1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 struct mem_cgroup *root_mem_cgroup __read_mostly;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 static const char * const mem_cgroup_stat_names[] = {
111 static const char * const mem_cgroup_events_names[] = {
118 static const char * const mem_cgroup_lru_names[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_node {
136 struct rb_root rb_root;
140 struct mem_cgroup_tree {
141 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
144 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
147 struct mem_cgroup_eventfd_list {
148 struct list_head list;
149 struct eventfd_ctx *eventfd;
153 * cgroup_event represents events which userspace want to receive.
155 struct mem_cgroup_event {
157 * memcg which the event belongs to.
159 struct mem_cgroup *memcg;
161 * eventfd to signal userspace about the event.
163 struct eventfd_ctx *eventfd;
165 * Each of these stored in a list by the cgroup.
167 struct list_head list;
169 * register_event() callback will be used to add new userspace
170 * waiter for changes related to this event. Use eventfd_signal()
171 * on eventfd to send notification to userspace.
173 int (*register_event)(struct mem_cgroup *memcg,
174 struct eventfd_ctx *eventfd, const char *args);
176 * unregister_event() callback will be called when userspace closes
177 * the eventfd or on cgroup removing. This callback must be set,
178 * if you want provide notification functionality.
180 void (*unregister_event)(struct mem_cgroup *memcg,
181 struct eventfd_ctx *eventfd);
183 * All fields below needed to unregister event when
184 * userspace closes eventfd.
187 wait_queue_head_t *wqh;
189 struct work_struct remove;
192 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
193 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
195 /* Stuffs for move charges at task migration. */
197 * Types of charges to be moved.
199 #define MOVE_ANON 0x1U
200 #define MOVE_FILE 0x2U
201 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
203 /* "mc" and its members are protected by cgroup_mutex */
204 static struct move_charge_struct {
205 spinlock_t lock; /* for from, to */
206 struct mm_struct *mm;
207 struct mem_cgroup *from;
208 struct mem_cgroup *to;
210 unsigned long precharge;
211 unsigned long moved_charge;
212 unsigned long moved_swap;
213 struct task_struct *moving_task; /* a task moving charges */
214 wait_queue_head_t waitq; /* a waitq for other context */
216 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
217 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
221 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
222 * limit reclaim to prevent infinite loops, if they ever occur.
224 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
225 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
228 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
229 MEM_CGROUP_CHARGE_TYPE_ANON,
230 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
231 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
235 /* for encoding cft->private value on file */
244 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
245 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
246 #define MEMFILE_ATTR(val) ((val) & 0xffff)
247 /* Used for OOM nofiier */
248 #define OOM_CONTROL (0)
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
254 memcg = root_mem_cgroup;
255 return &memcg->vmpressure;
258 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
260 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
263 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
265 return (memcg == root_mem_cgroup);
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida);
281 int memcg_nr_cache_ids;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key);
320 #endif /* !CONFIG_SLOB */
323 * mem_cgroup_css_from_page - css of the memcg associated with a page
324 * @page: page of interest
326 * If memcg is bound to the default hierarchy, css of the memcg associated
327 * with @page is returned. The returned css remains associated with @page
328 * until it is released.
330 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
333 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
335 struct mem_cgroup *memcg;
337 memcg = page->mem_cgroup;
339 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
340 memcg = root_mem_cgroup;
346 * page_cgroup_ino - return inode number of the memcg a page is charged to
349 * Look up the closest online ancestor of the memory cgroup @page is charged to
350 * and return its inode number or 0 if @page is not charged to any cgroup. It
351 * is safe to call this function without holding a reference to @page.
353 * Note, this function is inherently racy, because there is nothing to prevent
354 * the cgroup inode from getting torn down and potentially reallocated a moment
355 * after page_cgroup_ino() returns, so it only should be used by callers that
356 * do not care (such as procfs interfaces).
358 ino_t page_cgroup_ino(struct page *page)
360 struct mem_cgroup *memcg;
361 unsigned long ino = 0;
364 memcg = READ_ONCE(page->mem_cgroup);
365 while (memcg && !(memcg->css.flags & CSS_ONLINE))
366 memcg = parent_mem_cgroup(memcg);
368 ino = cgroup_ino(memcg->css.cgroup);
373 static struct mem_cgroup_per_node *
374 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
376 int nid = page_to_nid(page);
378 return memcg->nodeinfo[nid];
381 static struct mem_cgroup_tree_per_node *
382 soft_limit_tree_node(int nid)
384 return soft_limit_tree.rb_tree_per_node[nid];
387 static struct mem_cgroup_tree_per_node *
388 soft_limit_tree_from_page(struct page *page)
390 int nid = page_to_nid(page);
392 return soft_limit_tree.rb_tree_per_node[nid];
395 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
396 struct mem_cgroup_tree_per_node *mctz,
397 unsigned long new_usage_in_excess)
399 struct rb_node **p = &mctz->rb_root.rb_node;
400 struct rb_node *parent = NULL;
401 struct mem_cgroup_per_node *mz_node;
406 mz->usage_in_excess = new_usage_in_excess;
407 if (!mz->usage_in_excess)
411 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
413 if (mz->usage_in_excess < mz_node->usage_in_excess)
416 * We can't avoid mem cgroups that are over their soft
417 * limit by the same amount
419 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
422 rb_link_node(&mz->tree_node, parent, p);
423 rb_insert_color(&mz->tree_node, &mctz->rb_root);
427 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
428 struct mem_cgroup_tree_per_node *mctz)
432 rb_erase(&mz->tree_node, &mctz->rb_root);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
437 struct mem_cgroup_tree_per_node *mctz)
441 spin_lock_irqsave(&mctz->lock, flags);
442 __mem_cgroup_remove_exceeded(mz, mctz);
443 spin_unlock_irqrestore(&mctz->lock, flags);
446 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
448 unsigned long nr_pages = page_counter_read(&memcg->memory);
449 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
450 unsigned long excess = 0;
452 if (nr_pages > soft_limit)
453 excess = nr_pages - soft_limit;
458 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
460 unsigned long excess;
461 struct mem_cgroup_per_node *mz;
462 struct mem_cgroup_tree_per_node *mctz;
464 mctz = soft_limit_tree_from_page(page);
468 * Necessary to update all ancestors when hierarchy is used.
469 * because their event counter is not touched.
471 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
472 mz = mem_cgroup_page_nodeinfo(memcg, page);
473 excess = soft_limit_excess(memcg);
475 * We have to update the tree if mz is on RB-tree or
476 * mem is over its softlimit.
478 if (excess || mz->on_tree) {
481 spin_lock_irqsave(&mctz->lock, flags);
482 /* if on-tree, remove it */
484 __mem_cgroup_remove_exceeded(mz, mctz);
486 * Insert again. mz->usage_in_excess will be updated.
487 * If excess is 0, no tree ops.
489 __mem_cgroup_insert_exceeded(mz, mctz, excess);
490 spin_unlock_irqrestore(&mctz->lock, flags);
495 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
497 struct mem_cgroup_tree_per_node *mctz;
498 struct mem_cgroup_per_node *mz;
502 mz = mem_cgroup_nodeinfo(memcg, nid);
503 mctz = soft_limit_tree_node(nid);
505 mem_cgroup_remove_exceeded(mz, mctz);
509 static struct mem_cgroup_per_node *
510 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
512 struct rb_node *rightmost = NULL;
513 struct mem_cgroup_per_node *mz;
517 rightmost = rb_last(&mctz->rb_root);
519 goto done; /* Nothing to reclaim from */
521 mz = rb_entry(rightmost, struct mem_cgroup_per_node, tree_node);
523 * Remove the node now but someone else can add it back,
524 * we will to add it back at the end of reclaim to its correct
525 * position in the tree.
527 __mem_cgroup_remove_exceeded(mz, mctz);
528 if (!soft_limit_excess(mz->memcg) ||
529 !css_tryget_online(&mz->memcg->css))
535 static struct mem_cgroup_per_node *
536 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
538 struct mem_cgroup_per_node *mz;
540 spin_lock_irq(&mctz->lock);
541 mz = __mem_cgroup_largest_soft_limit_node(mctz);
542 spin_unlock_irq(&mctz->lock);
547 * Return page count for single (non recursive) @memcg.
549 * Implementation Note: reading percpu statistics for memcg.
551 * Both of vmstat[] and percpu_counter has threshold and do periodic
552 * synchronization to implement "quick" read. There are trade-off between
553 * reading cost and precision of value. Then, we may have a chance to implement
554 * a periodic synchronization of counter in memcg's counter.
556 * But this _read() function is used for user interface now. The user accounts
557 * memory usage by memory cgroup and he _always_ requires exact value because
558 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
559 * have to visit all online cpus and make sum. So, for now, unnecessary
560 * synchronization is not implemented. (just implemented for cpu hotplug)
562 * If there are kernel internal actions which can make use of some not-exact
563 * value, and reading all cpu value can be performance bottleneck in some
564 * common workload, threshold and synchronization as vmstat[] should be
568 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
573 /* Per-cpu values can be negative, use a signed accumulator */
574 for_each_possible_cpu(cpu)
575 val += per_cpu(memcg->stat->count[idx], cpu);
577 * Summing races with updates, so val may be negative. Avoid exposing
578 * transient negative values.
585 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
586 enum mem_cgroup_events_index idx)
588 unsigned long val = 0;
591 for_each_possible_cpu(cpu)
592 val += per_cpu(memcg->stat->events[idx], cpu);
596 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
598 bool compound, int nr_pages)
601 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
602 * counted as CACHE even if it's on ANON LRU.
605 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
608 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
612 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
613 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
617 /* pagein of a big page is an event. So, ignore page size */
619 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
621 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
622 nr_pages = -nr_pages; /* for event */
625 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
628 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
629 int nid, unsigned int lru_mask)
631 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
632 unsigned long nr = 0;
635 VM_BUG_ON((unsigned)nid >= nr_node_ids);
638 if (!(BIT(lru) & lru_mask))
640 nr += mem_cgroup_get_lru_size(lruvec, lru);
645 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
646 unsigned int lru_mask)
648 unsigned long nr = 0;
651 for_each_node_state(nid, N_MEMORY)
652 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
656 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
657 enum mem_cgroup_events_target target)
659 unsigned long val, next;
661 val = __this_cpu_read(memcg->stat->nr_page_events);
662 next = __this_cpu_read(memcg->stat->targets[target]);
663 /* from time_after() in jiffies.h */
664 if ((long)next - (long)val < 0) {
666 case MEM_CGROUP_TARGET_THRESH:
667 next = val + THRESHOLDS_EVENTS_TARGET;
669 case MEM_CGROUP_TARGET_SOFTLIMIT:
670 next = val + SOFTLIMIT_EVENTS_TARGET;
672 case MEM_CGROUP_TARGET_NUMAINFO:
673 next = val + NUMAINFO_EVENTS_TARGET;
678 __this_cpu_write(memcg->stat->targets[target], next);
685 * Check events in order.
688 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
690 /* threshold event is triggered in finer grain than soft limit */
691 if (unlikely(mem_cgroup_event_ratelimit(memcg,
692 MEM_CGROUP_TARGET_THRESH))) {
694 bool do_numainfo __maybe_unused;
696 do_softlimit = mem_cgroup_event_ratelimit(memcg,
697 MEM_CGROUP_TARGET_SOFTLIMIT);
699 do_numainfo = mem_cgroup_event_ratelimit(memcg,
700 MEM_CGROUP_TARGET_NUMAINFO);
702 mem_cgroup_threshold(memcg);
703 if (unlikely(do_softlimit))
704 mem_cgroup_update_tree(memcg, page);
706 if (unlikely(do_numainfo))
707 atomic_inc(&memcg->numainfo_events);
712 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
715 * mm_update_next_owner() may clear mm->owner to NULL
716 * if it races with swapoff, page migration, etc.
717 * So this can be called with p == NULL.
722 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
724 EXPORT_SYMBOL(mem_cgroup_from_task);
726 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
728 struct mem_cgroup *memcg = NULL;
733 * Page cache insertions can happen withou an
734 * actual mm context, e.g. during disk probing
735 * on boot, loopback IO, acct() writes etc.
738 memcg = root_mem_cgroup;
740 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
741 if (unlikely(!memcg))
742 memcg = root_mem_cgroup;
744 } while (!css_tryget_online(&memcg->css));
750 * mem_cgroup_iter - iterate over memory cgroup hierarchy
751 * @root: hierarchy root
752 * @prev: previously returned memcg, NULL on first invocation
753 * @reclaim: cookie for shared reclaim walks, NULL for full walks
755 * Returns references to children of the hierarchy below @root, or
756 * @root itself, or %NULL after a full round-trip.
758 * Caller must pass the return value in @prev on subsequent
759 * invocations for reference counting, or use mem_cgroup_iter_break()
760 * to cancel a hierarchy walk before the round-trip is complete.
762 * Reclaimers can specify a zone and a priority level in @reclaim to
763 * divide up the memcgs in the hierarchy among all concurrent
764 * reclaimers operating on the same zone and priority.
766 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
767 struct mem_cgroup *prev,
768 struct mem_cgroup_reclaim_cookie *reclaim)
770 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
771 struct cgroup_subsys_state *css = NULL;
772 struct mem_cgroup *memcg = NULL;
773 struct mem_cgroup *pos = NULL;
775 if (mem_cgroup_disabled())
779 root = root_mem_cgroup;
781 if (prev && !reclaim)
784 if (!root->use_hierarchy && root != root_mem_cgroup) {
793 struct mem_cgroup_per_node *mz;
795 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
796 iter = &mz->iter[reclaim->priority];
798 if (prev && reclaim->generation != iter->generation)
802 pos = READ_ONCE(iter->position);
803 if (!pos || css_tryget(&pos->css))
806 * css reference reached zero, so iter->position will
807 * be cleared by ->css_released. However, we should not
808 * rely on this happening soon, because ->css_released
809 * is called from a work queue, and by busy-waiting we
810 * might block it. So we clear iter->position right
813 (void)cmpxchg(&iter->position, pos, NULL);
821 css = css_next_descendant_pre(css, &root->css);
824 * Reclaimers share the hierarchy walk, and a
825 * new one might jump in right at the end of
826 * the hierarchy - make sure they see at least
827 * one group and restart from the beginning.
835 * Verify the css and acquire a reference. The root
836 * is provided by the caller, so we know it's alive
837 * and kicking, and don't take an extra reference.
