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 mem_cgroup *from;
211 struct mem_cgroup *to;
213 unsigned long precharge;
214 unsigned long moved_charge;
215 unsigned long moved_swap;
216 struct task_struct *moving_task; /* a task moving charges */
217 wait_queue_head_t waitq; /* a waitq for other context */
219 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
220 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
224 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
225 * limit reclaim to prevent infinite loops, if they ever occur.
227 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
228 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
231 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
232 MEM_CGROUP_CHARGE_TYPE_ANON,
233 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
234 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
238 /* for encoding cft->private value on file */
247 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
248 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
249 #define MEMFILE_ATTR(val) ((val) & 0xffff)
250 /* Used for OOM nofiier */
251 #define OOM_CONTROL (0)
253 /* Some nice accessors for the vmpressure. */
254 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
257 memcg = root_mem_cgroup;
258 return &memcg->vmpressure;
261 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
263 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
266 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
268 return (memcg == root_mem_cgroup);
273 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
274 * The main reason for not using cgroup id for this:
275 * this works better in sparse environments, where we have a lot of memcgs,
276 * but only a few kmem-limited. Or also, if we have, for instance, 200
277 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
278 * 200 entry array for that.
280 * The current size of the caches array is stored in memcg_nr_cache_ids. It
281 * will double each time we have to increase it.
283 static DEFINE_IDA(memcg_cache_ida);
284 int memcg_nr_cache_ids;
286 /* Protects memcg_nr_cache_ids */
287 static DECLARE_RWSEM(memcg_cache_ids_sem);
289 void memcg_get_cache_ids(void)
291 down_read(&memcg_cache_ids_sem);
294 void memcg_put_cache_ids(void)
296 up_read(&memcg_cache_ids_sem);
300 * MIN_SIZE is different than 1, because we would like to avoid going through
301 * the alloc/free process all the time. In a small machine, 4 kmem-limited
302 * cgroups is a reasonable guess. In the future, it could be a parameter or
303 * tunable, but that is strictly not necessary.
305 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
306 * this constant directly from cgroup, but it is understandable that this is
307 * better kept as an internal representation in cgroup.c. In any case, the
308 * cgrp_id space is not getting any smaller, and we don't have to necessarily
309 * increase ours as well if it increases.
311 #define MEMCG_CACHES_MIN_SIZE 4
312 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
315 * A lot of the calls to the cache allocation functions are expected to be
316 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
317 * conditional to this static branch, we'll have to allow modules that does
318 * kmem_cache_alloc and the such to see this symbol as well
320 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
321 EXPORT_SYMBOL(memcg_kmem_enabled_key);
323 #endif /* !CONFIG_SLOB */
325 static struct mem_cgroup_per_zone *
326 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
328 int nid = zone_to_nid(zone);
329 int zid = zone_idx(zone);
331 return &memcg->nodeinfo[nid]->zoneinfo[zid];
335 * mem_cgroup_css_from_page - css of the memcg associated with a page
336 * @page: page of interest
338 * If memcg is bound to the default hierarchy, css of the memcg associated
339 * with @page is returned. The returned css remains associated with @page
340 * until it is released.
342 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
345 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
347 struct mem_cgroup *memcg;
349 memcg = page->mem_cgroup;
351 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
352 memcg = root_mem_cgroup;
358 * page_cgroup_ino - return inode number of the memcg a page is charged to
361 * Look up the closest online ancestor of the memory cgroup @page is charged to
362 * and return its inode number or 0 if @page is not charged to any cgroup. It
363 * is safe to call this function without holding a reference to @page.
365 * Note, this function is inherently racy, because there is nothing to prevent
366 * the cgroup inode from getting torn down and potentially reallocated a moment
367 * after page_cgroup_ino() returns, so it only should be used by callers that
368 * do not care (such as procfs interfaces).
370 ino_t page_cgroup_ino(struct page *page)
372 struct mem_cgroup *memcg;
373 unsigned long ino = 0;
376 memcg = READ_ONCE(page->mem_cgroup);
377 while (memcg && !(memcg->css.flags & CSS_ONLINE))
378 memcg = parent_mem_cgroup(memcg);
380 ino = cgroup_ino(memcg->css.cgroup);
385 static struct mem_cgroup_per_zone *
386 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
388 int nid = page_to_nid(page);
389 int zid = page_zonenum(page);
391 return &memcg->nodeinfo[nid]->zoneinfo[zid];
394 static struct mem_cgroup_tree_per_zone *
395 soft_limit_tree_node_zone(int nid, int zid)
397 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
400 static struct mem_cgroup_tree_per_zone *
401 soft_limit_tree_from_page(struct page *page)
403 int nid = page_to_nid(page);
404 int zid = page_zonenum(page);
406 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
409 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
410 struct mem_cgroup_tree_per_zone *mctz,
411 unsigned long new_usage_in_excess)
413 struct rb_node **p = &mctz->rb_root.rb_node;
414 struct rb_node *parent = NULL;
415 struct mem_cgroup_per_zone *mz_node;
420 mz->usage_in_excess = new_usage_in_excess;
421 if (!mz->usage_in_excess)
425 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
427 if (mz->usage_in_excess < mz_node->usage_in_excess)
430 * We can't avoid mem cgroups that are over their soft
431 * limit by the same amount
433 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
436 rb_link_node(&mz->tree_node, parent, p);
437 rb_insert_color(&mz->tree_node, &mctz->rb_root);
441 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
442 struct mem_cgroup_tree_per_zone *mctz)
446 rb_erase(&mz->tree_node, &mctz->rb_root);
450 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
451 struct mem_cgroup_tree_per_zone *mctz)
455 spin_lock_irqsave(&mctz->lock, flags);
456 __mem_cgroup_remove_exceeded(mz, mctz);
457 spin_unlock_irqrestore(&mctz->lock, flags);
460 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
462 unsigned long nr_pages = page_counter_read(&memcg->memory);
463 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
464 unsigned long excess = 0;
466 if (nr_pages > soft_limit)
467 excess = nr_pages - soft_limit;
472 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
474 unsigned long excess;
475 struct mem_cgroup_per_zone *mz;
476 struct mem_cgroup_tree_per_zone *mctz;
478 mctz = soft_limit_tree_from_page(page);
480 * Necessary to update all ancestors when hierarchy is used.
481 * because their event counter is not touched.
483 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
484 mz = mem_cgroup_page_zoneinfo(memcg, page);
485 excess = soft_limit_excess(memcg);
487 * We have to update the tree if mz is on RB-tree or
488 * mem is over its softlimit.
490 if (excess || mz->on_tree) {
493 spin_lock_irqsave(&mctz->lock, flags);
494 /* if on-tree, remove it */
496 __mem_cgroup_remove_exceeded(mz, mctz);
498 * Insert again. mz->usage_in_excess will be updated.
499 * If excess is 0, no tree ops.
501 __mem_cgroup_insert_exceeded(mz, mctz, excess);
502 spin_unlock_irqrestore(&mctz->lock, flags);
507 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
509 struct mem_cgroup_tree_per_zone *mctz;
510 struct mem_cgroup_per_zone *mz;
514 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
515 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
516 mctz = soft_limit_tree_node_zone(nid, zid);
517 mem_cgroup_remove_exceeded(mz, mctz);
522 static struct mem_cgroup_per_zone *
523 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
525 struct rb_node *rightmost = NULL;
526 struct mem_cgroup_per_zone *mz;
530 rightmost = rb_last(&mctz->rb_root);
532 goto done; /* Nothing to reclaim from */
534 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
536 * Remove the node now but someone else can add it back,
537 * we will to add it back at the end of reclaim to its correct
538 * position in the tree.
540 __mem_cgroup_remove_exceeded(mz, mctz);
541 if (!soft_limit_excess(mz->memcg) ||
542 !css_tryget_online(&mz->memcg->css))
548 static struct mem_cgroup_per_zone *
549 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
551 struct mem_cgroup_per_zone *mz;
553 spin_lock_irq(&mctz->lock);
554 mz = __mem_cgroup_largest_soft_limit_node(mctz);
555 spin_unlock_irq(&mctz->lock);
560 * Return page count for single (non recursive) @memcg.
562 * Implementation Note: reading percpu statistics for memcg.
564 * Both of vmstat[] and percpu_counter has threshold and do periodic
565 * synchronization to implement "quick" read. There are trade-off between
566 * reading cost and precision of value. Then, we may have a chance to implement
567 * a periodic synchronization of counter in memcg's counter.
569 * But this _read() function is used for user interface now. The user accounts
570 * memory usage by memory cgroup and he _always_ requires exact value because
571 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
572 * have to visit all online cpus and make sum. So, for now, unnecessary
573 * synchronization is not implemented. (just implemented for cpu hotplug)
575 * If there are kernel internal actions which can make use of some not-exact
576 * value, and reading all cpu value can be performance bottleneck in some
577 * common workload, threshold and synchronization as vmstat[] should be
581 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
586 /* Per-cpu values can be negative, use a signed accumulator */
587 for_each_possible_cpu(cpu)
588 val += per_cpu(memcg->stat->count[idx], cpu);
590 * Summing races with updates, so val may be negative. Avoid exposing
591 * transient negative values.
598 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
599 enum mem_cgroup_events_index idx)
601 unsigned long val = 0;
604 for_each_possible_cpu(cpu)
605 val += per_cpu(memcg->stat->events[idx], cpu);
609 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
611 bool compound, int nr_pages)
614 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
615 * counted as CACHE even if it's on ANON LRU.
618 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
621 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
625 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
626 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
630 /* pagein of a big page is an event. So, ignore page size */
632 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
634 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
635 nr_pages = -nr_pages; /* for event */
638 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
641 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
643 unsigned int lru_mask)
645 unsigned long nr = 0;
648 VM_BUG_ON((unsigned)nid >= nr_node_ids);
650 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
651 struct mem_cgroup_per_zone *mz;
655 if (!(BIT(lru) & lru_mask))
657 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
658 nr += mz->lru_size[lru];
664 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
665 unsigned int lru_mask)
667 unsigned long nr = 0;
670 for_each_node_state(nid, N_MEMORY)
671 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
675 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
676 enum mem_cgroup_events_target target)
678 unsigned long val, next;
680 val = __this_cpu_read(memcg->stat->nr_page_events);
681 next = __this_cpu_read(memcg->stat->targets[target]);
682 /* from time_after() in jiffies.h */
683 if ((long)next - (long)val < 0) {
685 case MEM_CGROUP_TARGET_THRESH:
686 next = val + THRESHOLDS_EVENTS_TARGET;
688 case MEM_CGROUP_TARGET_SOFTLIMIT:
689 next = val + SOFTLIMIT_EVENTS_TARGET;
691 case MEM_CGROUP_TARGET_NUMAINFO:
692 next = val + NUMAINFO_EVENTS_TARGET;
697 __this_cpu_write(memcg->stat->targets[target], next);
704 * Check events in order.
707 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
709 /* threshold event is triggered in finer grain than soft limit */
710 if (unlikely(mem_cgroup_event_ratelimit(memcg,
711 MEM_CGROUP_TARGET_THRESH))) {
713 bool do_numainfo __maybe_unused;
715 do_softlimit = mem_cgroup_event_ratelimit(memcg,
716 MEM_CGROUP_TARGET_SOFTLIMIT);
718 do_numainfo = mem_cgroup_event_ratelimit(memcg,
719 MEM_CGROUP_TARGET_NUMAINFO);
721 mem_cgroup_threshold(memcg);
722 if (unlikely(do_softlimit))
723 mem_cgroup_update_tree(memcg, page);
725 if (unlikely(do_numainfo))
726 atomic_inc(&memcg->numainfo_events);
731 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
734 * mm_update_next_owner() may clear mm->owner to NULL
735 * if it races with swapoff, page migration, etc.
736 * So this can be called with p == NULL.
741 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
743 EXPORT_SYMBOL(mem_cgroup_from_task);
745 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
747 struct mem_cgroup *memcg = NULL;
752 * Page cache insertions can happen withou an
753 * actual mm context, e.g. during disk probing
754 * on boot, loopback IO, acct() writes etc.
757 memcg = root_mem_cgroup;
759 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
760 if (unlikely(!memcg))
761 memcg = root_mem_cgroup;
763 } while (!css_tryget_online(&memcg->css));
769 * mem_cgroup_iter - iterate over memory cgroup hierarchy
770 * @root: hierarchy root
771 * @prev: previously returned memcg, NULL on first invocation
772 * @reclaim: cookie for shared reclaim walks, NULL for full walks
774 * Returns references to children of the hierarchy below @root, or
775 * @root itself, or %NULL after a full round-trip.
777 * Caller must pass the return value in @prev on subsequent
778 * invocations for reference counting, or use mem_cgroup_iter_break()
779 * to cancel a hierarchy walk before the round-trip is complete.
781 * Reclaimers can specify a zone and a priority level in @reclaim to
782 * divide up the memcgs in the hierarchy among all concurrent
783 * reclaimers operating on the same zone and priority.
785 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
786 struct mem_cgroup *prev,
787 struct mem_cgroup_reclaim_cookie *reclaim)
789 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
790 struct cgroup_subsys_state *css = NULL;
791 struct mem_cgroup *memcg = NULL;
792 struct mem_cgroup *pos = NULL;
794 if (mem_cgroup_disabled())
798 root = root_mem_cgroup;
800 if (prev && !reclaim)
803 if (!root->use_hierarchy && root != root_mem_cgroup) {
812 struct mem_cgroup_per_zone *mz;
814 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
815 iter = &mz->iter[reclaim->priority];
817 if (prev && reclaim->generation != iter->generation)
821 pos = READ_ONCE(iter->position);
822 if (!pos || css_tryget(&pos->css))
825 * css reference reached zero, so iter->position will
826 * be cleared by ->css_released. However, we should not
827 * rely on this happening soon, because ->css_released
828 * is called from a work queue, and by busy-waiting we
829 * might block it. So we clear iter->position right
832 (void)cmpxchg(&iter->position, pos, NULL);
840 css = css_next_descendant_pre(css, &root->css);
843 * Reclaimers share the hierarchy walk, and a
844 * new one might jump in right at the end of
845 * the hierarchy - make sure they see at least
846 * one group and restart from the beginning.
854 * Verify the css and acquire a reference. The root
855 * is provided by the caller, so we know it's alive
856 * and kicking, and don't take an extra reference.
858 memcg = mem_cgroup_from_css(css);
860 if (css == &root->css)
871 * The position could have already been updated by a competing
872 * thread, so check that the value hasn't changed since we read
873 * it to avoid reclaiming from the same cgroup twice.
875 (void)cmpxchg(&iter->position, pos, memcg);
883 reclaim->generation = iter->generation;
889 if (prev && prev != root)
896 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
897 * @root: hierarchy root
898 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
900 void mem_cgroup_iter_break(struct mem_cgroup *root,
901 struct mem_cgroup *prev)
904 root = root_mem_cgroup;
905 if (prev && prev != root)
909 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
911 struct mem_cgroup *memcg = dead_memcg;
912 struct mem_cgroup_reclaim_iter *iter;
913 struct mem_cgroup_per_zone *mz;
917 while ((memcg = parent_mem_cgroup(memcg))) {
919 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
920 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
921 for (i = 0; i <= DEF_PRIORITY; i++) {
923 cmpxchg(&iter->position,
932 * Iteration constructs for visiting all cgroups (under a tree). If
933 * loops are exited prematurely (break), mem_cgroup_iter_break() must
934 * be used for reference counting.
