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/sched/mm.h>
39 #include <linux/shmem_fs.h>
40 #include <linux/hugetlb.h>
41 #include <linux/pagemap.h>
42 #include <linux/smp.h>
43 #include <linux/page-flags.h>
44 #include <linux/backing-dev.h>
45 #include <linux/bit_spinlock.h>
46 #include <linux/rcupdate.h>
47 #include <linux/limits.h>
48 #include <linux/export.h>
49 #include <linux/mutex.h>
50 #include <linux/rbtree.h>
51 #include <linux/slab.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/spinlock.h>
55 #include <linux/eventfd.h>
56 #include <linux/poll.h>
57 #include <linux/sort.h>
59 #include <linux/seq_file.h>
60 #include <linux/vmpressure.h>
61 #include <linux/mm_inline.h>
62 #include <linux/swap_cgroup.h>
63 #include <linux/cpu.h>
64 #include <linux/oom.h>
65 #include <linux/lockdep.h>
66 #include <linux/file.h>
67 #include <linux/tracehook.h>
73 #include <linux/uaccess.h>
75 #include <trace/events/vmscan.h>
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
78 EXPORT_SYMBOL(memory_cgrp_subsys);
80 struct mem_cgroup *root_mem_cgroup __read_mostly;
82 #define MEM_CGROUP_RECLAIM_RETRIES 5
84 /* Socket memory accounting disabled? */
85 static bool cgroup_memory_nosocket;
87 /* Kernel memory accounting disabled? */
88 static bool cgroup_memory_nokmem;
90 /* Whether the swap controller is active */
91 #ifdef CONFIG_MEMCG_SWAP
92 int do_swap_account __read_mostly;
94 #define do_swap_account 0
97 /* Whether legacy memory+swap accounting is active */
98 static bool do_memsw_account(void)
100 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
103 static const char * const mem_cgroup_stat_names[] = {
114 static const char * const mem_cgroup_events_names[] = {
121 static const char * const mem_cgroup_lru_names[] = {
129 #define THRESHOLDS_EVENTS_TARGET 128
130 #define SOFTLIMIT_EVENTS_TARGET 1024
131 #define NUMAINFO_EVENTS_TARGET 1024
134 * Cgroups above their limits are maintained in a RB-Tree, independent of
135 * their hierarchy representation
138 struct mem_cgroup_tree_per_node {
139 struct rb_root rb_root;
143 struct mem_cgroup_tree {
144 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
147 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
150 struct mem_cgroup_eventfd_list {
151 struct list_head list;
152 struct eventfd_ctx *eventfd;
156 * cgroup_event represents events which userspace want to receive.
158 struct mem_cgroup_event {
160 * memcg which the event belongs to.
162 struct mem_cgroup *memcg;
164 * eventfd to signal userspace about the event.
166 struct eventfd_ctx *eventfd;
168 * Each of these stored in a list by the cgroup.
170 struct list_head list;
172 * register_event() callback will be used to add new userspace
173 * waiter for changes related to this event. Use eventfd_signal()
174 * on eventfd to send notification to userspace.
176 int (*register_event)(struct mem_cgroup *memcg,
177 struct eventfd_ctx *eventfd, const char *args);
179 * unregister_event() callback will be called when userspace closes
180 * the eventfd or on cgroup removing. This callback must be set,
181 * if you want provide notification functionality.
183 void (*unregister_event)(struct mem_cgroup *memcg,
184 struct eventfd_ctx *eventfd);
186 * All fields below needed to unregister event when
187 * userspace closes eventfd.
190 wait_queue_head_t *wqh;
192 struct work_struct remove;
195 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
196 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
198 /* Stuffs for move charges at task migration. */
200 * Types of charges to be moved.
202 #define MOVE_ANON 0x1U
203 #define MOVE_FILE 0x2U
204 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
206 /* "mc" and its members are protected by cgroup_mutex */
207 static struct move_charge_struct {
208 spinlock_t lock; /* for from, to */
209 struct mm_struct *mm;
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 struct workqueue_struct *memcg_kmem_cache_wq;
325 #endif /* !CONFIG_SLOB */
328 * mem_cgroup_css_from_page - css of the memcg associated with a page
329 * @page: page of interest
331 * If memcg is bound to the default hierarchy, css of the memcg associated
332 * with @page is returned. The returned css remains associated with @page
333 * until it is released.
335 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
338 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
340 struct mem_cgroup *memcg;
342 memcg = page->mem_cgroup;
344 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
345 memcg = root_mem_cgroup;
351 * page_cgroup_ino - return inode number of the memcg a page is charged to
354 * Look up the closest online ancestor of the memory cgroup @page is charged to
355 * and return its inode number or 0 if @page is not charged to any cgroup. It
356 * is safe to call this function without holding a reference to @page.
358 * Note, this function is inherently racy, because there is nothing to prevent
359 * the cgroup inode from getting torn down and potentially reallocated a moment
360 * after page_cgroup_ino() returns, so it only should be used by callers that
361 * do not care (such as procfs interfaces).
363 ino_t page_cgroup_ino(struct page *page)
365 struct mem_cgroup *memcg;
366 unsigned long ino = 0;
369 memcg = READ_ONCE(page->mem_cgroup);
370 while (memcg && !(memcg->css.flags & CSS_ONLINE))
371 memcg = parent_mem_cgroup(memcg);
373 ino = cgroup_ino(memcg->css.cgroup);
378 static struct mem_cgroup_per_node *
379 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
381 int nid = page_to_nid(page);
383 return memcg->nodeinfo[nid];
386 static struct mem_cgroup_tree_per_node *
387 soft_limit_tree_node(int nid)
389 return soft_limit_tree.rb_tree_per_node[nid];
392 static struct mem_cgroup_tree_per_node *
393 soft_limit_tree_from_page(struct page *page)
395 int nid = page_to_nid(page);
397 return soft_limit_tree.rb_tree_per_node[nid];
400 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
401 struct mem_cgroup_tree_per_node *mctz,
402 unsigned long new_usage_in_excess)
404 struct rb_node **p = &mctz->rb_root.rb_node;
405 struct rb_node *parent = NULL;
406 struct mem_cgroup_per_node *mz_node;
411 mz->usage_in_excess = new_usage_in_excess;
412 if (!mz->usage_in_excess)
416 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
418 if (mz->usage_in_excess < mz_node->usage_in_excess)
421 * We can't avoid mem cgroups that are over their soft
422 * limit by the same amount
424 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
427 rb_link_node(&mz->tree_node, parent, p);
428 rb_insert_color(&mz->tree_node, &mctz->rb_root);
432 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
433 struct mem_cgroup_tree_per_node *mctz)
437 rb_erase(&mz->tree_node, &mctz->rb_root);
441 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
442 struct mem_cgroup_tree_per_node *mctz)
446 spin_lock_irqsave(&mctz->lock, flags);
447 __mem_cgroup_remove_exceeded(mz, mctz);
448 spin_unlock_irqrestore(&mctz->lock, flags);
451 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
453 unsigned long nr_pages = page_counter_read(&memcg->memory);
454 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
455 unsigned long excess = 0;
457 if (nr_pages > soft_limit)
458 excess = nr_pages - soft_limit;
463 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
465 unsigned long excess;
466 struct mem_cgroup_per_node *mz;
467 struct mem_cgroup_tree_per_node *mctz;
469 mctz = soft_limit_tree_from_page(page);
473 * Necessary to update all ancestors when hierarchy is used.
474 * because their event counter is not touched.
476 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
477 mz = mem_cgroup_page_nodeinfo(memcg, page);
478 excess = soft_limit_excess(memcg);
480 * We have to update the tree if mz is on RB-tree or
481 * mem is over its softlimit.
483 if (excess || mz->on_tree) {
486 spin_lock_irqsave(&mctz->lock, flags);
487 /* if on-tree, remove it */
489 __mem_cgroup_remove_exceeded(mz, mctz);
491 * Insert again. mz->usage_in_excess will be updated.
492 * If excess is 0, no tree ops.
494 __mem_cgroup_insert_exceeded(mz, mctz, excess);
495 spin_unlock_irqrestore(&mctz->lock, flags);
500 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
502 struct mem_cgroup_tree_per_node *mctz;
503 struct mem_cgroup_per_node *mz;
507 mz = mem_cgroup_nodeinfo(memcg, nid);
508 mctz = soft_limit_tree_node(nid);
510 mem_cgroup_remove_exceeded(mz, mctz);
514 static struct mem_cgroup_per_node *
515 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
517 struct rb_node *rightmost = NULL;
518 struct mem_cgroup_per_node *mz;
522 rightmost = rb_last(&mctz->rb_root);
524 goto done; /* Nothing to reclaim from */
526 mz = rb_entry(rightmost, struct mem_cgroup_per_node, tree_node);
528 * Remove the node now but someone else can add it back,
529 * we will to add it back at the end of reclaim to its correct
530 * position in the tree.
532 __mem_cgroup_remove_exceeded(mz, mctz);
533 if (!soft_limit_excess(mz->memcg) ||
534 !css_tryget_online(&mz->memcg->css))
540 static struct mem_cgroup_per_node *
541 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
543 struct mem_cgroup_per_node *mz;
545 spin_lock_irq(&mctz->lock);
546 mz = __mem_cgroup_largest_soft_limit_node(mctz);
547 spin_unlock_irq(&mctz->lock);
552 * Return page count for single (non recursive) @memcg.
554 * Implementation Note: reading percpu statistics for memcg.
556 * Both of vmstat[] and percpu_counter has threshold and do periodic
557 * synchronization to implement "quick" read. There are trade-off between
558 * reading cost and precision of value. Then, we may have a chance to implement
559 * a periodic synchronization of counter in memcg's counter.
561 * But this _read() function is used for user interface now. The user accounts
562 * memory usage by memory cgroup and he _always_ requires exact value because
563 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
564 * have to visit all online cpus and make sum. So, for now, unnecessary
565 * synchronization is not implemented. (just implemented for cpu hotplug)
567 * If there are kernel internal actions which can make use of some not-exact
568 * value, and reading all cpu value can be performance bottleneck in some
569 * common workload, threshold and synchronization as vmstat[] should be
573 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
578 /* Per-cpu values can be negative, use a signed accumulator */
579 for_each_possible_cpu(cpu)
580 val += per_cpu(memcg->stat->count[idx], cpu);
582 * Summing races with updates, so val may be negative. Avoid exposing
583 * transient negative values.
590 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
591 enum mem_cgroup_events_index idx)
593 unsigned long val = 0;
596 for_each_possible_cpu(cpu)
597 val += per_cpu(memcg->stat->events[idx], cpu);
601 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
603 bool compound, int nr_pages)
606 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
607 * counted as CACHE even if it's on ANON LRU.
610 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
613 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
615 if (PageSwapBacked(page))
616 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SHMEM],
621 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
622 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
626 /* pagein of a big page is an event. So, ignore page size */
628 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
630 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
631 nr_pages = -nr_pages; /* for event */
634 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
637 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
638 int nid, unsigned int lru_mask)
640 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
641 unsigned long nr = 0;
644 VM_BUG_ON((unsigned)nid >= nr_node_ids);
647 if (!(BIT(lru) & lru_mask))
649 nr += mem_cgroup_get_lru_size(lruvec, lru);
654 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
655 unsigned int lru_mask)
657 unsigned long nr = 0;
660 for_each_node_state(nid, N_MEMORY)
661 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
665 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
666 enum mem_cgroup_events_target target)
668 unsigned long val, next;
670 val = __this_cpu_read(memcg->stat->nr_page_events);
671 next = __this_cpu_read(memcg->stat->targets[target]);
672 /* from time_after() in jiffies.h */
673 if ((long)next - (long)val < 0) {
675 case MEM_CGROUP_TARGET_THRESH:
676 next = val + THRESHOLDS_EVENTS_TARGET;
678 case MEM_CGROUP_TARGET_SOFTLIMIT:
679 next = val + SOFTLIMIT_EVENTS_TARGET;
681 case MEM_CGROUP_TARGET_NUMAINFO:
682 next = val + NUMAINFO_EVENTS_TARGET;
687 __this_cpu_write(memcg->stat->targets[target], next);
694 * Check events in order.
697 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
699 /* threshold event is triggered in finer grain than soft limit */
700 if (unlikely(mem_cgroup_event_ratelimit(memcg,
701 MEM_CGROUP_TARGET_THRESH))) {
703 bool do_numainfo __maybe_unused;
705 do_softlimit = mem_cgroup_event_ratelimit(memcg,
706 MEM_CGROUP_TARGET_SOFTLIMIT);
708 do_numainfo = mem_cgroup_event_ratelimit(memcg,
709 MEM_CGROUP_TARGET_NUMAINFO);
711 mem_cgroup_threshold(memcg);
712 if (unlikely(do_softlimit))
713 mem_cgroup_update_tree(memcg, page);
715 if (unlikely(do_numainfo))
716 atomic_inc(&memcg->numainfo_events);
721 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
724 * mm_update_next_owner() may clear mm->owner to NULL
725 * if it races with swapoff, page migration, etc.
726 * So this can be called with p == NULL.
731 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
733 EXPORT_SYMBOL(mem_cgroup_from_task);
735 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
737 struct mem_cgroup *memcg = NULL;
742 * Page cache insertions can happen withou an
743 * actual mm context, e.g. during disk probing
744 * on boot, loopback IO, acct() writes etc.
747 memcg = root_mem_cgroup;
749 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
750 if (unlikely(!memcg))
751 memcg = root_mem_cgroup;
753 } while (!css_tryget_online(&memcg->css));
759 * mem_cgroup_iter - iterate over memory cgroup hierarchy
760 * @root: hierarchy root
761 * @prev: previously returned memcg, NULL on first invocation
762 * @reclaim: cookie for shared reclaim walks, NULL for full walks
764 * Returns references to children of the hierarchy below @root, or
765 * @root itself, or %NULL after a full round-trip.
767 * Caller must pass the return value in @prev on subsequent
768 * invocations for reference counting, or use mem_cgroup_iter_break()
769 * to cancel a hierarchy walk before the round-trip is complete.
771 * Reclaimers can specify a zone and a priority level in @reclaim to
772 * divide up the memcgs in the hierarchy among all concurrent
773 * reclaimers operating on the same zone and priority.
775 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
776 struct mem_cgroup *prev,
777 struct mem_cgroup_reclaim_cookie *reclaim)
779 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
780 struct cgroup_subsys_state *css = NULL;
781 struct mem_cgroup *memcg = NULL;
782 struct mem_cgroup *pos = NULL;
784 if (mem_cgroup_disabled())
788 root = root_mem_cgroup;
790 if (prev && !reclaim)
793 if (!root->use_hierarchy && root != root_mem_cgroup) {
802 struct mem_cgroup_per_node *mz;
804 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
805 iter = &mz->iter[reclaim->priority];
807 if (prev && reclaim->generation != iter->generation)
811 pos = READ_ONCE(iter->position);
812 if (!pos || css_tryget(&pos->css))
815 * css reference reached zero, so iter->position will
816 * be cleared by ->css_released. However, we should not
817 * rely on this happening soon, because ->css_released
818 * is called from a work queue, and by busy-waiting we
819 * might block it. So we clear iter->position right
822 (void)cmpxchg(&iter->position, pos, NULL);
830 css = css_next_descendant_pre(css, &root->css);
833 * Reclaimers share the hierarchy walk, and a
834 * new one might jump in right at the end of
835 * the hierarchy - make sure they see at least
836 * one group and restart from the beginning.
844 * Verify the css and acquire a reference. The root
845 * is provided by the caller, so we know it's alive
846 * and kicking, and don't take an extra reference.
848 memcg = mem_cgroup_from_css(css);
850 if (css == &root->css)
861 * The position could have already been updated by a competing
862 * thread, so check that the value hasn't changed since we read
863 * it to avoid reclaiming from the same cgroup twice.
865 (void)cmpxchg(&iter->position, pos, memcg);
873 reclaim->generation = iter->generation;
879 if (prev && prev != root)
886 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
887 * @root: hierarchy root
888 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
890 void mem_cgroup_iter_break(struct mem_cgroup *root,
891 struct mem_cgroup *prev)
894 root = root_mem_cgroup;
895 if (prev && prev != root)
899 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
901 struct mem_cgroup *memcg = dead_memcg;
902 struct mem_cgroup_reclaim_iter *iter;
903 struct mem_cgroup_per_node *mz;
907 while ((memcg = parent_mem_cgroup(memcg))) {
909 mz = mem_cgroup_nodeinfo(memcg, nid);
910 for (i = 0; i <= DEF_PRIORITY; i++) {
912 cmpxchg(&iter->position,
920 * Iteration constructs for visiting all cgroups (under a tree). If
921 * loops are exited prematurely (break), mem_cgroup_iter_break() must
922 * be used for reference counting.