839 memcg = mem_cgroup_from_css(css);
841 if (css == &root->css)
852 * The position could have already been updated by a competing
853 * thread, so check that the value hasn't changed since we read
854 * it to avoid reclaiming from the same cgroup twice.
856 (void)cmpxchg(&iter->position, pos, memcg);
864 reclaim->generation = iter->generation;
870 if (prev && prev != root)
877 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
878 * @root: hierarchy root
879 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
881 void mem_cgroup_iter_break(struct mem_cgroup *root,
882 struct mem_cgroup *prev)
885 root = root_mem_cgroup;
886 if (prev && prev != root)
890 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
891 struct mem_cgroup *dead_memcg)
893 struct mem_cgroup_reclaim_iter *iter;
894 struct mem_cgroup_per_node *mz;
899 mz = mem_cgroup_nodeinfo(from, nid);
900 for (i = 0; i <= DEF_PRIORITY; i++) {
902 cmpxchg(&iter->position,
908 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
910 struct mem_cgroup *memcg = dead_memcg;
911 struct mem_cgroup *last;
914 __invalidate_reclaim_iterators(memcg, dead_memcg);
916 } while ((memcg = parent_mem_cgroup(memcg)));
919 * When cgruop1 non-hierarchy mode is used,
920 * parent_mem_cgroup() does not walk all the way up to the
921 * cgroup root (root_mem_cgroup). So we have to handle
922 * dead_memcg from cgroup root separately.
924 if (last != root_mem_cgroup)
925 __invalidate_reclaim_iterators(root_mem_cgroup,
930 * Iteration constructs for visiting all cgroups (under a tree). If
931 * loops are exited prematurely (break), mem_cgroup_iter_break() must
932 * be used for reference counting.
934 #define for_each_mem_cgroup_tree(iter, root) \
935 for (iter = mem_cgroup_iter(root, NULL, NULL); \
937 iter = mem_cgroup_iter(root, iter, NULL))
939 #define for_each_mem_cgroup(iter) \
940 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
942 iter = mem_cgroup_iter(NULL, iter, NULL))
945 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
946 * @memcg: hierarchy root
947 * @fn: function to call for each task
948 * @arg: argument passed to @fn
950 * This function iterates over tasks attached to @memcg or to any of its
951 * descendants and calls @fn for each task. If @fn returns a non-zero
952 * value, the function breaks the iteration loop and returns the value.
953 * Otherwise, it will iterate over all tasks and return 0.
955 * This function must not be called for the root memory cgroup.
957 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
958 int (*fn)(struct task_struct *, void *), void *arg)
960 struct mem_cgroup *iter;
963 BUG_ON(memcg == root_mem_cgroup);
965 for_each_mem_cgroup_tree(iter, memcg) {
966 struct css_task_iter it;
967 struct task_struct *task;
969 css_task_iter_start(&iter->css, &it);
970 while (!ret && (task = css_task_iter_next(&it)))
972 css_task_iter_end(&it);
974 mem_cgroup_iter_break(memcg, iter);
982 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
984 * @zone: zone of the page
986 * This function is only safe when following the LRU page isolation
987 * and putback protocol: the LRU lock must be held, and the page must
988 * either be PageLRU() or the caller must have isolated/allocated it.
990 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
992 struct mem_cgroup_per_node *mz;
993 struct mem_cgroup *memcg;
994 struct lruvec *lruvec;
996 if (mem_cgroup_disabled()) {
997 lruvec = &pgdat->lruvec;
1001 memcg = page->mem_cgroup;
1003 * Swapcache readahead pages are added to the LRU - and
1004 * possibly migrated - before they are charged.
1007 memcg = root_mem_cgroup;
1009 mz = mem_cgroup_page_nodeinfo(memcg, page);
1010 lruvec = &mz->lruvec;
1013 * Since a node can be onlined after the mem_cgroup was created,
1014 * we have to be prepared to initialize lruvec->zone here;
1015 * and if offlined then reonlined, we need to reinitialize it.
1017 if (unlikely(lruvec->pgdat != pgdat))
1018 lruvec->pgdat = pgdat;
1023 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1024 * @lruvec: mem_cgroup per zone lru vector
1025 * @lru: index of lru list the page is sitting on
1026 * @zid: zone id of the accounted pages
1027 * @nr_pages: positive when adding or negative when removing
1029 * This function must be called under lru_lock, just before a page is added
1030 * to or just after a page is removed from an lru list (that ordering being
1031 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1033 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1034 int zid, int nr_pages)
1036 struct mem_cgroup_per_node *mz;
1037 unsigned long *lru_size;
1040 if (mem_cgroup_disabled())
1043 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1044 lru_size = &mz->lru_zone_size[zid][lru];
1047 *lru_size += nr_pages;
1050 if (WARN_ONCE(size < 0,
1051 "%s(%p, %d, %d): lru_size %ld\n",
1052 __func__, lruvec, lru, nr_pages, size)) {
1058 *lru_size += nr_pages;
1061 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1063 struct mem_cgroup *task_memcg;
1064 struct task_struct *p;
1067 p = find_lock_task_mm(task);
1069 task_memcg = get_mem_cgroup_from_mm(p->mm);
1073 * All threads may have already detached their mm's, but the oom
1074 * killer still needs to detect if they have already been oom
1075 * killed to prevent needlessly killing additional tasks.
1078 task_memcg = mem_cgroup_from_task(task);
1079 css_get(&task_memcg->css);
1082 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1083 css_put(&task_memcg->css);
1088 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1089 * @memcg: the memory cgroup
1091 * Returns the maximum amount of memory @mem can be charged with, in
1094 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1096 unsigned long margin = 0;
1097 unsigned long count;
1098 unsigned long limit;
1100 count = page_counter_read(&memcg->memory);
1101 limit = READ_ONCE(memcg->memory.limit);
1103 margin = limit - count;
1105 if (do_memsw_account()) {
1106 count = page_counter_read(&memcg->memsw);
1107 limit = READ_ONCE(memcg->memsw.limit);
1109 margin = min(margin, limit - count);
1118 * A routine for checking "mem" is under move_account() or not.
1120 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1121 * moving cgroups. This is for waiting at high-memory pressure
1124 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1126 struct mem_cgroup *from;
1127 struct mem_cgroup *to;
1130 * Unlike task_move routines, we access mc.to, mc.from not under
1131 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1133 spin_lock(&mc.lock);
1139 ret = mem_cgroup_is_descendant(from, memcg) ||
1140 mem_cgroup_is_descendant(to, memcg);
1142 spin_unlock(&mc.lock);
1146 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1148 if (mc.moving_task && current != mc.moving_task) {
1149 if (mem_cgroup_under_move(memcg)) {
1151 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1152 /* moving charge context might have finished. */
1155 finish_wait(&mc.waitq, &wait);
1162 #define K(x) ((x) << (PAGE_SHIFT-10))
1164 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1165 * @memcg: The memory cgroup that went over limit
1166 * @p: Task that is going to be killed
1168 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1171 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1173 struct mem_cgroup *iter;
1179 pr_info("Task in ");
1180 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1181 pr_cont(" killed as a result of limit of ");
1183 pr_info("Memory limit reached of cgroup ");
1186 pr_cont_cgroup_path(memcg->css.cgroup);
1191 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1192 K((u64)page_counter_read(&memcg->memory)),
1193 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1194 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1195 K((u64)page_counter_read(&memcg->memsw)),
1196 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1197 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1198 K((u64)page_counter_read(&memcg->kmem)),
1199 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1201 for_each_mem_cgroup_tree(iter, memcg) {
1202 pr_info("Memory cgroup stats for ");
1203 pr_cont_cgroup_path(iter->css.cgroup);
1206 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1207 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1209 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1210 K(mem_cgroup_read_stat(iter, i)));
1213 for (i = 0; i < NR_LRU_LISTS; i++)
1214 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1215 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1222 * This function returns the number of memcg under hierarchy tree. Returns
1223 * 1(self count) if no children.
1225 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1228 struct mem_cgroup *iter;
1230 for_each_mem_cgroup_tree(iter, memcg)
1236 * Return the memory (and swap, if configured) limit for a memcg.
1238 unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1240 unsigned long limit;
1242 limit = memcg->memory.limit;
1243 if (mem_cgroup_swappiness(memcg)) {
1244 unsigned long memsw_limit;
1245 unsigned long swap_limit;
1247 memsw_limit = memcg->memsw.limit;
1248 swap_limit = memcg->swap.limit;
1249 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1250 limit = min(limit + swap_limit, memsw_limit);
1255 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1258 struct oom_control oc = {
1262 .gfp_mask = gfp_mask,
1267 mutex_lock(&oom_lock);
1268 ret = out_of_memory(&oc);
1269 mutex_unlock(&oom_lock);
1273 #if MAX_NUMNODES > 1
1276 * test_mem_cgroup_node_reclaimable
1277 * @memcg: the target memcg
1278 * @nid: the node ID to be checked.
1279 * @noswap : specify true here if the user wants flle only information.
1281 * This function returns whether the specified memcg contains any
1282 * reclaimable pages on a node. Returns true if there are any reclaimable
1283 * pages in the node.
1285 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1286 int nid, bool noswap)
1288 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1290 if (noswap || !total_swap_pages)
1292 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1299 * Always updating the nodemask is not very good - even if we have an empty
1300 * list or the wrong list here, we can start from some node and traverse all
1301 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1304 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1308 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1309 * pagein/pageout changes since the last update.
1311 if (!atomic_read(&memcg->numainfo_events))
1313 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1316 /* make a nodemask where this memcg uses memory from */
1317 memcg->scan_nodes = node_states[N_MEMORY];
1319 for_each_node_mask(nid, node_states[N_MEMORY]) {
1321 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1322 node_clear(nid, memcg->scan_nodes);
1325 atomic_set(&memcg->numainfo_events, 0);
1326 atomic_set(&memcg->numainfo_updating, 0);
1330 * Selecting a node where we start reclaim from. Because what we need is just
1331 * reducing usage counter, start from anywhere is O,K. Considering
1332 * memory reclaim from current node, there are pros. and cons.
1334 * Freeing memory from current node means freeing memory from a node which
1335 * we'll use or we've used. So, it may make LRU bad. And if several threads
1336 * hit limits, it will see a contention on a node. But freeing from remote
1337 * node means more costs for memory reclaim because of memory latency.
1339 * Now, we use round-robin. Better algorithm is welcomed.
1341 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1345 mem_cgroup_may_update_nodemask(memcg);
1346 node = memcg->last_scanned_node;
1348 node = next_node_in(node, memcg->scan_nodes);
1350 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1351 * last time it really checked all the LRUs due to rate limiting.
1352 * Fallback to the current node in that case for simplicity.
1354 if (unlikely(node == MAX_NUMNODES))
1355 node = numa_node_id();
1357 memcg->last_scanned_node = node;
1361 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1367 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1370 unsigned long *total_scanned)
1372 struct mem_cgroup *victim = NULL;
1375 unsigned long excess;
1376 unsigned long nr_scanned;
1377 struct mem_cgroup_reclaim_cookie reclaim = {
1382 excess = soft_limit_excess(root_memcg);
1385 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1390 * If we have not been able to reclaim
1391 * anything, it might because there are
1392 * no reclaimable pages under this hierarchy
1397 * We want to do more targeted reclaim.
1398 * excess >> 2 is not to excessive so as to
1399 * reclaim too much, nor too less that we keep
1400 * coming back to reclaim from this cgroup
1402 if (total >= (excess >> 2) ||
1403 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1408 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1409 pgdat, &nr_scanned);
1410 *total_scanned += nr_scanned;
1411 if (!soft_limit_excess(root_memcg))
1414 mem_cgroup_iter_break(root_memcg, victim);
1418 #ifdef CONFIG_LOCKDEP
1419 static struct lockdep_map memcg_oom_lock_dep_map = {
1420 .name = "memcg_oom_lock",
1424 static DEFINE_SPINLOCK(memcg_oom_lock);
1427 * Check OOM-Killer is already running under our hierarchy.
1428 * If someone is running, return false.
1430 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1432 struct mem_cgroup *iter, *failed = NULL;
1434 spin_lock(&memcg_oom_lock);
1436 for_each_mem_cgroup_tree(iter, memcg) {
1437 if (iter->oom_lock) {
1439 * this subtree of our hierarchy is already locked
1440 * so we cannot give a lock.
1443 mem_cgroup_iter_break(memcg, iter);
1446 iter->oom_lock = true;
1451 * OK, we failed to lock the whole subtree so we have
1452 * to clean up what we set up to the failing subtree
1454 for_each_mem_cgroup_tree(iter, memcg) {
1455 if (iter == failed) {
1456 mem_cgroup_iter_break(memcg, iter);
1459 iter->oom_lock = false;
1462 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1464 spin_unlock(&memcg_oom_lock);
1469 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1471 struct mem_cgroup *iter;
1473 spin_lock(&memcg_oom_lock);
1474 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1475 for_each_mem_cgroup_tree(iter, memcg)
1476 iter->oom_lock = false;
1477 spin_unlock(&memcg_oom_lock);
1480 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1482 struct mem_cgroup *iter;
1484 spin_lock(&memcg_oom_lock);
1485 for_each_mem_cgroup_tree(iter, memcg)
1487 spin_unlock(&memcg_oom_lock);
1490 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1492 struct mem_cgroup *iter;
1495 * When a new child is created while the hierarchy is under oom,
1496 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1498 spin_lock(&memcg_oom_lock);
1499 for_each_mem_cgroup_tree(iter, memcg)
1500 if (iter->under_oom > 0)
1502 spin_unlock(&memcg_oom_lock);
1505 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1507 struct oom_wait_info {
1508 struct mem_cgroup *memcg;
1512 static int memcg_oom_wake_function(wait_queue_t *wait,
1513 unsigned mode, int sync, void *arg)
1515 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1516 struct mem_cgroup *oom_wait_memcg;
1517 struct oom_wait_info *oom_wait_info;
1519 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1520 oom_wait_memcg = oom_wait_info->memcg;
1522 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1523 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1525 return autoremove_wake_function(wait, mode, sync, arg);
1528 static void memcg_oom_recover(struct mem_cgroup *memcg)
1531 * For the following lockless ->under_oom test, the only required
1532 * guarantee is that it must see the state asserted by an OOM when
1533 * this function is called as a result of userland actions
1534 * triggered by the notification of the OOM. This is trivially
1535 * achieved by invoking mem_cgroup_mark_under_oom() before
1536 * triggering notification.