936 #define for_each_mem_cgroup_tree(iter, root) \
937 for (iter = mem_cgroup_iter(root, NULL, NULL); \
939 iter = mem_cgroup_iter(root, iter, NULL))
941 #define for_each_mem_cgroup(iter) \
942 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
944 iter = mem_cgroup_iter(NULL, iter, NULL))
947 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
948 * @zone: zone of the wanted lruvec
949 * @memcg: memcg of the wanted lruvec
951 * Returns the lru list vector holding pages for the given @zone and
952 * @mem. This can be the global zone lruvec, if the memory controller
955 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
956 struct mem_cgroup *memcg)
958 struct mem_cgroup_per_zone *mz;
959 struct lruvec *lruvec;
961 if (mem_cgroup_disabled()) {
962 lruvec = &zone->lruvec;
966 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
967 lruvec = &mz->lruvec;
970 * Since a node can be onlined after the mem_cgroup was created,
971 * we have to be prepared to initialize lruvec->zone here;
972 * and if offlined then reonlined, we need to reinitialize it.
974 if (unlikely(lruvec->zone != zone))
980 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
982 * @zone: zone of the page
984 * This function is only safe when following the LRU page isolation
985 * and putback protocol: the LRU lock must be held, and the page must
986 * either be PageLRU() or the caller must have isolated/allocated it.
988 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
990 struct mem_cgroup_per_zone *mz;
991 struct mem_cgroup *memcg;
992 struct lruvec *lruvec;
994 if (mem_cgroup_disabled()) {
995 lruvec = &zone->lruvec;
999 memcg = page->mem_cgroup;
1001 * Swapcache readahead pages are added to the LRU - and
1002 * possibly migrated - before they are charged.
1005 memcg = root_mem_cgroup;
1007 mz = mem_cgroup_page_zoneinfo(memcg, page);
1008 lruvec = &mz->lruvec;
1011 * Since a node can be onlined after the mem_cgroup was created,
1012 * we have to be prepared to initialize lruvec->zone here;
1013 * and if offlined then reonlined, we need to reinitialize it.
1015 if (unlikely(lruvec->zone != zone))
1016 lruvec->zone = zone;
1021 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1022 * @lruvec: mem_cgroup per zone lru vector
1023 * @lru: index of lru list the page is sitting on
1024 * @nr_pages: positive when adding or negative when removing
1026 * This function must be called when a page is added to or removed from an
1029 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1032 struct mem_cgroup_per_zone *mz;
1033 unsigned long *lru_size;
1035 if (mem_cgroup_disabled())
1038 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1039 lru_size = mz->lru_size + lru;
1040 *lru_size += nr_pages;
1041 VM_BUG_ON((long)(*lru_size) < 0);
1044 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1046 struct mem_cgroup *task_memcg;
1047 struct task_struct *p;
1050 p = find_lock_task_mm(task);
1052 task_memcg = get_mem_cgroup_from_mm(p->mm);
1056 * All threads may have already detached their mm's, but the oom
1057 * killer still needs to detect if they have already been oom
1058 * killed to prevent needlessly killing additional tasks.
1061 task_memcg = mem_cgroup_from_task(task);
1062 css_get(&task_memcg->css);
1065 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1066 css_put(&task_memcg->css);
1071 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1072 * @memcg: the memory cgroup
1074 * Returns the maximum amount of memory @mem can be charged with, in
1077 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1079 unsigned long margin = 0;
1080 unsigned long count;
1081 unsigned long limit;
1083 count = page_counter_read(&memcg->memory);
1084 limit = READ_ONCE(memcg->memory.limit);
1086 margin = limit - count;
1088 if (do_memsw_account()) {
1089 count = page_counter_read(&memcg->memsw);
1090 limit = READ_ONCE(memcg->memsw.limit);
1092 margin = min(margin, limit - count);
1099 * A routine for checking "mem" is under move_account() or not.
1101 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1102 * moving cgroups. This is for waiting at high-memory pressure
1105 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1107 struct mem_cgroup *from;
1108 struct mem_cgroup *to;
1111 * Unlike task_move routines, we access mc.to, mc.from not under
1112 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1114 spin_lock(&mc.lock);
1120 ret = mem_cgroup_is_descendant(from, memcg) ||
1121 mem_cgroup_is_descendant(to, memcg);
1123 spin_unlock(&mc.lock);
1127 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1129 if (mc.moving_task && current != mc.moving_task) {
1130 if (mem_cgroup_under_move(memcg)) {
1132 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1133 /* moving charge context might have finished. */
1136 finish_wait(&mc.waitq, &wait);
1143 #define K(x) ((x) << (PAGE_SHIFT-10))
1145 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1146 * @memcg: The memory cgroup that went over limit
1147 * @p: Task that is going to be killed
1149 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1152 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1154 /* oom_info_lock ensures that parallel ooms do not interleave */
1155 static DEFINE_MUTEX(oom_info_lock);
1156 struct mem_cgroup *iter;
1159 mutex_lock(&oom_info_lock);
1163 pr_info("Task in ");
1164 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1165 pr_cont(" killed as a result of limit of ");
1167 pr_info("Memory limit reached of cgroup ");
1170 pr_cont_cgroup_path(memcg->css.cgroup);
1175 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1176 K((u64)page_counter_read(&memcg->memory)),
1177 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1178 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1179 K((u64)page_counter_read(&memcg->memsw)),
1180 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1181 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1182 K((u64)page_counter_read(&memcg->kmem)),
1183 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1185 for_each_mem_cgroup_tree(iter, memcg) {
1186 pr_info("Memory cgroup stats for ");
1187 pr_cont_cgroup_path(iter->css.cgroup);
1190 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1191 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1193 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1194 K(mem_cgroup_read_stat(iter, i)));
1197 for (i = 0; i < NR_LRU_LISTS; i++)
1198 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1199 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1203 mutex_unlock(&oom_info_lock);
1207 * This function returns the number of memcg under hierarchy tree. Returns
1208 * 1(self count) if no children.
1210 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1213 struct mem_cgroup *iter;
1215 for_each_mem_cgroup_tree(iter, memcg)
1221 * Return the memory (and swap, if configured) limit for a memcg.
1223 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1225 unsigned long limit;
1227 limit = memcg->memory.limit;
1228 if (mem_cgroup_swappiness(memcg)) {
1229 unsigned long memsw_limit;
1230 unsigned long swap_limit;
1232 memsw_limit = memcg->memsw.limit;
1233 swap_limit = memcg->swap.limit;
1234 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1235 limit = min(limit + swap_limit, memsw_limit);
1240 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1243 struct oom_control oc = {
1246 .gfp_mask = gfp_mask,
1249 struct mem_cgroup *iter;
1250 unsigned long chosen_points = 0;
1251 unsigned long totalpages;
1252 unsigned int points = 0;
1253 struct task_struct *chosen = NULL;
1255 mutex_lock(&oom_lock);
1258 * If current has a pending SIGKILL or is exiting, then automatically
1259 * select it. The goal is to allow it to allocate so that it may
1260 * quickly exit and free its memory.
1262 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1263 mark_oom_victim(current);
1267 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1268 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1269 for_each_mem_cgroup_tree(iter, memcg) {
1270 struct css_task_iter it;
1271 struct task_struct *task;
1273 css_task_iter_start(&iter->css, &it);
1274 while ((task = css_task_iter_next(&it))) {
1275 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1276 case OOM_SCAN_SELECT:
1278 put_task_struct(chosen);
1280 chosen_points = ULONG_MAX;
1281 get_task_struct(chosen);
1283 case OOM_SCAN_CONTINUE:
1285 case OOM_SCAN_ABORT:
1286 css_task_iter_end(&it);
1287 mem_cgroup_iter_break(memcg, iter);
1289 put_task_struct(chosen);
1294 points = oom_badness(task, memcg, NULL, totalpages);
1295 if (!points || points < chosen_points)
1297 /* Prefer thread group leaders for display purposes */
1298 if (points == chosen_points &&
1299 thread_group_leader(chosen))
1303 put_task_struct(chosen);
1305 chosen_points = points;
1306 get_task_struct(chosen);
1308 css_task_iter_end(&it);
1312 points = chosen_points * 1000 / totalpages;
1313 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1314 "Memory cgroup out of memory");
1317 mutex_unlock(&oom_lock);
1320 #if MAX_NUMNODES > 1
1323 * test_mem_cgroup_node_reclaimable
1324 * @memcg: the target memcg
1325 * @nid: the node ID to be checked.
1326 * @noswap : specify true here if the user wants flle only information.
1328 * This function returns whether the specified memcg contains any
1329 * reclaimable pages on a node. Returns true if there are any reclaimable
1330 * pages in the node.
1332 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1333 int nid, bool noswap)
1335 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1337 if (noswap || !total_swap_pages)
1339 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1346 * Always updating the nodemask is not very good - even if we have an empty
1347 * list or the wrong list here, we can start from some node and traverse all
1348 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1351 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1355 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1356 * pagein/pageout changes since the last update.
1358 if (!atomic_read(&memcg->numainfo_events))
1360 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1363 /* make a nodemask where this memcg uses memory from */
1364 memcg->scan_nodes = node_states[N_MEMORY];
1366 for_each_node_mask(nid, node_states[N_MEMORY]) {
1368 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1369 node_clear(nid, memcg->scan_nodes);
1372 atomic_set(&memcg->numainfo_events, 0);
1373 atomic_set(&memcg->numainfo_updating, 0);
1377 * Selecting a node where we start reclaim from. Because what we need is just
1378 * reducing usage counter, start from anywhere is O,K. Considering
1379 * memory reclaim from current node, there are pros. and cons.
1381 * Freeing memory from current node means freeing memory from a node which
1382 * we'll use or we've used. So, it may make LRU bad. And if several threads
1383 * hit limits, it will see a contention on a node. But freeing from remote
1384 * node means more costs for memory reclaim because of memory latency.
1386 * Now, we use round-robin. Better algorithm is welcomed.
1388 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1392 mem_cgroup_may_update_nodemask(memcg);
1393 node = memcg->last_scanned_node;
1395 node = next_node(node, memcg->scan_nodes);
1396 if (node == MAX_NUMNODES)
1397 node = first_node(memcg->scan_nodes);
1399 * We call this when we hit limit, not when pages are added to LRU.
1400 * No LRU may hold pages because all pages are UNEVICTABLE or
1401 * memcg is too small and all pages are not on LRU. In that case,
1402 * we use curret node.
1404 if (unlikely(node == MAX_NUMNODES))
1405 node = numa_node_id();
1407 memcg->last_scanned_node = node;
1411 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1417 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1420 unsigned long *total_scanned)
1422 struct mem_cgroup *victim = NULL;
1425 unsigned long excess;
1426 unsigned long nr_scanned;
1427 struct mem_cgroup_reclaim_cookie reclaim = {
1432 excess = soft_limit_excess(root_memcg);
1435 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1440 * If we have not been able to reclaim
1441 * anything, it might because there are
1442 * no reclaimable pages under this hierarchy
1447 * We want to do more targeted reclaim.
1448 * excess >> 2 is not to excessive so as to
1449 * reclaim too much, nor too less that we keep
1450 * coming back to reclaim from this cgroup
1452 if (total >= (excess >> 2) ||
1453 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1458 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1460 *total_scanned += nr_scanned;
1461 if (!soft_limit_excess(root_memcg))
1464 mem_cgroup_iter_break(root_memcg, victim);
1468 #ifdef CONFIG_LOCKDEP
1469 static struct lockdep_map memcg_oom_lock_dep_map = {
1470 .name = "memcg_oom_lock",
1474 static DEFINE_SPINLOCK(memcg_oom_lock);
1477 * Check OOM-Killer is already running under our hierarchy.
1478 * If someone is running, return false.
1480 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1482 struct mem_cgroup *iter, *failed = NULL;
1484 spin_lock(&memcg_oom_lock);
1486 for_each_mem_cgroup_tree(iter, memcg) {
1487 if (iter->oom_lock) {
1489 * this subtree of our hierarchy is already locked
1490 * so we cannot give a lock.
1493 mem_cgroup_iter_break(memcg, iter);
1496 iter->oom_lock = true;
1501 * OK, we failed to lock the whole subtree so we have
1502 * to clean up what we set up to the failing subtree
1504 for_each_mem_cgroup_tree(iter, memcg) {
1505 if (iter == failed) {
1506 mem_cgroup_iter_break(memcg, iter);
1509 iter->oom_lock = false;
1512 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1514 spin_unlock(&memcg_oom_lock);
1519 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1521 struct mem_cgroup *iter;
1523 spin_lock(&memcg_oom_lock);
1524 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1525 for_each_mem_cgroup_tree(iter, memcg)
1526 iter->oom_lock = false;
1527 spin_unlock(&memcg_oom_lock);
1530 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1532 struct mem_cgroup *iter;
1534 spin_lock(&memcg_oom_lock);
1535 for_each_mem_cgroup_tree(iter, memcg)
1537 spin_unlock(&memcg_oom_lock);
1540 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1542 struct mem_cgroup *iter;
1545 * When a new child is created while the hierarchy is under oom,
1546 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1548 spin_lock(&memcg_oom_lock);
1549 for_each_mem_cgroup_tree(iter, memcg)
1550 if (iter->under_oom > 0)
1552 spin_unlock(&memcg_oom_lock);
1555 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1557 struct oom_wait_info {
1558 struct mem_cgroup *memcg;
1562 static int memcg_oom_wake_function(wait_queue_t *wait,
1563 unsigned mode, int sync, void *arg)
1565 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1566 struct mem_cgroup *oom_wait_memcg;
1567 struct oom_wait_info *oom_wait_info;
1569 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1570 oom_wait_memcg = oom_wait_info->memcg;
1572 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1573 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1575 return autoremove_wake_function(wait, mode, sync, arg);
1578 static void memcg_oom_recover(struct mem_cgroup *memcg)
1581 * For the following lockless ->under_oom test, the only required
1582 * guarantee is that it must see the state asserted by an OOM when
1583 * this function is called as a result of userland actions
1584 * triggered by the notification of the OOM. This is trivially
1585 * achieved by invoking mem_cgroup_mark_under_oom() before
1586 * triggering notification.
1588 if (memcg && memcg->under_oom)
1589 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1592 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1594 if (!current->memcg_may_oom)
1597 * We are in the middle of the charge context here, so we
1598 * don't want to block when potentially sitting on a callstack
1599 * that holds all kinds of filesystem and mm locks.
1601 * Also, the caller may handle a failed allocation gracefully
1602 * (like optional page cache readahead) and so an OOM killer
1603 * invocation might not even be necessary.
1605 * That's why we don't do anything here except remember the
1606 * OOM context and then deal with it at the end of the page
1607 * fault when the stack is unwound, the locks are released,
1608 * and when we know whether the fault was overall successful.
1610 css_get(&memcg->css);
1611 current->memcg_in_oom = memcg;
1612 current->memcg_oom_gfp_mask = mask;
1613 current->memcg_oom_order = order;
1617 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1618 * @handle: actually kill/wait or just clean up the OOM state
1620 * This has to be called at the end of a page fault if the memcg OOM
1621 * handler was enabled.
1623 * Memcg supports userspace OOM handling where failed allocations must
1624 * sleep on a waitqueue until the userspace task resolves the
1625 * situation. Sleeping directly in the charge context with all kinds
1626 * of locks held is not a good idea, instead we remember an OOM state
1627 * in the task and mem_cgroup_oom_synchronize() has to be called at
1628 * the end of the page fault to complete the OOM handling.
1630 * Returns %true if an ongoing memcg OOM situation was detected and
1631 * completed, %false otherwise.