924 #define for_each_mem_cgroup_tree(iter, root) \
925 for (iter = mem_cgroup_iter(root, NULL, NULL); \
927 iter = mem_cgroup_iter(root, iter, NULL))
929 #define for_each_mem_cgroup(iter) \
930 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
932 iter = mem_cgroup_iter(NULL, iter, NULL))
935 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
936 * @memcg: hierarchy root
937 * @fn: function to call for each task
938 * @arg: argument passed to @fn
940 * This function iterates over tasks attached to @memcg or to any of its
941 * descendants and calls @fn for each task. If @fn returns a non-zero
942 * value, the function breaks the iteration loop and returns the value.
943 * Otherwise, it will iterate over all tasks and return 0.
945 * This function must not be called for the root memory cgroup.
947 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
948 int (*fn)(struct task_struct *, void *), void *arg)
950 struct mem_cgroup *iter;
953 BUG_ON(memcg == root_mem_cgroup);
955 for_each_mem_cgroup_tree(iter, memcg) {
956 struct css_task_iter it;
957 struct task_struct *task;
959 css_task_iter_start(&iter->css, &it);
960 while (!ret && (task = css_task_iter_next(&it)))
962 css_task_iter_end(&it);
964 mem_cgroup_iter_break(memcg, iter);
972 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
974 * @zone: zone of the page
976 * This function is only safe when following the LRU page isolation
977 * and putback protocol: the LRU lock must be held, and the page must
978 * either be PageLRU() or the caller must have isolated/allocated it.
980 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
982 struct mem_cgroup_per_node *mz;
983 struct mem_cgroup *memcg;
984 struct lruvec *lruvec;
986 if (mem_cgroup_disabled()) {
987 lruvec = &pgdat->lruvec;
991 memcg = page->mem_cgroup;
993 * Swapcache readahead pages are added to the LRU - and
994 * possibly migrated - before they are charged.
997 memcg = root_mem_cgroup;
999 mz = mem_cgroup_page_nodeinfo(memcg, page);
1000 lruvec = &mz->lruvec;
1003 * Since a node can be onlined after the mem_cgroup was created,
1004 * we have to be prepared to initialize lruvec->zone here;
1005 * and if offlined then reonlined, we need to reinitialize it.
1007 if (unlikely(lruvec->pgdat != pgdat))
1008 lruvec->pgdat = pgdat;
1013 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1014 * @lruvec: mem_cgroup per zone lru vector
1015 * @lru: index of lru list the page is sitting on
1016 * @zid: zone id of the accounted pages
1017 * @nr_pages: positive when adding or negative when removing
1019 * This function must be called under lru_lock, just before a page is added
1020 * to or just after a page is removed from an lru list (that ordering being
1021 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1023 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1024 int zid, int nr_pages)
1026 struct mem_cgroup_per_node *mz;
1027 unsigned long *lru_size;
1030 if (mem_cgroup_disabled())
1033 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1034 lru_size = &mz->lru_zone_size[zid][lru];
1037 *lru_size += nr_pages;
1040 if (WARN_ONCE(size < 0,
1041 "%s(%p, %d, %d): lru_size %ld\n",
1042 __func__, lruvec, lru, nr_pages, size)) {
1048 *lru_size += nr_pages;
1051 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1053 struct mem_cgroup *task_memcg;
1054 struct task_struct *p;
1057 p = find_lock_task_mm(task);
1059 task_memcg = get_mem_cgroup_from_mm(p->mm);
1063 * All threads may have already detached their mm's, but the oom
1064 * killer still needs to detect if they have already been oom
1065 * killed to prevent needlessly killing additional tasks.
1068 task_memcg = mem_cgroup_from_task(task);
1069 css_get(&task_memcg->css);
1072 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1073 css_put(&task_memcg->css);
1078 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1079 * @memcg: the memory cgroup
1081 * Returns the maximum amount of memory @mem can be charged with, in
1084 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1086 unsigned long margin = 0;
1087 unsigned long count;
1088 unsigned long limit;
1090 count = page_counter_read(&memcg->memory);
1091 limit = READ_ONCE(memcg->memory.limit);
1093 margin = limit - count;
1095 if (do_memsw_account()) {
1096 count = page_counter_read(&memcg->memsw);
1097 limit = READ_ONCE(memcg->memsw.limit);
1099 margin = min(margin, limit - count);
1108 * A routine for checking "mem" is under move_account() or not.
1110 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1111 * moving cgroups. This is for waiting at high-memory pressure
1114 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1116 struct mem_cgroup *from;
1117 struct mem_cgroup *to;
1120 * Unlike task_move routines, we access mc.to, mc.from not under
1121 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1123 spin_lock(&mc.lock);
1129 ret = mem_cgroup_is_descendant(from, memcg) ||
1130 mem_cgroup_is_descendant(to, memcg);
1132 spin_unlock(&mc.lock);
1136 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1138 if (mc.moving_task && current != mc.moving_task) {
1139 if (mem_cgroup_under_move(memcg)) {
1141 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1142 /* moving charge context might have finished. */
1145 finish_wait(&mc.waitq, &wait);
1152 #define K(x) ((x) << (PAGE_SHIFT-10))
1154 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1155 * @memcg: The memory cgroup that went over limit
1156 * @p: Task that is going to be killed
1158 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1161 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1163 struct mem_cgroup *iter;
1169 pr_info("Task in ");
1170 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1171 pr_cont(" killed as a result of limit of ");
1173 pr_info("Memory limit reached of cgroup ");
1176 pr_cont_cgroup_path(memcg->css.cgroup);
1181 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1182 K((u64)page_counter_read(&memcg->memory)),
1183 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1184 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1185 K((u64)page_counter_read(&memcg->memsw)),
1186 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1187 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1188 K((u64)page_counter_read(&memcg->kmem)),
1189 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1191 for_each_mem_cgroup_tree(iter, memcg) {
1192 pr_info("Memory cgroup stats for ");
1193 pr_cont_cgroup_path(iter->css.cgroup);
1196 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1197 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1199 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1200 K(mem_cgroup_read_stat(iter, i)));
1203 for (i = 0; i < NR_LRU_LISTS; i++)
1204 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1205 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1212 * This function returns the number of memcg under hierarchy tree. Returns
1213 * 1(self count) if no children.
1215 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1218 struct mem_cgroup *iter;
1220 for_each_mem_cgroup_tree(iter, memcg)
1226 * Return the memory (and swap, if configured) limit for a memcg.
1228 unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1230 unsigned long limit;
1232 limit = memcg->memory.limit;
1233 if (mem_cgroup_swappiness(memcg)) {
1234 unsigned long memsw_limit;
1235 unsigned long swap_limit;
1237 memsw_limit = memcg->memsw.limit;
1238 swap_limit = memcg->swap.limit;
1239 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1240 limit = min(limit + swap_limit, memsw_limit);
1245 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1248 struct oom_control oc = {
1252 .gfp_mask = gfp_mask,
1257 mutex_lock(&oom_lock);
1258 ret = out_of_memory(&oc);
1259 mutex_unlock(&oom_lock);
1263 #if MAX_NUMNODES > 1
1266 * test_mem_cgroup_node_reclaimable
1267 * @memcg: the target memcg
1268 * @nid: the node ID to be checked.
1269 * @noswap : specify true here if the user wants flle only information.
1271 * This function returns whether the specified memcg contains any
1272 * reclaimable pages on a node. Returns true if there are any reclaimable
1273 * pages in the node.
1275 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1276 int nid, bool noswap)
1278 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1280 if (noswap || !total_swap_pages)
1282 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1289 * Always updating the nodemask is not very good - even if we have an empty
1290 * list or the wrong list here, we can start from some node and traverse all
1291 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1294 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1298 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1299 * pagein/pageout changes since the last update.
1301 if (!atomic_read(&memcg->numainfo_events))
1303 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1306 /* make a nodemask where this memcg uses memory from */
1307 memcg->scan_nodes = node_states[N_MEMORY];
1309 for_each_node_mask(nid, node_states[N_MEMORY]) {
1311 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1312 node_clear(nid, memcg->scan_nodes);
1315 atomic_set(&memcg->numainfo_events, 0);
1316 atomic_set(&memcg->numainfo_updating, 0);
1320 * Selecting a node where we start reclaim from. Because what we need is just
1321 * reducing usage counter, start from anywhere is O,K. Considering
1322 * memory reclaim from current node, there are pros. and cons.
1324 * Freeing memory from current node means freeing memory from a node which
1325 * we'll use or we've used. So, it may make LRU bad. And if several threads
1326 * hit limits, it will see a contention on a node. But freeing from remote
1327 * node means more costs for memory reclaim because of memory latency.
1329 * Now, we use round-robin. Better algorithm is welcomed.
1331 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1335 mem_cgroup_may_update_nodemask(memcg);
1336 node = memcg->last_scanned_node;
1338 node = next_node_in(node, memcg->scan_nodes);
1340 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1341 * last time it really checked all the LRUs due to rate limiting.
1342 * Fallback to the current node in that case for simplicity.
1344 if (unlikely(node == MAX_NUMNODES))
1345 node = numa_node_id();
1347 memcg->last_scanned_node = node;
1351 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1357 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1360 unsigned long *total_scanned)
1362 struct mem_cgroup *victim = NULL;
1365 unsigned long excess;
1366 unsigned long nr_scanned;
1367 struct mem_cgroup_reclaim_cookie reclaim = {
1372 excess = soft_limit_excess(root_memcg);
1375 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1380 * If we have not been able to reclaim
1381 * anything, it might because there are
1382 * no reclaimable pages under this hierarchy
1387 * We want to do more targeted reclaim.
1388 * excess >> 2 is not to excessive so as to
1389 * reclaim too much, nor too less that we keep
1390 * coming back to reclaim from this cgroup
1392 if (total >= (excess >> 2) ||
1393 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1398 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1399 pgdat, &nr_scanned);
1400 *total_scanned += nr_scanned;
1401 if (!soft_limit_excess(root_memcg))
1404 mem_cgroup_iter_break(root_memcg, victim);
1408 #ifdef CONFIG_LOCKDEP
1409 static struct lockdep_map memcg_oom_lock_dep_map = {
1410 .name = "memcg_oom_lock",
1414 static DEFINE_SPINLOCK(memcg_oom_lock);
1417 * Check OOM-Killer is already running under our hierarchy.
1418 * If someone is running, return false.
1420 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1422 struct mem_cgroup *iter, *failed = NULL;
1424 spin_lock(&memcg_oom_lock);
1426 for_each_mem_cgroup_tree(iter, memcg) {
1427 if (iter->oom_lock) {
1429 * this subtree of our hierarchy is already locked
1430 * so we cannot give a lock.
1433 mem_cgroup_iter_break(memcg, iter);
1436 iter->oom_lock = true;
1441 * OK, we failed to lock the whole subtree so we have
1442 * to clean up what we set up to the failing subtree
1444 for_each_mem_cgroup_tree(iter, memcg) {
1445 if (iter == failed) {
1446 mem_cgroup_iter_break(memcg, iter);
1449 iter->oom_lock = false;
1452 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1454 spin_unlock(&memcg_oom_lock);
1459 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1461 struct mem_cgroup *iter;
1463 spin_lock(&memcg_oom_lock);
1464 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1465 for_each_mem_cgroup_tree(iter, memcg)
1466 iter->oom_lock = false;
1467 spin_unlock(&memcg_oom_lock);
1470 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1472 struct mem_cgroup *iter;
1474 spin_lock(&memcg_oom_lock);
1475 for_each_mem_cgroup_tree(iter, memcg)
1477 spin_unlock(&memcg_oom_lock);
1480 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1482 struct mem_cgroup *iter;
1485 * When a new child is created while the hierarchy is under oom,
1486 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1488 spin_lock(&memcg_oom_lock);
1489 for_each_mem_cgroup_tree(iter, memcg)
1490 if (iter->under_oom > 0)
1492 spin_unlock(&memcg_oom_lock);
1495 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1497 struct oom_wait_info {
1498 struct mem_cgroup *memcg;
1502 static int memcg_oom_wake_function(wait_queue_t *wait,
1503 unsigned mode, int sync, void *arg)
1505 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1506 struct mem_cgroup *oom_wait_memcg;
1507 struct oom_wait_info *oom_wait_info;
1509 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1510 oom_wait_memcg = oom_wait_info->memcg;
1512 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1513 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1515 return autoremove_wake_function(wait, mode, sync, arg);
1518 static void memcg_oom_recover(struct mem_cgroup *memcg)
1521 * For the following lockless ->under_oom test, the only required
1522 * guarantee is that it must see the state asserted by an OOM when
1523 * this function is called as a result of userland actions
1524 * triggered by the notification of the OOM. This is trivially
1525 * achieved by invoking mem_cgroup_mark_under_oom() before
1526 * triggering notification.
1528 if (memcg && memcg->under_oom)
1529 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1532 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1534 if (!current->memcg_may_oom)
1537 * We are in the middle of the charge context here, so we
1538 * don't want to block when potentially sitting on a callstack
1539 * that holds all kinds of filesystem and mm locks.
1541 * Also, the caller may handle a failed allocation gracefully
1542 * (like optional page cache readahead) and so an OOM killer
1543 * invocation might not even be necessary.
1545 * That's why we don't do anything here except remember the
1546 * OOM context and then deal with it at the end of the page
1547 * fault when the stack is unwound, the locks are released,
1548 * and when we know whether the fault was overall successful.
1550 css_get(&memcg->css);
1551 current->memcg_in_oom = memcg;
1552 current->memcg_oom_gfp_mask = mask;
1553 current->memcg_oom_order = order;
1557 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1558 * @handle: actually kill/wait or just clean up the OOM state
1560 * This has to be called at the end of a page fault if the memcg OOM
1561 * handler was enabled.
1563 * Memcg supports userspace OOM handling where failed allocations must
1564 * sleep on a waitqueue until the userspace task resolves the
1565 * situation. Sleeping directly in the charge context with all kinds
1566 * of locks held is not a good idea, instead we remember an OOM state
1567 * in the task and mem_cgroup_oom_synchronize() has to be called at
1568 * the end of the page fault to complete the OOM handling.
1570 * Returns %true if an ongoing memcg OOM situation was detected and
1571 * completed, %false otherwise.
1573 bool mem_cgroup_oom_synchronize(bool handle)
1575 struct mem_cgroup *memcg = current->memcg_in_oom;
1576 struct oom_wait_info owait;
1579 /* OOM is global, do not handle */
1586 owait.memcg = memcg;
1587 owait.wait.flags = 0;
1588 owait.wait.func = memcg_oom_wake_function;
1589 owait.wait.private = current;
1590 INIT_LIST_HEAD(&owait.wait.task_list);
1592 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1593 mem_cgroup_mark_under_oom(memcg);
1595 locked = mem_cgroup_oom_trylock(memcg);
1598 mem_cgroup_oom_notify(memcg);
1600 if (locked && !memcg->oom_kill_disable) {
1601 mem_cgroup_unmark_under_oom(memcg);
1602 finish_wait(&memcg_oom_waitq, &owait.wait);
1603 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1604 current->memcg_oom_order);
1607 mem_cgroup_unmark_under_oom(memcg);
1608 finish_wait(&memcg_oom_waitq, &owait.wait);
1612 mem_cgroup_oom_unlock(memcg);
1614 * There is no guarantee that an OOM-lock contender
1615 * sees the wakeups triggered by the OOM kill
1616 * uncharges. Wake any sleepers explicitely.
1618 memcg_oom_recover(memcg);
1621 current->memcg_in_oom = NULL;
1622 css_put(&memcg->css);
1627 * lock_page_memcg - lock a page->mem_cgroup binding
1630 * This function protects unlocked LRU pages from being moved to
1631 * another cgroup and stabilizes their page->mem_cgroup binding.
1633 void lock_page_memcg(struct page *page)
1635 struct mem_cgroup *memcg;
1636 unsigned long flags;
1639 * The RCU lock is held throughout the transaction. The fast
1640 * path can get away without acquiring the memcg->move_lock
1641 * because page moving starts with an RCU grace period.