1538 if (memcg && memcg->under_oom)
1539 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1542 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1544 if (!current->memcg_may_oom)
1547 * We are in the middle of the charge context here, so we
1548 * don't want to block when potentially sitting on a callstack
1549 * that holds all kinds of filesystem and mm locks.
1551 * Also, the caller may handle a failed allocation gracefully
1552 * (like optional page cache readahead) and so an OOM killer
1553 * invocation might not even be necessary.
1555 * That's why we don't do anything here except remember the
1556 * OOM context and then deal with it at the end of the page
1557 * fault when the stack is unwound, the locks are released,
1558 * and when we know whether the fault was overall successful.
1560 css_get(&memcg->css);
1561 current->memcg_in_oom = memcg;
1562 current->memcg_oom_gfp_mask = mask;
1563 current->memcg_oom_order = order;
1567 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1568 * @handle: actually kill/wait or just clean up the OOM state
1570 * This has to be called at the end of a page fault if the memcg OOM
1571 * handler was enabled.
1573 * Memcg supports userspace OOM handling where failed allocations must
1574 * sleep on a waitqueue until the userspace task resolves the
1575 * situation. Sleeping directly in the charge context with all kinds
1576 * of locks held is not a good idea, instead we remember an OOM state
1577 * in the task and mem_cgroup_oom_synchronize() has to be called at
1578 * the end of the page fault to complete the OOM handling.
1580 * Returns %true if an ongoing memcg OOM situation was detected and
1581 * completed, %false otherwise.
1583 bool mem_cgroup_oom_synchronize(bool handle)
1585 struct mem_cgroup *memcg = current->memcg_in_oom;
1586 struct oom_wait_info owait;
1589 /* OOM is global, do not handle */
1596 owait.memcg = memcg;
1597 owait.wait.flags = 0;
1598 owait.wait.func = memcg_oom_wake_function;
1599 owait.wait.private = current;
1600 INIT_LIST_HEAD(&owait.wait.task_list);
1602 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1603 mem_cgroup_mark_under_oom(memcg);
1605 locked = mem_cgroup_oom_trylock(memcg);
1608 mem_cgroup_oom_notify(memcg);
1610 if (locked && !memcg->oom_kill_disable) {
1611 mem_cgroup_unmark_under_oom(memcg);
1612 finish_wait(&memcg_oom_waitq, &owait.wait);
1613 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1614 current->memcg_oom_order);
1617 mem_cgroup_unmark_under_oom(memcg);
1618 finish_wait(&memcg_oom_waitq, &owait.wait);
1622 mem_cgroup_oom_unlock(memcg);
1624 * There is no guarantee that an OOM-lock contender
1625 * sees the wakeups triggered by the OOM kill
1626 * uncharges. Wake any sleepers explicitely.
1628 memcg_oom_recover(memcg);
1631 current->memcg_in_oom = NULL;
1632 css_put(&memcg->css);
1637 * lock_page_memcg - lock a page->mem_cgroup binding
1640 * This function protects unlocked LRU pages from being moved to
1641 * another cgroup and stabilizes their page->mem_cgroup binding.
1643 void lock_page_memcg(struct page *page)
1645 struct mem_cgroup *memcg;
1646 unsigned long flags;
1649 * The RCU lock is held throughout the transaction. The fast
1650 * path can get away without acquiring the memcg->move_lock
1651 * because page moving starts with an RCU grace period.
1655 if (mem_cgroup_disabled())
1658 memcg = page->mem_cgroup;
1659 if (unlikely(!memcg))
1662 if (atomic_read(&memcg->moving_account) <= 0)
1665 spin_lock_irqsave(&memcg->move_lock, flags);
1666 if (memcg != page->mem_cgroup) {
1667 spin_unlock_irqrestore(&memcg->move_lock, flags);
1672 * When charge migration first begins, we can have locked and
1673 * unlocked page stat updates happening concurrently. Track
1674 * the task who has the lock for unlock_page_memcg().
1676 memcg->move_lock_task = current;
1677 memcg->move_lock_flags = flags;
1681 EXPORT_SYMBOL(lock_page_memcg);
1684 * unlock_page_memcg - unlock a page->mem_cgroup binding
1687 void unlock_page_memcg(struct page *page)
1689 struct mem_cgroup *memcg = page->mem_cgroup;
1691 if (memcg && memcg->move_lock_task == current) {
1692 unsigned long flags = memcg->move_lock_flags;
1694 memcg->move_lock_task = NULL;
1695 memcg->move_lock_flags = 0;
1697 spin_unlock_irqrestore(&memcg->move_lock, flags);
1702 EXPORT_SYMBOL(unlock_page_memcg);
1705 * size of first charge trial. "32" comes from vmscan.c's magic value.
1706 * TODO: maybe necessary to use big numbers in big irons.
1708 #define CHARGE_BATCH 32U
1709 struct memcg_stock_pcp {
1710 struct mem_cgroup *cached; /* this never be root cgroup */
1711 unsigned int nr_pages;
1712 struct work_struct work;
1713 unsigned long flags;
1714 #define FLUSHING_CACHED_CHARGE 0
1716 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1717 static DEFINE_MUTEX(percpu_charge_mutex);
1720 * consume_stock: Try to consume stocked charge on this cpu.
1721 * @memcg: memcg to consume from.
1722 * @nr_pages: how many pages to charge.
1724 * The charges will only happen if @memcg matches the current cpu's memcg
1725 * stock, and at least @nr_pages are available in that stock. Failure to
1726 * service an allocation will refill the stock.
1728 * returns true if successful, false otherwise.
1730 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1732 struct memcg_stock_pcp *stock;
1733 unsigned long flags;
1736 if (nr_pages > CHARGE_BATCH)
1739 local_irq_save(flags);
1741 stock = this_cpu_ptr(&memcg_stock);
1742 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1743 stock->nr_pages -= nr_pages;
1747 local_irq_restore(flags);
1753 * Returns stocks cached in percpu and reset cached information.
1755 static void drain_stock(struct memcg_stock_pcp *stock)
1757 struct mem_cgroup *old = stock->cached;
1759 if (stock->nr_pages) {
1760 page_counter_uncharge(&old->memory, stock->nr_pages);
1761 if (do_memsw_account())
1762 page_counter_uncharge(&old->memsw, stock->nr_pages);
1763 css_put_many(&old->css, stock->nr_pages);
1764 stock->nr_pages = 0;
1766 stock->cached = NULL;
1769 static void drain_local_stock(struct work_struct *dummy)
1771 struct memcg_stock_pcp *stock;
1772 unsigned long flags;
1774 local_irq_save(flags);
1776 stock = this_cpu_ptr(&memcg_stock);
1778 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1780 local_irq_restore(flags);
1784 * Cache charges(val) to local per_cpu area.
1785 * This will be consumed by consume_stock() function, later.
1787 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1789 struct memcg_stock_pcp *stock;
1790 unsigned long flags;
1792 local_irq_save(flags);
1794 stock = this_cpu_ptr(&memcg_stock);
1795 if (stock->cached != memcg) { /* reset if necessary */
1797 stock->cached = memcg;
1799 stock->nr_pages += nr_pages;
1801 local_irq_restore(flags);
1805 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1806 * of the hierarchy under it.
1808 static void drain_all_stock(struct mem_cgroup *root_memcg)
1812 /* If someone's already draining, avoid adding running more workers. */
1813 if (!mutex_trylock(&percpu_charge_mutex))
1815 /* Notify other cpus that system-wide "drain" is running */
1818 for_each_online_cpu(cpu) {
1819 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1820 struct mem_cgroup *memcg;
1822 memcg = stock->cached;
1823 if (!memcg || !stock->nr_pages)
1825 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1827 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1829 drain_local_stock(&stock->work);
1831 schedule_work_on(cpu, &stock->work);
1836 mutex_unlock(&percpu_charge_mutex);
1839 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1840 unsigned long action,
1843 int cpu = (unsigned long)hcpu;
1844 struct memcg_stock_pcp *stock;
1846 if (action == CPU_ONLINE)
1849 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1852 stock = &per_cpu(memcg_stock, cpu);
1857 static void reclaim_high(struct mem_cgroup *memcg,
1858 unsigned int nr_pages,
1862 if (page_counter_read(&memcg->memory) <= memcg->high)
1864 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1865 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1866 } while ((memcg = parent_mem_cgroup(memcg)));
1869 static void high_work_func(struct work_struct *work)
1871 struct mem_cgroup *memcg;
1873 memcg = container_of(work, struct mem_cgroup, high_work);
1874 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1878 * Scheduled by try_charge() to be executed from the userland return path
1879 * and reclaims memory over the high limit.
1881 void mem_cgroup_handle_over_high(void)
1883 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1884 struct mem_cgroup *memcg;
1886 if (likely(!nr_pages))
1889 memcg = get_mem_cgroup_from_mm(current->mm);
1890 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1891 css_put(&memcg->css);
1892 current->memcg_nr_pages_over_high = 0;
1895 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1896 unsigned int nr_pages)
1898 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1899 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1900 struct mem_cgroup *mem_over_limit;
1901 struct page_counter *counter;
1902 unsigned long nr_reclaimed;
1903 bool may_swap = true;
1904 bool drained = false;
1906 if (mem_cgroup_is_root(memcg))
1909 if (consume_stock(memcg, nr_pages))
1912 if (!do_memsw_account() ||
1913 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1914 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1916 if (do_memsw_account())
1917 page_counter_uncharge(&memcg->memsw, batch);
1918 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1920 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1924 if (batch > nr_pages) {
1930 * Unlike in global OOM situations, memcg is not in a physical
1931 * memory shortage. Allow dying and OOM-killed tasks to
1932 * bypass the last charges so that they can exit quickly and
1933 * free their memory.
1935 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1936 fatal_signal_pending(current) ||
1937 current->flags & PF_EXITING))
1941 * Prevent unbounded recursion when reclaim operations need to
1942 * allocate memory. This might exceed the limits temporarily,
1943 * but we prefer facilitating memory reclaim and getting back
1944 * under the limit over triggering OOM kills in these cases.
1946 if (unlikely(current->flags & PF_MEMALLOC))
1949 if (unlikely(task_in_memcg_oom(current)))
1952 if (!gfpflags_allow_blocking(gfp_mask))
1955 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1957 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1958 gfp_mask, may_swap);
1960 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1964 drain_all_stock(mem_over_limit);
1969 if (gfp_mask & __GFP_NORETRY)
1972 * Even though the limit is exceeded at this point, reclaim
1973 * may have been able to free some pages. Retry the charge
1974 * before killing the task.
1976 * Only for regular pages, though: huge pages are rather
1977 * unlikely to succeed so close to the limit, and we fall back
1978 * to regular pages anyway in case of failure.
1980 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
1983 * At task move, charge accounts can be doubly counted. So, it's
1984 * better to wait until the end of task_move if something is going on.
1986 if (mem_cgroup_wait_acct_move(mem_over_limit))
1992 if (gfp_mask & __GFP_NOFAIL)
1995 if (fatal_signal_pending(current))
1998 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2000 mem_cgroup_oom(mem_over_limit, gfp_mask,
2001 get_order(nr_pages * PAGE_SIZE));
2003 if (!(gfp_mask & __GFP_NOFAIL))
2007 * The allocation either can't fail or will lead to more memory
2008 * being freed very soon. Allow memory usage go over the limit
2009 * temporarily by force charging it.
2011 page_counter_charge(&memcg->memory, nr_pages);
2012 if (do_memsw_account())
2013 page_counter_charge(&memcg->memsw, nr_pages);
2014 css_get_many(&memcg->css, nr_pages);
2019 css_get_many(&memcg->css, batch);
2020 if (batch > nr_pages)
2021 refill_stock(memcg, batch - nr_pages);
2024 * If the hierarchy is above the normal consumption range, schedule
2025 * reclaim on returning to userland. We can perform reclaim here
2026 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2027 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2028 * not recorded as it most likely matches current's and won't
2029 * change in the meantime. As high limit is checked again before
2030 * reclaim, the cost of mismatch is negligible.
2033 if (page_counter_read(&memcg->memory) > memcg->high) {
2034 /* Don't bother a random interrupted task */
2035 if (in_interrupt()) {
2036 schedule_work(&memcg->high_work);
2039 current->memcg_nr_pages_over_high += batch;
2040 set_notify_resume(current);
2043 } while ((memcg = parent_mem_cgroup(memcg)));
2048 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2050 if (mem_cgroup_is_root(memcg))
2053 page_counter_uncharge(&memcg->memory, nr_pages);
2054 if (do_memsw_account())
2055 page_counter_uncharge(&memcg->memsw, nr_pages);
2057 css_put_many(&memcg->css, nr_pages);
2060 static void lock_page_lru(struct page *page, int *isolated)
2062 struct zone *zone = page_zone(page);
2064 spin_lock_irq(zone_lru_lock(zone));
2065 if (PageLRU(page)) {
2066 struct lruvec *lruvec;
2068 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2070 del_page_from_lru_list(page, lruvec, page_lru(page));
2076 static void unlock_page_lru(struct page *page, int isolated)
2078 struct zone *zone = page_zone(page);
2081 struct lruvec *lruvec;
2083 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2084 VM_BUG_ON_PAGE(PageLRU(page), page);
2086 add_page_to_lru_list(page, lruvec, page_lru(page));
2088 spin_unlock_irq(zone_lru_lock(zone));
2091 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2096 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2099 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2100 * may already be on some other mem_cgroup's LRU. Take care of it.
2103 lock_page_lru(page, &isolated);
2106 * Nobody should be changing or seriously looking at
2107 * page->mem_cgroup at this point:
2109 * - the page is uncharged
2111 * - the page is off-LRU
2113 * - an anonymous fault has exclusive page access, except for
2114 * a locked page table
2116 * - a page cache insertion, a swapin fault, or a migration
2117 * have the page locked
2119 page->mem_cgroup = memcg;
2122 unlock_page_lru(page, isolated);
2126 static int memcg_alloc_cache_id(void)
2131 id = ida_simple_get(&memcg_cache_ida,
2132 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2136 if (id < memcg_nr_cache_ids)
2140 * There's no space for the new id in memcg_caches arrays,
2141 * so we have to grow them.