1633 bool mem_cgroup_oom_synchronize(bool handle)
1635 struct mem_cgroup *memcg = current->memcg_in_oom;
1636 struct oom_wait_info owait;
1639 /* OOM is global, do not handle */
1643 if (!handle || oom_killer_disabled)
1646 owait.memcg = memcg;
1647 owait.wait.flags = 0;
1648 owait.wait.func = memcg_oom_wake_function;
1649 owait.wait.private = current;
1650 INIT_LIST_HEAD(&owait.wait.task_list);
1652 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1653 mem_cgroup_mark_under_oom(memcg);
1655 locked = mem_cgroup_oom_trylock(memcg);
1658 mem_cgroup_oom_notify(memcg);
1660 if (locked && !memcg->oom_kill_disable) {
1661 mem_cgroup_unmark_under_oom(memcg);
1662 finish_wait(&memcg_oom_waitq, &owait.wait);
1663 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1664 current->memcg_oom_order);
1667 mem_cgroup_unmark_under_oom(memcg);
1668 finish_wait(&memcg_oom_waitq, &owait.wait);
1672 mem_cgroup_oom_unlock(memcg);
1674 * There is no guarantee that an OOM-lock contender
1675 * sees the wakeups triggered by the OOM kill
1676 * uncharges. Wake any sleepers explicitely.
1678 memcg_oom_recover(memcg);
1681 current->memcg_in_oom = NULL;
1682 css_put(&memcg->css);
1687 * lock_page_memcg - lock a page->mem_cgroup binding
1690 * This function protects unlocked LRU pages from being moved to
1691 * another cgroup and stabilizes their page->mem_cgroup binding.
1693 void lock_page_memcg(struct page *page)
1695 struct mem_cgroup *memcg;
1696 unsigned long flags;
1699 * The RCU lock is held throughout the transaction. The fast
1700 * path can get away without acquiring the memcg->move_lock
1701 * because page moving starts with an RCU grace period.
1705 if (mem_cgroup_disabled())
1708 memcg = page->mem_cgroup;
1709 if (unlikely(!memcg))
1712 if (atomic_read(&memcg->moving_account) <= 0)
1715 spin_lock_irqsave(&memcg->move_lock, flags);
1716 if (memcg != page->mem_cgroup) {
1717 spin_unlock_irqrestore(&memcg->move_lock, flags);
1722 * When charge migration first begins, we can have locked and
1723 * unlocked page stat updates happening concurrently. Track
1724 * the task who has the lock for unlock_page_memcg().
1726 memcg->move_lock_task = current;
1727 memcg->move_lock_flags = flags;
1731 EXPORT_SYMBOL(lock_page_memcg);
1734 * unlock_page_memcg - unlock a page->mem_cgroup binding
1737 void unlock_page_memcg(struct page *page)
1739 struct mem_cgroup *memcg = page->mem_cgroup;
1741 if (memcg && memcg->move_lock_task == current) {
1742 unsigned long flags = memcg->move_lock_flags;
1744 memcg->move_lock_task = NULL;
1745 memcg->move_lock_flags = 0;
1747 spin_unlock_irqrestore(&memcg->move_lock, flags);
1752 EXPORT_SYMBOL(unlock_page_memcg);
1755 * size of first charge trial. "32" comes from vmscan.c's magic value.
1756 * TODO: maybe necessary to use big numbers in big irons.
1758 #define CHARGE_BATCH 32U
1759 struct memcg_stock_pcp {
1760 struct mem_cgroup *cached; /* this never be root cgroup */
1761 unsigned int nr_pages;
1762 struct work_struct work;
1763 unsigned long flags;
1764 #define FLUSHING_CACHED_CHARGE 0
1766 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1767 static DEFINE_MUTEX(percpu_charge_mutex);
1770 * consume_stock: Try to consume stocked charge on this cpu.
1771 * @memcg: memcg to consume from.
1772 * @nr_pages: how many pages to charge.
1774 * The charges will only happen if @memcg matches the current cpu's memcg
1775 * stock, and at least @nr_pages are available in that stock. Failure to
1776 * service an allocation will refill the stock.
1778 * returns true if successful, false otherwise.
1780 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1782 struct memcg_stock_pcp *stock;
1785 if (nr_pages > CHARGE_BATCH)
1788 stock = &get_cpu_var(memcg_stock);
1789 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1790 stock->nr_pages -= nr_pages;
1793 put_cpu_var(memcg_stock);
1798 * Returns stocks cached in percpu and reset cached information.
1800 static void drain_stock(struct memcg_stock_pcp *stock)
1802 struct mem_cgroup *old = stock->cached;
1804 if (stock->nr_pages) {
1805 page_counter_uncharge(&old->memory, stock->nr_pages);
1806 if (do_memsw_account())
1807 page_counter_uncharge(&old->memsw, stock->nr_pages);
1808 css_put_many(&old->css, stock->nr_pages);
1809 stock->nr_pages = 0;
1811 stock->cached = NULL;
1815 * This must be called under preempt disabled or must be called by
1816 * a thread which is pinned to local cpu.
1818 static void drain_local_stock(struct work_struct *dummy)
1820 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1822 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1826 * Cache charges(val) to local per_cpu area.
1827 * This will be consumed by consume_stock() function, later.
1829 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1831 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1833 if (stock->cached != memcg) { /* reset if necessary */
1835 stock->cached = memcg;
1837 stock->nr_pages += nr_pages;
1838 put_cpu_var(memcg_stock);
1842 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1843 * of the hierarchy under it.
1845 static void drain_all_stock(struct mem_cgroup *root_memcg)
1849 /* If someone's already draining, avoid adding running more workers. */
1850 if (!mutex_trylock(&percpu_charge_mutex))
1852 /* Notify other cpus that system-wide "drain" is running */
1855 for_each_online_cpu(cpu) {
1856 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1857 struct mem_cgroup *memcg;
1859 memcg = stock->cached;
1860 if (!memcg || !stock->nr_pages)
1862 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1864 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1866 drain_local_stock(&stock->work);
1868 schedule_work_on(cpu, &stock->work);
1873 mutex_unlock(&percpu_charge_mutex);
1876 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1877 unsigned long action,
1880 int cpu = (unsigned long)hcpu;
1881 struct memcg_stock_pcp *stock;
1883 if (action == CPU_ONLINE)
1886 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1889 stock = &per_cpu(memcg_stock, cpu);
1894 static void reclaim_high(struct mem_cgroup *memcg,
1895 unsigned int nr_pages,
1899 if (page_counter_read(&memcg->memory) <= memcg->high)
1901 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1902 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1903 } while ((memcg = parent_mem_cgroup(memcg)));
1906 static void high_work_func(struct work_struct *work)
1908 struct mem_cgroup *memcg;
1910 memcg = container_of(work, struct mem_cgroup, high_work);
1911 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1915 * Scheduled by try_charge() to be executed from the userland return path
1916 * and reclaims memory over the high limit.
1918 void mem_cgroup_handle_over_high(void)
1920 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1921 struct mem_cgroup *memcg;
1923 if (likely(!nr_pages))
1926 memcg = get_mem_cgroup_from_mm(current->mm);
1927 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1928 css_put(&memcg->css);
1929 current->memcg_nr_pages_over_high = 0;
1932 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1933 unsigned int nr_pages)
1935 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1936 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1937 struct mem_cgroup *mem_over_limit;
1938 struct page_counter *counter;
1939 unsigned long nr_reclaimed;
1940 bool may_swap = true;
1941 bool drained = false;
1943 if (mem_cgroup_is_root(memcg))
1946 if (consume_stock(memcg, nr_pages))
1949 if (!do_memsw_account() ||
1950 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1951 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1953 if (do_memsw_account())
1954 page_counter_uncharge(&memcg->memsw, batch);
1955 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1957 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1961 if (batch > nr_pages) {
1967 * Unlike in global OOM situations, memcg is not in a physical
1968 * memory shortage. Allow dying and OOM-killed tasks to
1969 * bypass the last charges so that they can exit quickly and
1970 * free their memory.
1972 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1973 fatal_signal_pending(current) ||
1974 current->flags & PF_EXITING))
1977 if (unlikely(task_in_memcg_oom(current)))
1980 if (!gfpflags_allow_blocking(gfp_mask))
1983 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1985 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1986 gfp_mask, may_swap);
1988 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1992 drain_all_stock(mem_over_limit);
1997 if (gfp_mask & __GFP_NORETRY)
2000 * Even though the limit is exceeded at this point, reclaim
2001 * may have been able to free some pages. Retry the charge
2002 * before killing the task.
2004 * Only for regular pages, though: huge pages are rather
2005 * unlikely to succeed so close to the limit, and we fall back
2006 * to regular pages anyway in case of failure.
2008 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2011 * At task move, charge accounts can be doubly counted. So, it's
2012 * better to wait until the end of task_move if something is going on.
2014 if (mem_cgroup_wait_acct_move(mem_over_limit))
2020 if (gfp_mask & __GFP_NOFAIL)
2023 if (fatal_signal_pending(current))
2026 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2028 mem_cgroup_oom(mem_over_limit, gfp_mask,
2029 get_order(nr_pages * PAGE_SIZE));
2031 if (!(gfp_mask & __GFP_NOFAIL))
2035 * The allocation either can't fail or will lead to more memory
2036 * being freed very soon. Allow memory usage go over the limit
2037 * temporarily by force charging it.
2039 page_counter_charge(&memcg->memory, nr_pages);
2040 if (do_memsw_account())
2041 page_counter_charge(&memcg->memsw, nr_pages);
2042 css_get_many(&memcg->css, nr_pages);
2047 css_get_many(&memcg->css, batch);
2048 if (batch > nr_pages)
2049 refill_stock(memcg, batch - nr_pages);
2052 * If the hierarchy is above the normal consumption range, schedule
2053 * reclaim on returning to userland. We can perform reclaim here
2054 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2055 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2056 * not recorded as it most likely matches current's and won't
2057 * change in the meantime. As high limit is checked again before
2058 * reclaim, the cost of mismatch is negligible.
2061 if (page_counter_read(&memcg->memory) > memcg->high) {
2062 /* Don't bother a random interrupted task */
2063 if (in_interrupt()) {
2064 schedule_work(&memcg->high_work);
2067 current->memcg_nr_pages_over_high += batch;
2068 set_notify_resume(current);
2071 } while ((memcg = parent_mem_cgroup(memcg)));
2076 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2078 if (mem_cgroup_is_root(memcg))
2081 page_counter_uncharge(&memcg->memory, nr_pages);
2082 if (do_memsw_account())
2083 page_counter_uncharge(&memcg->memsw, nr_pages);
2085 css_put_many(&memcg->css, nr_pages);
2088 static void lock_page_lru(struct page *page, int *isolated)
2090 struct zone *zone = page_zone(page);
2092 spin_lock_irq(&zone->lru_lock);
2093 if (PageLRU(page)) {
2094 struct lruvec *lruvec;
2096 lruvec = mem_cgroup_page_lruvec(page, zone);
2098 del_page_from_lru_list(page, lruvec, page_lru(page));
2104 static void unlock_page_lru(struct page *page, int isolated)
2106 struct zone *zone = page_zone(page);
2109 struct lruvec *lruvec;
2111 lruvec = mem_cgroup_page_lruvec(page, zone);
2112 VM_BUG_ON_PAGE(PageLRU(page), page);
2114 add_page_to_lru_list(page, lruvec, page_lru(page));
2116 spin_unlock_irq(&zone->lru_lock);
2119 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2124 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2127 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2128 * may already be on some other mem_cgroup's LRU. Take care of it.
2131 lock_page_lru(page, &isolated);
2134 * Nobody should be changing or seriously looking at
2135 * page->mem_cgroup at this point:
2137 * - the page is uncharged
2139 * - the page is off-LRU
2141 * - an anonymous fault has exclusive page access, except for
2142 * a locked page table
2144 * - a page cache insertion, a swapin fault, or a migration
2145 * have the page locked
2147 page->mem_cgroup = memcg;
2150 unlock_page_lru(page, isolated);
2154 static int memcg_alloc_cache_id(void)
2159 id = ida_simple_get(&memcg_cache_ida,
2160 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2164 if (id < memcg_nr_cache_ids)
2168 * There's no space for the new id in memcg_caches arrays,
2169 * so we have to grow them.
2171 down_write(&memcg_cache_ids_sem);
2173 size = 2 * (id + 1);
2174 if (size < MEMCG_CACHES_MIN_SIZE)
2175 size = MEMCG_CACHES_MIN_SIZE;
2176 else if (size > MEMCG_CACHES_MAX_SIZE)
2177 size = MEMCG_CACHES_MAX_SIZE;
2179 err = memcg_update_all_caches(size);
2181 err = memcg_update_all_list_lrus(size);
2183 memcg_nr_cache_ids = size;
2185 up_write(&memcg_cache_ids_sem);
2188 ida_simple_remove(&memcg_cache_ida, id);
2194 static void memcg_free_cache_id(int id)
2196 ida_simple_remove(&memcg_cache_ida, id);
2199 struct memcg_kmem_cache_create_work {
2200 struct mem_cgroup *memcg;
2201 struct kmem_cache *cachep;
2202 struct work_struct work;
2205 static void memcg_kmem_cache_create_func(struct work_struct *w)
2207 struct memcg_kmem_cache_create_work *cw =
2208 container_of(w, struct memcg_kmem_cache_create_work, work);
2209 struct mem_cgroup *memcg = cw->memcg;
2210 struct kmem_cache *cachep = cw->cachep;
2212 memcg_create_kmem_cache(memcg, cachep);
2214 css_put(&memcg->css);
2219 * Enqueue the creation of a per-memcg kmem_cache.
2221 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2222 struct kmem_cache *cachep)
2224 struct memcg_kmem_cache_create_work *cw;
2226 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2230 css_get(&memcg->css);
2233 cw->cachep = cachep;
2234 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2236 schedule_work(&cw->work);
2239 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2240 struct kmem_cache *cachep)
2243 * We need to stop accounting when we kmalloc, because if the
2244 * corresponding kmalloc cache is not yet created, the first allocation
2245 * in __memcg_schedule_kmem_cache_create will recurse.
2247 * However, it is better to enclose the whole function. Depending on
2248 * the debugging options enabled, INIT_WORK(), for instance, can
2249 * trigger an allocation. This too, will make us recurse. Because at
2250 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2251 * the safest choice is to do it like this, wrapping the whole function.
2253 current->memcg_kmem_skip_account = 1;
2254 __memcg_schedule_kmem_cache_create(memcg, cachep);
2255 current->memcg_kmem_skip_account = 0;
2259 * Return the kmem_cache we're supposed to use for a slab allocation.
2260 * We try to use the current memcg's version of the cache.
2262 * If the cache does not exist yet, if we are the first user of it,
2263 * we either create it immediately, if possible, or create it asynchronously
2265 * In the latter case, we will let the current allocation go through with
2266 * the original cache.
2268 * Can't be called in interrupt context or from kernel threads.
2269 * This function needs to be called with rcu_read_lock() held.
2271 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp)
2273 struct mem_cgroup *memcg;
2274 struct kmem_cache *memcg_cachep;
2277 VM_BUG_ON(!is_root_cache(cachep));
2279 if (cachep->flags & SLAB_ACCOUNT)
2280 gfp |= __GFP_ACCOUNT;
2282 if (!(gfp & __GFP_ACCOUNT))
2285 if (current->memcg_kmem_skip_account)
2288 memcg = get_mem_cgroup_from_mm(current->mm);
2289 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2293 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2294 if (likely(memcg_cachep))
2295 return memcg_cachep;
2298 * If we are in a safe context (can wait, and not in interrupt
2299 * context), we could be be predictable and return right away.
2300 * This would guarantee that the allocation being performed
2301 * already belongs in the new cache.
2303 * However, there are some clashes that can arrive from locking.