1645 if (mem_cgroup_disabled())
1648 memcg = page->mem_cgroup;
1649 if (unlikely(!memcg))
1652 if (atomic_read(&memcg->moving_account) <= 0)
1655 spin_lock_irqsave(&memcg->move_lock, flags);
1656 if (memcg != page->mem_cgroup) {
1657 spin_unlock_irqrestore(&memcg->move_lock, flags);
1662 * When charge migration first begins, we can have locked and
1663 * unlocked page stat updates happening concurrently. Track
1664 * the task who has the lock for unlock_page_memcg().
1666 memcg->move_lock_task = current;
1667 memcg->move_lock_flags = flags;
1671 EXPORT_SYMBOL(lock_page_memcg);
1674 * unlock_page_memcg - unlock a page->mem_cgroup binding
1677 void unlock_page_memcg(struct page *page)
1679 struct mem_cgroup *memcg = page->mem_cgroup;
1681 if (memcg && memcg->move_lock_task == current) {
1682 unsigned long flags = memcg->move_lock_flags;
1684 memcg->move_lock_task = NULL;
1685 memcg->move_lock_flags = 0;
1687 spin_unlock_irqrestore(&memcg->move_lock, flags);
1692 EXPORT_SYMBOL(unlock_page_memcg);
1695 * size of first charge trial. "32" comes from vmscan.c's magic value.
1696 * TODO: maybe necessary to use big numbers in big irons.
1698 #define CHARGE_BATCH 32U
1699 struct memcg_stock_pcp {
1700 struct mem_cgroup *cached; /* this never be root cgroup */
1701 unsigned int nr_pages;
1702 struct work_struct work;
1703 unsigned long flags;
1704 #define FLUSHING_CACHED_CHARGE 0
1706 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1707 static DEFINE_MUTEX(percpu_charge_mutex);
1710 * consume_stock: Try to consume stocked charge on this cpu.
1711 * @memcg: memcg to consume from.
1712 * @nr_pages: how many pages to charge.
1714 * The charges will only happen if @memcg matches the current cpu's memcg
1715 * stock, and at least @nr_pages are available in that stock. Failure to
1716 * service an allocation will refill the stock.
1718 * returns true if successful, false otherwise.
1720 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1722 struct memcg_stock_pcp *stock;
1723 unsigned long flags;
1726 if (nr_pages > CHARGE_BATCH)
1729 local_irq_save(flags);
1731 stock = this_cpu_ptr(&memcg_stock);
1732 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1733 stock->nr_pages -= nr_pages;
1737 local_irq_restore(flags);
1743 * Returns stocks cached in percpu and reset cached information.
1745 static void drain_stock(struct memcg_stock_pcp *stock)
1747 struct mem_cgroup *old = stock->cached;
1749 if (stock->nr_pages) {
1750 page_counter_uncharge(&old->memory, stock->nr_pages);
1751 if (do_memsw_account())
1752 page_counter_uncharge(&old->memsw, stock->nr_pages);
1753 css_put_many(&old->css, stock->nr_pages);
1754 stock->nr_pages = 0;
1756 stock->cached = NULL;
1759 static void drain_local_stock(struct work_struct *dummy)
1761 struct memcg_stock_pcp *stock;
1762 unsigned long flags;
1764 local_irq_save(flags);
1766 stock = this_cpu_ptr(&memcg_stock);
1768 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1770 local_irq_restore(flags);
1774 * Cache charges(val) to local per_cpu area.
1775 * This will be consumed by consume_stock() function, later.
1777 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1779 struct memcg_stock_pcp *stock;
1780 unsigned long flags;
1782 local_irq_save(flags);
1784 stock = this_cpu_ptr(&memcg_stock);
1785 if (stock->cached != memcg) { /* reset if necessary */
1787 stock->cached = memcg;
1789 stock->nr_pages += nr_pages;
1791 local_irq_restore(flags);
1795 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1796 * of the hierarchy under it.
1798 static void drain_all_stock(struct mem_cgroup *root_memcg)
1802 /* If someone's already draining, avoid adding running more workers. */
1803 if (!mutex_trylock(&percpu_charge_mutex))
1805 /* Notify other cpus that system-wide "drain" is running */
1808 for_each_online_cpu(cpu) {
1809 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1810 struct mem_cgroup *memcg;
1812 memcg = stock->cached;
1813 if (!memcg || !stock->nr_pages)
1815 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1817 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1819 drain_local_stock(&stock->work);
1821 schedule_work_on(cpu, &stock->work);
1826 mutex_unlock(&percpu_charge_mutex);
1829 static int memcg_hotplug_cpu_dead(unsigned int cpu)
1831 struct memcg_stock_pcp *stock;
1833 stock = &per_cpu(memcg_stock, cpu);
1838 static void reclaim_high(struct mem_cgroup *memcg,
1839 unsigned int nr_pages,
1843 if (page_counter_read(&memcg->memory) <= memcg->high)
1845 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1846 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1847 } while ((memcg = parent_mem_cgroup(memcg)));
1850 static void high_work_func(struct work_struct *work)
1852 struct mem_cgroup *memcg;
1854 memcg = container_of(work, struct mem_cgroup, high_work);
1855 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1859 * Scheduled by try_charge() to be executed from the userland return path
1860 * and reclaims memory over the high limit.
1862 void mem_cgroup_handle_over_high(void)
1864 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1865 struct mem_cgroup *memcg;
1867 if (likely(!nr_pages))
1870 memcg = get_mem_cgroup_from_mm(current->mm);
1871 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1872 css_put(&memcg->css);
1873 current->memcg_nr_pages_over_high = 0;
1876 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1877 unsigned int nr_pages)
1879 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1880 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1881 struct mem_cgroup *mem_over_limit;
1882 struct page_counter *counter;
1883 unsigned long nr_reclaimed;
1884 bool may_swap = true;
1885 bool drained = false;
1887 if (mem_cgroup_is_root(memcg))
1890 if (consume_stock(memcg, nr_pages))
1893 if (!do_memsw_account() ||
1894 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1895 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1897 if (do_memsw_account())
1898 page_counter_uncharge(&memcg->memsw, batch);
1899 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1901 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1905 if (batch > nr_pages) {
1911 * Unlike in global OOM situations, memcg is not in a physical
1912 * memory shortage. Allow dying and OOM-killed tasks to
1913 * bypass the last charges so that they can exit quickly and
1914 * free their memory.
1916 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1917 fatal_signal_pending(current) ||
1918 current->flags & PF_EXITING))
1922 * Prevent unbounded recursion when reclaim operations need to
1923 * allocate memory. This might exceed the limits temporarily,
1924 * but we prefer facilitating memory reclaim and getting back
1925 * under the limit over triggering OOM kills in these cases.
1927 if (unlikely(current->flags & PF_MEMALLOC))
1930 if (unlikely(task_in_memcg_oom(current)))
1933 if (!gfpflags_allow_blocking(gfp_mask))
1936 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1938 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1939 gfp_mask, may_swap);
1941 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1945 drain_all_stock(mem_over_limit);
1950 if (gfp_mask & __GFP_NORETRY)
1953 * Even though the limit is exceeded at this point, reclaim
1954 * may have been able to free some pages. Retry the charge
1955 * before killing the task.
1957 * Only for regular pages, though: huge pages are rather
1958 * unlikely to succeed so close to the limit, and we fall back
1959 * to regular pages anyway in case of failure.
1961 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
1964 * At task move, charge accounts can be doubly counted. So, it's
1965 * better to wait until the end of task_move if something is going on.
1967 if (mem_cgroup_wait_acct_move(mem_over_limit))
1973 if (gfp_mask & __GFP_NOFAIL)
1976 if (fatal_signal_pending(current))
1979 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
1981 mem_cgroup_oom(mem_over_limit, gfp_mask,
1982 get_order(nr_pages * PAGE_SIZE));
1984 if (!(gfp_mask & __GFP_NOFAIL))
1988 * The allocation either can't fail or will lead to more memory
1989 * being freed very soon. Allow memory usage go over the limit
1990 * temporarily by force charging it.
1992 page_counter_charge(&memcg->memory, nr_pages);
1993 if (do_memsw_account())
1994 page_counter_charge(&memcg->memsw, nr_pages);
1995 css_get_many(&memcg->css, nr_pages);
2000 css_get_many(&memcg->css, batch);
2001 if (batch > nr_pages)
2002 refill_stock(memcg, batch - nr_pages);
2005 * If the hierarchy is above the normal consumption range, schedule
2006 * reclaim on returning to userland. We can perform reclaim here
2007 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2008 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2009 * not recorded as it most likely matches current's and won't
2010 * change in the meantime. As high limit is checked again before
2011 * reclaim, the cost of mismatch is negligible.
2014 if (page_counter_read(&memcg->memory) > memcg->high) {
2015 /* Don't bother a random interrupted task */
2016 if (in_interrupt()) {
2017 schedule_work(&memcg->high_work);
2020 current->memcg_nr_pages_over_high += batch;
2021 set_notify_resume(current);
2024 } while ((memcg = parent_mem_cgroup(memcg)));
2029 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2031 if (mem_cgroup_is_root(memcg))
2034 page_counter_uncharge(&memcg->memory, nr_pages);
2035 if (do_memsw_account())
2036 page_counter_uncharge(&memcg->memsw, nr_pages);
2038 css_put_many(&memcg->css, nr_pages);
2041 static void lock_page_lru(struct page *page, int *isolated)
2043 struct zone *zone = page_zone(page);
2045 spin_lock_irq(zone_lru_lock(zone));
2046 if (PageLRU(page)) {
2047 struct lruvec *lruvec;
2049 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2051 del_page_from_lru_list(page, lruvec, page_lru(page));
2057 static void unlock_page_lru(struct page *page, int isolated)
2059 struct zone *zone = page_zone(page);
2062 struct lruvec *lruvec;
2064 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2065 VM_BUG_ON_PAGE(PageLRU(page), page);
2067 add_page_to_lru_list(page, lruvec, page_lru(page));
2069 spin_unlock_irq(zone_lru_lock(zone));
2072 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2077 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2080 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2081 * may already be on some other mem_cgroup's LRU. Take care of it.
2084 lock_page_lru(page, &isolated);
2087 * Nobody should be changing or seriously looking at
2088 * page->mem_cgroup at this point:
2090 * - the page is uncharged
2092 * - the page is off-LRU
2094 * - an anonymous fault has exclusive page access, except for
2095 * a locked page table
2097 * - a page cache insertion, a swapin fault, or a migration
2098 * have the page locked
2100 page->mem_cgroup = memcg;
2103 unlock_page_lru(page, isolated);
2107 static int memcg_alloc_cache_id(void)
2112 id = ida_simple_get(&memcg_cache_ida,
2113 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2117 if (id < memcg_nr_cache_ids)
2121 * There's no space for the new id in memcg_caches arrays,
2122 * so we have to grow them.
2124 down_write(&memcg_cache_ids_sem);
2126 size = 2 * (id + 1);
2127 if (size < MEMCG_CACHES_MIN_SIZE)
2128 size = MEMCG_CACHES_MIN_SIZE;
2129 else if (size > MEMCG_CACHES_MAX_SIZE)
2130 size = MEMCG_CACHES_MAX_SIZE;
2132 err = memcg_update_all_caches(size);
2134 err = memcg_update_all_list_lrus(size);
2136 memcg_nr_cache_ids = size;
2138 up_write(&memcg_cache_ids_sem);
2141 ida_simple_remove(&memcg_cache_ida, id);
2147 static void memcg_free_cache_id(int id)
2149 ida_simple_remove(&memcg_cache_ida, id);
2152 struct memcg_kmem_cache_create_work {
2153 struct mem_cgroup *memcg;
2154 struct kmem_cache *cachep;
2155 struct work_struct work;
2158 static void memcg_kmem_cache_create_func(struct work_struct *w)
2160 struct memcg_kmem_cache_create_work *cw =
2161 container_of(w, struct memcg_kmem_cache_create_work, work);
2162 struct mem_cgroup *memcg = cw->memcg;
2163 struct kmem_cache *cachep = cw->cachep;
2165 memcg_create_kmem_cache(memcg, cachep);
2167 css_put(&memcg->css);
2172 * Enqueue the creation of a per-memcg kmem_cache.
2174 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2175 struct kmem_cache *cachep)
2177 struct memcg_kmem_cache_create_work *cw;
2179 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2183 css_get(&memcg->css);
2186 cw->cachep = cachep;
2187 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2189 queue_work(memcg_kmem_cache_wq, &cw->work);
2192 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2193 struct kmem_cache *cachep)
2196 * We need to stop accounting when we kmalloc, because if the
2197 * corresponding kmalloc cache is not yet created, the first allocation
2198 * in __memcg_schedule_kmem_cache_create will recurse.
2200 * However, it is better to enclose the whole function. Depending on
2201 * the debugging options enabled, INIT_WORK(), for instance, can
2202 * trigger an allocation. This too, will make us recurse. Because at
2203 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2204 * the safest choice is to do it like this, wrapping the whole function.
2206 current->memcg_kmem_skip_account = 1;
2207 __memcg_schedule_kmem_cache_create(memcg, cachep);
2208 current->memcg_kmem_skip_account = 0;
2211 static inline bool memcg_kmem_bypass(void)
2213 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2219 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2220 * @cachep: the original global kmem cache
2222 * Return the kmem_cache we're supposed to use for a slab allocation.
2223 * We try to use the current memcg's version of the cache.
2225 * If the cache does not exist yet, if we are the first user of it, we
2226 * create it asynchronously in a workqueue and let the current allocation
2227 * go through with the original cache.
2229 * This function takes a reference to the cache it returns to assure it
2230 * won't get destroyed while we are working with it. Once the caller is
2231 * done with it, memcg_kmem_put_cache() must be called to release the
2234 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2236 struct mem_cgroup *memcg;
2237 struct kmem_cache *memcg_cachep;
2240 VM_BUG_ON(!is_root_cache(cachep));
2242 if (memcg_kmem_bypass())
2245 if (current->memcg_kmem_skip_account)
2248 memcg = get_mem_cgroup_from_mm(current->mm);
2249 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2253 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2254 if (likely(memcg_cachep))
2255 return memcg_cachep;
2258 * If we are in a safe context (can wait, and not in interrupt
2259 * context), we could be be predictable and return right away.
2260 * This would guarantee that the allocation being performed
2261 * already belongs in the new cache.
2263 * However, there are some clashes that can arrive from locking.
2264 * For instance, because we acquire the slab_mutex while doing
2265 * memcg_create_kmem_cache, this means no further allocation
2266 * could happen with the slab_mutex held. So it's better to
2269 memcg_schedule_kmem_cache_create(memcg, cachep);
2271 css_put(&memcg->css);
2276 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2277 * @cachep: the cache returned by memcg_kmem_get_cache
2279 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2281 if (!is_root_cache(cachep))
2282 css_put(&cachep->memcg_params.memcg->css);
2286 * memcg_kmem_charge: charge a kmem page
2287 * @page: page to charge
2288 * @gfp: reclaim mode
2289 * @order: allocation order
2290 * @memcg: memory cgroup to charge
2292 * Returns 0 on success, an error code on failure.
2294 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2295 struct mem_cgroup *memcg)
2297 unsigned int nr_pages = 1 << order;
2298 struct page_counter *counter;
2301 ret = try_charge(memcg, gfp, nr_pages);
2305 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2306 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2307 cancel_charge(memcg, nr_pages);
2311 page->mem_cgroup = memcg;
2317 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2318 * @page: page to charge
2319 * @gfp: reclaim mode
2320 * @order: allocation order
2322 * Returns 0 on success, an error code on failure.