2143 down_write(&memcg_cache_ids_sem);
2145 size = 2 * (id + 1);
2146 if (size < MEMCG_CACHES_MIN_SIZE)
2147 size = MEMCG_CACHES_MIN_SIZE;
2148 else if (size > MEMCG_CACHES_MAX_SIZE)
2149 size = MEMCG_CACHES_MAX_SIZE;
2151 err = memcg_update_all_caches(size);
2153 err = memcg_update_all_list_lrus(size);
2155 memcg_nr_cache_ids = size;
2157 up_write(&memcg_cache_ids_sem);
2160 ida_simple_remove(&memcg_cache_ida, id);
2166 static void memcg_free_cache_id(int id)
2168 ida_simple_remove(&memcg_cache_ida, id);
2171 struct memcg_kmem_cache_create_work {
2172 struct mem_cgroup *memcg;
2173 struct kmem_cache *cachep;
2174 struct work_struct work;
2177 static struct workqueue_struct *memcg_kmem_cache_create_wq;
2179 static void memcg_kmem_cache_create_func(struct work_struct *w)
2181 struct memcg_kmem_cache_create_work *cw =
2182 container_of(w, struct memcg_kmem_cache_create_work, work);
2183 struct mem_cgroup *memcg = cw->memcg;
2184 struct kmem_cache *cachep = cw->cachep;
2186 memcg_create_kmem_cache(memcg, cachep);
2188 css_put(&memcg->css);
2193 * Enqueue the creation of a per-memcg kmem_cache.
2195 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2196 struct kmem_cache *cachep)
2198 struct memcg_kmem_cache_create_work *cw;
2200 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2204 css_get(&memcg->css);
2207 cw->cachep = cachep;
2208 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2210 queue_work(memcg_kmem_cache_create_wq, &cw->work);
2213 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2214 struct kmem_cache *cachep)
2217 * We need to stop accounting when we kmalloc, because if the
2218 * corresponding kmalloc cache is not yet created, the first allocation
2219 * in __memcg_schedule_kmem_cache_create will recurse.
2221 * However, it is better to enclose the whole function. Depending on
2222 * the debugging options enabled, INIT_WORK(), for instance, can
2223 * trigger an allocation. This too, will make us recurse. Because at
2224 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2225 * the safest choice is to do it like this, wrapping the whole function.
2227 current->memcg_kmem_skip_account = 1;
2228 __memcg_schedule_kmem_cache_create(memcg, cachep);
2229 current->memcg_kmem_skip_account = 0;
2232 static inline bool memcg_kmem_bypass(void)
2234 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2240 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2241 * @cachep: the original global kmem cache
2243 * Return the kmem_cache we're supposed to use for a slab allocation.
2244 * We try to use the current memcg's version of the cache.
2246 * If the cache does not exist yet, if we are the first user of it, we
2247 * create it asynchronously in a workqueue and let the current allocation
2248 * go through with the original cache.
2250 * This function takes a reference to the cache it returns to assure it
2251 * won't get destroyed while we are working with it. Once the caller is
2252 * done with it, memcg_kmem_put_cache() must be called to release the
2255 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2257 struct mem_cgroup *memcg;
2258 struct kmem_cache *memcg_cachep;
2261 VM_BUG_ON(!is_root_cache(cachep));
2263 if (memcg_kmem_bypass())
2266 if (current->memcg_kmem_skip_account)
2269 memcg = get_mem_cgroup_from_mm(current->mm);
2270 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2274 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2275 if (likely(memcg_cachep))
2276 return memcg_cachep;
2279 * If we are in a safe context (can wait, and not in interrupt
2280 * context), we could be be predictable and return right away.
2281 * This would guarantee that the allocation being performed
2282 * already belongs in the new cache.
2284 * However, there are some clashes that can arrive from locking.
2285 * For instance, because we acquire the slab_mutex while doing
2286 * memcg_create_kmem_cache, this means no further allocation
2287 * could happen with the slab_mutex held. So it's better to
2290 memcg_schedule_kmem_cache_create(memcg, cachep);
2292 css_put(&memcg->css);
2297 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2298 * @cachep: the cache returned by memcg_kmem_get_cache
2300 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2302 if (!is_root_cache(cachep))
2303 css_put(&cachep->memcg_params.memcg->css);
2307 * memcg_kmem_charge: charge a kmem page
2308 * @page: page to charge
2309 * @gfp: reclaim mode
2310 * @order: allocation order
2311 * @memcg: memory cgroup to charge
2313 * Returns 0 on success, an error code on failure.
2315 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2316 struct mem_cgroup *memcg)
2318 unsigned int nr_pages = 1 << order;
2319 struct page_counter *counter;
2322 ret = try_charge(memcg, gfp, nr_pages);
2326 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2327 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2328 cancel_charge(memcg, nr_pages);
2332 page->mem_cgroup = memcg;
2338 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2339 * @page: page to charge
2340 * @gfp: reclaim mode
2341 * @order: allocation order
2343 * Returns 0 on success, an error code on failure.
2345 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2347 struct mem_cgroup *memcg;
2350 if (memcg_kmem_bypass())
2353 memcg = get_mem_cgroup_from_mm(current->mm);
2354 if (!mem_cgroup_is_root(memcg)) {
2355 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2357 __SetPageKmemcg(page);
2359 css_put(&memcg->css);
2363 * memcg_kmem_uncharge: uncharge a kmem page
2364 * @page: page to uncharge
2365 * @order: allocation order
2367 void memcg_kmem_uncharge(struct page *page, int order)
2369 struct mem_cgroup *memcg = page->mem_cgroup;
2370 unsigned int nr_pages = 1 << order;
2375 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2377 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2378 page_counter_uncharge(&memcg->kmem, nr_pages);
2380 page_counter_uncharge(&memcg->memory, nr_pages);
2381 if (do_memsw_account())
2382 page_counter_uncharge(&memcg->memsw, nr_pages);
2384 page->mem_cgroup = NULL;
2386 /* slab pages do not have PageKmemcg flag set */
2387 if (PageKmemcg(page))
2388 __ClearPageKmemcg(page);
2390 css_put_many(&memcg->css, nr_pages);
2392 #endif /* !CONFIG_SLOB */
2394 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2397 * Because tail pages are not marked as "used", set it. We're under
2398 * zone_lru_lock and migration entries setup in all page mappings.
2400 void mem_cgroup_split_huge_fixup(struct page *head)
2404 if (mem_cgroup_disabled())
2407 for (i = 1; i < HPAGE_PMD_NR; i++)
2408 head[i].mem_cgroup = head->mem_cgroup;
2410 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2413 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2415 #ifdef CONFIG_MEMCG_SWAP
2416 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2419 int val = (charge) ? 1 : -1;
2420 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2424 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2425 * @entry: swap entry to be moved
2426 * @from: mem_cgroup which the entry is moved from
2427 * @to: mem_cgroup which the entry is moved to
2429 * It succeeds only when the swap_cgroup's record for this entry is the same
2430 * as the mem_cgroup's id of @from.
2432 * Returns 0 on success, -EINVAL on failure.
2434 * The caller must have charged to @to, IOW, called page_counter_charge() about
2435 * both res and memsw, and called css_get().
2437 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2438 struct mem_cgroup *from, struct mem_cgroup *to)
2440 unsigned short old_id, new_id;
2442 old_id = mem_cgroup_id(from);
2443 new_id = mem_cgroup_id(to);
2445 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2446 mem_cgroup_swap_statistics(from, false);
2447 mem_cgroup_swap_statistics(to, true);
2453 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2454 struct mem_cgroup *from, struct mem_cgroup *to)
2460 static DEFINE_MUTEX(memcg_limit_mutex);
2462 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2463 unsigned long limit)
2465 unsigned long curusage;
2466 unsigned long oldusage;
2467 bool enlarge = false;
2472 * For keeping hierarchical_reclaim simple, how long we should retry
2473 * is depends on callers. We set our retry-count to be function
2474 * of # of children which we should visit in this loop.
2476 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2477 mem_cgroup_count_children(memcg);
2479 oldusage = page_counter_read(&memcg->memory);
2482 if (signal_pending(current)) {
2487 mutex_lock(&memcg_limit_mutex);
2488 if (limit > memcg->memsw.limit) {
2489 mutex_unlock(&memcg_limit_mutex);
2493 if (limit > memcg->memory.limit)
2495 ret = page_counter_limit(&memcg->memory, limit);
2496 mutex_unlock(&memcg_limit_mutex);
2501 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2503 curusage = page_counter_read(&memcg->memory);
2504 /* Usage is reduced ? */
2505 if (curusage >= oldusage)
2508 oldusage = curusage;
2509 } while (retry_count);
2511 if (!ret && enlarge)
2512 memcg_oom_recover(memcg);
2517 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2518 unsigned long limit)
2520 unsigned long curusage;
2521 unsigned long oldusage;
2522 bool enlarge = false;
2526 /* see mem_cgroup_resize_res_limit */
2527 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2528 mem_cgroup_count_children(memcg);
2530 oldusage = page_counter_read(&memcg->memsw);
2533 if (signal_pending(current)) {
2538 mutex_lock(&memcg_limit_mutex);
2539 if (limit < memcg->memory.limit) {
2540 mutex_unlock(&memcg_limit_mutex);
2544 if (limit > memcg->memsw.limit)
2546 ret = page_counter_limit(&memcg->memsw, limit);
2547 mutex_unlock(&memcg_limit_mutex);
2552 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2554 curusage = page_counter_read(&memcg->memsw);
2555 /* Usage is reduced ? */
2556 if (curusage >= oldusage)
2559 oldusage = curusage;
2560 } while (retry_count);
2562 if (!ret && enlarge)
2563 memcg_oom_recover(memcg);
2568 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2570 unsigned long *total_scanned)
2572 unsigned long nr_reclaimed = 0;
2573 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2574 unsigned long reclaimed;
2576 struct mem_cgroup_tree_per_node *mctz;
2577 unsigned long excess;
2578 unsigned long nr_scanned;
2583 mctz = soft_limit_tree_node(pgdat->node_id);
2586 * Do not even bother to check the largest node if the root
2587 * is empty. Do it lockless to prevent lock bouncing. Races
2588 * are acceptable as soft limit is best effort anyway.
2590 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2594 * This loop can run a while, specially if mem_cgroup's continuously
2595 * keep exceeding their soft limit and putting the system under
2602 mz = mem_cgroup_largest_soft_limit_node(mctz);
2607 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2608 gfp_mask, &nr_scanned);
2609 nr_reclaimed += reclaimed;
2610 *total_scanned += nr_scanned;
2611 spin_lock_irq(&mctz->lock);
2612 __mem_cgroup_remove_exceeded(mz, mctz);
2615 * If we failed to reclaim anything from this memory cgroup
2616 * it is time to move on to the next cgroup
2620 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2622 excess = soft_limit_excess(mz->memcg);
2624 * One school of thought says that we should not add
2625 * back the node to the tree if reclaim returns 0.
2626 * But our reclaim could return 0, simply because due
2627 * to priority we are exposing a smaller subset of
2628 * memory to reclaim from. Consider this as a longer
2631 /* If excess == 0, no tree ops */
2632 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2633 spin_unlock_irq(&mctz->lock);
2634 css_put(&mz->memcg->css);
2637 * Could not reclaim anything and there are no more
2638 * mem cgroups to try or we seem to be looping without
2639 * reclaiming anything.
2641 if (!nr_reclaimed &&
2643 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2645 } while (!nr_reclaimed);
2647 css_put(&next_mz->memcg->css);
2648 return nr_reclaimed;
2652 * Test whether @memcg has children, dead or alive. Note that this
2653 * function doesn't care whether @memcg has use_hierarchy enabled and
2654 * returns %true if there are child csses according to the cgroup
2655 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2657 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2662 ret = css_next_child(NULL, &memcg->css);
2668 * Reclaims as many pages from the given memcg as possible.
2670 * Caller is responsible for holding css reference for memcg.
2672 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2674 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2676 /* we call try-to-free pages for make this cgroup empty */
2677 lru_add_drain_all();
2678 /* try to free all pages in this cgroup */
2679 while (nr_retries && page_counter_read(&memcg->memory)) {
2682 if (signal_pending(current))
2685 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2689 /* maybe some writeback is necessary */
2690 congestion_wait(BLK_RW_ASYNC, HZ/10);
2698 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2699 char *buf, size_t nbytes,
2702 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2704 if (mem_cgroup_is_root(memcg))
2706 return mem_cgroup_force_empty(memcg) ?: nbytes;
2709 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2712 return mem_cgroup_from_css(css)->use_hierarchy;
2715 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2716 struct cftype *cft, u64 val)
2719 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2720 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2722 if (memcg->use_hierarchy == val)
2726 * If parent's use_hierarchy is set, we can't make any modifications
2727 * in the child subtrees. If it is unset, then the change can
2728 * occur, provided the current cgroup has no children.
2730 * For the root cgroup, parent_mem is NULL, we allow value to be
2731 * set if there are no children.
2733 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2734 (val == 1 || val == 0)) {
2735 if (!memcg_has_children(memcg))
2736 memcg->use_hierarchy = val;
2745 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2747 struct mem_cgroup *iter;
2750 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2752 for_each_mem_cgroup_tree(iter, memcg) {
2753 for (i = 0; i < MEMCG_NR_STAT; i++)
2754 stat[i] += mem_cgroup_read_stat(iter, i);
2758 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2760 struct mem_cgroup *iter;
2763 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2765 for_each_mem_cgroup_tree(iter, memcg) {
2766 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2767 events[i] += mem_cgroup_read_events(iter, i);
2771 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2773 unsigned long val = 0;
2775 if (mem_cgroup_is_root(memcg)) {
2776 struct mem_cgroup *iter;
2778 for_each_mem_cgroup_tree(iter, memcg) {
2779 val += mem_cgroup_read_stat(iter,
2780 MEM_CGROUP_STAT_CACHE);
2781 val += mem_cgroup_read_stat(iter,
2782 MEM_CGROUP_STAT_RSS);
2784 val += mem_cgroup_read_stat(iter,
2785 MEM_CGROUP_STAT_SWAP);
2789 val = page_counter_read(&memcg->memory);
2791 val = page_counter_read(&memcg->memsw);
2804 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2807 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2808 struct page_counter *counter;
2810 switch (MEMFILE_TYPE(cft->private)) {
2812 counter = &memcg->memory;
2815 counter = &memcg->memsw;
2818 counter = &memcg->kmem;
2821 counter = &memcg->tcpmem;
2827 switch (MEMFILE_ATTR(cft->private)) {
2829 if (counter == &memcg->memory)
2830 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2831 if (counter == &memcg->memsw)
2832 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2833 return (u64)page_counter_read(counter) * PAGE_SIZE;
2835 return (u64)counter->limit * PAGE_SIZE;
2837 return (u64)counter->watermark * PAGE_SIZE;
2839 return counter->failcnt;
2840 case RES_SOFT_LIMIT:
2841 return (u64)memcg->soft_limit * PAGE_SIZE;
2848 static int memcg_online_kmem(struct mem_cgroup *memcg)
2852 if (cgroup_memory_nokmem)
2855 BUG_ON(memcg->kmemcg_id >= 0);
2856 BUG_ON(memcg->kmem_state);
2858 memcg_id = memcg_alloc_cache_id();
2862 static_branch_inc(&memcg_kmem_enabled_key);
2864 * A memory cgroup is considered kmem-online as soon as it gets
2865 * kmemcg_id. Setting the id after enabling static branching will
2866 * guarantee no one starts accounting before all call sites are
2869 memcg->kmemcg_id = memcg_id;
2870 memcg->kmem_state = KMEM_ONLINE;
2875 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2877 struct cgroup_subsys_state *css;
2878 struct mem_cgroup *parent, *child;
2881 if (memcg->kmem_state != KMEM_ONLINE)
2884 * Clear the online state before clearing memcg_caches array
2885 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2886 * guarantees that no cache will be created for this cgroup
2887 * after we are done (see memcg_create_kmem_cache()).