2304 * For instance, because we acquire the slab_mutex while doing
2305 * memcg_create_kmem_cache, this means no further allocation
2306 * could happen with the slab_mutex held. So it's better to
2309 memcg_schedule_kmem_cache_create(memcg, cachep);
2311 css_put(&memcg->css);
2315 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2317 if (!is_root_cache(cachep))
2318 css_put(&cachep->memcg_params.memcg->css);
2321 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2322 struct mem_cgroup *memcg)
2324 unsigned int nr_pages = 1 << order;
2325 struct page_counter *counter;
2328 ret = try_charge(memcg, gfp, nr_pages);
2332 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2333 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2334 cancel_charge(memcg, nr_pages);
2338 page->mem_cgroup = memcg;
2343 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2345 struct mem_cgroup *memcg;
2348 memcg = get_mem_cgroup_from_mm(current->mm);
2349 if (memcg_kmem_online(memcg))
2350 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2351 css_put(&memcg->css);
2355 void __memcg_kmem_uncharge(struct page *page, int order)
2357 struct mem_cgroup *memcg = page->mem_cgroup;
2358 unsigned int nr_pages = 1 << order;
2363 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2365 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2366 page_counter_uncharge(&memcg->kmem, nr_pages);
2368 page_counter_uncharge(&memcg->memory, nr_pages);
2369 if (do_memsw_account())
2370 page_counter_uncharge(&memcg->memsw, nr_pages);
2372 page->mem_cgroup = NULL;
2373 css_put_many(&memcg->css, nr_pages);
2375 #endif /* !CONFIG_SLOB */
2377 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2380 * Because tail pages are not marked as "used", set it. We're under
2381 * zone->lru_lock and migration entries setup in all page mappings.
2383 void mem_cgroup_split_huge_fixup(struct page *head)
2387 if (mem_cgroup_disabled())
2390 for (i = 1; i < HPAGE_PMD_NR; i++)
2391 head[i].mem_cgroup = head->mem_cgroup;
2393 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2396 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2398 #ifdef CONFIG_MEMCG_SWAP
2399 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2402 int val = (charge) ? 1 : -1;
2403 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2407 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2408 * @entry: swap entry to be moved
2409 * @from: mem_cgroup which the entry is moved from
2410 * @to: mem_cgroup which the entry is moved to
2412 * It succeeds only when the swap_cgroup's record for this entry is the same
2413 * as the mem_cgroup's id of @from.
2415 * Returns 0 on success, -EINVAL on failure.
2417 * The caller must have charged to @to, IOW, called page_counter_charge() about
2418 * both res and memsw, and called css_get().
2420 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2421 struct mem_cgroup *from, struct mem_cgroup *to)
2423 unsigned short old_id, new_id;
2425 old_id = mem_cgroup_id(from);
2426 new_id = mem_cgroup_id(to);
2428 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2429 mem_cgroup_swap_statistics(from, false);
2430 mem_cgroup_swap_statistics(to, true);
2436 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2437 struct mem_cgroup *from, struct mem_cgroup *to)
2443 static DEFINE_MUTEX(memcg_limit_mutex);
2445 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2446 unsigned long limit)
2448 unsigned long curusage;
2449 unsigned long oldusage;
2450 bool enlarge = false;
2455 * For keeping hierarchical_reclaim simple, how long we should retry
2456 * is depends on callers. We set our retry-count to be function
2457 * of # of children which we should visit in this loop.
2459 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2460 mem_cgroup_count_children(memcg);
2462 oldusage = page_counter_read(&memcg->memory);
2465 if (signal_pending(current)) {
2470 mutex_lock(&memcg_limit_mutex);
2471 if (limit > memcg->memsw.limit) {
2472 mutex_unlock(&memcg_limit_mutex);
2476 if (limit > memcg->memory.limit)
2478 ret = page_counter_limit(&memcg->memory, limit);
2479 mutex_unlock(&memcg_limit_mutex);
2484 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2486 curusage = page_counter_read(&memcg->memory);
2487 /* Usage is reduced ? */
2488 if (curusage >= oldusage)
2491 oldusage = curusage;
2492 } while (retry_count);
2494 if (!ret && enlarge)
2495 memcg_oom_recover(memcg);
2500 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2501 unsigned long limit)
2503 unsigned long curusage;
2504 unsigned long oldusage;
2505 bool enlarge = false;
2509 /* see mem_cgroup_resize_res_limit */
2510 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2511 mem_cgroup_count_children(memcg);
2513 oldusage = page_counter_read(&memcg->memsw);
2516 if (signal_pending(current)) {
2521 mutex_lock(&memcg_limit_mutex);
2522 if (limit < memcg->memory.limit) {
2523 mutex_unlock(&memcg_limit_mutex);
2527 if (limit > memcg->memsw.limit)
2529 ret = page_counter_limit(&memcg->memsw, limit);
2530 mutex_unlock(&memcg_limit_mutex);
2535 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2537 curusage = page_counter_read(&memcg->memsw);
2538 /* Usage is reduced ? */
2539 if (curusage >= oldusage)
2542 oldusage = curusage;
2543 } while (retry_count);
2545 if (!ret && enlarge)
2546 memcg_oom_recover(memcg);
2551 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2553 unsigned long *total_scanned)
2555 unsigned long nr_reclaimed = 0;
2556 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2557 unsigned long reclaimed;
2559 struct mem_cgroup_tree_per_zone *mctz;
2560 unsigned long excess;
2561 unsigned long nr_scanned;
2566 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2568 * This loop can run a while, specially if mem_cgroup's continuously
2569 * keep exceeding their soft limit and putting the system under
2576 mz = mem_cgroup_largest_soft_limit_node(mctz);
2581 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2582 gfp_mask, &nr_scanned);
2583 nr_reclaimed += reclaimed;
2584 *total_scanned += nr_scanned;
2585 spin_lock_irq(&mctz->lock);
2586 __mem_cgroup_remove_exceeded(mz, mctz);
2589 * If we failed to reclaim anything from this memory cgroup
2590 * it is time to move on to the next cgroup
2594 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2596 excess = soft_limit_excess(mz->memcg);
2598 * One school of thought says that we should not add
2599 * back the node to the tree if reclaim returns 0.
2600 * But our reclaim could return 0, simply because due
2601 * to priority we are exposing a smaller subset of
2602 * memory to reclaim from. Consider this as a longer
2605 /* If excess == 0, no tree ops */
2606 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2607 spin_unlock_irq(&mctz->lock);
2608 css_put(&mz->memcg->css);
2611 * Could not reclaim anything and there are no more
2612 * mem cgroups to try or we seem to be looping without
2613 * reclaiming anything.
2615 if (!nr_reclaimed &&
2617 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2619 } while (!nr_reclaimed);
2621 css_put(&next_mz->memcg->css);
2622 return nr_reclaimed;
2626 * Test whether @memcg has children, dead or alive. Note that this
2627 * function doesn't care whether @memcg has use_hierarchy enabled and
2628 * returns %true if there are child csses according to the cgroup
2629 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2631 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2636 ret = css_next_child(NULL, &memcg->css);
2642 * Reclaims as many pages from the given memcg as possible and moves
2643 * the rest to the parent.
2645 * Caller is responsible for holding css reference for memcg.
2647 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2649 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2651 /* we call try-to-free pages for make this cgroup empty */
2652 lru_add_drain_all();
2653 /* try to free all pages in this cgroup */
2654 while (nr_retries && page_counter_read(&memcg->memory)) {
2657 if (signal_pending(current))
2660 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2664 /* maybe some writeback is necessary */
2665 congestion_wait(BLK_RW_ASYNC, HZ/10);
2673 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2674 char *buf, size_t nbytes,
2677 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2679 if (mem_cgroup_is_root(memcg))
2681 return mem_cgroup_force_empty(memcg) ?: nbytes;
2684 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2687 return mem_cgroup_from_css(css)->use_hierarchy;
2690 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2691 struct cftype *cft, u64 val)
2694 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2695 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2697 if (memcg->use_hierarchy == val)
2701 * If parent's use_hierarchy is set, we can't make any modifications
2702 * in the child subtrees. If it is unset, then the change can
2703 * occur, provided the current cgroup has no children.
2705 * For the root cgroup, parent_mem is NULL, we allow value to be
2706 * set if there are no children.
2708 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2709 (val == 1 || val == 0)) {
2710 if (!memcg_has_children(memcg))
2711 memcg->use_hierarchy = val;
2720 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2722 struct mem_cgroup *iter;
2725 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2727 for_each_mem_cgroup_tree(iter, memcg) {
2728 for (i = 0; i < MEMCG_NR_STAT; i++)
2729 stat[i] += mem_cgroup_read_stat(iter, i);
2733 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2735 struct mem_cgroup *iter;
2738 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2740 for_each_mem_cgroup_tree(iter, memcg) {
2741 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2742 events[i] += mem_cgroup_read_events(iter, i);
2746 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2748 unsigned long val = 0;
2750 if (mem_cgroup_is_root(memcg)) {
2751 struct mem_cgroup *iter;
2753 for_each_mem_cgroup_tree(iter, memcg) {
2754 val += mem_cgroup_read_stat(iter,
2755 MEM_CGROUP_STAT_CACHE);
2756 val += mem_cgroup_read_stat(iter,
2757 MEM_CGROUP_STAT_RSS);
2759 val += mem_cgroup_read_stat(iter,
2760 MEM_CGROUP_STAT_SWAP);
2764 val = page_counter_read(&memcg->memory);
2766 val = page_counter_read(&memcg->memsw);
2779 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2782 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2783 struct page_counter *counter;
2785 switch (MEMFILE_TYPE(cft->private)) {
2787 counter = &memcg->memory;
2790 counter = &memcg->memsw;
2793 counter = &memcg->kmem;
2796 counter = &memcg->tcpmem;
2802 switch (MEMFILE_ATTR(cft->private)) {
2804 if (counter == &memcg->memory)
2805 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2806 if (counter == &memcg->memsw)
2807 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2808 return (u64)page_counter_read(counter) * PAGE_SIZE;
2810 return (u64)counter->limit * PAGE_SIZE;
2812 return (u64)counter->watermark * PAGE_SIZE;
2814 return counter->failcnt;
2815 case RES_SOFT_LIMIT:
2816 return (u64)memcg->soft_limit * PAGE_SIZE;
2823 static int memcg_online_kmem(struct mem_cgroup *memcg)
2827 BUG_ON(memcg->kmemcg_id >= 0);
2828 BUG_ON(memcg->kmem_state);
2830 memcg_id = memcg_alloc_cache_id();
2834 static_branch_inc(&memcg_kmem_enabled_key);
2836 * A memory cgroup is considered kmem-online as soon as it gets
2837 * kmemcg_id. Setting the id after enabling static branching will
2838 * guarantee no one starts accounting before all call sites are
2841 memcg->kmemcg_id = memcg_id;
2842 memcg->kmem_state = KMEM_ONLINE;
2847 static int memcg_propagate_kmem(struct mem_cgroup *parent,
2848 struct mem_cgroup *memcg)
2852 mutex_lock(&memcg_limit_mutex);
2854 * If the parent cgroup is not kmem-online now, it cannot be
2855 * onlined after this point, because it has at least one child
2858 if (memcg_kmem_online(parent) ||
2859 (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nokmem))
2860 ret = memcg_online_kmem(memcg);
2861 mutex_unlock(&memcg_limit_mutex);
2865 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2867 struct cgroup_subsys_state *css;
2868 struct mem_cgroup *parent, *child;
2871 if (memcg->kmem_state != KMEM_ONLINE)
2874 * Clear the online state before clearing memcg_caches array
2875 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2876 * guarantees that no cache will be created for this cgroup
2877 * after we are done (see memcg_create_kmem_cache()).
2879 memcg->kmem_state = KMEM_ALLOCATED;
2881 memcg_deactivate_kmem_caches(memcg);
2883 kmemcg_id = memcg->kmemcg_id;
2884 BUG_ON(kmemcg_id < 0);
2886 parent = parent_mem_cgroup(memcg);
2888 parent = root_mem_cgroup;
2891 * Change kmemcg_id of this cgroup and all its descendants to the
2892 * parent's id, and then move all entries from this cgroup's list_lrus
2893 * to ones of the parent. After we have finished, all list_lrus
2894 * corresponding to this cgroup are guaranteed to remain empty. The
2895 * ordering is imposed by list_lru_node->lock taken by
2896 * memcg_drain_all_list_lrus().
2898 css_for_each_descendant_pre(css, &memcg->css) {
2899 child = mem_cgroup_from_css(css);
2900 BUG_ON(child->kmemcg_id != kmemcg_id);
2901 child->kmemcg_id = parent->kmemcg_id;
2902 if (!memcg->use_hierarchy)
2905 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2907 memcg_free_cache_id(kmemcg_id);
2910 static void memcg_free_kmem(struct mem_cgroup *memcg)
2912 /* css_alloc() failed, offlining didn't happen */
2913 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2914 memcg_offline_kmem(memcg);
2916 if (memcg->kmem_state == KMEM_ALLOCATED) {
2917 memcg_destroy_kmem_caches(memcg);
2918 static_branch_dec(&memcg_kmem_enabled_key);
2919 WARN_ON(page_counter_read(&memcg->kmem));
2923 static int memcg_propagate_kmem(struct mem_cgroup *parent, struct mem_cgroup *memcg)
2927 static int memcg_online_kmem(struct mem_cgroup *memcg)
2931 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2934 static void memcg_free_kmem(struct mem_cgroup *memcg)
2937 #endif /* !CONFIG_SLOB */
2939 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2940 unsigned long limit)
2944 mutex_lock(&memcg_limit_mutex);
2945 /* Top-level cgroup doesn't propagate from root */
2946 if (!memcg_kmem_online(memcg)) {
2947 if (cgroup_is_populated(memcg->css.cgroup) ||
2948 (memcg->use_hierarchy && memcg_has_children(memcg)))
2952 ret = memcg_online_kmem(memcg);
2956 ret = page_counter_limit(&memcg->kmem, limit);
2958 mutex_unlock(&memcg_limit_mutex);
2962 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2966 mutex_lock(&memcg_limit_mutex);
2968 ret = page_counter_limit(&memcg->tcpmem, limit);
2972 if (!memcg->tcpmem_active) {
2974 * The active flag needs to be written after the static_key
2975 * update. This is what guarantees that the socket activation
2976 * function is the last one to run. See sock_update_memcg() for
2977 * details, and note that we don't mark any socket as belonging
2978 * to this memcg until that flag is up.
2980 * We need to do this, because static_keys will span multiple
2981 * sites, but we can't control their order. If we mark a socket
2982 * as accounted, but the accounting functions are not patched in
2983 * yet, we'll lose accounting.
2985 * We never race with the readers in sock_update_memcg(),
2986 * because when this value change, the code to process it is not
2989 static_branch_inc(&memcg_sockets_enabled_key);
2990 memcg->tcpmem_active = true;
2993 mutex_unlock(&memcg_limit_mutex);
2998 * The user of this function is...
3001 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3002 char *buf, size_t nbytes, loff_t off)
3004 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3005 unsigned long nr_pages;
3008 buf = strstrip(buf);
3009 ret = page_counter_memparse(buf, "-1", &nr_pages);
3013 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3015 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3019 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3021 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3024 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3027 ret = memcg_update_kmem_limit(memcg, nr_pages);
3030 ret = memcg_update_tcp_limit(memcg, nr_pages);
3034 case RES_SOFT_LIMIT:
3035 memcg->soft_limit = nr_pages;
3039 return ret ?: nbytes;
3042 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3043 size_t nbytes, loff_t off)
3045 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3046 struct page_counter *counter;
3048 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3050 counter = &memcg->memory;
3053 counter = &memcg->memsw;
3056 counter = &memcg->kmem;
3059 counter = &memcg->tcpmem;
3065 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3067 page_counter_reset_watermark(counter);
3070 counter->failcnt = 0;
3079 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3082 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3086 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3087 struct cftype *cft, u64 val)
3089 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3091 if (val & ~MOVE_MASK)
3095 * No kind of locking is needed in here, because ->can_attach() will
3096 * check this value once in the beginning of the process, and then carry
3097 * on with stale data. This means that changes to this value will only
3098 * affect task migrations starting after the change.