2324 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2326 struct mem_cgroup *memcg;
2329 if (memcg_kmem_bypass())
2332 memcg = get_mem_cgroup_from_mm(current->mm);
2333 if (!mem_cgroup_is_root(memcg)) {
2334 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2336 __SetPageKmemcg(page);
2338 css_put(&memcg->css);
2342 * memcg_kmem_uncharge: uncharge a kmem page
2343 * @page: page to uncharge
2344 * @order: allocation order
2346 void memcg_kmem_uncharge(struct page *page, int order)
2348 struct mem_cgroup *memcg = page->mem_cgroup;
2349 unsigned int nr_pages = 1 << order;
2354 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2356 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2357 page_counter_uncharge(&memcg->kmem, nr_pages);
2359 page_counter_uncharge(&memcg->memory, nr_pages);
2360 if (do_memsw_account())
2361 page_counter_uncharge(&memcg->memsw, nr_pages);
2363 page->mem_cgroup = NULL;
2365 /* slab pages do not have PageKmemcg flag set */
2366 if (PageKmemcg(page))
2367 __ClearPageKmemcg(page);
2369 css_put_many(&memcg->css, nr_pages);
2371 #endif /* !CONFIG_SLOB */
2373 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2376 * Because tail pages are not marked as "used", set it. We're under
2377 * zone_lru_lock and migration entries setup in all page mappings.
2379 void mem_cgroup_split_huge_fixup(struct page *head)
2383 if (mem_cgroup_disabled())
2386 for (i = 1; i < HPAGE_PMD_NR; i++)
2387 head[i].mem_cgroup = head->mem_cgroup;
2389 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2392 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2394 #ifdef CONFIG_MEMCG_SWAP
2395 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2398 int val = (charge) ? 1 : -1;
2399 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2403 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2404 * @entry: swap entry to be moved
2405 * @from: mem_cgroup which the entry is moved from
2406 * @to: mem_cgroup which the entry is moved to
2408 * It succeeds only when the swap_cgroup's record for this entry is the same
2409 * as the mem_cgroup's id of @from.
2411 * Returns 0 on success, -EINVAL on failure.
2413 * The caller must have charged to @to, IOW, called page_counter_charge() about
2414 * both res and memsw, and called css_get().
2416 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2417 struct mem_cgroup *from, struct mem_cgroup *to)
2419 unsigned short old_id, new_id;
2421 old_id = mem_cgroup_id(from);
2422 new_id = mem_cgroup_id(to);
2424 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2425 mem_cgroup_swap_statistics(from, false);
2426 mem_cgroup_swap_statistics(to, true);
2432 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2433 struct mem_cgroup *from, struct mem_cgroup *to)
2439 static DEFINE_MUTEX(memcg_limit_mutex);
2441 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2442 unsigned long limit)
2444 unsigned long curusage;
2445 unsigned long oldusage;
2446 bool enlarge = false;
2451 * For keeping hierarchical_reclaim simple, how long we should retry
2452 * is depends on callers. We set our retry-count to be function
2453 * of # of children which we should visit in this loop.
2455 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2456 mem_cgroup_count_children(memcg);
2458 oldusage = page_counter_read(&memcg->memory);
2461 if (signal_pending(current)) {
2466 mutex_lock(&memcg_limit_mutex);
2467 if (limit > memcg->memsw.limit) {
2468 mutex_unlock(&memcg_limit_mutex);
2472 if (limit > memcg->memory.limit)
2474 ret = page_counter_limit(&memcg->memory, limit);
2475 mutex_unlock(&memcg_limit_mutex);
2480 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2482 curusage = page_counter_read(&memcg->memory);
2483 /* Usage is reduced ? */
2484 if (curusage >= oldusage)
2487 oldusage = curusage;
2488 } while (retry_count);
2490 if (!ret && enlarge)
2491 memcg_oom_recover(memcg);
2496 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2497 unsigned long limit)
2499 unsigned long curusage;
2500 unsigned long oldusage;
2501 bool enlarge = false;
2505 /* see mem_cgroup_resize_res_limit */
2506 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2507 mem_cgroup_count_children(memcg);
2509 oldusage = page_counter_read(&memcg->memsw);
2512 if (signal_pending(current)) {
2517 mutex_lock(&memcg_limit_mutex);
2518 if (limit < memcg->memory.limit) {
2519 mutex_unlock(&memcg_limit_mutex);
2523 if (limit > memcg->memsw.limit)
2525 ret = page_counter_limit(&memcg->memsw, limit);
2526 mutex_unlock(&memcg_limit_mutex);
2531 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2533 curusage = page_counter_read(&memcg->memsw);
2534 /* Usage is reduced ? */
2535 if (curusage >= oldusage)
2538 oldusage = curusage;
2539 } while (retry_count);
2541 if (!ret && enlarge)
2542 memcg_oom_recover(memcg);
2547 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2549 unsigned long *total_scanned)
2551 unsigned long nr_reclaimed = 0;
2552 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2553 unsigned long reclaimed;
2555 struct mem_cgroup_tree_per_node *mctz;
2556 unsigned long excess;
2557 unsigned long nr_scanned;
2562 mctz = soft_limit_tree_node(pgdat->node_id);
2565 * Do not even bother to check the largest node if the root
2566 * is empty. Do it lockless to prevent lock bouncing. Races
2567 * are acceptable as soft limit is best effort anyway.
2569 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2573 * This loop can run a while, specially if mem_cgroup's continuously
2574 * keep exceeding their soft limit and putting the system under
2581 mz = mem_cgroup_largest_soft_limit_node(mctz);
2586 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2587 gfp_mask, &nr_scanned);
2588 nr_reclaimed += reclaimed;
2589 *total_scanned += nr_scanned;
2590 spin_lock_irq(&mctz->lock);
2591 __mem_cgroup_remove_exceeded(mz, mctz);
2594 * If we failed to reclaim anything from this memory cgroup
2595 * it is time to move on to the next cgroup
2599 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2601 excess = soft_limit_excess(mz->memcg);
2603 * One school of thought says that we should not add
2604 * back the node to the tree if reclaim returns 0.
2605 * But our reclaim could return 0, simply because due
2606 * to priority we are exposing a smaller subset of
2607 * memory to reclaim from. Consider this as a longer
2610 /* If excess == 0, no tree ops */
2611 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2612 spin_unlock_irq(&mctz->lock);
2613 css_put(&mz->memcg->css);
2616 * Could not reclaim anything and there are no more
2617 * mem cgroups to try or we seem to be looping without
2618 * reclaiming anything.
2620 if (!nr_reclaimed &&
2622 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2624 } while (!nr_reclaimed);
2626 css_put(&next_mz->memcg->css);
2627 return nr_reclaimed;
2631 * Test whether @memcg has children, dead or alive. Note that this
2632 * function doesn't care whether @memcg has use_hierarchy enabled and
2633 * returns %true if there are child csses according to the cgroup
2634 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2636 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2641 ret = css_next_child(NULL, &memcg->css);
2647 * Reclaims as many pages from the given memcg as possible.
2649 * Caller is responsible for holding css reference for memcg.
2651 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2653 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2655 /* we call try-to-free pages for make this cgroup empty */
2656 lru_add_drain_all();
2657 /* try to free all pages in this cgroup */
2658 while (nr_retries && page_counter_read(&memcg->memory)) {
2661 if (signal_pending(current))
2664 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2668 /* maybe some writeback is necessary */
2669 congestion_wait(BLK_RW_ASYNC, HZ/10);
2677 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2678 char *buf, size_t nbytes,
2681 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2683 if (mem_cgroup_is_root(memcg))
2685 return mem_cgroup_force_empty(memcg) ?: nbytes;
2688 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2691 return mem_cgroup_from_css(css)->use_hierarchy;
2694 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2695 struct cftype *cft, u64 val)
2698 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2699 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2701 if (memcg->use_hierarchy == val)
2705 * If parent's use_hierarchy is set, we can't make any modifications
2706 * in the child subtrees. If it is unset, then the change can
2707 * occur, provided the current cgroup has no children.
2709 * For the root cgroup, parent_mem is NULL, we allow value to be
2710 * set if there are no children.
2712 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2713 (val == 1 || val == 0)) {
2714 if (!memcg_has_children(memcg))
2715 memcg->use_hierarchy = val;
2724 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2726 struct mem_cgroup *iter;
2729 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2731 for_each_mem_cgroup_tree(iter, memcg) {
2732 for (i = 0; i < MEMCG_NR_STAT; i++)
2733 stat[i] += mem_cgroup_read_stat(iter, i);
2737 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2739 struct mem_cgroup *iter;
2742 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2744 for_each_mem_cgroup_tree(iter, memcg) {
2745 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2746 events[i] += mem_cgroup_read_events(iter, i);
2750 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2752 unsigned long val = 0;
2754 if (mem_cgroup_is_root(memcg)) {
2755 struct mem_cgroup *iter;
2757 for_each_mem_cgroup_tree(iter, memcg) {
2758 val += mem_cgroup_read_stat(iter,
2759 MEM_CGROUP_STAT_CACHE);
2760 val += mem_cgroup_read_stat(iter,
2761 MEM_CGROUP_STAT_RSS);
2763 val += mem_cgroup_read_stat(iter,
2764 MEM_CGROUP_STAT_SWAP);
2768 val = page_counter_read(&memcg->memory);
2770 val = page_counter_read(&memcg->memsw);
2783 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2786 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2787 struct page_counter *counter;
2789 switch (MEMFILE_TYPE(cft->private)) {
2791 counter = &memcg->memory;
2794 counter = &memcg->memsw;
2797 counter = &memcg->kmem;
2800 counter = &memcg->tcpmem;
2806 switch (MEMFILE_ATTR(cft->private)) {
2808 if (counter == &memcg->memory)
2809 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2810 if (counter == &memcg->memsw)
2811 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2812 return (u64)page_counter_read(counter) * PAGE_SIZE;
2814 return (u64)counter->limit * PAGE_SIZE;
2816 return (u64)counter->watermark * PAGE_SIZE;
2818 return counter->failcnt;
2819 case RES_SOFT_LIMIT:
2820 return (u64)memcg->soft_limit * PAGE_SIZE;
2827 static int memcg_online_kmem(struct mem_cgroup *memcg)
2831 if (cgroup_memory_nokmem)
2834 BUG_ON(memcg->kmemcg_id >= 0);
2835 BUG_ON(memcg->kmem_state);
2837 memcg_id = memcg_alloc_cache_id();
2841 static_branch_inc(&memcg_kmem_enabled_key);
2843 * A memory cgroup is considered kmem-online as soon as it gets
2844 * kmemcg_id. Setting the id after enabling static branching will
2845 * guarantee no one starts accounting before all call sites are
2848 memcg->kmemcg_id = memcg_id;
2849 memcg->kmem_state = KMEM_ONLINE;
2850 INIT_LIST_HEAD(&memcg->kmem_caches);
2855 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2857 struct cgroup_subsys_state *css;
2858 struct mem_cgroup *parent, *child;
2861 if (memcg->kmem_state != KMEM_ONLINE)
2864 * Clear the online state before clearing memcg_caches array
2865 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2866 * guarantees that no cache will be created for this cgroup
2867 * after we are done (see memcg_create_kmem_cache()).
2869 memcg->kmem_state = KMEM_ALLOCATED;
2871 memcg_deactivate_kmem_caches(memcg);
2873 kmemcg_id = memcg->kmemcg_id;
2874 BUG_ON(kmemcg_id < 0);
2876 parent = parent_mem_cgroup(memcg);
2878 parent = root_mem_cgroup;
2881 * Change kmemcg_id of this cgroup and all its descendants to the
2882 * parent's id, and then move all entries from this cgroup's list_lrus
2883 * to ones of the parent. After we have finished, all list_lrus
2884 * corresponding to this cgroup are guaranteed to remain empty. The
2885 * ordering is imposed by list_lru_node->lock taken by
2886 * memcg_drain_all_list_lrus().
2888 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2889 css_for_each_descendant_pre(css, &memcg->css) {
2890 child = mem_cgroup_from_css(css);
2891 BUG_ON(child->kmemcg_id != kmemcg_id);
2892 child->kmemcg_id = parent->kmemcg_id;
2893 if (!memcg->use_hierarchy)
2898 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2900 memcg_free_cache_id(kmemcg_id);
2903 static void memcg_free_kmem(struct mem_cgroup *memcg)
2905 /* css_alloc() failed, offlining didn't happen */
2906 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2907 memcg_offline_kmem(memcg);
2909 if (memcg->kmem_state == KMEM_ALLOCATED) {
2910 memcg_destroy_kmem_caches(memcg);
2911 static_branch_dec(&memcg_kmem_enabled_key);
2912 WARN_ON(page_counter_read(&memcg->kmem));
2916 static int memcg_online_kmem(struct mem_cgroup *memcg)
2920 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2923 static void memcg_free_kmem(struct mem_cgroup *memcg)
2926 #endif /* !CONFIG_SLOB */
2928 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2929 unsigned long limit)
2933 mutex_lock(&memcg_limit_mutex);
2934 ret = page_counter_limit(&memcg->kmem, limit);
2935 mutex_unlock(&memcg_limit_mutex);
2939 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2943 mutex_lock(&memcg_limit_mutex);
2945 ret = page_counter_limit(&memcg->tcpmem, limit);
2949 if (!memcg->tcpmem_active) {
2951 * The active flag needs to be written after the static_key
2952 * update. This is what guarantees that the socket activation
2953 * function is the last one to run. See mem_cgroup_sk_alloc()
2954 * for details, and note that we don't mark any socket as
2955 * belonging to this memcg until that flag is up.
2957 * We need to do this, because static_keys will span multiple
2958 * sites, but we can't control their order. If we mark a socket
2959 * as accounted, but the accounting functions are not patched in
2960 * yet, we'll lose accounting.
2962 * We never race with the readers in mem_cgroup_sk_alloc(),
2963 * because when this value change, the code to process it is not
2966 static_branch_inc(&memcg_sockets_enabled_key);
2967 memcg->tcpmem_active = true;
2970 mutex_unlock(&memcg_limit_mutex);
2975 * The user of this function is...
2978 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2979 char *buf, size_t nbytes, loff_t off)
2981 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2982 unsigned long nr_pages;
2985 buf = strstrip(buf);
2986 ret = page_counter_memparse(buf, "-1", &nr_pages);
2990 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2992 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2996 switch (MEMFILE_TYPE(of_cft(of)->private)) {
2998 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3001 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3004 ret = memcg_update_kmem_limit(memcg, nr_pages);
3007 ret = memcg_update_tcp_limit(memcg, nr_pages);
3011 case RES_SOFT_LIMIT:
3012 memcg->soft_limit = nr_pages;
3016 return ret ?: nbytes;
3019 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3020 size_t nbytes, loff_t off)
3022 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3023 struct page_counter *counter;
3025 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3027 counter = &memcg->memory;
3030 counter = &memcg->memsw;
3033 counter = &memcg->kmem;
3036 counter = &memcg->tcpmem;
3042 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3044 page_counter_reset_watermark(counter);
3047 counter->failcnt = 0;
3056 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3059 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3063 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3064 struct cftype *cft, u64 val)
3066 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3068 if (val & ~MOVE_MASK)
3072 * No kind of locking is needed in here, because ->can_attach() will
3073 * check this value once in the beginning of the process, and then carry
3074 * on with stale data. This means that changes to this value will only
3075 * affect task migrations starting after the change.
