limit_in_bytes",
6038 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6039 .write = mem_cgroup_write,
6040 .read_u64 = mem_cgroup_read_u64,
6043 .name = "memsw.failcnt",
6044 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6045 .write = mem_cgroup_reset,
6046 .read_u64 = mem_cgroup_read_u64,
6048 { }, /* terminate */
6051 static int __init mem_cgroup_swap_init(void)
6053 if (!mem_cgroup_disabled() && really_do_swap_account) {
6054 do_swap_account = 1;
6055 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6057 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6058 memsw_cgroup_files));
6062 subsys_initcall(mem_cgroup_swap_init);
6064 #endif /* CONFIG_MEMCG_SWAP */