3100 memcg->move_charge_at_immigrate = val;
3104 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3105 struct cftype *cft, u64 val)
3112 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3116 unsigned int lru_mask;
3119 static const struct numa_stat stats[] = {
3120 { "total", LRU_ALL },
3121 { "file", LRU_ALL_FILE },
3122 { "anon", LRU_ALL_ANON },
3123 { "unevictable", BIT(LRU_UNEVICTABLE) },
3125 const struct numa_stat *stat;
3128 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3130 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3131 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3132 seq_printf(m, "%s=%lu", stat->name, nr);
3133 for_each_node_state(nid, N_MEMORY) {
3134 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3136 seq_printf(m, " N%d=%lu", nid, nr);
3141 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3142 struct mem_cgroup *iter;
3145 for_each_mem_cgroup_tree(iter, memcg)
3146 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3147 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3148 for_each_node_state(nid, N_MEMORY) {
3150 for_each_mem_cgroup_tree(iter, memcg)
3151 nr += mem_cgroup_node_nr_lru_pages(
3152 iter, nid, stat->lru_mask);
3153 seq_printf(m, " N%d=%lu", nid, nr);
3160 #endif /* CONFIG_NUMA */
3162 static int memcg_stat_show(struct seq_file *m, void *v)
3164 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3165 unsigned long memory, memsw;
3166 struct mem_cgroup *mi;
3169 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3170 MEM_CGROUP_STAT_NSTATS);
3171 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3172 MEM_CGROUP_EVENTS_NSTATS);
3173 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3175 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3176 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3178 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3179 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3182 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3183 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3184 mem_cgroup_read_events(memcg, i));
3186 for (i = 0; i < NR_LRU_LISTS; i++)
3187 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3188 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3190 /* Hierarchical information */
3191 memory = memsw = PAGE_COUNTER_MAX;
3192 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3193 memory = min(memory, mi->memory.limit);
3194 memsw = min(memsw, mi->memsw.limit);
3196 seq_printf(m, "hierarchical_memory_limit %llu\n",
3197 (u64)memory * PAGE_SIZE);
3198 if (do_memsw_account())
3199 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3200 (u64)memsw * PAGE_SIZE);
3202 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3203 unsigned long long val = 0;
3205 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3207 for_each_mem_cgroup_tree(mi, memcg)
3208 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3209 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3212 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3213 unsigned long long val = 0;
3215 for_each_mem_cgroup_tree(mi, memcg)
3216 val += mem_cgroup_read_events(mi, i);
3217 seq_printf(m, "total_%s %llu\n",
3218 mem_cgroup_events_names[i], val);
3221 for (i = 0; i < NR_LRU_LISTS; i++) {
3222 unsigned long long val = 0;
3224 for_each_mem_cgroup_tree(mi, memcg)
3225 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3226 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3229 #ifdef CONFIG_DEBUG_VM
3232 struct mem_cgroup_per_zone *mz;
3233 struct zone_reclaim_stat *rstat;
3234 unsigned long recent_rotated[2] = {0, 0};
3235 unsigned long recent_scanned[2] = {0, 0};
3237 for_each_online_node(nid)
3238 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3239 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3240 rstat = &mz->lruvec.reclaim_stat;
3242 recent_rotated[0] += rstat->recent_rotated[0];
3243 recent_rotated[1] += rstat->recent_rotated[1];
3244 recent_scanned[0] += rstat->recent_scanned[0];
3245 recent_scanned[1] += rstat->recent_scanned[1];
3247 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3248 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3249 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3250 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3257 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3260 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3262 return mem_cgroup_swappiness(memcg);
3265 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3266 struct cftype *cft, u64 val)
3268 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3274 memcg->swappiness = val;
3276 vm_swappiness = val;
3281 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3283 struct mem_cgroup_threshold_ary *t;
3284 unsigned long usage;
3289 t = rcu_dereference(memcg->thresholds.primary);
3291 t = rcu_dereference(memcg->memsw_thresholds.primary);
3296 usage = mem_cgroup_usage(memcg, swap);
3299 * current_threshold points to threshold just below or equal to usage.
3300 * If it's not true, a threshold was crossed after last
3301 * call of __mem_cgroup_threshold().
3303 i = t->current_threshold;
3306 * Iterate backward over array of thresholds starting from
3307 * current_threshold and check if a threshold is crossed.
3308 * If none of thresholds below usage is crossed, we read
3309 * only one element of the array here.
3311 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3312 eventfd_signal(t->entries[i].eventfd, 1);
3314 /* i = current_threshold + 1 */
3318 * Iterate forward over array of thresholds starting from
3319 * current_threshold+1 and check if a threshold is crossed.
3320 * If none of thresholds above usage is crossed, we read
3321 * only one element of the array here.
3323 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3324 eventfd_signal(t->entries[i].eventfd, 1);
3326 /* Update current_threshold */
3327 t->current_threshold = i - 1;
3332 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3335 __mem_cgroup_threshold(memcg, false);
3336 if (do_memsw_account())
3337 __mem_cgroup_threshold(memcg, true);
3339 memcg = parent_mem_cgroup(memcg);
3343 static int compare_thresholds(const void *a, const void *b)
3345 const struct mem_cgroup_threshold *_a = a;
3346 const struct mem_cgroup_threshold *_b = b;
3348 if (_a->threshold > _b->threshold)
3351 if (_a->threshold < _b->threshold)
3357 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3359 struct mem_cgroup_eventfd_list *ev;
3361 spin_lock(&memcg_oom_lock);
3363 list_for_each_entry(ev, &memcg->oom_notify, list)
3364 eventfd_signal(ev->eventfd, 1);
3366 spin_unlock(&memcg_oom_lock);
3370 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3372 struct mem_cgroup *iter;
3374 for_each_mem_cgroup_tree(iter, memcg)
3375 mem_cgroup_oom_notify_cb(iter);
3378 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3379 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3381 struct mem_cgroup_thresholds *thresholds;
3382 struct mem_cgroup_threshold_ary *new;
3383 unsigned long threshold;
3384 unsigned long usage;
3387 ret = page_counter_memparse(args, "-1", &threshold);
3391 mutex_lock(&memcg->thresholds_lock);
3394 thresholds = &memcg->thresholds;
3395 usage = mem_cgroup_usage(memcg, false);
3396 } else if (type == _MEMSWAP) {
3397 thresholds = &memcg->memsw_thresholds;
3398 usage = mem_cgroup_usage(memcg, true);
3402 /* Check if a threshold crossed before adding a new one */
3403 if (thresholds->primary)
3404 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3406 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3408 /* Allocate memory for new array of thresholds */
3409 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3417 /* Copy thresholds (if any) to new array */
3418 if (thresholds->primary) {
3419 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3420 sizeof(struct mem_cgroup_threshold));
3423 /* Add new threshold */
3424 new->entries[size - 1].eventfd = eventfd;
3425 new->entries[size - 1].threshold = threshold;
3427 /* Sort thresholds. Registering of new threshold isn't time-critical */
3428 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3429 compare_thresholds, NULL);
3431 /* Find current threshold */
3432 new->current_threshold = -1;
3433 for (i = 0; i < size; i++) {
3434 if (new->entries[i].threshold <= usage) {
3436 * new->current_threshold will not be used until
3437 * rcu_assign_pointer(), so it's safe to increment
3440 ++new->current_threshold;
3445 /* Free old spare buffer and save old primary buffer as spare */
3446 kfree(thresholds->spare);
3447 thresholds->spare = thresholds->primary;
3449 rcu_assign_pointer(thresholds->primary, new);
3451 /* To be sure that nobody uses thresholds */
3455 mutex_unlock(&memcg->thresholds_lock);
3460 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3461 struct eventfd_ctx *eventfd, const char *args)
3463 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3466 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3467 struct eventfd_ctx *eventfd, const char *args)
3469 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3472 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3473 struct eventfd_ctx *eventfd, enum res_type type)
3475 struct mem_cgroup_thresholds *thresholds;
3476 struct mem_cgroup_threshold_ary *new;
3477 unsigned long usage;
3480 mutex_lock(&memcg->thresholds_lock);
3483 thresholds = &memcg->thresholds;
3484 usage = mem_cgroup_usage(memcg, false);
3485 } else if (type == _MEMSWAP) {
3486 thresholds = &memcg->memsw_thresholds;
3487 usage = mem_cgroup_usage(memcg, true);
3491 if (!thresholds->primary)
3494 /* Check if a threshold crossed before removing */
3495 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3497 /* Calculate new number of threshold */
3499 for (i = 0; i < thresholds->primary->size; i++) {
3500 if (thresholds->primary->entries[i].eventfd != eventfd)
3504 new = thresholds->spare;
3506 /* Set thresholds array to NULL if we don't have thresholds */
3515 /* Copy thresholds and find current threshold */
3516 new->current_threshold = -1;
3517 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3518 if (thresholds->primary->entries[i].eventfd == eventfd)
3521 new->entries[j] = thresholds->primary->entries[i];
3522 if (new->entries[j].threshold <= usage) {
3524 * new->current_threshold will not be used
3525 * until rcu_assign_pointer(), so it's safe to increment
3528 ++new->current_threshold;
3534 /* Swap primary and spare array */
3535 thresholds->spare = thresholds->primary;
3537 rcu_assign_pointer(thresholds->primary, new);
3539 /* To be sure that nobody uses thresholds */
3542 /* If all events are unregistered, free the spare array */
3544 kfree(thresholds->spare);
3545 thresholds->spare = NULL;
3548 mutex_unlock(&memcg->thresholds_lock);
3551 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3552 struct eventfd_ctx *eventfd)
3554 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3557 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3558 struct eventfd_ctx *eventfd)
3560 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3563 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3564 struct eventfd_ctx *eventfd, const char *args)
3566 struct mem_cgroup_eventfd_list *event;
3568 event = kmalloc(sizeof(*event), GFP_KERNEL);
3572 spin_lock(&memcg_oom_lock);
3574 event->eventfd = eventfd;
3575 list_add(&event->list, &memcg->oom_notify);
3577 /* already in OOM ? */
3578 if (memcg->under_oom)
3579 eventfd_signal(eventfd, 1);
3580 spin_unlock(&memcg_oom_lock);
3585 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3586 struct eventfd_ctx *eventfd)
3588 struct mem_cgroup_eventfd_list *ev, *tmp;
3590 spin_lock(&memcg_oom_lock);
3592 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3593 if (ev->eventfd == eventfd) {
3594 list_del(&ev->list);
3599 spin_unlock(&memcg_oom_lock);
3602 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3604 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3606 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3607 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3611 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3612 struct cftype *cft, u64 val)
3614 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3616 /* cannot set to root cgroup and only 0 and 1 are allowed */
3617 if (!css->parent || !((val == 0) || (val == 1)))
3620 memcg->oom_kill_disable = val;
3622 memcg_oom_recover(memcg);
3627 #ifdef CONFIG_CGROUP_WRITEBACK
3629 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3631 return &memcg->cgwb_list;
3634 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3636 return wb_domain_init(&memcg->cgwb_domain, gfp);
3639 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3641 wb_domain_exit(&memcg->cgwb_domain);
3644 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3646 wb_domain_size_changed(&memcg->cgwb_domain);
3649 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3651 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3653 if (!memcg->css.parent)
3656 return &memcg->cgwb_domain;
3660 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3661 * @wb: bdi_writeback in question
3662 * @pfilepages: out parameter for number of file pages
3663 * @pheadroom: out parameter for number of allocatable pages according to memcg
3664 * @pdirty: out parameter for number of dirty pages
3665 * @pwriteback: out parameter for number of pages under writeback
3667 * Determine the numbers of file, headroom, dirty, and writeback pages in
3668 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3669 * is a bit more involved.
3671 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3672 * headroom is calculated as the lowest headroom of itself and the
3673 * ancestors. Note that this doesn't consider the actual amount of
3674 * available memory in the system. The caller should further cap
3675 * *@pheadroom accordingly.
3677 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3678 unsigned long *pheadroom, unsigned long *pdirty,
3679 unsigned long *pwriteback)
3681 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3682 struct mem_cgroup *parent;
3684 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3686 /* this should eventually include NR_UNSTABLE_NFS */
3687 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3688 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3689 (1 << LRU_ACTIVE_FILE));
3690 *pheadroom = PAGE_COUNTER_MAX;
3692 while ((parent = parent_mem_cgroup(memcg))) {
3693 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3694 unsigned long used = page_counter_read(&memcg->memory);
3696 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3701 #else /* CONFIG_CGROUP_WRITEBACK */
3703 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3708 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3712 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3716 #endif /* CONFIG_CGROUP_WRITEBACK */
3719 * DO NOT USE IN NEW FILES.
3721 * "cgroup.event_control" implementation.
3723 * This is way over-engineered. It tries to support fully configurable
3724 * events for each user. Such level of flexibility is completely
3725 * unnecessary especially in the light of the planned unified hierarchy.
3727 * Please deprecate this and replace with something simpler if at all
3732 * Unregister event and free resources.
3734 * Gets called from workqueue.
3736 static void memcg_event_remove(struct work_struct *work)
3738 struct mem_cgroup_event *event =
3739 container_of(work, struct mem_cgroup_event, remove);
3740 struct mem_cgroup *memcg = event->memcg;
3742 remove_wait_queue(event->wqh, &event->wait);
3744 event->unregister_event(memcg, event->eventfd);
3746 /* Notify userspace the event is going away. */
3747 eventfd_signal(event->eventfd, 1);
3749 eventfd_ctx_put(event->eventfd);
3751 css_put(&memcg->css);
3755 * Gets called on POLLHUP on eventfd when user closes it.
3757 * Called with wqh->lock held and interrupts disabled.
3759 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3760 int sync, void *key)
3762 struct mem_cgroup_event *event =
3763 container_of(wait, struct mem_cgroup_event, wait);
3764 struct mem_cgroup *memcg = event->memcg;
3765 unsigned long flags = (unsigned long)key;
3767 if (flags & POLLHUP) {
3769 * If the event has been detached at cgroup removal, we
3770 * can simply return knowing the other side will cleanup
3773 * We can't race against event freeing since the other
3774 * side will require wqh->lock via remove_wait_queue(),
3777 spin_lock(&memcg->event_list_lock);
3778 if (!list_empty(&event->list)) {
3779 list_del_init(&event->list);
3781 * We are in atomic context, but cgroup_event_remove()
3782 * may sleep, so we have to call it in workqueue.
3784 schedule_work(&event->remove);
3786 spin_unlock(&memcg->event_list_lock);
3792 static void memcg_event_ptable_queue_proc(struct file *file,
3793 wait_queue_head_t *wqh, poll_table *pt)
3795 struct mem_cgroup_event *event =
3796 container_of(pt, struct mem_cgroup_event, pt);
3799 add_wait_queue(wqh, &event->wait);
3803 * DO NOT USE IN NEW FILES.
3805 * Parse input and register new cgroup event handler.
3807 * Input must be in format '<event_fd> <control_fd> <args>'.
3808 * Interpretation of args is defined by control file implementation.