3077 memcg->move_charge_at_immigrate = val;
3081 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3082 struct cftype *cft, u64 val)
3089 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3093 unsigned int lru_mask;
3096 static const struct numa_stat stats[] = {
3097 { "total", LRU_ALL },
3098 { "file", LRU_ALL_FILE },
3099 { "anon", LRU_ALL_ANON },
3100 { "unevictable", BIT(LRU_UNEVICTABLE) },
3102 const struct numa_stat *stat;
3105 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3107 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3108 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3109 seq_printf(m, "%s=%lu", stat->name, nr);
3110 for_each_node_state(nid, N_MEMORY) {
3111 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3113 seq_printf(m, " N%d=%lu", nid, nr);
3118 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3119 struct mem_cgroup *iter;
3122 for_each_mem_cgroup_tree(iter, memcg)
3123 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3124 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3125 for_each_node_state(nid, N_MEMORY) {
3127 for_each_mem_cgroup_tree(iter, memcg)
3128 nr += mem_cgroup_node_nr_lru_pages(
3129 iter, nid, stat->lru_mask);
3130 seq_printf(m, " N%d=%lu", nid, nr);
3137 #endif /* CONFIG_NUMA */
3139 static int memcg_stat_show(struct seq_file *m, void *v)
3141 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3142 unsigned long memory, memsw;
3143 struct mem_cgroup *mi;
3146 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3147 MEM_CGROUP_STAT_NSTATS);
3148 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3149 MEM_CGROUP_EVENTS_NSTATS);
3150 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3152 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3153 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3155 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3156 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3159 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3160 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3161 mem_cgroup_read_events(memcg, i));
3163 for (i = 0; i < NR_LRU_LISTS; i++)
3164 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3165 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3167 /* Hierarchical information */
3168 memory = memsw = PAGE_COUNTER_MAX;
3169 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3170 memory = min(memory, mi->memory.limit);
3171 memsw = min(memsw, mi->memsw.limit);
3173 seq_printf(m, "hierarchical_memory_limit %llu\n",
3174 (u64)memory * PAGE_SIZE);
3175 if (do_memsw_account())
3176 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3177 (u64)memsw * PAGE_SIZE);
3179 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3180 unsigned long long val = 0;
3182 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3184 for_each_mem_cgroup_tree(mi, memcg)
3185 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3186 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3189 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3190 unsigned long long val = 0;
3192 for_each_mem_cgroup_tree(mi, memcg)
3193 val += mem_cgroup_read_events(mi, i);
3194 seq_printf(m, "total_%s %llu\n",
3195 mem_cgroup_events_names[i], val);
3198 for (i = 0; i < NR_LRU_LISTS; i++) {
3199 unsigned long long val = 0;
3201 for_each_mem_cgroup_tree(mi, memcg)
3202 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3203 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3206 #ifdef CONFIG_DEBUG_VM
3209 struct mem_cgroup_per_node *mz;
3210 struct zone_reclaim_stat *rstat;
3211 unsigned long recent_rotated[2] = {0, 0};
3212 unsigned long recent_scanned[2] = {0, 0};
3214 for_each_online_pgdat(pgdat) {
3215 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3216 rstat = &mz->lruvec.reclaim_stat;
3218 recent_rotated[0] += rstat->recent_rotated[0];
3219 recent_rotated[1] += rstat->recent_rotated[1];
3220 recent_scanned[0] += rstat->recent_scanned[0];
3221 recent_scanned[1] += rstat->recent_scanned[1];
3223 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3224 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3225 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3226 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3233 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3236 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3238 return mem_cgroup_swappiness(memcg);
3241 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3242 struct cftype *cft, u64 val)
3244 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3250 memcg->swappiness = val;
3252 vm_swappiness = val;
3257 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3259 struct mem_cgroup_threshold_ary *t;
3260 unsigned long usage;
3265 t = rcu_dereference(memcg->thresholds.primary);
3267 t = rcu_dereference(memcg->memsw_thresholds.primary);
3272 usage = mem_cgroup_usage(memcg, swap);
3275 * current_threshold points to threshold just below or equal to usage.
3276 * If it's not true, a threshold was crossed after last
3277 * call of __mem_cgroup_threshold().
3279 i = t->current_threshold;
3282 * Iterate backward over array of thresholds starting from
3283 * current_threshold and check if a threshold is crossed.
3284 * If none of thresholds below usage is crossed, we read
3285 * only one element of the array here.
3287 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3288 eventfd_signal(t->entries[i].eventfd, 1);
3290 /* i = current_threshold + 1 */
3294 * Iterate forward over array of thresholds starting from
3295 * current_threshold+1 and check if a threshold is crossed.
3296 * If none of thresholds above usage is crossed, we read
3297 * only one element of the array here.
3299 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3300 eventfd_signal(t->entries[i].eventfd, 1);
3302 /* Update current_threshold */
3303 t->current_threshold = i - 1;
3308 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3311 __mem_cgroup_threshold(memcg, false);
3312 if (do_memsw_account())
3313 __mem_cgroup_threshold(memcg, true);
3315 memcg = parent_mem_cgroup(memcg);
3319 static int compare_thresholds(const void *a, const void *b)
3321 const struct mem_cgroup_threshold *_a = a;
3322 const struct mem_cgroup_threshold *_b = b;
3324 if (_a->threshold > _b->threshold)
3327 if (_a->threshold < _b->threshold)
3333 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3335 struct mem_cgroup_eventfd_list *ev;
3337 spin_lock(&memcg_oom_lock);
3339 list_for_each_entry(ev, &memcg->oom_notify, list)
3340 eventfd_signal(ev->eventfd, 1);
3342 spin_unlock(&memcg_oom_lock);
3346 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3348 struct mem_cgroup *iter;
3350 for_each_mem_cgroup_tree(iter, memcg)
3351 mem_cgroup_oom_notify_cb(iter);
3354 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3355 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3357 struct mem_cgroup_thresholds *thresholds;
3358 struct mem_cgroup_threshold_ary *new;
3359 unsigned long threshold;
3360 unsigned long usage;
3363 ret = page_counter_memparse(args, "-1", &threshold);
3367 mutex_lock(&memcg->thresholds_lock);
3370 thresholds = &memcg->thresholds;
3371 usage = mem_cgroup_usage(memcg, false);
3372 } else if (type == _MEMSWAP) {
3373 thresholds = &memcg->memsw_thresholds;
3374 usage = mem_cgroup_usage(memcg, true);
3378 /* Check if a threshold crossed before adding a new one */
3379 if (thresholds->primary)
3380 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3382 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3384 /* Allocate memory for new array of thresholds */
3385 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3393 /* Copy thresholds (if any) to new array */
3394 if (thresholds->primary) {
3395 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3396 sizeof(struct mem_cgroup_threshold));
3399 /* Add new threshold */
3400 new->entries[size - 1].eventfd = eventfd;
3401 new->entries[size - 1].threshold = threshold;
3403 /* Sort thresholds. Registering of new threshold isn't time-critical */
3404 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3405 compare_thresholds, NULL);
3407 /* Find current threshold */
3408 new->current_threshold = -1;
3409 for (i = 0; i < size; i++) {
3410 if (new->entries[i].threshold <= usage) {
3412 * new->current_threshold will not be used until
3413 * rcu_assign_pointer(), so it's safe to increment
3416 ++new->current_threshold;
3421 /* Free old spare buffer and save old primary buffer as spare */
3422 kfree(thresholds->spare);
3423 thresholds->spare = thresholds->primary;
3425 rcu_assign_pointer(thresholds->primary, new);
3427 /* To be sure that nobody uses thresholds */
3431 mutex_unlock(&memcg->thresholds_lock);
3436 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3437 struct eventfd_ctx *eventfd, const char *args)
3439 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3442 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3443 struct eventfd_ctx *eventfd, const char *args)
3445 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3448 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3449 struct eventfd_ctx *eventfd, enum res_type type)
3451 struct mem_cgroup_thresholds *thresholds;
3452 struct mem_cgroup_threshold_ary *new;
3453 unsigned long usage;
3456 mutex_lock(&memcg->thresholds_lock);
3459 thresholds = &memcg->thresholds;
3460 usage = mem_cgroup_usage(memcg, false);
3461 } else if (type == _MEMSWAP) {
3462 thresholds = &memcg->memsw_thresholds;
3463 usage = mem_cgroup_usage(memcg, true);
3467 if (!thresholds->primary)
3470 /* Check if a threshold crossed before removing */
3471 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3473 /* Calculate new number of threshold */
3475 for (i = 0; i < thresholds->primary->size; i++) {
3476 if (thresholds->primary->entries[i].eventfd != eventfd)
3480 new = thresholds->spare;
3482 /* Set thresholds array to NULL if we don't have thresholds */
3491 /* Copy thresholds and find current threshold */
3492 new->current_threshold = -1;
3493 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3494 if (thresholds->primary->entries[i].eventfd == eventfd)
3497 new->entries[j] = thresholds->primary->entries[i];
3498 if (new->entries[j].threshold <= usage) {
3500 * new->current_threshold will not be used
3501 * until rcu_assign_pointer(), so it's safe to increment
3504 ++new->current_threshold;
3510 /* Swap primary and spare array */
3511 thresholds->spare = thresholds->primary;
3513 rcu_assign_pointer(thresholds->primary, new);
3515 /* To be sure that nobody uses thresholds */
3518 /* If all events are unregistered, free the spare array */
3520 kfree(thresholds->spare);
3521 thresholds->spare = NULL;
3524 mutex_unlock(&memcg->thresholds_lock);
3527 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3528 struct eventfd_ctx *eventfd)
3530 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3533 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3534 struct eventfd_ctx *eventfd)
3536 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3539 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3540 struct eventfd_ctx *eventfd, const char *args)
3542 struct mem_cgroup_eventfd_list *event;
3544 event = kmalloc(sizeof(*event), GFP_KERNEL);
3548 spin_lock(&memcg_oom_lock);
3550 event->eventfd = eventfd;
3551 list_add(&event->list, &memcg->oom_notify);
3553 /* already in OOM ? */
3554 if (memcg->under_oom)
3555 eventfd_signal(eventfd, 1);
3556 spin_unlock(&memcg_oom_lock);
3561 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3562 struct eventfd_ctx *eventfd)
3564 struct mem_cgroup_eventfd_list *ev, *tmp;
3566 spin_lock(&memcg_oom_lock);
3568 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3569 if (ev->eventfd == eventfd) {
3570 list_del(&ev->list);
3575 spin_unlock(&memcg_oom_lock);
3578 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3580 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3582 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3583 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3587 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3588 struct cftype *cft, u64 val)
3590 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3592 /* cannot set to root cgroup and only 0 and 1 are allowed */
3593 if (!css->parent || !((val == 0) || (val == 1)))
3596 memcg->oom_kill_disable = val;
3598 memcg_oom_recover(memcg);
3603 #ifdef CONFIG_CGROUP_WRITEBACK
3605 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3607 return &memcg->cgwb_list;
3610 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3612 return wb_domain_init(&memcg->cgwb_domain, gfp);
3615 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3617 wb_domain_exit(&memcg->cgwb_domain);
3620 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3622 wb_domain_size_changed(&memcg->cgwb_domain);
3625 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3627 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3629 if (!memcg->css.parent)
3632 return &memcg->cgwb_domain;
3636 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3637 * @wb: bdi_writeback in question
3638 * @pfilepages: out parameter for number of file pages
3639 * @pheadroom: out parameter for number of allocatable pages according to memcg
3640 * @pdirty: out parameter for number of dirty pages
3641 * @pwriteback: out parameter for number of pages under writeback
3643 * Determine the numbers of file, headroom, dirty, and writeback pages in
3644 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3645 * is a bit more involved.
3647 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3648 * headroom is calculated as the lowest headroom of itself and the
3649 * ancestors. Note that this doesn't consider the actual amount of
3650 * available memory in the system. The caller should further cap
3651 * *@pheadroom accordingly.
3653 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3654 unsigned long *pheadroom, unsigned long *pdirty,
3655 unsigned long *pwriteback)
3657 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3658 struct mem_cgroup *parent;
3660 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3662 /* this should eventually include NR_UNSTABLE_NFS */
3663 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3664 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3665 (1 << LRU_ACTIVE_FILE));
3666 *pheadroom = PAGE_COUNTER_MAX;
3668 while ((parent = parent_mem_cgroup(memcg))) {
3669 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3670 unsigned long used = page_counter_read(&memcg->memory);
3672 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3677 #else /* CONFIG_CGROUP_WRITEBACK */
3679 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3684 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3688 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3692 #endif /* CONFIG_CGROUP_WRITEBACK */
3695 * DO NOT USE IN NEW FILES.
3697 * "cgroup.event_control" implementation.
3699 * This is way over-engineered. It tries to support fully configurable
3700 * events for each user. Such level of flexibility is completely
3701 * unnecessary especially in the light of the planned unified hierarchy.
3703 * Please deprecate this and replace with something simpler if at all
3708 * Unregister event and free resources.
3710 * Gets called from workqueue.
3712 static void memcg_event_remove(struct work_struct *work)
3714 struct mem_cgroup_event *event =
3715 container_of(work, struct mem_cgroup_event, remove);
3716 struct mem_cgroup *memcg = event->memcg;
3718 remove_wait_queue(event->wqh, &event->wait);
3720 event->unregister_event(memcg, event->eventfd);
3722 /* Notify userspace the event is going away. */
3723 eventfd_signal(event->eventfd, 1);
3725 eventfd_ctx_put(event->eventfd);
3727 css_put(&memcg->css);
3731 * Gets called on POLLHUP on eventfd when user closes it.
3733 * Called with wqh->lock held and interrupts disabled.
3735 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3736 int sync, void *key)
3738 struct mem_cgroup_event *event =
3739 container_of(wait, struct mem_cgroup_event, wait);
3740 struct mem_cgroup *memcg = event->memcg;
3741 unsigned long flags = (unsigned long)key;
3743 if (flags & POLLHUP) {
3745 * If the event has been detached at cgroup removal, we
3746 * can simply return knowing the other side will cleanup
3749 * We can't race against event freeing since the other
3750 * side will require wqh->lock via remove_wait_queue(),
3753 spin_lock(&memcg->event_list_lock);
3754 if (!list_empty(&event->list)) {
3755 list_del_init(&event->list);
3757 * We are in atomic context, but cgroup_event_remove()
3758 * may sleep, so we have to call it in workqueue.
3760 schedule_work(&event->remove);
3762 spin_unlock(&memcg->event_list_lock);
3768 static void memcg_event_ptable_queue_proc(struct file *file,
3769 wait_queue_head_t *wqh, poll_table *pt)
3771 struct mem_cgroup_event *event =
3772 container_of(pt, struct mem_cgroup_event, pt);
3775 add_wait_queue(wqh, &event->wait);
3779 * DO NOT USE IN NEW FILES.
3781 * Parse input and register new cgroup event handler.
3783 * Input must be in format '<event_fd> <control_fd> <args>'.
3784 * Interpretation of args is defined by control file implementation.
3786 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3787 char *buf, size_t nbytes, loff_t off)
3789 struct cgroup_subsys_state *css = of_css(of);
3790 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3791 struct mem_cgroup_event *event;
3792 struct cgroup_subsys_state *cfile_css;
3793 unsigned int efd, cfd;
3800 buf = strstrip(buf);
3802 efd = simple_strtoul(buf, &endp, 10);
3807 cfd = simple_strtoul(buf, &endp, 10);
3808 if ((*endp != ' ') && (*endp != '\0'))
3812 event = kzalloc(sizeof(*event), GFP_KERNEL);
3816 event->memcg = memcg;
3817 INIT_LIST_HEAD(&event->list);
3818 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3819 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3820 INIT_WORK(&event->remove, memcg_event_remove);
3828 event->eventfd = eventfd_ctx_fileget(efile.file);
3829 if (IS_ERR(event->eventfd)) {
3830 ret = PTR_ERR(event->eventfd);
3837 goto out_put_eventfd;
3840 /* the process need read permission on control file */
3841 /* AV: shouldn't we check that it's been opened for read instead? */
3842 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3847 * Determine the event callbacks and set them in @event. This used
3848 * to be done via struct cftype but cgroup core no longer knows
3849 * about these events. The following is crude but the whole thing
3850 * is for compatibility anyway.
3852 * DO NOT ADD NEW FILES.
3854 name = cfile.file->f_path.dentry->d_name.name;
3856 if (!strcmp(name, "memory.usage_in_bytes")) {
3857 event->register_event = mem_cgroup_usage_register_event;
3858 event->unregister_event = mem_cgroup_usage_unregister_event;
3859 } else if (!strcmp(name, "memory.oom_control")) {
3860 event->register_event = mem_cgroup_oom_register_event;
3861 event->unregister_event = mem_cgroup_oom_unregister_event;
3862 } else if (!strcmp(name, "memory.pressure_level")) {
3863 event->register_event = vmpressure_register_event;
3864 event->unregister_event = vmpressure_unregister_event;
3865 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3866 event->register_event = memsw_cgroup_usage_register_event;
3867 event->unregister_event = memsw_cgroup_usage_unregister_event;
3874 * Verify @cfile should belong to @css. Also, remaining events are
3875 * automatically removed on cgroup destruction but the removal is
3876 * asynchronous, so take an extra ref on @css.