2889 memcg->kmem_state = KMEM_ALLOCATED;
2891 memcg_deactivate_kmem_caches(memcg);
2893 kmemcg_id = memcg->kmemcg_id;
2894 BUG_ON(kmemcg_id < 0);
2896 parent = parent_mem_cgroup(memcg);
2898 parent = root_mem_cgroup;
2901 * Change kmemcg_id of this cgroup and all its descendants to the
2902 * parent's id, and then move all entries from this cgroup's list_lrus
2903 * to ones of the parent. After we have finished, all list_lrus
2904 * corresponding to this cgroup are guaranteed to remain empty. The
2905 * ordering is imposed by list_lru_node->lock taken by
2906 * memcg_drain_all_list_lrus().
2908 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2909 css_for_each_descendant_pre(css, &memcg->css) {
2910 child = mem_cgroup_from_css(css);
2911 BUG_ON(child->kmemcg_id != kmemcg_id);
2912 child->kmemcg_id = parent->kmemcg_id;
2913 if (!memcg->use_hierarchy)
2918 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2920 memcg_free_cache_id(kmemcg_id);
2923 static void memcg_free_kmem(struct mem_cgroup *memcg)
2925 /* css_alloc() failed, offlining didn't happen */
2926 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2927 memcg_offline_kmem(memcg);
2929 if (memcg->kmem_state == KMEM_ALLOCATED) {
2930 memcg_destroy_kmem_caches(memcg);
2931 static_branch_dec(&memcg_kmem_enabled_key);
2932 WARN_ON(page_counter_read(&memcg->kmem));
2936 static int memcg_online_kmem(struct mem_cgroup *memcg)
2940 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2943 static void memcg_free_kmem(struct mem_cgroup *memcg)
2946 #endif /* !CONFIG_SLOB */
2948 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2949 unsigned long limit)
2953 mutex_lock(&memcg_limit_mutex);
2954 ret = page_counter_limit(&memcg->kmem, limit);
2955 mutex_unlock(&memcg_limit_mutex);
2959 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2963 mutex_lock(&memcg_limit_mutex);
2965 ret = page_counter_limit(&memcg->tcpmem, limit);
2969 if (!memcg->tcpmem_active) {
2971 * The active flag needs to be written after the static_key
2972 * update. This is what guarantees that the socket activation
2973 * function is the last one to run. See mem_cgroup_sk_alloc()
2974 * for details, and note that we don't mark any socket as
2975 * belonging to this memcg until that flag is up.
2977 * We need to do this, because static_keys will span multiple
2978 * sites, but we can't control their order. If we mark a socket
2979 * as accounted, but the accounting functions are not patched in
2980 * yet, we'll lose accounting.
2982 * We never race with the readers in mem_cgroup_sk_alloc(),
2983 * because when this value change, the code to process it is not
2986 static_branch_inc(&memcg_sockets_enabled_key);
2987 memcg->tcpmem_active = true;
2990 mutex_unlock(&memcg_limit_mutex);
2995 * The user of this function is...
2998 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2999 char *buf, size_t nbytes, loff_t off)
3001 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3002 unsigned long nr_pages;
3005 buf = strstrip(buf);
3006 ret = page_counter_memparse(buf, "-1", &nr_pages);
3010 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3012 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3016 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3018 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3021 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3024 ret = memcg_update_kmem_limit(memcg, nr_pages);
3027 ret = memcg_update_tcp_limit(memcg, nr_pages);
3031 case RES_SOFT_LIMIT:
3032 memcg->soft_limit = nr_pages;
3036 return ret ?: nbytes;
3039 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3040 size_t nbytes, loff_t off)
3042 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3043 struct page_counter *counter;
3045 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3047 counter = &memcg->memory;
3050 counter = &memcg->memsw;
3053 counter = &memcg->kmem;
3056 counter = &memcg->tcpmem;
3062 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3064 page_counter_reset_watermark(counter);
3067 counter->failcnt = 0;
3076 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3079 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3083 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3084 struct cftype *cft, u64 val)
3086 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3088 if (val & ~MOVE_MASK)
3092 * No kind of locking is needed in here, because ->can_attach() will
3093 * check this value once in the beginning of the process, and then carry
3094 * on with stale data. This means that changes to this value will only
3095 * affect task migrations starting after the change.
3097 memcg->move_charge_at_immigrate = val;
3101 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3102 struct cftype *cft, u64 val)
3109 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3113 unsigned int lru_mask;
3116 static const struct numa_stat stats[] = {
3117 { "total", LRU_ALL },
3118 { "file", LRU_ALL_FILE },
3119 { "anon", LRU_ALL_ANON },
3120 { "unevictable", BIT(LRU_UNEVICTABLE) },
3122 const struct numa_stat *stat;
3125 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3127 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3128 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3129 seq_printf(m, "%s=%lu", stat->name, nr);
3130 for_each_node_state(nid, N_MEMORY) {
3131 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3133 seq_printf(m, " N%d=%lu", nid, nr);
3138 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3139 struct mem_cgroup *iter;
3142 for_each_mem_cgroup_tree(iter, memcg)
3143 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3144 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3145 for_each_node_state(nid, N_MEMORY) {
3147 for_each_mem_cgroup_tree(iter, memcg)
3148 nr += mem_cgroup_node_nr_lru_pages(
3149 iter, nid, stat->lru_mask);
3150 seq_printf(m, " N%d=%lu", nid, nr);
3157 #endif /* CONFIG_NUMA */
3159 static int memcg_stat_show(struct seq_file *m, void *v)
3161 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3162 unsigned long memory, memsw;
3163 struct mem_cgroup *mi;
3166 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3167 MEM_CGROUP_STAT_NSTATS);
3168 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3169 MEM_CGROUP_EVENTS_NSTATS);
3170 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3172 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3173 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3175 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3176 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3179 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3180 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3181 mem_cgroup_read_events(memcg, i));
3183 for (i = 0; i < NR_LRU_LISTS; i++)
3184 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3185 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3187 /* Hierarchical information */
3188 memory = memsw = PAGE_COUNTER_MAX;
3189 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3190 memory = min(memory, mi->memory.limit);
3191 memsw = min(memsw, mi->memsw.limit);
3193 seq_printf(m, "hierarchical_memory_limit %llu\n",
3194 (u64)memory * PAGE_SIZE);
3195 if (do_memsw_account())
3196 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3197 (u64)memsw * PAGE_SIZE);
3199 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3200 unsigned long long val = 0;
3202 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3204 for_each_mem_cgroup_tree(mi, memcg)
3205 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3206 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3209 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3210 unsigned long long val = 0;
3212 for_each_mem_cgroup_tree(mi, memcg)
3213 val += mem_cgroup_read_events(mi, i);
3214 seq_printf(m, "total_%s %llu\n",
3215 mem_cgroup_events_names[i], val);
3218 for (i = 0; i < NR_LRU_LISTS; i++) {
3219 unsigned long long val = 0;
3221 for_each_mem_cgroup_tree(mi, memcg)
3222 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3223 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3226 #ifdef CONFIG_DEBUG_VM
3229 struct mem_cgroup_per_node *mz;
3230 struct zone_reclaim_stat *rstat;
3231 unsigned long recent_rotated[2] = {0, 0};
3232 unsigned long recent_scanned[2] = {0, 0};
3234 for_each_online_pgdat(pgdat) {
3235 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3236 rstat = &mz->lruvec.reclaim_stat;
3238 recent_rotated[0] += rstat->recent_rotated[0];
3239 recent_rotated[1] += rstat->recent_rotated[1];
3240 recent_scanned[0] += rstat->recent_scanned[0];
3241 recent_scanned[1] += rstat->recent_scanned[1];
3243 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3244 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3245 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3246 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3253 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3256 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3258 return mem_cgroup_swappiness(memcg);
3261 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3262 struct cftype *cft, u64 val)
3264 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3270 memcg->swappiness = val;
3272 vm_swappiness = val;
3277 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3279 struct mem_cgroup_threshold_ary *t;
3280 unsigned long usage;
3285 t = rcu_dereference(memcg->thresholds.primary);
3287 t = rcu_dereference(memcg->memsw_thresholds.primary);
3292 usage = mem_cgroup_usage(memcg, swap);
3295 * current_threshold points to threshold just below or equal to usage.
3296 * If it's not true, a threshold was crossed after last
3297 * call of __mem_cgroup_threshold().
3299 i = t->current_threshold;
3302 * Iterate backward over array of thresholds starting from
3303 * current_threshold and check if a threshold is crossed.
3304 * If none of thresholds below usage is crossed, we read
3305 * only one element of the array here.
3307 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3308 eventfd_signal(t->entries[i].eventfd, 1);
3310 /* i = current_threshold + 1 */
3314 * Iterate forward over array of thresholds starting from
3315 * current_threshold+1 and check if a threshold is crossed.
3316 * If none of thresholds above usage is crossed, we read
3317 * only one element of the array here.
3319 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3320 eventfd_signal(t->entries[i].eventfd, 1);
3322 /* Update current_threshold */
3323 t->current_threshold = i - 1;
3328 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3331 __mem_cgroup_threshold(memcg, false);
3332 if (do_memsw_account())
3333 __mem_cgroup_threshold(memcg, true);
3335 memcg = parent_mem_cgroup(memcg);
3339 static int compare_thresholds(const void *a, const void *b)
3341 const struct mem_cgroup_threshold *_a = a;
3342 const struct mem_cgroup_threshold *_b = b;
3344 if (_a->threshold > _b->threshold)
3347 if (_a->threshold < _b->threshold)
3353 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3355 struct mem_cgroup_eventfd_list *ev;
3357 spin_lock(&memcg_oom_lock);
3359 list_for_each_entry(ev, &memcg->oom_notify, list)
3360 eventfd_signal(ev->eventfd, 1);
3362 spin_unlock(&memcg_oom_lock);
3366 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3368 struct mem_cgroup *iter;
3370 for_each_mem_cgroup_tree(iter, memcg)
3371 mem_cgroup_oom_notify_cb(iter);
3374 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3375 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3377 struct mem_cgroup_thresholds *thresholds;
3378 struct mem_cgroup_threshold_ary *new;
3379 unsigned long threshold;
3380 unsigned long usage;
3383 ret = page_counter_memparse(args, "-1", &threshold);
3387 mutex_lock(&memcg->thresholds_lock);
3390 thresholds = &memcg->thresholds;
3391 usage = mem_cgroup_usage(memcg, false);
3392 } else if (type == _MEMSWAP) {
3393 thresholds = &memcg->memsw_thresholds;
3394 usage = mem_cgroup_usage(memcg, true);
3398 /* Check if a threshold crossed before adding a new one */
3399 if (thresholds->primary)
3400 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3402 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3404 /* Allocate memory for new array of thresholds */
3405 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3413 /* Copy thresholds (if any) to new array */
3414 if (thresholds->primary) {
3415 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3416 sizeof(struct mem_cgroup_threshold));
3419 /* Add new threshold */
3420 new->entries[size - 1].eventfd = eventfd;
3421 new->entries[size - 1].threshold = threshold;
3423 /* Sort thresholds. Registering of new threshold isn't time-critical */
3424 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3425 compare_thresholds, NULL);
3427 /* Find current threshold */
3428 new->current_threshold = -1;
3429 for (i = 0; i < size; i++) {
3430 if (new->entries[i].threshold <= usage) {
3432 * new->current_threshold will not be used until
3433 * rcu_assign_pointer(), so it's safe to increment
3436 ++new->current_threshold;
3441 /* Free old spare buffer and save old primary buffer as spare */
3442 kfree(thresholds->spare);
3443 thresholds->spare = thresholds->primary;
3445 rcu_assign_pointer(thresholds->primary, new);
3447 /* To be sure that nobody uses thresholds */
3451 mutex_unlock(&memcg->thresholds_lock);
3456 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3457 struct eventfd_ctx *eventfd, const char *args)
3459 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3462 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3463 struct eventfd_ctx *eventfd, const char *args)
3465 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3468 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3469 struct eventfd_ctx *eventfd, enum res_type type)
3471 struct mem_cgroup_thresholds *thresholds;
3472 struct mem_cgroup_threshold_ary *new;
3473 unsigned long usage;
3476 mutex_lock(&memcg->thresholds_lock);
3479 thresholds = &memcg->thresholds;
3480 usage = mem_cgroup_usage(memcg, false);
3481 } else if (type == _MEMSWAP) {
3482 thresholds = &memcg->memsw_thresholds;
3483 usage = mem_cgroup_usage(memcg, true);
3487 if (!thresholds->primary)
3490 /* Check if a threshold crossed before removing */
3491 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3493 /* Calculate new number of threshold */
3495 for (i = 0; i < thresholds->primary->size; i++) {
3496 if (thresholds->primary->entries[i].eventfd != eventfd)
3500 new = thresholds->spare;
3502 /* Set thresholds array to NULL if we don't have thresholds */
3511 /* Copy thresholds and find current threshold */
3512 new->current_threshold = -1;
3513 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3514 if (thresholds->primary->entries[i].eventfd == eventfd)
3517 new->entries[j] = thresholds->primary->entries[i];
3518 if (new->entries[j].threshold <= usage) {
3520 * new->current_threshold will not be used
3521 * until rcu_assign_pointer(), so it's safe to increment
3524 ++new->current_threshold;
3530 /* Swap primary and spare array */
3531 thresholds->spare = thresholds->primary;
3533 rcu_assign_pointer(thresholds->primary, new);
3535 /* To be sure that nobody uses thresholds */
3538 /* If all events are unregistered, free the spare array */
3540 kfree(thresholds->spare);
3541 thresholds->spare = NULL;
3544 mutex_unlock(&memcg->thresholds_lock);
3547 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3548 struct eventfd_ctx *eventfd)
3550 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3553 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3554 struct eventfd_ctx *eventfd)
3556 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3559 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3560 struct eventfd_ctx *eventfd, const char *args)
3562 struct mem_cgroup_eventfd_list *event;
3564 event = kmalloc(sizeof(*event), GFP_KERNEL);
3568 spin_lock(&memcg_oom_lock);
3570 event->eventfd = eventfd;
3571 list_add(&event->list, &memcg->oom_notify);
3573 /* already in OOM ? */
3574 if (memcg->under_oom)
3575 eventfd_signal(eventfd, 1);
3576 spin_unlock(&memcg_oom_lock);
3581 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3582 struct eventfd_ctx *eventfd)
3584 struct mem_cgroup_eventfd_list *ev, *tmp;
3586 spin_lock(&memcg_oom_lock);
3588 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3589 if (ev->eventfd == eventfd) {
3590 list_del(&ev->list);
3595 spin_unlock(&memcg_oom_lock);
3598 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3600 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3602 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3603 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3607 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3608 struct cftype *cft, u64 val)
3610 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3612 /* cannot set to root cgroup and only 0 and 1 are allowed */
3613 if (!css->parent || !((val == 0) || (val == 1)))
3616 memcg->oom_kill_disable = val;
3618 memcg_oom_recover(memcg);
3623 #ifdef CONFIG_CGROUP_WRITEBACK
3625 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3627 return &memcg->cgwb_list;
3630 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3632 return wb_domain_init(&memcg->cgwb_domain, gfp);
3635 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3637 wb_domain_exit(&memcg->cgwb_domain);
3640 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3642 wb_domain_size_changed(&memcg->cgwb_domain);
3645 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3647 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3649 if (!memcg->css.parent)
3652 return &memcg->cgwb_domain;
3656 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3657 * @wb: bdi_writeback in question
3658 * @pfilepages: out parameter for number of file pages
3659 * @pheadroom: out parameter for number of allocatable pages according to memcg
3660 * @pdirty: out parameter for number of dirty pages
3661 * @pwriteback: out parameter for number of pages under writeback
3663 * Determine the numbers of file, headroom, dirty, and writeback pages in
3664 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3665 * is a bit more involved.