3810 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3811 char *buf, size_t nbytes, loff_t off)
3813 struct cgroup_subsys_state *css = of_css(of);
3814 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3815 struct mem_cgroup_event *event;
3816 struct cgroup_subsys_state *cfile_css;
3817 unsigned int efd, cfd;
3824 buf = strstrip(buf);
3826 efd = simple_strtoul(buf, &endp, 10);
3831 cfd = simple_strtoul(buf, &endp, 10);
3832 if ((*endp != ' ') && (*endp != '\0'))
3836 event = kzalloc(sizeof(*event), GFP_KERNEL);
3840 event->memcg = memcg;
3841 INIT_LIST_HEAD(&event->list);
3842 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3843 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3844 INIT_WORK(&event->remove, memcg_event_remove);
3852 event->eventfd = eventfd_ctx_fileget(efile.file);
3853 if (IS_ERR(event->eventfd)) {
3854 ret = PTR_ERR(event->eventfd);
3861 goto out_put_eventfd;
3864 /* the process need read permission on control file */
3865 /* AV: shouldn't we check that it's been opened for read instead? */
3866 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3871 * Determine the event callbacks and set them in @event. This used
3872 * to be done via struct cftype but cgroup core no longer knows
3873 * about these events. The following is crude but the whole thing
3874 * is for compatibility anyway.
3876 * DO NOT ADD NEW FILES.
3878 name = cfile.file->f_path.dentry->d_name.name;
3880 if (!strcmp(name, "memory.usage_in_bytes")) {
3881 event->register_event = mem_cgroup_usage_register_event;
3882 event->unregister_event = mem_cgroup_usage_unregister_event;
3883 } else if (!strcmp(name, "memory.oom_control")) {
3884 event->register_event = mem_cgroup_oom_register_event;
3885 event->unregister_event = mem_cgroup_oom_unregister_event;
3886 } else if (!strcmp(name, "memory.pressure_level")) {
3887 event->register_event = vmpressure_register_event;
3888 event->unregister_event = vmpressure_unregister_event;
3889 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3890 event->register_event = memsw_cgroup_usage_register_event;
3891 event->unregister_event = memsw_cgroup_usage_unregister_event;
3898 * Verify @cfile should belong to @css. Also, remaining events are
3899 * automatically removed on cgroup destruction but the removal is
3900 * asynchronous, so take an extra ref on @css.
3902 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3903 &memory_cgrp_subsys);
3905 if (IS_ERR(cfile_css))
3907 if (cfile_css != css) {
3912 ret = event->register_event(memcg, event->eventfd, buf);
3916 efile.file->f_op->poll(efile.file, &event->pt);
3918 spin_lock(&memcg->event_list_lock);
3919 list_add(&event->list, &memcg->event_list);
3920 spin_unlock(&memcg->event_list_lock);
3932 eventfd_ctx_put(event->eventfd);
3941 static struct cftype mem_cgroup_legacy_files[] = {
3943 .name = "usage_in_bytes",
3944 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3945 .read_u64 = mem_cgroup_read_u64,
3948 .name = "max_usage_in_bytes",
3949 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3950 .write = mem_cgroup_reset,
3951 .read_u64 = mem_cgroup_read_u64,
3954 .name = "limit_in_bytes",
3955 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3956 .write = mem_cgroup_write,
3957 .read_u64 = mem_cgroup_read_u64,
3960 .name = "soft_limit_in_bytes",
3961 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3962 .write = mem_cgroup_write,
3963 .read_u64 = mem_cgroup_read_u64,
3967 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3968 .write = mem_cgroup_reset,
3969 .read_u64 = mem_cgroup_read_u64,
3973 .seq_show = memcg_stat_show,
3976 .name = "force_empty",
3977 .write = mem_cgroup_force_empty_write,
3980 .name = "use_hierarchy",
3981 .write_u64 = mem_cgroup_hierarchy_write,
3982 .read_u64 = mem_cgroup_hierarchy_read,
3985 .name = "cgroup.event_control", /* XXX: for compat */
3986 .write = memcg_write_event_control,
3987 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3990 .name = "swappiness",
3991 .read_u64 = mem_cgroup_swappiness_read,
3992 .write_u64 = mem_cgroup_swappiness_write,
3995 .name = "move_charge_at_immigrate",
3996 .read_u64 = mem_cgroup_move_charge_read,
3997 .write_u64 = mem_cgroup_move_charge_write,
4000 .name = "oom_control",
4001 .seq_show = mem_cgroup_oom_control_read,
4002 .write_u64 = mem_cgroup_oom_control_write,
4003 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4006 .name = "pressure_level",
4010 .name = "numa_stat",
4011 .seq_show = memcg_numa_stat_show,
4015 .name = "kmem.limit_in_bytes",
4016 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4017 .write = mem_cgroup_write,
4018 .read_u64 = mem_cgroup_read_u64,
4021 .name = "kmem.usage_in_bytes",
4022 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4023 .read_u64 = mem_cgroup_read_u64,
4026 .name = "kmem.failcnt",
4027 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4028 .write = mem_cgroup_reset,
4029 .read_u64 = mem_cgroup_read_u64,
4032 .name = "kmem.max_usage_in_bytes",
4033 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4034 .write = mem_cgroup_reset,
4035 .read_u64 = mem_cgroup_read_u64,
4037 #ifdef CONFIG_SLABINFO
4039 .name = "kmem.slabinfo",
4040 .seq_start = slab_start,
4041 .seq_next = slab_next,
4042 .seq_stop = slab_stop,
4043 .seq_show = memcg_slab_show,
4047 .name = "kmem.tcp.limit_in_bytes",
4048 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4049 .write = mem_cgroup_write,
4050 .read_u64 = mem_cgroup_read_u64,
4053 .name = "kmem.tcp.usage_in_bytes",
4054 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4055 .read_u64 = mem_cgroup_read_u64,
4058 .name = "kmem.tcp.failcnt",
4059 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4060 .write = mem_cgroup_reset,
4061 .read_u64 = mem_cgroup_read_u64,
4064 .name = "kmem.tcp.max_usage_in_bytes",
4065 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4066 .write = mem_cgroup_reset,
4067 .read_u64 = mem_cgroup_read_u64,
4069 { }, /* terminate */
4072 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4074 struct mem_cgroup_per_node *pn;
4075 struct mem_cgroup_per_zone *mz;
4076 int zone, tmp = node;
4078 * This routine is called against possible nodes.
4079 * But it's BUG to call kmalloc() against offline node.
4081 * TODO: this routine can waste much memory for nodes which will
4082 * never be onlined. It's better to use memory hotplug callback
4085 if (!node_state(node, N_NORMAL_MEMORY))
4087 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4091 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4092 mz = &pn->zoneinfo[zone];
4093 lruvec_init(&mz->lruvec);
4094 mz->usage_in_excess = 0;
4095 mz->on_tree = false;
4098 memcg->nodeinfo[node] = pn;
4102 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4104 kfree(memcg->nodeinfo[node]);
4107 static void mem_cgroup_free(struct mem_cgroup *memcg)
4111 memcg_wb_domain_exit(memcg);
4113 free_mem_cgroup_per_zone_info(memcg, node);
4114 free_percpu(memcg->stat);
4118 static struct mem_cgroup *mem_cgroup_alloc(void)
4120 struct mem_cgroup *memcg;
4124 size = sizeof(struct mem_cgroup);
4125 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4127 memcg = kzalloc(size, GFP_KERNEL);
4131 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4136 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4139 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4142 INIT_WORK(&memcg->high_work, high_work_func);
4143 memcg->last_scanned_node = MAX_NUMNODES;
4144 INIT_LIST_HEAD(&memcg->oom_notify);
4145 mutex_init(&memcg->thresholds_lock);
4146 spin_lock_init(&memcg->move_lock);
4147 vmpressure_init(&memcg->vmpressure);
4148 INIT_LIST_HEAD(&memcg->event_list);
4149 spin_lock_init(&memcg->event_list_lock);
4150 memcg->socket_pressure = jiffies;
4152 memcg->kmemcg_id = -1;
4154 #ifdef CONFIG_CGROUP_WRITEBACK
4155 INIT_LIST_HEAD(&memcg->cgwb_list);
4159 mem_cgroup_free(memcg);
4163 static struct cgroup_subsys_state * __ref
4164 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4166 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4167 struct mem_cgroup *memcg;
4168 long error = -ENOMEM;
4170 memcg = mem_cgroup_alloc();
4172 return ERR_PTR(error);
4174 memcg->high = PAGE_COUNTER_MAX;
4175 memcg->soft_limit = PAGE_COUNTER_MAX;
4177 memcg->swappiness = mem_cgroup_swappiness(parent);
4178 memcg->oom_kill_disable = parent->oom_kill_disable;
4180 if (parent && parent->use_hierarchy) {
4181 memcg->use_hierarchy = true;
4182 page_counter_init(&memcg->memory, &parent->memory);
4183 page_counter_init(&memcg->swap, &parent->swap);
4184 page_counter_init(&memcg->memsw, &parent->memsw);
4185 page_counter_init(&memcg->kmem, &parent->kmem);
4186 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4188 page_counter_init(&memcg->memory, NULL);
4189 page_counter_init(&memcg->swap, NULL);
4190 page_counter_init(&memcg->memsw, NULL);
4191 page_counter_init(&memcg->kmem, NULL);
4192 page_counter_init(&memcg->tcpmem, NULL);
4194 * Deeper hierachy with use_hierarchy == false doesn't make
4195 * much sense so let cgroup subsystem know about this
4196 * unfortunate state in our controller.
4198 if (parent != root_mem_cgroup)
4199 memory_cgrp_subsys.broken_hierarchy = true;
4202 /* The following stuff does not apply to the root */
4204 root_mem_cgroup = memcg;
4208 error = memcg_propagate_kmem(parent, memcg);
4212 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4213 static_branch_inc(&memcg_sockets_enabled_key);
4217 mem_cgroup_free(memcg);
4222 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4224 if (css->id > MEM_CGROUP_ID_MAX)
4230 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4232 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4233 struct mem_cgroup_event *event, *tmp;
4236 * Unregister events and notify userspace.
4237 * Notify userspace about cgroup removing only after rmdir of cgroup
4238 * directory to avoid race between userspace and kernelspace.
4240 spin_lock(&memcg->event_list_lock);
4241 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4242 list_del_init(&event->list);
4243 schedule_work(&event->remove);
4245 spin_unlock(&memcg->event_list_lock);
4247 memcg_offline_kmem(memcg);
4248 wb_memcg_offline(memcg);
4251 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4253 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4255 invalidate_reclaim_iterators(memcg);
4258 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4260 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4262 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4263 static_branch_dec(&memcg_sockets_enabled_key);
4265 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4266 static_branch_dec(&memcg_sockets_enabled_key);
4268 vmpressure_cleanup(&memcg->vmpressure);
4269 cancel_work_sync(&memcg->high_work);
4270 mem_cgroup_remove_from_trees(memcg);
4271 memcg_free_kmem(memcg);
4272 mem_cgroup_free(memcg);
4276 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4277 * @css: the target css
4279 * Reset the states of the mem_cgroup associated with @css. This is
4280 * invoked when the userland requests disabling on the default hierarchy
4281 * but the memcg is pinned through dependency. The memcg should stop
4282 * applying policies and should revert to the vanilla state as it may be
4283 * made visible again.
4285 * The current implementation only resets the essential configurations.
4286 * This needs to be expanded to cover all the visible parts.
4288 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4290 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4292 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4293 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4294 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4296 memcg->high = PAGE_COUNTER_MAX;
4297 memcg->soft_limit = PAGE_COUNTER_MAX;
4298 memcg_wb_domain_size_changed(memcg);
4302 /* Handlers for move charge at task migration. */
4303 static int mem_cgroup_do_precharge(unsigned long count)
4307 /* Try a single bulk charge without reclaim first, kswapd may wake */
4308 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4310 mc.precharge += count;
4314 /* Try charges one by one with reclaim */
4316 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4326 * get_mctgt_type - get target type of moving charge
4327 * @vma: the vma the pte to be checked belongs
4328 * @addr: the address corresponding to the pte to be checked
4329 * @ptent: the pte to be checked
4330 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4333 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4334 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4335 * move charge. if @target is not NULL, the page is stored in target->page
4336 * with extra refcnt got(Callers should handle it).
4337 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4338 * target for charge migration. if @target is not NULL, the entry is stored
4341 * Called with pte lock held.
4348 enum mc_target_type {
4354 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4355 unsigned long addr, pte_t ptent)
4357 struct page *page = vm_normal_page(vma, addr, ptent);
4359 if (!page || !page_mapped(page))
4361 if (PageAnon(page)) {
4362 if (!(mc.flags & MOVE_ANON))
4365 if (!(mc.flags & MOVE_FILE))
4368 if (!get_page_unless_zero(page))
4375 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4376 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4378 struct page *page = NULL;
4379 swp_entry_t ent = pte_to_swp_entry(ptent);
4381 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4384 * Because lookup_swap_cache() updates some statistics counter,
4385 * we call find_get_page() with swapper_space directly.
4387 page = find_get_page(swap_address_space(ent), ent.val);
4388 if (do_memsw_account())
4389 entry->val = ent.val;
4394 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4395 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4401 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4402 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4404 struct page *page = NULL;
4405 struct address_space *mapping;
4408 if (!vma->vm_file) /* anonymous vma */
4410 if (!(mc.flags & MOVE_FILE))
4413 mapping = vma->vm_file->f_mapping;
4414 pgoff = linear_page_index(vma, addr);
4416 /* page is moved even if it's not RSS of this task(page-faulted). */
4418 /* shmem/tmpfs may report page out on swap: account for that too. */
4419 if (shmem_mapping(mapping)) {
4420 page = find_get_entry(mapping, pgoff);
4421 if (radix_tree_exceptional_entry(page)) {
4422 swp_entry_t swp = radix_to_swp_entry(page);
4423 if (do_memsw_account())
4425 page = find_get_page(swap_address_space(swp), swp.val);
4428 page = find_get_page(mapping, pgoff);
4430 page = find_get_page(mapping, pgoff);
4436 * mem_cgroup_move_account - move account of the page
4438 * @nr_pages: number of regular pages (>1 for huge pages)
4439 * @from: mem_cgroup which the page is moved from.
4440 * @to: mem_cgroup which the page is moved to. @from != @to.
4442 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4444 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4447 static int mem_cgroup_move_account(struct page *page,
4449 struct mem_cgroup *from,
4450 struct mem_cgroup *to)
4452 unsigned long flags;
4453 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4457 VM_BUG_ON(from == to);
4458 VM_BUG_ON_PAGE(PageLRU(page), page);
4459 VM_BUG_ON(compound && !PageTransHuge(page));
4462 * Prevent mem_cgroup_migrate() from looking at
4463 * page->mem_cgroup of its source page while we change it.
4466 if (!trylock_page(page))
4470 if (page->mem_cgroup != from)
4473 anon = PageAnon(page);
4475 spin_lock_irqsave(&from->move_lock, flags);
4477 if (!anon && page_mapped(page)) {
4478 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4480 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4485 * move_lock grabbed above and caller set from->moving_account, so
4486 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4487 * So mapping should be stable for dirty pages.
4489 if (!anon && PageDirty(page)) {
4490 struct address_space *mapping = page_mapping(page);
4492 if (mapping_cap_account_dirty(mapping)) {
4493 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4495 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4500 if (PageWriteback(page)) {
4501 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4503 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4508 * It is safe to change page->mem_cgroup here because the page
4509 * is referenced, charged, and isolated - we can't race with
4510 * uncharging, charging, migration, or LRU putback.