3878 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3879 &memory_cgrp_subsys);
3881 if (IS_ERR(cfile_css))
3883 if (cfile_css != css) {
3888 ret = event->register_event(memcg, event->eventfd, buf);
3892 efile.file->f_op->poll(efile.file, &event->pt);
3894 spin_lock(&memcg->event_list_lock);
3895 list_add(&event->list, &memcg->event_list);
3896 spin_unlock(&memcg->event_list_lock);
3908 eventfd_ctx_put(event->eventfd);
3917 static struct cftype mem_cgroup_legacy_files[] = {
3919 .name = "usage_in_bytes",
3920 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3921 .read_u64 = mem_cgroup_read_u64,
3924 .name = "max_usage_in_bytes",
3925 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3926 .write = mem_cgroup_reset,
3927 .read_u64 = mem_cgroup_read_u64,
3930 .name = "limit_in_bytes",
3931 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3932 .write = mem_cgroup_write,
3933 .read_u64 = mem_cgroup_read_u64,
3936 .name = "soft_limit_in_bytes",
3937 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3938 .write = mem_cgroup_write,
3939 .read_u64 = mem_cgroup_read_u64,
3943 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3944 .write = mem_cgroup_reset,
3945 .read_u64 = mem_cgroup_read_u64,
3949 .seq_show = memcg_stat_show,
3952 .name = "force_empty",
3953 .write = mem_cgroup_force_empty_write,
3956 .name = "use_hierarchy",
3957 .write_u64 = mem_cgroup_hierarchy_write,
3958 .read_u64 = mem_cgroup_hierarchy_read,
3961 .name = "cgroup.event_control", /* XXX: for compat */
3962 .write = memcg_write_event_control,
3963 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3966 .name = "swappiness",
3967 .read_u64 = mem_cgroup_swappiness_read,
3968 .write_u64 = mem_cgroup_swappiness_write,
3971 .name = "move_charge_at_immigrate",
3972 .read_u64 = mem_cgroup_move_charge_read,
3973 .write_u64 = mem_cgroup_move_charge_write,
3976 .name = "oom_control",
3977 .seq_show = mem_cgroup_oom_control_read,
3978 .write_u64 = mem_cgroup_oom_control_write,
3979 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3982 .name = "pressure_level",
3986 .name = "numa_stat",
3987 .seq_show = memcg_numa_stat_show,
3991 .name = "kmem.limit_in_bytes",
3992 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
3993 .write = mem_cgroup_write,
3994 .read_u64 = mem_cgroup_read_u64,
3997 .name = "kmem.usage_in_bytes",
3998 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
3999 .read_u64 = mem_cgroup_read_u64,
4002 .name = "kmem.failcnt",
4003 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4004 .write = mem_cgroup_reset,
4005 .read_u64 = mem_cgroup_read_u64,
4008 .name = "kmem.max_usage_in_bytes",
4009 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4010 .write = mem_cgroup_reset,
4011 .read_u64 = mem_cgroup_read_u64,
4013 #ifdef CONFIG_SLABINFO
4015 .name = "kmem.slabinfo",
4016 .seq_start = memcg_slab_start,
4017 .seq_next = memcg_slab_next,
4018 .seq_stop = memcg_slab_stop,
4019 .seq_show = memcg_slab_show,
4023 .name = "kmem.tcp.limit_in_bytes",
4024 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4025 .write = mem_cgroup_write,
4026 .read_u64 = mem_cgroup_read_u64,
4029 .name = "kmem.tcp.usage_in_bytes",
4030 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4031 .read_u64 = mem_cgroup_read_u64,
4034 .name = "kmem.tcp.failcnt",
4035 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4036 .write = mem_cgroup_reset,
4037 .read_u64 = mem_cgroup_read_u64,
4040 .name = "kmem.tcp.max_usage_in_bytes",
4041 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4042 .write = mem_cgroup_reset,
4043 .read_u64 = mem_cgroup_read_u64,
4045 { }, /* terminate */
4049 * Private memory cgroup IDR
4051 * Swap-out records and page cache shadow entries need to store memcg
4052 * references in constrained space, so we maintain an ID space that is
4053 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4054 * memory-controlled cgroups to 64k.
4056 * However, there usually are many references to the oflline CSS after
4057 * the cgroup has been destroyed, such as page cache or reclaimable
4058 * slab objects, that don't need to hang on to the ID. We want to keep
4059 * those dead CSS from occupying IDs, or we might quickly exhaust the
4060 * relatively small ID space and prevent the creation of new cgroups
4061 * even when there are much fewer than 64k cgroups - possibly none.
4063 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4064 * be freed and recycled when it's no longer needed, which is usually
4065 * when the CSS is offlined.
4067 * The only exception to that are records of swapped out tmpfs/shmem
4068 * pages that need to be attributed to live ancestors on swapin. But
4069 * those references are manageable from userspace.
4072 static DEFINE_IDR(mem_cgroup_idr);
4074 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4076 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4077 atomic_add(n, &memcg->id.ref);
4080 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4082 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4083 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4084 idr_remove(&mem_cgroup_idr, memcg->id.id);
4087 /* Memcg ID pins CSS */
4088 css_put(&memcg->css);
4092 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4094 mem_cgroup_id_get_many(memcg, 1);
4097 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4099 mem_cgroup_id_put_many(memcg, 1);
4103 * mem_cgroup_from_id - look up a memcg from a memcg id
4104 * @id: the memcg id to look up
4106 * Caller must hold rcu_read_lock().
4108 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4110 WARN_ON_ONCE(!rcu_read_lock_held());
4111 return idr_find(&mem_cgroup_idr, id);
4114 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4116 struct mem_cgroup_per_node *pn;
4119 * This routine is called against possible nodes.
4120 * But it's BUG to call kmalloc() against offline node.
4122 * TODO: this routine can waste much memory for nodes which will
4123 * never be onlined. It's better to use memory hotplug callback
4126 if (!node_state(node, N_NORMAL_MEMORY))
4128 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4132 lruvec_init(&pn->lruvec);
4133 pn->usage_in_excess = 0;
4134 pn->on_tree = false;
4137 memcg->nodeinfo[node] = pn;
4141 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4143 kfree(memcg->nodeinfo[node]);
4146 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4151 free_mem_cgroup_per_node_info(memcg, node);
4152 free_percpu(memcg->stat);
4156 static void mem_cgroup_free(struct mem_cgroup *memcg)
4158 memcg_wb_domain_exit(memcg);
4159 __mem_cgroup_free(memcg);
4162 static struct mem_cgroup *mem_cgroup_alloc(void)
4164 struct mem_cgroup *memcg;
4168 size = sizeof(struct mem_cgroup);
4169 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4171 memcg = kzalloc(size, GFP_KERNEL);
4175 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4176 1, MEM_CGROUP_ID_MAX,
4178 if (memcg->id.id < 0)
4181 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4186 if (alloc_mem_cgroup_per_node_info(memcg, node))
4189 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4192 INIT_WORK(&memcg->high_work, high_work_func);
4193 memcg->last_scanned_node = MAX_NUMNODES;
4194 INIT_LIST_HEAD(&memcg->oom_notify);
4195 mutex_init(&memcg->thresholds_lock);
4196 spin_lock_init(&memcg->move_lock);
4197 vmpressure_init(&memcg->vmpressure);
4198 INIT_LIST_HEAD(&memcg->event_list);
4199 spin_lock_init(&memcg->event_list_lock);
4200 memcg->socket_pressure = jiffies;
4202 memcg->kmemcg_id = -1;
4204 #ifdef CONFIG_CGROUP_WRITEBACK
4205 INIT_LIST_HEAD(&memcg->cgwb_list);
4207 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4210 if (memcg->id.id > 0)
4211 idr_remove(&mem_cgroup_idr, memcg->id.id);
4212 __mem_cgroup_free(memcg);
4216 static struct cgroup_subsys_state * __ref
4217 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4219 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4220 struct mem_cgroup *memcg;
4221 long error = -ENOMEM;
4223 memcg = mem_cgroup_alloc();
4225 return ERR_PTR(error);
4227 memcg->high = PAGE_COUNTER_MAX;
4228 memcg->soft_limit = PAGE_COUNTER_MAX;
4230 memcg->swappiness = mem_cgroup_swappiness(parent);
4231 memcg->oom_kill_disable = parent->oom_kill_disable;
4233 if (parent && parent->use_hierarchy) {
4234 memcg->use_hierarchy = true;
4235 page_counter_init(&memcg->memory, &parent->memory);
4236 page_counter_init(&memcg->swap, &parent->swap);
4237 page_counter_init(&memcg->memsw, &parent->memsw);
4238 page_counter_init(&memcg->kmem, &parent->kmem);
4239 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4241 page_counter_init(&memcg->memory, NULL);
4242 page_counter_init(&memcg->swap, NULL);
4243 page_counter_init(&memcg->memsw, NULL);
4244 page_counter_init(&memcg->kmem, NULL);
4245 page_counter_init(&memcg->tcpmem, NULL);
4247 * Deeper hierachy with use_hierarchy == false doesn't make
4248 * much sense so let cgroup subsystem know about this
4249 * unfortunate state in our controller.
4251 if (parent != root_mem_cgroup)
4252 memory_cgrp_subsys.broken_hierarchy = true;
4255 /* The following stuff does not apply to the root */
4257 root_mem_cgroup = memcg;
4261 error = memcg_online_kmem(memcg);
4265 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4266 static_branch_inc(&memcg_sockets_enabled_key);
4270 mem_cgroup_free(memcg);
4271 return ERR_PTR(-ENOMEM);
4274 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4276 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4278 /* Online state pins memcg ID, memcg ID pins CSS */
4279 atomic_set(&memcg->id.ref, 1);
4284 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4286 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4287 struct mem_cgroup_event *event, *tmp;
4290 * Unregister events and notify userspace.
4291 * Notify userspace about cgroup removing only after rmdir of cgroup
4292 * directory to avoid race between userspace and kernelspace.
4294 spin_lock(&memcg->event_list_lock);
4295 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4296 list_del_init(&event->list);
4297 schedule_work(&event->remove);
4299 spin_unlock(&memcg->event_list_lock);
4301 memcg_offline_kmem(memcg);
4302 wb_memcg_offline(memcg);
4304 mem_cgroup_id_put(memcg);
4307 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4309 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4311 invalidate_reclaim_iterators(memcg);
4314 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4316 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4318 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4319 static_branch_dec(&memcg_sockets_enabled_key);
4321 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4322 static_branch_dec(&memcg_sockets_enabled_key);
4324 vmpressure_cleanup(&memcg->vmpressure);
4325 cancel_work_sync(&memcg->high_work);
4326 mem_cgroup_remove_from_trees(memcg);
4327 memcg_free_kmem(memcg);
4328 mem_cgroup_free(memcg);
4332 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4333 * @css: the target css
4335 * Reset the states of the mem_cgroup associated with @css. This is
4336 * invoked when the userland requests disabling on the default hierarchy
4337 * but the memcg is pinned through dependency. The memcg should stop
4338 * applying policies and should revert to the vanilla state as it may be
4339 * made visible again.
4341 * The current implementation only resets the essential configurations.
4342 * This needs to be expanded to cover all the visible parts.
4344 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4346 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4348 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4349 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4350 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4351 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4352 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4354 memcg->high = PAGE_COUNTER_MAX;
4355 memcg->soft_limit = PAGE_COUNTER_MAX;
4356 memcg_wb_domain_size_changed(memcg);
4360 /* Handlers for move charge at task migration. */
4361 static int mem_cgroup_do_precharge(unsigned long count)
4365 /* Try a single bulk charge without reclaim first, kswapd may wake */
4366 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4368 mc.precharge += count;
4372 /* Try charges one by one with reclaim, but do not retry */
4374 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4388 enum mc_target_type {
4394 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4395 unsigned long addr, pte_t ptent)
4397 struct page *page = vm_normal_page(vma, addr, ptent);
4399 if (!page || !page_mapped(page))
4401 if (PageAnon(page)) {
4402 if (!(mc.flags & MOVE_ANON))
4405 if (!(mc.flags & MOVE_FILE))
4408 if (!get_page_unless_zero(page))
4415 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4416 pte_t ptent, swp_entry_t *entry)
4418 struct page *page = NULL;
4419 swp_entry_t ent = pte_to_swp_entry(ptent);
4421 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4424 * Because lookup_swap_cache() updates some statistics counter,
4425 * we call find_get_page() with swapper_space directly.
4427 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4428 if (do_memsw_account())
4429 entry->val = ent.val;
4434 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4435 pte_t ptent, swp_entry_t *entry)
4441 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4442 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4444 struct page *page = NULL;
4445 struct address_space *mapping;
4448 if (!vma->vm_file) /* anonymous vma */
4450 if (!(mc.flags & MOVE_FILE))
4453 mapping = vma->vm_file->f_mapping;
4454 pgoff = linear_page_index(vma, addr);
4456 /* page is moved even if it's not RSS of this task(page-faulted). */
4458 /* shmem/tmpfs may report page out on swap: account for that too. */
4459 if (shmem_mapping(mapping)) {
4460 page = find_get_entry(mapping, pgoff);
4461 if (radix_tree_exceptional_entry(page)) {
4462 swp_entry_t swp = radix_to_swp_entry(page);
4463 if (do_memsw_account())
4465 page = find_get_page(swap_address_space(swp),
4469 page = find_get_page(mapping, pgoff);
4471 page = find_get_page(mapping, pgoff);
4477 * mem_cgroup_move_account - move account of the page
4479 * @compound: charge the page as compound or small page
4480 * @from: mem_cgroup which the page is moved from.
4481 * @to: mem_cgroup which the page is moved to. @from != @to.
4483 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4485 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4488 static int mem_cgroup_move_account(struct page *page,
4490 struct mem_cgroup *from,
4491 struct mem_cgroup *to)
4493 unsigned long flags;
4494 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4498 VM_BUG_ON(from == to);
4499 VM_BUG_ON_PAGE(PageLRU(page), page);
4500 VM_BUG_ON(compound && !PageTransHuge(page));
4503 * Prevent mem_cgroup_migrate() from looking at
4504 * page->mem_cgroup of its source page while we change it.
4507 if (!trylock_page(page))
4511 if (page->mem_cgroup != from)
4514 anon = PageAnon(page);
4516 spin_lock_irqsave(&from->move_lock, flags);
4518 if (!anon && page_mapped(page)) {
4519 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4521 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4526 * move_lock grabbed above and caller set from->moving_account, so
4527 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4528 * So mapping should be stable for dirty pages.
4530 if (!anon && PageDirty(page)) {
4531 struct address_space *mapping = page_mapping(page);
4533 if (mapping_cap_account_dirty(mapping)) {
4534 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4536 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4541 if (PageWriteback(page)) {
4542 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4544 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4549 * It is safe to change page->mem_cgroup here because the page
4550 * is referenced, charged, and isolated - we can't race with
4551 * uncharging, charging, migration, or LRU putback.
4554 /* caller should have done css_get */
4555 page->mem_cgroup = to;
4556 spin_unlock_irqrestore(&from->move_lock, flags);
4560 local_irq_disable();
4561 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4562 memcg_check_events(to, page);
4563 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4564 memcg_check_events(from, page);
4573 * get_mctgt_type - get target type of moving charge
4574 * @vma: the vma the pte to be checked belongs
4575 * @addr: the address corresponding to the pte to be checked
4576 * @ptent: the pte to be checked
4577 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4580 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4581 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4582 * move charge. if @target is not NULL, the page is stored in target->page
4583 * with extra refcnt got(Callers should handle it).
4584 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4585 * target for charge migration. if @target is not NULL, the entry is stored
4588 * Called with pte lock held.
4591 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4592 unsigned long addr, pte_t ptent, union mc_target *target)
4594 struct page *page = NULL;
4595 enum mc_target_type ret = MC_TARGET_NONE;
4596 swp_entry_t ent = { .val = 0 };
4598 if (pte_present(ptent))
4599 page = mc_handle_present_pte(vma, addr, ptent);
4600 else if (is_swap_pte(ptent))
4601 page = mc_handle_swap_pte(vma, ptent, &ent);
4602 else if (pte_none(ptent))
4603 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4605 if (!page && !ent.val)
4609 * Do only loose check w/o serialization.
4610 * mem_cgroup_move_account() checks the page is valid or
4611 * not under LRU exclusion.
4613 if (page->mem_cgroup == mc.from) {
4614 ret = MC_TARGET_PAGE;
4616 target->page = page;
4618 if (!ret || !target)
4621 /* There is a swap entry and a page doesn't exist or isn't charged */
4622 if (ent.val && !ret &&
4623 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4624 ret = MC_TARGET_SWAP;
4631 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4633 * We don't consider swapping or file mapped pages because THP does not
4634 * support them for now.
4635 * Caller should make sure that pmd_trans_huge(pmd) is true.