3667 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3668 * headroom is calculated as the lowest headroom of itself and the
3669 * ancestors. Note that this doesn't consider the actual amount of
3670 * available memory in the system. The caller should further cap
3671 * *@pheadroom accordingly.
3673 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3674 unsigned long *pheadroom, unsigned long *pdirty,
3675 unsigned long *pwriteback)
3677 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3678 struct mem_cgroup *parent;
3680 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3682 /* this should eventually include NR_UNSTABLE_NFS */
3683 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3684 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3685 (1 << LRU_ACTIVE_FILE));
3686 *pheadroom = PAGE_COUNTER_MAX;
3688 while ((parent = parent_mem_cgroup(memcg))) {
3689 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3690 unsigned long used = page_counter_read(&memcg->memory);
3692 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3697 #else /* CONFIG_CGROUP_WRITEBACK */
3699 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3704 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3708 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3712 #endif /* CONFIG_CGROUP_WRITEBACK */
3715 * DO NOT USE IN NEW FILES.
3717 * "cgroup.event_control" implementation.
3719 * This is way over-engineered. It tries to support fully configurable
3720 * events for each user. Such level of flexibility is completely
3721 * unnecessary especially in the light of the planned unified hierarchy.
3723 * Please deprecate this and replace with something simpler if at all
3728 * Unregister event and free resources.
3730 * Gets called from workqueue.
3732 static void memcg_event_remove(struct work_struct *work)
3734 struct mem_cgroup_event *event =
3735 container_of(work, struct mem_cgroup_event, remove);
3736 struct mem_cgroup *memcg = event->memcg;
3738 remove_wait_queue(event->wqh, &event->wait);
3740 event->unregister_event(memcg, event->eventfd);
3742 /* Notify userspace the event is going away. */
3743 eventfd_signal(event->eventfd, 1);
3745 eventfd_ctx_put(event->eventfd);
3747 css_put(&memcg->css);
3751 * Gets called on POLLHUP on eventfd when user closes it.
3753 * Called with wqh->lock held and interrupts disabled.
3755 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3756 int sync, void *key)
3758 struct mem_cgroup_event *event =
3759 container_of(wait, struct mem_cgroup_event, wait);
3760 struct mem_cgroup *memcg = event->memcg;
3761 unsigned long flags = (unsigned long)key;
3763 if (flags & POLLHUP) {
3765 * If the event has been detached at cgroup removal, we
3766 * can simply return knowing the other side will cleanup
3769 * We can't race against event freeing since the other
3770 * side will require wqh->lock via remove_wait_queue(),
3773 spin_lock(&memcg->event_list_lock);
3774 if (!list_empty(&event->list)) {
3775 list_del_init(&event->list);
3777 * We are in atomic context, but cgroup_event_remove()
3778 * may sleep, so we have to call it in workqueue.
3780 schedule_work(&event->remove);
3782 spin_unlock(&memcg->event_list_lock);
3788 static void memcg_event_ptable_queue_proc(struct file *file,
3789 wait_queue_head_t *wqh, poll_table *pt)
3791 struct mem_cgroup_event *event =
3792 container_of(pt, struct mem_cgroup_event, pt);
3795 add_wait_queue(wqh, &event->wait);
3799 * DO NOT USE IN NEW FILES.
3801 * Parse input and register new cgroup event handler.
3803 * Input must be in format '<event_fd> <control_fd> <args>'.
3804 * Interpretation of args is defined by control file implementation.
3806 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3807 char *buf, size_t nbytes, loff_t off)
3809 struct cgroup_subsys_state *css = of_css(of);
3810 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3811 struct mem_cgroup_event *event;
3812 struct cgroup_subsys_state *cfile_css;
3813 unsigned int efd, cfd;
3820 buf = strstrip(buf);
3822 efd = simple_strtoul(buf, &endp, 10);
3827 cfd = simple_strtoul(buf, &endp, 10);
3828 if ((*endp != ' ') && (*endp != '\0'))
3832 event = kzalloc(sizeof(*event), GFP_KERNEL);
3836 event->memcg = memcg;
3837 INIT_LIST_HEAD(&event->list);
3838 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3839 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3840 INIT_WORK(&event->remove, memcg_event_remove);
3848 event->eventfd = eventfd_ctx_fileget(efile.file);
3849 if (IS_ERR(event->eventfd)) {
3850 ret = PTR_ERR(event->eventfd);
3857 goto out_put_eventfd;
3860 /* the process need read permission on control file */
3861 /* AV: shouldn't we check that it's been opened for read instead? */
3862 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3867 * Determine the event callbacks and set them in @event. This used
3868 * to be done via struct cftype but cgroup core no longer knows
3869 * about these events. The following is crude but the whole thing
3870 * is for compatibility anyway.
3872 * DO NOT ADD NEW FILES.
3874 name = cfile.file->f_path.dentry->d_name.name;
3876 if (!strcmp(name, "memory.usage_in_bytes")) {
3877 event->register_event = mem_cgroup_usage_register_event;
3878 event->unregister_event = mem_cgroup_usage_unregister_event;
3879 } else if (!strcmp(name, "memory.oom_control")) {
3880 event->register_event = mem_cgroup_oom_register_event;
3881 event->unregister_event = mem_cgroup_oom_unregister_event;
3882 } else if (!strcmp(name, "memory.pressure_level")) {
3883 event->register_event = vmpressure_register_event;
3884 event->unregister_event = vmpressure_unregister_event;
3885 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3886 event->register_event = memsw_cgroup_usage_register_event;
3887 event->unregister_event = memsw_cgroup_usage_unregister_event;
3894 * Verify @cfile should belong to @css. Also, remaining events are
3895 * automatically removed on cgroup destruction but the removal is
3896 * asynchronous, so take an extra ref on @css.
3898 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3899 &memory_cgrp_subsys);
3901 if (IS_ERR(cfile_css))
3903 if (cfile_css != css) {
3908 ret = event->register_event(memcg, event->eventfd, buf);
3912 efile.file->f_op->poll(efile.file, &event->pt);
3914 spin_lock(&memcg->event_list_lock);
3915 list_add(&event->list, &memcg->event_list);
3916 spin_unlock(&memcg->event_list_lock);
3928 eventfd_ctx_put(event->eventfd);
3937 static struct cftype mem_cgroup_legacy_files[] = {
3939 .name = "usage_in_bytes",
3940 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3941 .read_u64 = mem_cgroup_read_u64,
3944 .name = "max_usage_in_bytes",
3945 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3946 .write = mem_cgroup_reset,
3947 .read_u64 = mem_cgroup_read_u64,
3950 .name = "limit_in_bytes",
3951 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3952 .write = mem_cgroup_write,
3953 .read_u64 = mem_cgroup_read_u64,
3956 .name = "soft_limit_in_bytes",
3957 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3958 .write = mem_cgroup_write,
3959 .read_u64 = mem_cgroup_read_u64,
3963 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3964 .write = mem_cgroup_reset,
3965 .read_u64 = mem_cgroup_read_u64,
3969 .seq_show = memcg_stat_show,
3972 .name = "force_empty",
3973 .write = mem_cgroup_force_empty_write,
3976 .name = "use_hierarchy",
3977 .write_u64 = mem_cgroup_hierarchy_write,
3978 .read_u64 = mem_cgroup_hierarchy_read,
3981 .name = "cgroup.event_control", /* XXX: for compat */
3982 .write = memcg_write_event_control,
3983 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3986 .name = "swappiness",
3987 .read_u64 = mem_cgroup_swappiness_read,
3988 .write_u64 = mem_cgroup_swappiness_write,
3991 .name = "move_charge_at_immigrate",
3992 .read_u64 = mem_cgroup_move_charge_read,
3993 .write_u64 = mem_cgroup_move_charge_write,
3996 .name = "oom_control",
3997 .seq_show = mem_cgroup_oom_control_read,
3998 .write_u64 = mem_cgroup_oom_control_write,
3999 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4002 .name = "pressure_level",
4006 .name = "numa_stat",
4007 .seq_show = memcg_numa_stat_show,
4011 .name = "kmem.limit_in_bytes",
4012 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4013 .write = mem_cgroup_write,
4014 .read_u64 = mem_cgroup_read_u64,
4017 .name = "kmem.usage_in_bytes",
4018 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4019 .read_u64 = mem_cgroup_read_u64,
4022 .name = "kmem.failcnt",
4023 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4024 .write = mem_cgroup_reset,
4025 .read_u64 = mem_cgroup_read_u64,
4028 .name = "kmem.max_usage_in_bytes",
4029 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4030 .write = mem_cgroup_reset,
4031 .read_u64 = mem_cgroup_read_u64,
4033 #ifdef CONFIG_SLABINFO
4035 .name = "kmem.slabinfo",
4036 .seq_start = slab_start,
4037 .seq_next = slab_next,
4038 .seq_stop = slab_stop,
4039 .seq_show = memcg_slab_show,
4043 .name = "kmem.tcp.limit_in_bytes",
4044 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4045 .write = mem_cgroup_write,
4046 .read_u64 = mem_cgroup_read_u64,
4049 .name = "kmem.tcp.usage_in_bytes",
4050 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4051 .read_u64 = mem_cgroup_read_u64,
4054 .name = "kmem.tcp.failcnt",
4055 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4056 .write = mem_cgroup_reset,
4057 .read_u64 = mem_cgroup_read_u64,
4060 .name = "kmem.tcp.max_usage_in_bytes",
4061 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4062 .write = mem_cgroup_reset,
4063 .read_u64 = mem_cgroup_read_u64,
4065 { }, /* terminate */
4069 * Private memory cgroup IDR
4071 * Swap-out records and page cache shadow entries need to store memcg
4072 * references in constrained space, so we maintain an ID space that is
4073 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4074 * memory-controlled cgroups to 64k.
4076 * However, there usually are many references to the oflline CSS after
4077 * the cgroup has been destroyed, such as page cache or reclaimable
4078 * slab objects, that don't need to hang on to the ID. We want to keep
4079 * those dead CSS from occupying IDs, or we might quickly exhaust the
4080 * relatively small ID space and prevent the creation of new cgroups
4081 * even when there are much fewer than 64k cgroups - possibly none.
4083 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4084 * be freed and recycled when it's no longer needed, which is usually
4085 * when the CSS is offlined.
4087 * The only exception to that are records of swapped out tmpfs/shmem
4088 * pages that need to be attributed to live ancestors on swapin. But
4089 * those references are manageable from userspace.
4092 static DEFINE_IDR(mem_cgroup_idr);
4094 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4096 if (memcg->id.id > 0) {
4097 idr_remove(&mem_cgroup_idr, memcg->id.id);
4102 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4104 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4105 atomic_add(n, &memcg->id.ref);
4108 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4110 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4111 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4112 mem_cgroup_id_remove(memcg);
4114 /* Memcg ID pins CSS */
4115 css_put(&memcg->css);
4119 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4121 mem_cgroup_id_get_many(memcg, 1);
4124 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4126 mem_cgroup_id_put_many(memcg, 1);
4130 * mem_cgroup_from_id - look up a memcg from a memcg id
4131 * @id: the memcg id to look up
4133 * Caller must hold rcu_read_lock().
4135 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4137 WARN_ON_ONCE(!rcu_read_lock_held());
4138 return idr_find(&mem_cgroup_idr, id);
4141 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4143 struct mem_cgroup_per_node *pn;
4146 * This routine is called against possible nodes.
4147 * But it's BUG to call kmalloc() against offline node.