4513 /* caller should have done css_get */
4514 page->mem_cgroup = to;
4515 spin_unlock_irqrestore(&from->move_lock, flags);
4519 local_irq_disable();
4520 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4521 memcg_check_events(to, page);
4522 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4523 memcg_check_events(from, page);
4531 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4532 unsigned long addr, pte_t ptent, union mc_target *target)
4534 struct page *page = NULL;
4535 enum mc_target_type ret = MC_TARGET_NONE;
4536 swp_entry_t ent = { .val = 0 };
4538 if (pte_present(ptent))
4539 page = mc_handle_present_pte(vma, addr, ptent);
4540 else if (is_swap_pte(ptent))
4541 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4542 else if (pte_none(ptent))
4543 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4545 if (!page && !ent.val)
4549 * Do only loose check w/o serialization.
4550 * mem_cgroup_move_account() checks the page is valid or
4551 * not under LRU exclusion.
4553 if (page->mem_cgroup == mc.from) {
4554 ret = MC_TARGET_PAGE;
4556 target->page = page;
4558 if (!ret || !target)
4561 /* There is a swap entry and a page doesn't exist or isn't charged */
4562 if (ent.val && !ret &&
4563 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4564 ret = MC_TARGET_SWAP;
4571 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4573 * We don't consider swapping or file mapped pages because THP does not
4574 * support them for now.
4575 * Caller should make sure that pmd_trans_huge(pmd) is true.
4577 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4578 unsigned long addr, pmd_t pmd, union mc_target *target)
4580 struct page *page = NULL;
4581 enum mc_target_type ret = MC_TARGET_NONE;
4583 page = pmd_page(pmd);
4584 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4585 if (!(mc.flags & MOVE_ANON))
4587 if (page->mem_cgroup == mc.from) {
4588 ret = MC_TARGET_PAGE;
4591 target->page = page;
4597 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4598 unsigned long addr, pmd_t pmd, union mc_target *target)
4600 return MC_TARGET_NONE;
4604 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4605 unsigned long addr, unsigned long end,
4606 struct mm_walk *walk)
4608 struct vm_area_struct *vma = walk->vma;
4612 ptl = pmd_trans_huge_lock(pmd, vma);
4614 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4615 mc.precharge += HPAGE_PMD_NR;
4620 if (pmd_trans_unstable(pmd))
4622 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4623 for (; addr != end; pte++, addr += PAGE_SIZE)
4624 if (get_mctgt_type(vma, addr, *pte, NULL))
4625 mc.precharge++; /* increment precharge temporarily */
4626 pte_unmap_unlock(pte - 1, ptl);
4632 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4634 unsigned long precharge;
4636 struct mm_walk mem_cgroup_count_precharge_walk = {
4637 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4640 down_read(&mm->mmap_sem);
4641 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4642 up_read(&mm->mmap_sem);
4644 precharge = mc.precharge;
4650 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4652 unsigned long precharge = mem_cgroup_count_precharge(mm);
4654 VM_BUG_ON(mc.moving_task);
4655 mc.moving_task = current;
4656 return mem_cgroup_do_precharge(precharge);
4659 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4660 static void __mem_cgroup_clear_mc(void)
4662 struct mem_cgroup *from = mc.from;
4663 struct mem_cgroup *to = mc.to;
4665 /* we must uncharge all the leftover precharges from mc.to */
4667 cancel_charge(mc.to, mc.precharge);
4671 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4672 * we must uncharge here.
4674 if (mc.moved_charge) {
4675 cancel_charge(mc.from, mc.moved_charge);
4676 mc.moved_charge = 0;
4678 /* we must fixup refcnts and charges */
4679 if (mc.moved_swap) {
4680 /* uncharge swap account from the old cgroup */
4681 if (!mem_cgroup_is_root(mc.from))
4682 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4685 * we charged both to->memory and to->memsw, so we
4686 * should uncharge to->memory.
4688 if (!mem_cgroup_is_root(mc.to))
4689 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4691 css_put_many(&mc.from->css, mc.moved_swap);
4693 /* we've already done css_get(mc.to) */
4696 memcg_oom_recover(from);
4697 memcg_oom_recover(to);
4698 wake_up_all(&mc.waitq);
4701 static void mem_cgroup_clear_mc(void)
4704 * we must clear moving_task before waking up waiters at the end of
4707 mc.moving_task = NULL;
4708 __mem_cgroup_clear_mc();
4709 spin_lock(&mc.lock);
4712 spin_unlock(&mc.lock);
4715 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4717 struct cgroup_subsys_state *css;
4718 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4719 struct mem_cgroup *from;
4720 struct task_struct *leader, *p;
4721 struct mm_struct *mm;
4722 unsigned long move_flags;
4725 /* charge immigration isn't supported on the default hierarchy */
4726 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4730 * Multi-process migrations only happen on the default hierarchy
4731 * where charge immigration is not used. Perform charge
4732 * immigration if @tset contains a leader and whine if there are
4736 cgroup_taskset_for_each_leader(leader, css, tset) {
4739 memcg = mem_cgroup_from_css(css);
4745 * We are now commited to this value whatever it is. Changes in this
4746 * tunable will only affect upcoming migrations, not the current one.
4747 * So we need to save it, and keep it going.
4749 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4753 from = mem_cgroup_from_task(p);
4755 VM_BUG_ON(from == memcg);
4757 mm = get_task_mm(p);
4760 /* We move charges only when we move a owner of the mm */
4761 if (mm->owner == p) {
4764 VM_BUG_ON(mc.precharge);
4765 VM_BUG_ON(mc.moved_charge);
4766 VM_BUG_ON(mc.moved_swap);
4768 spin_lock(&mc.lock);
4771 mc.flags = move_flags;
4772 spin_unlock(&mc.lock);
4773 /* We set mc.moving_task later */
4775 ret = mem_cgroup_precharge_mc(mm);
4777 mem_cgroup_clear_mc();
4783 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4786 mem_cgroup_clear_mc();
4789 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4790 unsigned long addr, unsigned long end,
4791 struct mm_walk *walk)
4794 struct vm_area_struct *vma = walk->vma;
4797 enum mc_target_type target_type;
4798 union mc_target target;
4801 ptl = pmd_trans_huge_lock(pmd, vma);
4803 if (mc.precharge < HPAGE_PMD_NR) {
4807 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4808 if (target_type == MC_TARGET_PAGE) {
4810 if (!isolate_lru_page(page)) {
4811 if (!mem_cgroup_move_account(page, true,
4813 mc.precharge -= HPAGE_PMD_NR;
4814 mc.moved_charge += HPAGE_PMD_NR;
4816 putback_lru_page(page);
4824 if (pmd_trans_unstable(pmd))
4827 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4828 for (; addr != end; addr += PAGE_SIZE) {
4829 pte_t ptent = *(pte++);
4835 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4836 case MC_TARGET_PAGE:
4839 * We can have a part of the split pmd here. Moving it
4840 * can be done but it would be too convoluted so simply
4841 * ignore such a partial THP and keep it in original
4842 * memcg. There should be somebody mapping the head.
4844 if (PageTransCompound(page))
4846 if (isolate_lru_page(page))
4848 if (!mem_cgroup_move_account(page, false,
4851 /* we uncharge from mc.from later. */
4854 putback_lru_page(page);
4855 put: /* get_mctgt_type() gets the page */
4858 case MC_TARGET_SWAP:
4860 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4862 /* we fixup refcnts and charges later. */
4870 pte_unmap_unlock(pte - 1, ptl);
4875 * We have consumed all precharges we got in can_attach().
4876 * We try charge one by one, but don't do any additional
4877 * charges to mc.to if we have failed in charge once in attach()
4880 ret = mem_cgroup_do_precharge(1);
4888 static void mem_cgroup_move_charge(struct mm_struct *mm)
4890 struct mm_walk mem_cgroup_move_charge_walk = {
4891 .pmd_entry = mem_cgroup_move_charge_pte_range,
4895 lru_add_drain_all();
4897 * Signal lock_page_memcg() to take the memcg's move_lock
4898 * while we're moving its pages to another memcg. Then wait
4899 * for already started RCU-only updates to finish.
4901 atomic_inc(&mc.from->moving_account);
4904 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4906 * Someone who are holding the mmap_sem might be waiting in
4907 * waitq. So we cancel all extra charges, wake up all waiters,
4908 * and retry. Because we cancel precharges, we might not be able
4909 * to move enough charges, but moving charge is a best-effort
4910 * feature anyway, so it wouldn't be a big problem.
4912 __mem_cgroup_clear_mc();
4917 * When we have consumed all precharges and failed in doing
4918 * additional charge, the page walk just aborts.
4920 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
4921 up_read(&mm->mmap_sem);
4922 atomic_dec(&mc.from->moving_account);
4925 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
4927 struct cgroup_subsys_state *css;
4928 struct task_struct *p = cgroup_taskset_first(tset, &css);
4929 struct mm_struct *mm = get_task_mm(p);
4933 mem_cgroup_move_charge(mm);
4937 mem_cgroup_clear_mc();
4939 #else /* !CONFIG_MMU */
4940 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4944 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4947 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
4953 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4954 * to verify whether we're attached to the default hierarchy on each mount
4957 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
4960 * use_hierarchy is forced on the default hierarchy. cgroup core
4961 * guarantees that @root doesn't have any children, so turning it
4962 * on for the root memcg is enough.
4964 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4965 root_mem_cgroup->use_hierarchy = true;
4967 root_mem_cgroup->use_hierarchy = false;
4970 static u64 memory_current_read(struct cgroup_subsys_state *css,
4973 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4975 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
4978 static int memory_low_show(struct seq_file *m, void *v)
4980 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4981 unsigned long low = READ_ONCE(memcg->low);
4983 if (low == PAGE_COUNTER_MAX)
4984 seq_puts(m, "max\n");
4986 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
4991 static ssize_t memory_low_write(struct kernfs_open_file *of,
4992 char *buf, size_t nbytes, loff_t off)
4994 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4998 buf = strstrip(buf);
4999 err = page_counter_memparse(buf, "max", &low);
5008 static int memory_high_show(struct seq_file *m, void *v)
5010 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5011 unsigned long high = READ_ONCE(memcg->high);
5013 if (high == PAGE_COUNTER_MAX)
5014 seq_puts(m, "max\n");
5016 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5021 static ssize_t memory_high_write(struct kernfs_open_file *of,
5022 char *buf, size_t nbytes, loff_t off)
5024 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5028 buf = strstrip(buf);
5029 err = page_counter_memparse(buf, "max", &high);
5035 memcg_wb_domain_size_changed(memcg);
5039 static int memory_max_show(struct seq_file *m, void *v)
5041 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5042 unsigned long max = READ_ONCE(memcg->memory.limit);
5044 if (max == PAGE_COUNTER_MAX)
5045 seq_puts(m, "max\n");
5047 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5052 static ssize_t memory_max_write(struct kernfs_open_file *of,
5053 char *buf, size_t nbytes, loff_t off)
5055 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5059 buf = strstrip(buf);
5060 err = page_counter_memparse(buf, "max", &max);
5064 err = mem_cgroup_resize_limit(memcg, max);
5068 memcg_wb_domain_size_changed(memcg);
5072 static int memory_events_show(struct seq_file *m, void *v)
5074 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5076 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5077 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5078 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5079 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5084 static int memory_stat_show(struct seq_file *m, void *v)
5086 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5087 unsigned long stat[MEMCG_NR_STAT];
5088 unsigned long events[MEMCG_NR_EVENTS];
5092 * Provide statistics on the state of the memory subsystem as
5093 * well as cumulative event counters that show past behavior.
5095 * This list is ordered following a combination of these gradients:
5096 * 1) generic big picture -> specifics and details
5097 * 2) reflecting userspace activity -> reflecting kernel heuristics
5099 * Current memory state:
5102 tree_stat(memcg, stat);
5103 tree_events(memcg, events);
5105 seq_printf(m, "anon %llu\n",
5106 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5107 seq_printf(m, "file %llu\n",
5108 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5109 seq_printf(m, "kernel_stack %llu\n",
5110 (u64)stat[MEMCG_KERNEL_STACK] * PAGE_SIZE);
5111 seq_printf(m, "slab %llu\n",
5112 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5113 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5114 seq_printf(m, "sock %llu\n",
5115 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5117 seq_printf(m, "file_mapped %llu\n",
5118 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5119 seq_printf(m, "file_dirty %llu\n",
5120 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5121 seq_printf(m, "file_writeback %llu\n",
5122 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5124 for (i = 0; i < NR_LRU_LISTS; i++) {
5125 struct mem_cgroup *mi;
5126 unsigned long val = 0;
5128 for_each_mem_cgroup_tree(mi, memcg)
5129 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5130 seq_printf(m, "%s %llu\n",
5131 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5134 seq_printf(m, "slab_reclaimable %llu\n",
5135 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5136 seq_printf(m, "slab_unreclaimable %llu\n",
5137 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5139 /* Accumulated memory events */
5141 seq_printf(m, "pgfault %lu\n",
5142 events[MEM_CGROUP_EVENTS_PGFAULT]);
5143 seq_printf(m, "pgmajfault %lu\n",
5144 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5149 static struct cftype memory_files[] = {
5152 .flags = CFTYPE_NOT_ON_ROOT,
5153 .read_u64 = memory_current_read,
5157 .flags = CFTYPE_NOT_ON_ROOT,
5158 .seq_show = memory_low_show,
5159 .write = memory_low_write,
5163 .flags = CFTYPE_NOT_ON_ROOT,
5164 .seq_show = memory_high_show,
5165 .write = memory_high_write,
5169 .flags = CFTYPE_NOT_ON_ROOT,
5170 .seq_show = memory_max_show,
5171 .write = memory_max_write,
5175 .flags = CFTYPE_NOT_ON_ROOT,
5176 .file_offset = offsetof(struct mem_cgroup, events_file),
5177 .seq_show = memory_events_show,
5181 .flags = CFTYPE_NOT_ON_ROOT,
5182 .seq_show = memory_stat_show,
5187 struct cgroup_subsys memory_cgrp_subsys = {
5188 .css_alloc = mem_cgroup_css_alloc,
5189 .css_online = mem_cgroup_css_online,
5190 .css_offline = mem_cgroup_css_offline,
5191 .css_released = mem_cgroup_css_released,
5192 .css_free = mem_cgroup_css_free,
5193 .css_reset = mem_cgroup_css_reset,
5194 .can_attach = mem_cgroup_can_attach,
5195 .cancel_attach = mem_cgroup_cancel_attach,
5196 .attach = mem_cgroup_move_task,
5197 .bind = mem_cgroup_bind,
5198 .dfl_cftypes = memory_files,
5199 .legacy_cftypes = mem_cgroup_legacy_files,
5204 * mem_cgroup_low - check if memory consumption is below the normal range
5205 * @root: the highest ancestor to consider
5206 * @memcg: the memory cgroup to check
5208 * Returns %true if memory consumption of @memcg, and that of all
5209 * configurable ancestors up to @root, is below the normal range.
5211 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5213 if (mem_cgroup_disabled())
5217 * The toplevel group doesn't have a configurable range, so
5218 * it's never low when looked at directly, and it is not
5219 * considered an ancestor when assessing the hierarchy.
5222 if (memcg == root_mem_cgroup)
5225 if (page_counter_read(&memcg->memory) >= memcg->low)
5228 while (memcg != root) {
5229 memcg = parent_mem_cgroup(memcg);
5231 if (memcg == root_mem_cgroup)
5234 if (page_counter_read(&memcg->memory) >= memcg->low)
5241 * mem_cgroup_try_charge - try charging a page
5242 * @page: page to charge
5243 * @mm: mm context of the victim
5244 * @gfp_mask: reclaim mode
5245 * @memcgp: charged memcg return
5247 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5248 * pages according to @gfp_mask if necessary.
5250 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5251 * Otherwise, an error code is returned.