4637 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4638 unsigned long addr, pmd_t pmd, union mc_target *target)
4640 struct page *page = NULL;
4641 enum mc_target_type ret = MC_TARGET_NONE;
4643 page = pmd_page(pmd);
4644 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4645 if (!(mc.flags & MOVE_ANON))
4647 if (page->mem_cgroup == mc.from) {
4648 ret = MC_TARGET_PAGE;
4651 target->page = page;
4657 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4658 unsigned long addr, pmd_t pmd, union mc_target *target)
4660 return MC_TARGET_NONE;
4664 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4665 unsigned long addr, unsigned long end,
4666 struct mm_walk *walk)
4668 struct vm_area_struct *vma = walk->vma;
4672 ptl = pmd_trans_huge_lock(pmd, vma);
4674 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4675 mc.precharge += HPAGE_PMD_NR;
4680 if (pmd_trans_unstable(pmd))
4682 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4683 for (; addr != end; pte++, addr += PAGE_SIZE)
4684 if (get_mctgt_type(vma, addr, *pte, NULL))
4685 mc.precharge++; /* increment precharge temporarily */
4686 pte_unmap_unlock(pte - 1, ptl);
4692 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4694 unsigned long precharge;
4696 struct mm_walk mem_cgroup_count_precharge_walk = {
4697 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4700 down_read(&mm->mmap_sem);
4701 walk_page_range(0, mm->highest_vm_end,
4702 &mem_cgroup_count_precharge_walk);
4703 up_read(&mm->mmap_sem);
4705 precharge = mc.precharge;
4711 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4713 unsigned long precharge = mem_cgroup_count_precharge(mm);
4715 VM_BUG_ON(mc.moving_task);
4716 mc.moving_task = current;
4717 return mem_cgroup_do_precharge(precharge);
4720 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4721 static void __mem_cgroup_clear_mc(void)
4723 struct mem_cgroup *from = mc.from;
4724 struct mem_cgroup *to = mc.to;
4726 /* we must uncharge all the leftover precharges from mc.to */
4728 cancel_charge(mc.to, mc.precharge);
4732 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4733 * we must uncharge here.
4735 if (mc.moved_charge) {
4736 cancel_charge(mc.from, mc.moved_charge);
4737 mc.moved_charge = 0;
4739 /* we must fixup refcnts and charges */
4740 if (mc.moved_swap) {
4741 /* uncharge swap account from the old cgroup */
4742 if (!mem_cgroup_is_root(mc.from))
4743 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4745 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4748 * we charged both to->memory and to->memsw, so we
4749 * should uncharge to->memory.
4751 if (!mem_cgroup_is_root(mc.to))
4752 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4754 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4755 css_put_many(&mc.to->css, mc.moved_swap);
4759 memcg_oom_recover(from);
4760 memcg_oom_recover(to);
4761 wake_up_all(&mc.waitq);
4764 static void mem_cgroup_clear_mc(void)
4766 struct mm_struct *mm = mc.mm;
4769 * we must clear moving_task before waking up waiters at the end of
4772 mc.moving_task = NULL;
4773 __mem_cgroup_clear_mc();
4774 spin_lock(&mc.lock);
4778 spin_unlock(&mc.lock);
4783 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4785 struct cgroup_subsys_state *css;
4786 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4787 struct mem_cgroup *from;
4788 struct task_struct *leader, *p;
4789 struct mm_struct *mm;
4790 unsigned long move_flags;
4793 /* charge immigration isn't supported on the default hierarchy */
4794 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4798 * Multi-process migrations only happen on the default hierarchy
4799 * where charge immigration is not used. Perform charge
4800 * immigration if @tset contains a leader and whine if there are
4804 cgroup_taskset_for_each_leader(leader, css, tset) {
4807 memcg = mem_cgroup_from_css(css);
4813 * We are now commited to this value whatever it is. Changes in this
4814 * tunable will only affect upcoming migrations, not the current one.
4815 * So we need to save it, and keep it going.
4817 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4821 from = mem_cgroup_from_task(p);
4823 VM_BUG_ON(from == memcg);
4825 mm = get_task_mm(p);
4828 /* We move charges only when we move a owner of the mm */
4829 if (mm->owner == p) {
4832 VM_BUG_ON(mc.precharge);
4833 VM_BUG_ON(mc.moved_charge);
4834 VM_BUG_ON(mc.moved_swap);
4836 spin_lock(&mc.lock);
4840 mc.flags = move_flags;
4841 spin_unlock(&mc.lock);
4842 /* We set mc.moving_task later */
4844 ret = mem_cgroup_precharge_mc(mm);
4846 mem_cgroup_clear_mc();
4853 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4856 mem_cgroup_clear_mc();
4859 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4860 unsigned long addr, unsigned long end,
4861 struct mm_walk *walk)
4864 struct vm_area_struct *vma = walk->vma;
4867 enum mc_target_type target_type;
4868 union mc_target target;
4871 ptl = pmd_trans_huge_lock(pmd, vma);
4873 if (mc.precharge < HPAGE_PMD_NR) {
4877 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4878 if (target_type == MC_TARGET_PAGE) {
4880 if (!isolate_lru_page(page)) {
4881 if (!mem_cgroup_move_account(page, true,
4883 mc.precharge -= HPAGE_PMD_NR;
4884 mc.moved_charge += HPAGE_PMD_NR;
4886 putback_lru_page(page);
4894 if (pmd_trans_unstable(pmd))
4897 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4898 for (; addr != end; addr += PAGE_SIZE) {
4899 pte_t ptent = *(pte++);
4905 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4906 case MC_TARGET_PAGE:
4909 * We can have a part of the split pmd here. Moving it
4910 * can be done but it would be too convoluted so simply
4911 * ignore such a partial THP and keep it in original
4912 * memcg. There should be somebody mapping the head.
4914 if (PageTransCompound(page))
4916 if (isolate_lru_page(page))
4918 if (!mem_cgroup_move_account(page, false,
4921 /* we uncharge from mc.from later. */
4924 putback_lru_page(page);
4925 put: /* get_mctgt_type() gets the page */
4928 case MC_TARGET_SWAP:
4930 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4932 /* we fixup refcnts and charges later. */
4940 pte_unmap_unlock(pte - 1, ptl);
4945 * We have consumed all precharges we got in can_attach().
4946 * We try charge one by one, but don't do any additional
4947 * charges to mc.to if we have failed in charge once in attach()
4950 ret = mem_cgroup_do_precharge(1);
4958 static void mem_cgroup_move_charge(void)
4960 struct mm_walk mem_cgroup_move_charge_walk = {
4961 .pmd_entry = mem_cgroup_move_charge_pte_range,
4965 lru_add_drain_all();
4967 * Signal lock_page_memcg() to take the memcg's move_lock
4968 * while we're moving its pages to another memcg. Then wait
4969 * for already started RCU-only updates to finish.
4971 atomic_inc(&mc.from->moving_account);
4974 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
4976 * Someone who are holding the mmap_sem might be waiting in
4977 * waitq. So we cancel all extra charges, wake up all waiters,
4978 * and retry. Because we cancel precharges, we might not be able
4979 * to move enough charges, but moving charge is a best-effort
4980 * feature anyway, so it wouldn't be a big problem.
4982 __mem_cgroup_clear_mc();
4987 * When we have consumed all precharges and failed in doing
4988 * additional charge, the page walk just aborts.
4990 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
4992 up_read(&mc.mm->mmap_sem);
4993 atomic_dec(&mc.from->moving_account);
4996 static void mem_cgroup_move_task(void)
4999 mem_cgroup_move_charge();
5000 mem_cgroup_clear_mc();
5003 #else /* !CONFIG_MMU */
5004 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5008 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5011 static void mem_cgroup_move_task(void)
5017 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5018 * to verify whether we're attached to the default hierarchy on each mount
5021 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5024 * use_hierarchy is forced on the default hierarchy. cgroup core
5025 * guarantees that @root doesn't have any children, so turning it
5026 * on for the root memcg is enough.
5028 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5029 root_mem_cgroup->use_hierarchy = true;
5031 root_mem_cgroup->use_hierarchy = false;
5034 static u64 memory_current_read(struct cgroup_subsys_state *css,
5037 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5039 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5042 static int memory_low_show(struct seq_file *m, void *v)
5044 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5045 unsigned long low = READ_ONCE(memcg->low);
5047 if (low == PAGE_COUNTER_MAX)
5048 seq_puts(m, "max\n");
5050 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5055 static ssize_t memory_low_write(struct kernfs_open_file *of,
5056 char *buf, size_t nbytes, loff_t off)
5058 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5062 buf = strstrip(buf);
5063 err = page_counter_memparse(buf, "max", &low);
5072 static int memory_high_show(struct seq_file *m, void *v)
5074 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5075 unsigned long high = READ_ONCE(memcg->high);
5077 if (high == PAGE_COUNTER_MAX)
5078 seq_puts(m, "max\n");
5080 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5085 static ssize_t memory_high_write(struct kernfs_open_file *of,
5086 char *buf, size_t nbytes, loff_t off)
5088 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5089 unsigned long nr_pages;
5093 buf = strstrip(buf);
5094 err = page_counter_memparse(buf, "max", &high);
5100 nr_pages = page_counter_read(&memcg->memory);
5101 if (nr_pages > high)
5102 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5105 memcg_wb_domain_size_changed(memcg);
5109 static int memory_max_show(struct seq_file *m, void *v)
5111 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5112 unsigned long max = READ_ONCE(memcg->memory.limit);
5114 if (max == PAGE_COUNTER_MAX)
5115 seq_puts(m, "max\n");
5117 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5122 static ssize_t memory_max_write(struct kernfs_open_file *of,
5123 char *buf, size_t nbytes, loff_t off)
5125 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5126 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5127 bool drained = false;
5131 buf = strstrip(buf);
5132 err = page_counter_memparse(buf, "max", &max);
5136 xchg(&memcg->memory.limit, max);
5139 unsigned long nr_pages = page_counter_read(&memcg->memory);
5141 if (nr_pages <= max)
5144 if (signal_pending(current)) {
5150 drain_all_stock(memcg);
5156 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5162 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5163 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5167 memcg_wb_domain_size_changed(memcg);
5171 static int memory_events_show(struct seq_file *m, void *v)
5173 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5175 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5176 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5177 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5178 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5183 static int memory_stat_show(struct seq_file *m, void *v)
5185 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5186 unsigned long stat[MEMCG_NR_STAT];
5187 unsigned long events[MEMCG_NR_EVENTS];
5191 * Provide statistics on the state of the memory subsystem as
5192 * well as cumulative event counters that show past behavior.
5194 * This list is ordered following a combination of these gradients:
5195 * 1) generic big picture -> specifics and details
5196 * 2) reflecting userspace activity -> reflecting kernel heuristics
5198 * Current memory state:
5201 tree_stat(memcg, stat);
5202 tree_events(memcg, events);
5204 seq_printf(m, "anon %llu\n",
5205 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5206 seq_printf(m, "file %llu\n",
5207 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5208 seq_printf(m, "kernel_stack %llu\n",
5209 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5210 seq_printf(m, "slab %llu\n",
5211 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5212 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5213 seq_printf(m, "sock %llu\n",
5214 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5216 seq_printf(m, "shmem %llu\n",
5217 (u64)stat[MEM_CGROUP_STAT_SHMEM] * PAGE_SIZE);
5218 seq_printf(m, "file_mapped %llu\n",
5219 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5220 seq_printf(m, "file_dirty %llu\n",
5221 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5222 seq_printf(m, "file_writeback %llu\n",
5223 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5225 for (i = 0; i < NR_LRU_LISTS; i++) {
5226 struct mem_cgroup *mi;
5227 unsigned long val = 0;
5229 for_each_mem_cgroup_tree(mi, memcg)
5230 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5231 seq_printf(m, "%s %llu\n",
5232 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5235 seq_printf(m, "slab_reclaimable %llu\n",
5236 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5237 seq_printf(m, "slab_unreclaimable %llu\n",
5238 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5240 /* Accumulated memory events */
5242 seq_printf(m, "pgfault %lu\n",
5243 events[MEM_CGROUP_EVENTS_PGFAULT]);
5244 seq_printf(m, "pgmajfault %lu\n",
5245 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5250 static struct cftype memory_files[] = {
5253 .flags = CFTYPE_NOT_ON_ROOT,
5254 .read_u64 = memory_current_read,
5258 .flags = CFTYPE_NOT_ON_ROOT,
5259 .seq_show = memory_low_show,
5260 .write = memory_low_write,
5264 .flags = CFTYPE_NOT_ON_ROOT,
5265 .seq_show = memory_high_show,
5266 .write = memory_high_write,
5270 .flags = CFTYPE_NOT_ON_ROOT,
5271 .seq_show = memory_max_show,
5272 .write = memory_max_write,
5276 .flags = CFTYPE_NOT_ON_ROOT,
5277 .file_offset = offsetof(struct mem_cgroup, events_file),
5278 .seq_show = memory_events_show,
5282 .flags = CFTYPE_NOT_ON_ROOT,
5283 .seq_show = memory_stat_show,
5288 struct cgroup_subsys memory_cgrp_subsys = {
5289 .css_alloc = mem_cgroup_css_alloc,
5290 .css_online = mem_cgroup_css_online,
5291 .css_offline = mem_cgroup_css_offline,
5292 .css_released = mem_cgroup_css_released,
5293 .css_free = mem_cgroup_css_free,
5294 .css_reset = mem_cgroup_css_reset,
5295 .can_attach = mem_cgroup_can_attach,
5296 .cancel_attach = mem_cgroup_cancel_attach,
5297 .post_attach = mem_cgroup_move_task,
5298 .bind = mem_cgroup_bind,
5299 .dfl_cftypes = memory_files,
5300 .legacy_cftypes = mem_cgroup_legacy_files,
5305 * mem_cgroup_low - check if memory consumption is below the normal range
5306 * @root: the highest ancestor to consider
5307 * @memcg: the memory cgroup to check
5309 * Returns %true if memory consumption of @memcg, and that of all
5310 * configurable ancestors up to @root, is below the normal range.
5312 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5314 if (mem_cgroup_disabled())
5318 * The toplevel group doesn't have a configurable range, so
5319 * it's never low when looked at directly, and it is not
5320 * considered an ancestor when assessing the hierarchy.
5323 if (memcg == root_mem_cgroup)
5326 if (page_counter_read(&memcg->memory) >= memcg->low)
5329 while (memcg != root) {
5330 memcg = parent_mem_cgroup(memcg);
5332 if (memcg == root_mem_cgroup)
5335 if (page_counter_read(&memcg->memory) >= memcg->low)
5342 * mem_cgroup_try_charge - try charging a page
5343 * @page: page to charge
5344 * @mm: mm context of the victim
5345 * @gfp_mask: reclaim mode
5346 * @memcgp: charged memcg return
5347 * @compound: charge the page as compound or small page
5349 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5350 * pages according to @gfp_mask if necessary.
5352 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5353 * Otherwise, an error code is returned.
5355 * After page->mapping has been set up, the caller must finalize the
5356 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5357 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5359 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5360 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5363 struct mem_cgroup *memcg = NULL;
5364 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5367 if (mem_cgroup_disabled())
5370 if (PageSwapCache(page)) {
5372 * Every swap fault against a single page tries to charge the
5373 * page, bail as early as possible. shmem_unuse() encounters
5374 * already charged pages, too. The USED bit is protected by
5375 * the page lock, which serializes swap cache removal, which
5376 * in turn serializes uncharging.
5378 VM_BUG_ON_PAGE(!PageLocked(page), page);
5379 if (page->mem_cgroup)
5382 if (do_swap_account) {
5383 swp_entry_t ent = { .val = page_private(page), };
5384 unsigned short id = lookup_swap_cgroup_id(ent);
5387 memcg = mem_cgroup_from_id(id);
5388 if (memcg && !css_tryget_online(&memcg->css))
5395 memcg = get_mem_cgroup_from_mm(mm);
5397 ret = try_charge(memcg, gfp_mask, nr_pages);
5399 css_put(&memcg->css);
5406 * mem_cgroup_commit_charge - commit a page charge
5407 * @page: page to charge
5408 * @memcg: memcg to charge the page to
5409 * @lrucare: page might be on LRU already
5410 * @compound: charge the page as compound or small page
5412 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5413 * after page->mapping has been set up. This must happen atomically
5414 * as part of the page instantiation, i.e. under the page table lock
5415 * for anonymous pages, under the page lock for page and swap cache.
5417 * In addition, the page must not be on the LRU during the commit, to
5418 * prevent racing with task migration. If it might be, use @lrucare.