4149 * TODO: this routine can waste much memory for nodes which will
4150 * never be onlined. It's better to use memory hotplug callback
4153 if (!node_state(node, N_NORMAL_MEMORY))
4155 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4159 lruvec_init(&pn->lruvec);
4160 pn->usage_in_excess = 0;
4161 pn->on_tree = false;
4164 memcg->nodeinfo[node] = pn;
4168 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4170 kfree(memcg->nodeinfo[node]);
4173 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4178 free_mem_cgroup_per_node_info(memcg, node);
4179 free_percpu(memcg->stat);
4183 static void mem_cgroup_free(struct mem_cgroup *memcg)
4185 memcg_wb_domain_exit(memcg);
4186 __mem_cgroup_free(memcg);
4189 static struct mem_cgroup *mem_cgroup_alloc(void)
4191 struct mem_cgroup *memcg;
4195 size = sizeof(struct mem_cgroup);
4196 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4198 memcg = kzalloc(size, GFP_KERNEL);
4202 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4203 1, MEM_CGROUP_ID_MAX,
4205 if (memcg->id.id < 0)
4208 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4213 if (alloc_mem_cgroup_per_node_info(memcg, node))
4216 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4219 INIT_WORK(&memcg->high_work, high_work_func);
4220 memcg->last_scanned_node = MAX_NUMNODES;
4221 INIT_LIST_HEAD(&memcg->oom_notify);
4222 mutex_init(&memcg->thresholds_lock);
4223 spin_lock_init(&memcg->move_lock);
4224 vmpressure_init(&memcg->vmpressure);
4225 INIT_LIST_HEAD(&memcg->event_list);
4226 spin_lock_init(&memcg->event_list_lock);
4227 memcg->socket_pressure = jiffies;
4229 memcg->kmemcg_id = -1;
4231 #ifdef CONFIG_CGROUP_WRITEBACK
4232 INIT_LIST_HEAD(&memcg->cgwb_list);
4234 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4237 mem_cgroup_id_remove(memcg);
4238 __mem_cgroup_free(memcg);
4242 static struct cgroup_subsys_state * __ref
4243 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4245 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4246 struct mem_cgroup *memcg;
4247 long error = -ENOMEM;
4249 memcg = mem_cgroup_alloc();
4251 return ERR_PTR(error);
4253 memcg->high = PAGE_COUNTER_MAX;
4254 memcg->soft_limit = PAGE_COUNTER_MAX;
4256 memcg->swappiness = mem_cgroup_swappiness(parent);
4257 memcg->oom_kill_disable = parent->oom_kill_disable;
4259 if (parent && parent->use_hierarchy) {
4260 memcg->use_hierarchy = true;
4261 page_counter_init(&memcg->memory, &parent->memory);
4262 page_counter_init(&memcg->swap, &parent->swap);
4263 page_counter_init(&memcg->memsw, &parent->memsw);
4264 page_counter_init(&memcg->kmem, &parent->kmem);
4265 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4267 page_counter_init(&memcg->memory, NULL);
4268 page_counter_init(&memcg->swap, NULL);
4269 page_counter_init(&memcg->memsw, NULL);
4270 page_counter_init(&memcg->kmem, NULL);
4271 page_counter_init(&memcg->tcpmem, NULL);
4273 * Deeper hierachy with use_hierarchy == false doesn't make
4274 * much sense so let cgroup subsystem know about this
4275 * unfortunate state in our controller.
4277 if (parent != root_mem_cgroup)
4278 memory_cgrp_subsys.broken_hierarchy = true;
4281 /* The following stuff does not apply to the root */
4283 root_mem_cgroup = memcg;
4287 error = memcg_online_kmem(memcg);
4291 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4292 static_branch_inc(&memcg_sockets_enabled_key);
4296 mem_cgroup_id_remove(memcg);
4297 mem_cgroup_free(memcg);
4298 return ERR_PTR(-ENOMEM);
4301 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4303 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4305 /* Online state pins memcg ID, memcg ID pins CSS */
4306 atomic_set(&memcg->id.ref, 1);
4311 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4313 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4314 struct mem_cgroup_event *event, *tmp;
4317 * Unregister events and notify userspace.
4318 * Notify userspace about cgroup removing only after rmdir of cgroup
4319 * directory to avoid race between userspace and kernelspace.
4321 spin_lock(&memcg->event_list_lock);
4322 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4323 list_del_init(&event->list);
4324 schedule_work(&event->remove);
4326 spin_unlock(&memcg->event_list_lock);
4328 memcg_offline_kmem(memcg);
4329 wb_memcg_offline(memcg);
4331 mem_cgroup_id_put(memcg);
4334 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4336 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4338 invalidate_reclaim_iterators(memcg);
4341 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4343 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4345 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4346 static_branch_dec(&memcg_sockets_enabled_key);
4348 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4349 static_branch_dec(&memcg_sockets_enabled_key);
4351 vmpressure_cleanup(&memcg->vmpressure);
4352 cancel_work_sync(&memcg->high_work);
4353 mem_cgroup_remove_from_trees(memcg);
4354 memcg_free_kmem(memcg);
4355 mem_cgroup_free(memcg);
4359 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4360 * @css: the target css
4362 * Reset the states of the mem_cgroup associated with @css. This is
4363 * invoked when the userland requests disabling on the default hierarchy
4364 * but the memcg is pinned through dependency. The memcg should stop
4365 * applying policies and should revert to the vanilla state as it may be
4366 * made visible again.
4368 * The current implementation only resets the essential configurations.
4369 * This needs to be expanded to cover all the visible parts.
4371 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4373 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4375 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4376 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4377 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4378 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4379 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4381 memcg->high = PAGE_COUNTER_MAX;
4382 memcg->soft_limit = PAGE_COUNTER_MAX;
4383 memcg_wb_domain_size_changed(memcg);
4387 /* Handlers for move charge at task migration. */
4388 static int mem_cgroup_do_precharge(unsigned long count)
4392 /* Try a single bulk charge without reclaim first, kswapd may wake */
4393 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4395 mc.precharge += count;
4399 /* Try charges one by one with reclaim, but do not retry */
4401 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4415 enum mc_target_type {
4421 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4422 unsigned long addr, pte_t ptent)
4424 struct page *page = vm_normal_page(vma, addr, ptent);
4426 if (!page || !page_mapped(page))
4428 if (PageAnon(page)) {
4429 if (!(mc.flags & MOVE_ANON))
4432 if (!(mc.flags & MOVE_FILE))
4435 if (!get_page_unless_zero(page))
4442 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4443 pte_t ptent, swp_entry_t *entry)
4445 struct page *page = NULL;
4446 swp_entry_t ent = pte_to_swp_entry(ptent);
4448 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4451 * Because lookup_swap_cache() updates some statistics counter,
4452 * we call find_get_page() with swapper_space directly.
4454 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4455 if (do_memsw_account())
4456 entry->val = ent.val;
4461 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4462 pte_t ptent, swp_entry_t *entry)
4468 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4469 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4471 struct page *page = NULL;
4472 struct address_space *mapping;
4475 if (!vma->vm_file) /* anonymous vma */
4477 if (!(mc.flags & MOVE_FILE))
4480 mapping = vma->vm_file->f_mapping;
4481 pgoff = linear_page_index(vma, addr);
4483 /* page is moved even if it's not RSS of this task(page-faulted). */
4485 /* shmem/tmpfs may report page out on swap: account for that too. */
4486 if (shmem_mapping(mapping)) {
4487 page = find_get_entry(mapping, pgoff);
4488 if (radix_tree_exceptional_entry(page)) {
4489 swp_entry_t swp = radix_to_swp_entry(page);
4490 if (do_memsw_account())
4492 page = find_get_page(swap_address_space(swp),
4496 page = find_get_page(mapping, pgoff);
4498 page = find_get_page(mapping, pgoff);
4504 * mem_cgroup_move_account - move account of the page
4506 * @compound: charge the page as compound or small page
4507 * @from: mem_cgroup which the page is moved from.
4508 * @to: mem_cgroup which the page is moved to. @from != @to.
4510 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4512 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4515 static int mem_cgroup_move_account(struct page *page,
4517 struct mem_cgroup *from,
4518 struct mem_cgroup *to)
4520 unsigned long flags;
4521 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4525 VM_BUG_ON(from == to);
4526 VM_BUG_ON_PAGE(PageLRU(page), page);
4527 VM_BUG_ON(compound && !PageTransHuge(page));
4530 * Prevent mem_cgroup_migrate() from looking at
4531 * page->mem_cgroup of its source page while we change it.
4534 if (!trylock_page(page))
4538 if (page->mem_cgroup != from)
4541 anon = PageAnon(page);
4543 spin_lock_irqsave(&from->move_lock, flags);
4545 if (!anon && page_mapped(page)) {
4546 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4548 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4553 * move_lock grabbed above and caller set from->moving_account, so
4554 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4555 * So mapping should be stable for dirty pages.
4557 if (!anon && PageDirty(page)) {
4558 struct address_space *mapping = page_mapping(page);
4560 if (mapping_cap_account_dirty(mapping)) {
4561 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4563 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4568 if (PageWriteback(page)) {
4569 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4571 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4576 * It is safe to change page->mem_cgroup here because the page
4577 * is referenced, charged, and isolated - we can't race with
4578 * uncharging, charging, migration, or LRU putback.
4581 /* caller should have done css_get */
4582 page->mem_cgroup = to;
4583 spin_unlock_irqrestore(&from->move_lock, flags);
4587 local_irq_disable();
4588 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4589 memcg_check_events(to, page);
4590 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4591 memcg_check_events(from, page);
4600 * get_mctgt_type - get target type of moving charge
4601 * @vma: the vma the pte to be checked belongs
4602 * @addr: the address corresponding to the pte to be checked
4603 * @ptent: the pte to be checked
4604 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4607 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4608 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4609 * move charge. if @target is not NULL, the page is stored in target->page
4610 * with extra refcnt got(Callers should handle it).
4611 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4612 * target for charge migration. if @target is not NULL, the entry is stored
4615 * Called with pte lock held.
4618 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4619 unsigned long addr, pte_t ptent, union mc_target *target)
4621 struct page *page = NULL;
4622 enum mc_target_type ret = MC_TARGET_NONE;
4623 swp_entry_t ent = { .val = 0 };
4625 if (pte_present(ptent))
4626 page = mc_handle_present_pte(vma, addr, ptent);
4627 else if (is_swap_pte(ptent))
4628 page = mc_handle_swap_pte(vma, ptent, &ent);
4629 else if (pte_none(ptent))
4630 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4632 if (!page && !ent.val)
4636 * Do only loose check w/o serialization.
4637 * mem_cgroup_move_account() checks the page is valid or
4638 * not under LRU exclusion.
4640 if (page->mem_cgroup == mc.from) {
4641 ret = MC_TARGET_PAGE;
4643 target->page = page;
4645 if (!ret || !target)
4648 /* There is a swap entry and a page doesn't exist or isn't charged */
4649 if (ent.val && !ret &&
4650 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4651 ret = MC_TARGET_SWAP;
4658 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4660 * We don't consider swapping or file mapped pages because THP does not
4661 * support them for now.
4662 * Caller should make sure that pmd_trans_huge(pmd) is true.
4664 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4665 unsigned long addr, pmd_t pmd, union mc_target *target)
4667 struct page *page = NULL;
4668 enum mc_target_type ret = MC_TARGET_NONE;
4670 page = pmd_page(pmd);
4671 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4672 if (!(mc.flags & MOVE_ANON))
4674 if (page->mem_cgroup == mc.from) {
4675 ret = MC_TARGET_PAGE;
4678 target->page = page;
4684 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4685 unsigned long addr, pmd_t pmd, union mc_target *target)
4687 return MC_TARGET_NONE;
4691 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4692 unsigned long addr, unsigned long end,
4693 struct mm_walk *walk)
4695 struct vm_area_struct *vma = walk->vma;
4699 ptl = pmd_trans_huge_lock(pmd, vma);
4701 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4702 mc.precharge += HPAGE_PMD_NR;
4707 if (pmd_trans_unstable(pmd))
4709 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4710 for (; addr != end; pte++, addr += PAGE_SIZE)
4711 if (get_mctgt_type(vma, addr, *pte, NULL))
4712 mc.precharge++; /* increment precharge temporarily */
4713 pte_unmap_unlock(pte - 1, ptl);
4719 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4721 unsigned long precharge;
4723 struct mm_walk mem_cgroup_count_precharge_walk = {
4724 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4727 down_read(&mm->mmap_sem);
4728 walk_page_range(0, mm->highest_vm_end,
4729 &mem_cgroup_count_precharge_walk);
4730 up_read(&mm->mmap_sem);
4732 precharge = mc.precharge;
4738 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4740 unsigned long precharge = mem_cgroup_count_precharge(mm);
4742 VM_BUG_ON(mc.moving_task);
4743 mc.moving_task = current;
4744 return mem_cgroup_do_precharge(precharge);
4747 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4748 static void __mem_cgroup_clear_mc(void)
4750 struct mem_cgroup *from = mc.from;
4751 struct mem_cgroup *to = mc.to;
4753 /* we must uncharge all the leftover precharges from mc.to */
4755 cancel_charge(mc.to, mc.precharge);
4759 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4760 * we must uncharge here.
4762 if (mc.moved_charge) {
4763 cancel_charge(mc.from, mc.moved_charge);
4764 mc.moved_charge = 0;
4766 /* we must fixup refcnts and charges */
4767 if (mc.moved_swap) {
4768 /* uncharge swap account from the old cgroup */
4769 if (!mem_cgroup_is_root(mc.from))
4770 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4772 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4775 * we charged both to->memory and to->memsw, so we
4776 * should uncharge to->memory.
4778 if (!mem_cgroup_is_root(mc.to))
4779 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4781 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4782 css_put_many(&mc.to->css, mc.moved_swap);
4786 memcg_oom_recover(from);
4787 memcg_oom_recover(to);
4788 wake_up_all(&mc.waitq);
4791 static void mem_cgroup_clear_mc(void)
4793 struct mm_struct *mm = mc.mm;
4796 * we must clear moving_task before waking up waiters at the end of
4799 mc.moving_task = NULL;
4800 __mem_cgroup_clear_mc();
4801 spin_lock(&mc.lock);
4805 spin_unlock(&mc.lock);
4810 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4812 struct cgroup_subsys_state *css;
4813 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4814 struct mem_cgroup *from;
4815 struct task_struct *leader, *p;
4816 struct mm_struct *mm;
4817 unsigned long move_flags;
4820 /* charge immigration isn't supported on the default hierarchy */
4821 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4825 * Multi-process migrations only happen on the default hierarchy
4826 * where charge immigration is not used. Perform charge
4827 * immigration if @tset contains a leader and whine if there are
4831 cgroup_taskset_for_each_leader(leader, css, tset) {
4834 memcg = mem_cgroup_from_css(css);
4840 * We are now commited to this value whatever it is. Changes in this
4841 * tunable will only affect upcoming migrations, not the current one.