5253 * After page->mapping has been set up, the caller must finalize the
5254 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5255 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5257 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5258 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5261 struct mem_cgroup *memcg = NULL;
5262 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5265 if (mem_cgroup_disabled())
5268 if (PageSwapCache(page)) {
5270 * Every swap fault against a single page tries to charge the
5271 * page, bail as early as possible. shmem_unuse() encounters
5272 * already charged pages, too. The USED bit is protected by
5273 * the page lock, which serializes swap cache removal, which
5274 * in turn serializes uncharging.
5276 VM_BUG_ON_PAGE(!PageLocked(page), page);
5277 if (page->mem_cgroup)
5280 if (do_swap_account) {
5281 swp_entry_t ent = { .val = page_private(page), };
5282 unsigned short id = lookup_swap_cgroup_id(ent);
5285 memcg = mem_cgroup_from_id(id);
5286 if (memcg && !css_tryget_online(&memcg->css))
5293 memcg = get_mem_cgroup_from_mm(mm);
5295 ret = try_charge(memcg, gfp_mask, nr_pages);
5297 css_put(&memcg->css);
5304 * mem_cgroup_commit_charge - commit a page charge
5305 * @page: page to charge
5306 * @memcg: memcg to charge the page to
5307 * @lrucare: page might be on LRU already
5309 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5310 * after page->mapping has been set up. This must happen atomically
5311 * as part of the page instantiation, i.e. under the page table lock
5312 * for anonymous pages, under the page lock for page and swap cache.
5314 * In addition, the page must not be on the LRU during the commit, to
5315 * prevent racing with task migration. If it might be, use @lrucare.
5317 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5319 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5320 bool lrucare, bool compound)
5322 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5324 VM_BUG_ON_PAGE(!page->mapping, page);
5325 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5327 if (mem_cgroup_disabled())
5330 * Swap faults will attempt to charge the same page multiple
5331 * times. But reuse_swap_page() might have removed the page
5332 * from swapcache already, so we can't check PageSwapCache().
5337 commit_charge(page, memcg, lrucare);
5339 local_irq_disable();
5340 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5341 memcg_check_events(memcg, page);
5344 if (do_memsw_account() && PageSwapCache(page)) {
5345 swp_entry_t entry = { .val = page_private(page) };
5347 * The swap entry might not get freed for a long time,
5348 * let's not wait for it. The page already received a
5349 * memory+swap charge, drop the swap entry duplicate.
5351 mem_cgroup_uncharge_swap(entry);
5356 * mem_cgroup_cancel_charge - cancel a page charge
5357 * @page: page to charge
5358 * @memcg: memcg to charge the page to
5360 * Cancel a charge transaction started by mem_cgroup_try_charge().
5362 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5365 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5367 if (mem_cgroup_disabled())
5370 * Swap faults will attempt to charge the same page multiple
5371 * times. But reuse_swap_page() might have removed the page
5372 * from swapcache already, so we can't check PageSwapCache().
5377 cancel_charge(memcg, nr_pages);
5380 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5381 unsigned long nr_anon, unsigned long nr_file,
5382 unsigned long nr_huge, struct page *dummy_page)
5384 unsigned long nr_pages = nr_anon + nr_file;
5385 unsigned long flags;
5387 if (!mem_cgroup_is_root(memcg)) {
5388 page_counter_uncharge(&memcg->memory, nr_pages);
5389 if (do_memsw_account())
5390 page_counter_uncharge(&memcg->memsw, nr_pages);
5391 memcg_oom_recover(memcg);
5394 local_irq_save(flags);
5395 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5396 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5397 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5398 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5399 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5400 memcg_check_events(memcg, dummy_page);
5401 local_irq_restore(flags);
5403 if (!mem_cgroup_is_root(memcg))
5404 css_put_many(&memcg->css, nr_pages);
5407 static void uncharge_list(struct list_head *page_list)
5409 struct mem_cgroup *memcg = NULL;
5410 unsigned long nr_anon = 0;
5411 unsigned long nr_file = 0;
5412 unsigned long nr_huge = 0;
5413 unsigned long pgpgout = 0;
5414 struct list_head *next;
5417 next = page_list->next;
5419 unsigned int nr_pages = 1;
5421 page = list_entry(next, struct page, lru);
5422 next = page->lru.next;
5424 VM_BUG_ON_PAGE(PageLRU(page), page);
5425 VM_BUG_ON_PAGE(page_count(page), page);
5427 if (!page->mem_cgroup)
5431 * Nobody should be changing or seriously looking at
5432 * page->mem_cgroup at this point, we have fully
5433 * exclusive access to the page.
5436 if (memcg != page->mem_cgroup) {
5438 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5440 pgpgout = nr_anon = nr_file = nr_huge = 0;
5442 memcg = page->mem_cgroup;
5445 if (PageTransHuge(page)) {
5446 nr_pages <<= compound_order(page);
5447 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5448 nr_huge += nr_pages;
5452 nr_anon += nr_pages;
5454 nr_file += nr_pages;
5456 page->mem_cgroup = NULL;
5459 } while (next != page_list);
5462 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5467 * mem_cgroup_uncharge - uncharge a page
5468 * @page: page to uncharge
5470 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5471 * mem_cgroup_commit_charge().
5473 void mem_cgroup_uncharge(struct page *page)
5475 if (mem_cgroup_disabled())
5478 /* Don't touch page->lru of any random page, pre-check: */
5479 if (!page->mem_cgroup)
5482 INIT_LIST_HEAD(&page->lru);
5483 uncharge_list(&page->lru);
5487 * mem_cgroup_uncharge_list - uncharge a list of page
5488 * @page_list: list of pages to uncharge
5490 * Uncharge a list of pages previously charged with
5491 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5493 void mem_cgroup_uncharge_list(struct list_head *page_list)
5495 if (mem_cgroup_disabled())
5498 if (!list_empty(page_list))
5499 uncharge_list(page_list);
5503 * mem_cgroup_migrate - charge a page's replacement
5504 * @oldpage: currently circulating page
5505 * @newpage: replacement page
5507 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5508 * be uncharged upon free.
5510 * Both pages must be locked, @newpage->mapping must be set up.
5512 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5514 struct mem_cgroup *memcg;
5515 unsigned int nr_pages;
5518 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5519 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5520 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5521 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5524 if (mem_cgroup_disabled())
5527 /* Page cache replacement: new page already charged? */
5528 if (newpage->mem_cgroup)
5531 /* Swapcache readahead pages can get replaced before being charged */
5532 memcg = oldpage->mem_cgroup;
5536 /* Force-charge the new page. The old one will be freed soon */
5537 compound = PageTransHuge(newpage);
5538 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5540 page_counter_charge(&memcg->memory, nr_pages);
5541 if (do_memsw_account())
5542 page_counter_charge(&memcg->memsw, nr_pages);
5543 css_get_many(&memcg->css, nr_pages);
5545 commit_charge(newpage, memcg, false);
5547 local_irq_disable();
5548 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5549 memcg_check_events(memcg, newpage);
5553 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5554 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5556 void sock_update_memcg(struct sock *sk)
5558 struct mem_cgroup *memcg;
5560 /* Socket cloning can throw us here with sk_cgrp already
5561 * filled. It won't however, necessarily happen from
5562 * process context. So the test for root memcg given
5563 * the current task's memcg won't help us in this case.
5565 * Respecting the original socket's memcg is a better
5566 * decision in this case.
5569 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5570 css_get(&sk->sk_memcg->css);
5575 memcg = mem_cgroup_from_task(current);
5576 if (memcg == root_mem_cgroup)
5578 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5580 if (css_tryget_online(&memcg->css))
5581 sk->sk_memcg = memcg;
5585 EXPORT_SYMBOL(sock_update_memcg);
5587 void sock_release_memcg(struct sock *sk)
5589 WARN_ON(!sk->sk_memcg);
5590 css_put(&sk->sk_memcg->css);
5594 * mem_cgroup_charge_skmem - charge socket memory
5595 * @memcg: memcg to charge
5596 * @nr_pages: number of pages to charge
5598 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5599 * @memcg's configured limit, %false if the charge had to be forced.
5601 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5603 gfp_t gfp_mask = GFP_KERNEL;
5605 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5606 struct page_counter *fail;
5608 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5609 memcg->tcpmem_pressure = 0;
5612 page_counter_charge(&memcg->tcpmem, nr_pages);
5613 memcg->tcpmem_pressure = 1;
5617 /* Don't block in the packet receive path */
5619 gfp_mask = GFP_NOWAIT;
5621 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5623 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5626 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5631 * mem_cgroup_uncharge_skmem - uncharge socket memory
5632 * @memcg - memcg to uncharge
5633 * @nr_pages - number of pages to uncharge
5635 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5637 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5638 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5642 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5644 page_counter_uncharge(&memcg->memory, nr_pages);
5645 css_put_many(&memcg->css, nr_pages);
5648 static int __init cgroup_memory(char *s)
5652 while ((token = strsep(&s, ",")) != NULL) {
5655 if (!strcmp(token, "nosocket"))
5656 cgroup_memory_nosocket = true;
5657 if (!strcmp(token, "nokmem"))
5658 cgroup_memory_nokmem = true;
5662 __setup("cgroup.memory=", cgroup_memory);
5665 * subsys_initcall() for memory controller.
5667 * Some parts like hotcpu_notifier() have to be initialized from this context
5668 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5669 * everything that doesn't depend on a specific mem_cgroup structure should
5670 * be initialized from here.
5672 static int __init mem_cgroup_init(void)
5676 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5678 for_each_possible_cpu(cpu)
5679 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5682 for_each_node(node) {
5683 struct mem_cgroup_tree_per_node *rtpn;
5686 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5687 node_online(node) ? node : NUMA_NO_NODE);
5689 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5690 struct mem_cgroup_tree_per_zone *rtpz;
5692 rtpz = &rtpn->rb_tree_per_zone[zone];
5693 rtpz->rb_root = RB_ROOT;
5694 spin_lock_init(&rtpz->lock);
5696 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5701 subsys_initcall(mem_cgroup_init);
5703 #ifdef CONFIG_MEMCG_SWAP
5705 * mem_cgroup_swapout - transfer a memsw charge to swap
5706 * @page: page whose memsw charge to transfer
5707 * @entry: swap entry to move the charge to
5709 * Transfer the memsw charge of @page to @entry.
5711 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5713 struct mem_cgroup *memcg;
5714 unsigned short oldid;
5716 VM_BUG_ON_PAGE(PageLRU(page), page);
5717 VM_BUG_ON_PAGE(page_count(page), page);
5719 if (!do_memsw_account())
5722 memcg = page->mem_cgroup;
5724 /* Readahead page, never charged */
5728 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5729 VM_BUG_ON_PAGE(oldid, page);
5730 mem_cgroup_swap_statistics(memcg, true);
5732 page->mem_cgroup = NULL;
5734 if (!mem_cgroup_is_root(memcg))
5735 page_counter_uncharge(&memcg->memory, 1);
5738 * Interrupts should be disabled here because the caller holds the
5739 * mapping->tree_lock lock which is taken with interrupts-off. It is
5740 * important here to have the interrupts disabled because it is the
5741 * only synchronisation we have for udpating the per-CPU variables.
5743 VM_BUG_ON(!irqs_disabled());
5744 mem_cgroup_charge_statistics(memcg, page, false, -1);
5745 memcg_check_events(memcg, page);
5749 * mem_cgroup_try_charge_swap - try charging a swap entry
5750 * @page: page being added to swap
5751 * @entry: swap entry to charge
5753 * Try to charge @entry to the memcg that @page belongs to.
5755 * Returns 0 on success, -ENOMEM on failure.
5757 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5759 struct mem_cgroup *memcg;
5760 struct page_counter *counter;
5761 unsigned short oldid;
5763 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5766 memcg = page->mem_cgroup;
5768 /* Readahead page, never charged */
5772 if (!mem_cgroup_is_root(memcg) &&
5773 !page_counter_try_charge(&memcg->swap, 1, &counter))
5776 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5777 VM_BUG_ON_PAGE(oldid, page);
5778 mem_cgroup_swap_statistics(memcg, true);
5780 css_get(&memcg->css);
5785 * mem_cgroup_uncharge_swap - uncharge a swap entry
5786 * @entry: swap entry to uncharge
5788 * Drop the swap charge associated with @entry.
5790 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5792 struct mem_cgroup *memcg;
5795 if (!do_swap_account)
5798 id = swap_cgroup_record(entry, 0);
5800 memcg = mem_cgroup_from_id(id);
5802 if (!mem_cgroup_is_root(memcg)) {
5803 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5804 page_counter_uncharge(&memcg->swap, 1);
5806 page_counter_uncharge(&memcg->memsw, 1);
5808 mem_cgroup_swap_statistics(memcg, false);
5809 css_put(&memcg->css);
5814 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
5816 long nr_swap_pages = get_nr_swap_pages();
5818 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5819 return nr_swap_pages;
5820 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5821 nr_swap_pages = min_t(long, nr_swap_pages,
5822 READ_ONCE(memcg->swap.limit) -
5823 page_counter_read(&memcg->swap));
5824 return nr_swap_pages;
5827 bool mem_cgroup_swap_full(struct page *page)
5829 struct mem_cgroup *memcg;
5831 VM_BUG_ON_PAGE(!PageLocked(page), page);
5835 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5838 memcg = page->mem_cgroup;
5842 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5843 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
5849 /* for remember boot option*/
5850 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5851 static int really_do_swap_account __initdata = 1;
5853 static int really_do_swap_account __initdata;
5856 static int __init enable_swap_account(char *s)
5858 if (!strcmp(s, "1"))
5859 really_do_swap_account = 1;
5860 else if (!strcmp(s, "0"))
5861 really_do_swap_account = 0;
5864 __setup("swapaccount=", enable_swap_account);
5866 static u64 swap_current_read(struct cgroup_subsys_state *css,
5869 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5871 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
5874 static int swap_max_show(struct seq_file *m, void *v)
5876 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5877 unsigned long max = READ_ONCE(memcg->swap.limit);
5879 if (max == PAGE_COUNTER_MAX)
5880 seq_puts(m, "max\n");
5882 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5887 static ssize_t swap_max_write(struct kernfs_open_file *of,
5888 char *buf, size_t nbytes, loff_t off)
5890 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5894 buf = strstrip(buf);
5895 err = page_counter_memparse(buf, "max", &max);
5899 mutex_lock(&memcg_limit_mutex);
5900 err = page_counter_limit(&memcg->swap, max);
5901 mutex_unlock(&memcg_limit_mutex);
5908 static struct cftype swap_files[] = {
5910 .name = "swap.current",
5911 .flags = CFTYPE_NOT_ON_ROOT,
5912 .read_u64 = swap_current_read,
5916 .flags = CFTYPE_NOT_ON_ROOT,
5917 .seq_show = swap_max_show,
5918 .write = swap_max_write,
5923 static struct cftype memsw_cgroup_files[] = {
5925 .name = "memsw.usage_in_bytes",
5926 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5927 .read_u64 = mem_cgroup_read_u64,
5930 .name = "memsw.max_usage_in_bytes",
5931 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5932 .write = mem_cgroup_reset,
5933 .read_u64 = mem_cgroup_read_u64,
5936 .name = "memsw.limit_in_bytes",
5937 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5938 .write = mem_cgroup_write,
5939 .read_u64 = mem_cgroup_read_u64,
5942 .name = "memsw.failcnt",
5943 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5944 .write = mem_cgroup_reset,
5945 .read_u64 = mem_cgroup_read_u64,
5947 { }, /* terminate */
5950 static int __init mem_cgroup_swap_init(void)
5952 if (!mem_cgroup_disabled() && really_do_swap_account) {
5953 do_swap_account = 1;
5954 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
5956 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5957 memsw_cgroup_files));
5961 subsys_initcall(mem_cgroup_swap_init);
5963 #endif /* CONFIG_MEMCG_SWAP */