5420 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5422 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5423 bool lrucare, bool compound)
5425 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5427 VM_BUG_ON_PAGE(!page->mapping, page);
5428 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5430 if (mem_cgroup_disabled())
5433 * Swap faults will attempt to charge the same page multiple
5434 * times. But reuse_swap_page() might have removed the page
5435 * from swapcache already, so we can't check PageSwapCache().
5440 commit_charge(page, memcg, lrucare);
5442 local_irq_disable();
5443 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5444 memcg_check_events(memcg, page);
5447 if (do_memsw_account() && PageSwapCache(page)) {
5448 swp_entry_t entry = { .val = page_private(page) };
5450 * The swap entry might not get freed for a long time,
5451 * let's not wait for it. The page already received a
5452 * memory+swap charge, drop the swap entry duplicate.
5454 mem_cgroup_uncharge_swap(entry);
5459 * mem_cgroup_cancel_charge - cancel a page charge
5460 * @page: page to charge
5461 * @memcg: memcg to charge the page to
5462 * @compound: charge the page as compound or small page
5464 * Cancel a charge transaction started by mem_cgroup_try_charge().
5466 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5469 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5471 if (mem_cgroup_disabled())
5474 * Swap faults will attempt to charge the same page multiple
5475 * times. But reuse_swap_page() might have removed the page
5476 * from swapcache already, so we can't check PageSwapCache().
5481 cancel_charge(memcg, nr_pages);
5484 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5485 unsigned long nr_anon, unsigned long nr_file,
5486 unsigned long nr_kmem, unsigned long nr_huge,
5487 unsigned long nr_shmem, struct page *dummy_page)
5489 unsigned long nr_pages = nr_anon + nr_file + nr_kmem;
5490 unsigned long flags;
5492 if (!mem_cgroup_is_root(memcg)) {
5493 page_counter_uncharge(&memcg->memory, nr_pages);
5494 if (do_memsw_account())
5495 page_counter_uncharge(&memcg->memsw, nr_pages);
5496 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && nr_kmem)
5497 page_counter_uncharge(&memcg->kmem, nr_kmem);
5498 memcg_oom_recover(memcg);
5501 local_irq_save(flags);
5502 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5503 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5504 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5505 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_SHMEM], nr_shmem);
5506 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5507 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5508 memcg_check_events(memcg, dummy_page);
5509 local_irq_restore(flags);
5511 if (!mem_cgroup_is_root(memcg))
5512 css_put_many(&memcg->css, nr_pages);
5515 static void uncharge_list(struct list_head *page_list)
5517 struct mem_cgroup *memcg = NULL;
5518 unsigned long nr_shmem = 0;
5519 unsigned long nr_anon = 0;
5520 unsigned long nr_file = 0;
5521 unsigned long nr_huge = 0;
5522 unsigned long nr_kmem = 0;
5523 unsigned long pgpgout = 0;
5524 struct list_head *next;
5528 * Note that the list can be a single page->lru; hence the
5529 * do-while loop instead of a simple list_for_each_entry().
5531 next = page_list->next;
5533 page = list_entry(next, struct page, lru);
5534 next = page->lru.next;
5536 VM_BUG_ON_PAGE(PageLRU(page), page);
5537 VM_BUG_ON_PAGE(page_count(page), page);
5539 if (!page->mem_cgroup)
5543 * Nobody should be changing or seriously looking at
5544 * page->mem_cgroup at this point, we have fully
5545 * exclusive access to the page.
5548 if (memcg != page->mem_cgroup) {
5550 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5551 nr_kmem, nr_huge, nr_shmem, page);
5552 pgpgout = nr_anon = nr_file = nr_kmem = 0;
5553 nr_huge = nr_shmem = 0;
5555 memcg = page->mem_cgroup;
5558 if (!PageKmemcg(page)) {
5559 unsigned int nr_pages = 1;
5561 if (PageTransHuge(page)) {
5562 nr_pages <<= compound_order(page);
5563 nr_huge += nr_pages;
5566 nr_anon += nr_pages;
5568 nr_file += nr_pages;
5569 if (PageSwapBacked(page))
5570 nr_shmem += nr_pages;
5574 nr_kmem += 1 << compound_order(page);
5575 __ClearPageKmemcg(page);
5578 page->mem_cgroup = NULL;
5579 } while (next != page_list);
5582 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5583 nr_kmem, nr_huge, nr_shmem, page);
5587 * mem_cgroup_uncharge - uncharge a page
5588 * @page: page to uncharge
5590 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5591 * mem_cgroup_commit_charge().
5593 void mem_cgroup_uncharge(struct page *page)
5595 if (mem_cgroup_disabled())
5598 /* Don't touch page->lru of any random page, pre-check: */
5599 if (!page->mem_cgroup)
5602 INIT_LIST_HEAD(&page->lru);
5603 uncharge_list(&page->lru);
5607 * mem_cgroup_uncharge_list - uncharge a list of page
5608 * @page_list: list of pages to uncharge
5610 * Uncharge a list of pages previously charged with
5611 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5613 void mem_cgroup_uncharge_list(struct list_head *page_list)
5615 if (mem_cgroup_disabled())
5618 if (!list_empty(page_list))
5619 uncharge_list(page_list);
5623 * mem_cgroup_migrate - charge a page's replacement
5624 * @oldpage: currently circulating page
5625 * @newpage: replacement page
5627 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5628 * be uncharged upon free.
5630 * Both pages must be locked, @newpage->mapping must be set up.
5632 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5634 struct mem_cgroup *memcg;
5635 unsigned int nr_pages;
5637 unsigned long flags;
5639 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5640 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5641 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5642 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5645 if (mem_cgroup_disabled())
5648 /* Page cache replacement: new page already charged? */
5649 if (newpage->mem_cgroup)
5652 /* Swapcache readahead pages can get replaced before being charged */
5653 memcg = oldpage->mem_cgroup;
5657 /* Force-charge the new page. The old one will be freed soon */
5658 compound = PageTransHuge(newpage);
5659 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5661 page_counter_charge(&memcg->memory, nr_pages);
5662 if (do_memsw_account())
5663 page_counter_charge(&memcg->memsw, nr_pages);
5664 css_get_many(&memcg->css, nr_pages);
5666 commit_charge(newpage, memcg, false);
5668 local_irq_save(flags);
5669 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5670 memcg_check_events(memcg, newpage);
5671 local_irq_restore(flags);
5674 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5675 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5677 void mem_cgroup_sk_alloc(struct sock *sk)
5679 struct mem_cgroup *memcg;
5681 if (!mem_cgroup_sockets_enabled)
5685 * Socket cloning can throw us here with sk_memcg already
5686 * filled. It won't however, necessarily happen from
5687 * process context. So the test for root memcg given
5688 * the current task's memcg won't help us in this case.
5690 * Respecting the original socket's memcg is a better
5691 * decision in this case.
5694 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5695 css_get(&sk->sk_memcg->css);
5700 memcg = mem_cgroup_from_task(current);
5701 if (memcg == root_mem_cgroup)
5703 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5705 if (css_tryget_online(&memcg->css))
5706 sk->sk_memcg = memcg;
5711 void mem_cgroup_sk_free(struct sock *sk)
5714 css_put(&sk->sk_memcg->css);
5718 * mem_cgroup_charge_skmem - charge socket memory
5719 * @memcg: memcg to charge
5720 * @nr_pages: number of pages to charge
5722 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5723 * @memcg's configured limit, %false if the charge had to be forced.
5725 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5727 gfp_t gfp_mask = GFP_KERNEL;
5729 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5730 struct page_counter *fail;
5732 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5733 memcg->tcpmem_pressure = 0;
5736 page_counter_charge(&memcg->tcpmem, nr_pages);
5737 memcg->tcpmem_pressure = 1;
5741 /* Don't block in the packet receive path */
5743 gfp_mask = GFP_NOWAIT;
5745 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5747 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5750 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5755 * mem_cgroup_uncharge_skmem - uncharge socket memory
5756 * @memcg - memcg to uncharge
5757 * @nr_pages - number of pages to uncharge
5759 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5761 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5762 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5766 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5768 page_counter_uncharge(&memcg->memory, nr_pages);
5769 css_put_many(&memcg->css, nr_pages);
5772 static int __init cgroup_memory(char *s)
5776 while ((token = strsep(&s, ",")) != NULL) {
5779 if (!strcmp(token, "nosocket"))
5780 cgroup_memory_nosocket = true;
5781 if (!strcmp(token, "nokmem"))
5782 cgroup_memory_nokmem = true;
5786 __setup("cgroup.memory=", cgroup_memory);
5789 * subsys_initcall() for memory controller.
5791 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
5792 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
5793 * basically everything that doesn't depend on a specific mem_cgroup structure
5794 * should be initialized from here.
5796 static int __init mem_cgroup_init(void)
5802 * Kmem cache creation is mostly done with the slab_mutex held,
5803 * so use a workqueue with limited concurrency to avoid stalling
5804 * all worker threads in case lots of cgroups are created and
5805 * destroyed simultaneously.
5807 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
5808 BUG_ON(!memcg_kmem_cache_wq);
5811 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
5812 memcg_hotplug_cpu_dead);
5814 for_each_possible_cpu(cpu)
5815 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5818 for_each_node(node) {
5819 struct mem_cgroup_tree_per_node *rtpn;
5821 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5822 node_online(node) ? node : NUMA_NO_NODE);
5824 rtpn->rb_root = RB_ROOT;
5825 spin_lock_init(&rtpn->lock);
5826 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5831 subsys_initcall(mem_cgroup_init);
5833 #ifdef CONFIG_MEMCG_SWAP
5834 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
5836 while (!atomic_inc_not_zero(&memcg->id.ref)) {
5838 * The root cgroup cannot be destroyed, so it's refcount must
5841 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
5845 memcg = parent_mem_cgroup(memcg);
5847 memcg = root_mem_cgroup;
5853 * mem_cgroup_swapout - transfer a memsw charge to swap
5854 * @page: page whose memsw charge to transfer
5855 * @entry: swap entry to move the charge to
5857 * Transfer the memsw charge of @page to @entry.
5859 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5861 struct mem_cgroup *memcg, *swap_memcg;
5862 unsigned short oldid;
5864 VM_BUG_ON_PAGE(PageLRU(page), page);
5865 VM_BUG_ON_PAGE(page_count(page), page);
5867 if (!do_memsw_account())
5870 memcg = page->mem_cgroup;
5872 /* Readahead page, never charged */
5877 * In case the memcg owning these pages has been offlined and doesn't
5878 * have an ID allocated to it anymore, charge the closest online
5879 * ancestor for the swap instead and transfer the memory+swap charge.
5881 swap_memcg = mem_cgroup_id_get_online(memcg);
5882 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg));
5883 VM_BUG_ON_PAGE(oldid, page);
5884 mem_cgroup_swap_statistics(swap_memcg, true);
5886 page->mem_cgroup = NULL;
5888 if (!mem_cgroup_is_root(memcg))
5889 page_counter_uncharge(&memcg->memory, 1);
5891 if (memcg != swap_memcg) {
5892 if (!mem_cgroup_is_root(swap_memcg))
5893 page_counter_charge(&swap_memcg->memsw, 1);
5894 page_counter_uncharge(&memcg->memsw, 1);
5898 * Interrupts should be disabled here because the caller holds the
5899 * mapping->tree_lock lock which is taken with interrupts-off. It is
5900 * important here to have the interrupts disabled because it is the
5901 * only synchronisation we have for udpating the per-CPU variables.
5903 VM_BUG_ON(!irqs_disabled());
5904 mem_cgroup_charge_statistics(memcg, page, false, -1);
5905 memcg_check_events(memcg, page);
5907 if (!mem_cgroup_is_root(memcg))
5908 css_put(&memcg->css);
5912 * mem_cgroup_try_charge_swap - try charging a swap entry
5913 * @page: page being added to swap
5914 * @entry: swap entry to charge
5916 * Try to charge @entry to the memcg that @page belongs to.
5918 * Returns 0 on success, -ENOMEM on failure.
5920 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5922 struct mem_cgroup *memcg;
5923 struct page_counter *counter;
5924 unsigned short oldid;
5926 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5929 memcg = page->mem_cgroup;
5931 /* Readahead page, never charged */
5935 memcg = mem_cgroup_id_get_online(memcg);
5937 if (!mem_cgroup_is_root(memcg) &&
5938 !page_counter_try_charge(&memcg->swap, 1, &counter)) {
5939 mem_cgroup_id_put(memcg);
5943 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5944 VM_BUG_ON_PAGE(oldid, page);
5945 mem_cgroup_swap_statistics(memcg, true);
5951 * mem_cgroup_uncharge_swap - uncharge a swap entry
5952 * @entry: swap entry to uncharge
5954 * Drop the swap charge associated with @entry.
5956 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5958 struct mem_cgroup *memcg;
5961 if (!do_swap_account)
5964 id = swap_cgroup_record(entry, 0);
5966 memcg = mem_cgroup_from_id(id);
5968 if (!mem_cgroup_is_root(memcg)) {
5969 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5970 page_counter_uncharge(&memcg->swap, 1);
5972 page_counter_uncharge(&memcg->memsw, 1);
5974 mem_cgroup_swap_statistics(memcg, false);
5975 mem_cgroup_id_put(memcg);
5980 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
5982 long nr_swap_pages = get_nr_swap_pages();
5984 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5985 return nr_swap_pages;
5986 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5987 nr_swap_pages = min_t(long, nr_swap_pages,
5988 READ_ONCE(memcg->swap.limit) -
5989 page_counter_read(&memcg->swap));
5990 return nr_swap_pages;
5993 bool mem_cgroup_swap_full(struct page *page)
5995 struct mem_cgroup *memcg;
5997 VM_BUG_ON_PAGE(!PageLocked(page), page);
6001 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6004 memcg = page->mem_cgroup;
6008 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6009 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
6015 /* for remember boot option*/
6016 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6017 static int really_do_swap_account __initdata = 1;
6019 static int really_do_swap_account __initdata;
6022 static int __init enable_swap_account(char *s)
6024 if (!strcmp(s, "1"))
6025 really_do_swap_account = 1;
6026 else if (!strcmp(s, "0"))
6027 really_do_swap_account = 0;
6030 __setup("swapaccount=", enable_swap_account);
6032 static u64 swap_current_read(struct cgroup_subsys_state *css,
6035 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6037 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6040 static int swap_max_show(struct seq_file *m, void *v)
6042 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6043 unsigned long max = READ_ONCE(memcg->swap.limit);
6045 if (max == PAGE_COUNTER_MAX)
6046 seq_puts(m, "max\n");
6048 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6053 static ssize_t swap_max_write(struct kernfs_open_file *of,
6054 char *buf, size_t nbytes, loff_t off)
6056 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6060 buf = strstrip(buf);
6061 err = page_counter_memparse(buf, "max", &max);
6065 mutex_lock(&memcg_limit_mutex);
6066 err = page_counter_limit(&memcg->swap, max);
6067 mutex_unlock(&memcg_limit_mutex);
6074 static struct cftype swap_files[] = {
6076 .name = "swap.current",
6077 .flags = CFTYPE_NOT_ON_ROOT,
6078 .read_u64 = swap_current_read,
6082 .flags = CFTYPE_NOT_ON_ROOT,
6083 .seq_show = swap_max_show,
6084 .write = swap_max_write,
6089 static struct cftype memsw_cgroup_files[] = {
6091 .name = "memsw.usage_in_bytes",
6092 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6093 .read_u64 = mem_cgroup_read_u64,
6096 .name = "memsw.max_usage_in_bytes",
6097 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6098 .write = mem_cgroup_reset,
6099 .read_u64 = mem_cgroup_read_u64,
6102 .name = "memsw.limit_in_bytes",
6103 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6104 .write = mem_cgroup_write,
6105 .read_u64 = mem_cgroup_read_u64,
6108 .name = "memsw.failcnt",
6109 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6110 .write = mem_cgroup_reset,
6111 .read_u64 = mem_cgroup_read_u64,
6113 { }, /* terminate */
6116 static int __init mem_cgroup_swap_init(void)
6118 if (!mem_cgroup_disabled() && really_do_swap_account) {
6119 do_swap_account = 1;
6120 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6122 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6123 memsw_cgroup_files));
6127 subsys_initcall(mem_cgroup_swap_init);
6129 #endif /* CONFIG_MEMCG_SWAP */