4842 * So we need to save it, and keep it going.
4844 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4848 from = mem_cgroup_from_task(p);
4850 VM_BUG_ON(from == memcg);
4852 mm = get_task_mm(p);
4855 /* We move charges only when we move a owner of the mm */
4856 if (mm->owner == p) {
4859 VM_BUG_ON(mc.precharge);
4860 VM_BUG_ON(mc.moved_charge);
4861 VM_BUG_ON(mc.moved_swap);
4863 spin_lock(&mc.lock);
4867 mc.flags = move_flags;
4868 spin_unlock(&mc.lock);
4869 /* We set mc.moving_task later */
4871 ret = mem_cgroup_precharge_mc(mm);
4873 mem_cgroup_clear_mc();
4880 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4883 mem_cgroup_clear_mc();
4886 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4887 unsigned long addr, unsigned long end,
4888 struct mm_walk *walk)
4891 struct vm_area_struct *vma = walk->vma;
4894 enum mc_target_type target_type;
4895 union mc_target target;
4898 ptl = pmd_trans_huge_lock(pmd, vma);
4900 if (mc.precharge < HPAGE_PMD_NR) {
4904 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4905 if (target_type == MC_TARGET_PAGE) {
4907 if (!isolate_lru_page(page)) {
4908 if (!mem_cgroup_move_account(page, true,
4910 mc.precharge -= HPAGE_PMD_NR;
4911 mc.moved_charge += HPAGE_PMD_NR;
4913 putback_lru_page(page);
4921 if (pmd_trans_unstable(pmd))
4924 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4925 for (; addr != end; addr += PAGE_SIZE) {
4926 pte_t ptent = *(pte++);
4932 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4933 case MC_TARGET_PAGE:
4936 * We can have a part of the split pmd here. Moving it
4937 * can be done but it would be too convoluted so simply
4938 * ignore such a partial THP and keep it in original
4939 * memcg. There should be somebody mapping the head.
4941 if (PageTransCompound(page))
4943 if (isolate_lru_page(page))
4945 if (!mem_cgroup_move_account(page, false,
4948 /* we uncharge from mc.from later. */
4951 putback_lru_page(page);
4952 put: /* get_mctgt_type() gets the page */
4955 case MC_TARGET_SWAP:
4957 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4959 /* we fixup refcnts and charges later. */
4967 pte_unmap_unlock(pte - 1, ptl);
4972 * We have consumed all precharges we got in can_attach().
4973 * We try charge one by one, but don't do any additional
4974 * charges to mc.to if we have failed in charge once in attach()
4977 ret = mem_cgroup_do_precharge(1);
4985 static void mem_cgroup_move_charge(void)
4987 struct mm_walk mem_cgroup_move_charge_walk = {
4988 .pmd_entry = mem_cgroup_move_charge_pte_range,
4992 lru_add_drain_all();
4994 * Signal lock_page_memcg() to take the memcg's move_lock
4995 * while we're moving its pages to another memcg. Then wait
4996 * for already started RCU-only updates to finish.
4998 atomic_inc(&mc.from->moving_account);
5001 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5003 * Someone who are holding the mmap_sem might be waiting in
5004 * waitq. So we cancel all extra charges, wake up all waiters,
5005 * and retry. Because we cancel precharges, we might not be able
5006 * to move enough charges, but moving charge is a best-effort
5007 * feature anyway, so it wouldn't be a big problem.
5009 __mem_cgroup_clear_mc();
5014 * When we have consumed all precharges and failed in doing
5015 * additional charge, the page walk just aborts.
5017 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5019 up_read(&mc.mm->mmap_sem);
5020 atomic_dec(&mc.from->moving_account);
5023 static void mem_cgroup_move_task(void)
5026 mem_cgroup_move_charge();
5027 mem_cgroup_clear_mc();
5030 #else /* !CONFIG_MMU */
5031 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5035 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5038 static void mem_cgroup_move_task(void)
5044 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5045 * to verify whether we're attached to the default hierarchy on each mount
5048 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5051 * use_hierarchy is forced on the default hierarchy. cgroup core
5052 * guarantees that @root doesn't have any children, so turning it
5053 * on for the root memcg is enough.
5055 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5056 root_mem_cgroup->use_hierarchy = true;
5058 root_mem_cgroup->use_hierarchy = false;
5061 static u64 memory_current_read(struct cgroup_subsys_state *css,
5064 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5066 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5069 static int memory_low_show(struct seq_file *m, void *v)
5071 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5072 unsigned long low = READ_ONCE(memcg->low);
5074 if (low == PAGE_COUNTER_MAX)
5075 seq_puts(m, "max\n");
5077 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5082 static ssize_t memory_low_write(struct kernfs_open_file *of,
5083 char *buf, size_t nbytes, loff_t off)
5085 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5089 buf = strstrip(buf);
5090 err = page_counter_memparse(buf, "max", &low);
5099 static int memory_high_show(struct seq_file *m, void *v)
5101 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5102 unsigned long high = READ_ONCE(memcg->high);
5104 if (high == PAGE_COUNTER_MAX)
5105 seq_puts(m, "max\n");
5107 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5112 static ssize_t memory_high_write(struct kernfs_open_file *of,
5113 char *buf, size_t nbytes, loff_t off)
5115 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5116 unsigned long nr_pages;
5120 buf = strstrip(buf);
5121 err = page_counter_memparse(buf, "max", &high);
5127 nr_pages = page_counter_read(&memcg->memory);
5128 if (nr_pages > high)
5129 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5132 memcg_wb_domain_size_changed(memcg);
5136 static int memory_max_show(struct seq_file *m, void *v)
5138 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5139 unsigned long max = READ_ONCE(memcg->memory.limit);
5141 if (max == PAGE_COUNTER_MAX)
5142 seq_puts(m, "max\n");
5144 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5149 static ssize_t memory_max_write(struct kernfs_open_file *of,
5150 char *buf, size_t nbytes, loff_t off)
5152 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5153 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5154 bool drained = false;
5158 buf = strstrip(buf);
5159 err = page_counter_memparse(buf, "max", &max);
5163 xchg(&memcg->memory.limit, max);
5166 unsigned long nr_pages = page_counter_read(&memcg->memory);
5168 if (nr_pages <= max)
5171 if (signal_pending(current)) {
5177 drain_all_stock(memcg);
5183 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5189 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5190 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5194 memcg_wb_domain_size_changed(memcg);
5198 static int memory_events_show(struct seq_file *m, void *v)
5200 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5202 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5203 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5204 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5205 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5210 static int memory_stat_show(struct seq_file *m, void *v)
5212 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5213 unsigned long stat[MEMCG_NR_STAT];
5214 unsigned long events[MEMCG_NR_EVENTS];
5218 * Provide statistics on the state of the memory subsystem as
5219 * well as cumulative event counters that show past behavior.
5221 * This list is ordered following a combination of these gradients:
5222 * 1) generic big picture -> specifics and details
5223 * 2) reflecting userspace activity -> reflecting kernel heuristics
5225 * Current memory state:
5228 tree_stat(memcg, stat);
5229 tree_events(memcg, events);
5231 seq_printf(m, "anon %llu\n",
5232 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5233 seq_printf(m, "file %llu\n",
5234 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5235 seq_printf(m, "kernel_stack %llu\n",
5236 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5237 seq_printf(m, "slab %llu\n",
5238 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5239 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5240 seq_printf(m, "sock %llu\n",
5241 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5243 seq_printf(m, "file_mapped %llu\n",
5244 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5245 seq_printf(m, "file_dirty %llu\n",
5246 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5247 seq_printf(m, "file_writeback %llu\n",
5248 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5250 for (i = 0; i < NR_LRU_LISTS; i++) {
5251 struct mem_cgroup *mi;
5252 unsigned long val = 0;
5254 for_each_mem_cgroup_tree(mi, memcg)
5255 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5256 seq_printf(m, "%s %llu\n",
5257 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5260 seq_printf(m, "slab_reclaimable %llu\n",
5261 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5262 seq_printf(m, "slab_unreclaimable %llu\n",
5263 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5265 /* Accumulated memory events */
5267 seq_printf(m, "pgfault %lu\n",
5268 events[MEM_CGROUP_EVENTS_PGFAULT]);
5269 seq_printf(m, "pgmajfault %lu\n",
5270 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5275 static struct cftype memory_files[] = {
5278 .flags = CFTYPE_NOT_ON_ROOT,
5279 .read_u64 = memory_current_read,
5283 .flags = CFTYPE_NOT_ON_ROOT,
5284 .seq_show = memory_low_show,
5285 .write = memory_low_write,
5289 .flags = CFTYPE_NOT_ON_ROOT,
5290 .seq_show = memory_high_show,
5291 .write = memory_high_write,
5295 .flags = CFTYPE_NOT_ON_ROOT,
5296 .seq_show = memory_max_show,
5297 .write = memory_max_write,
5301 .flags = CFTYPE_NOT_ON_ROOT,
5302 .file_offset = offsetof(struct mem_cgroup, events_file),
5303 .seq_show = memory_events_show,
5307 .flags = CFTYPE_NOT_ON_ROOT,
5308 .seq_show = memory_stat_show,
5313 struct cgroup_subsys memory_cgrp_subsys = {
5314 .css_alloc = mem_cgroup_css_alloc,
5315 .css_online = mem_cgroup_css_online,
5316 .css_offline = mem_cgroup_css_offline,
5317 .css_released = mem_cgroup_css_released,
5318 .css_free = mem_cgroup_css_free,
5319 .css_reset = mem_cgroup_css_reset,
5320 .can_attach = mem_cgroup_can_attach,
5321 .cancel_attach = mem_cgroup_cancel_attach,
5322 .post_attach = mem_cgroup_move_task,
5323 .bind = mem_cgroup_bind,
5324 .dfl_cftypes = memory_files,
5325 .legacy_cftypes = mem_cgroup_legacy_files,
5330 * mem_cgroup_low - check if memory consumption is below the normal range
5331 * @root: the highest ancestor to consider
5332 * @memcg: the memory cgroup to check
5334 * Returns %true if memory consumption of @memcg, and that of all
5335 * configurable ancestors up to @root, is below the normal range.
5337 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5339 if (mem_cgroup_disabled())
5343 * The toplevel group doesn't have a configurable range, so
5344 * it's never low when looked at directly, and it is not
5345 * considered an ancestor when assessing the hierarchy.
5348 if (memcg == root_mem_cgroup)
5351 if (page_counter_read(&memcg->memory) >= memcg->low)
5354 while (memcg != root) {
5355 memcg = parent_mem_cgroup(memcg);
5357 if (memcg == root_mem_cgroup)
5360 if (page_counter_read(&memcg->memory) >= memcg->low)
5367 * mem_cgroup_try_charge - try charging a page
5368 * @page: page to charge
5369 * @mm: mm context of the victim
5370 * @gfp_mask: reclaim mode
5371 * @memcgp: charged memcg return
5372 * @compound: charge the page as compound or small page
5374 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5375 * pages according to @gfp_mask if necessary.
5377 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5378 * Otherwise, an error code is returned.
5380 * After page->mapping has been set up, the caller must finalize the
5381 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5382 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5384 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5385 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5388 struct mem_cgroup *memcg = NULL;
5389 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5392 if (mem_cgroup_disabled())
5395 if (PageSwapCache(page)) {
5397 * Every swap fault against a single page tries to charge the
5398 * page, bail as early as possible. shmem_unuse() encounters
5399 * already charged pages, too. The USED bit is protected by
5400 * the page lock, which serializes swap cache removal, which
5401 * in turn serializes uncharging.
5403 VM_BUG_ON_PAGE(!PageLocked(page), page);
5404 if (page->mem_cgroup)
5407 if (do_swap_account) {
5408 swp_entry_t ent = { .val = page_private(page), };
5409 unsigned short id = lookup_swap_cgroup_id(ent);
5412 memcg = mem_cgroup_from_id(id);
5413 if (memcg && !css_tryget_online(&memcg->css))
5420 memcg = get_mem_cgroup_from_mm(mm);
5422 ret = try_charge(memcg, gfp_mask, nr_pages);
5424 css_put(&memcg->css);
5431 * mem_cgroup_commit_charge - commit a page charge
5432 * @page: page to charge
5433 * @memcg: memcg to charge the page to
5434 * @lrucare: page might be on LRU already
5435 * @compound: charge the page as compound or small page
5437 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5438 * after page->mapping has been set up. This must happen atomically
5439 * as part of the page instantiation, i.e. under the page table lock
5440 * for anonymous pages, under the page lock for page and swap cache.
5442 * In addition, the page must not be on the LRU during the commit, to
5443 * prevent racing with task migration. If it might be, use @lrucare.
5445 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5447 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5448 bool lrucare, bool compound)
5450 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5452 VM_BUG_ON_PAGE(!page->mapping, page);
5453 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5455 if (mem_cgroup_disabled())
5458 * Swap faults will attempt to charge the same page multiple
5459 * times. But reuse_swap_page() might have removed the page
5460 * from swapcache already, so we can't check PageSwapCache().
5465 commit_charge(page, memcg, lrucare);
5467 local_irq_disable();
5468 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5469 memcg_check_events(memcg, page);
5472 if (do_memsw_account() && PageSwapCache(page)) {
5473 swp_entry_t entry = { .val = page_private(page) };
5475 * The swap entry might not get freed for a long time,
5476 * let's not wait for it. The page already received a
5477 * memory+swap charge, drop the swap entry duplicate.
5479 mem_cgroup_uncharge_swap(entry);
5484 * mem_cgroup_cancel_charge - cancel a page charge
5485 * @page: page to charge
5486 * @memcg: memcg to charge the page to
5487 * @compound: charge the page as compound or small page
5489 * Cancel a charge transaction started by mem_cgroup_try_charge().
5491 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5494 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5496 if (mem_cgroup_disabled())
5499 * Swap faults will attempt to charge the same page multiple
5500 * times. But reuse_swap_page() might have removed the page
5501 * from swapcache already, so we can't check PageSwapCache().
5506 cancel_charge(memcg, nr_pages);
5509 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5510 unsigned long nr_anon, unsigned long nr_file,
5511 unsigned long nr_huge, unsigned long nr_kmem,
5512 struct page *dummy_page)
5514 unsigned long nr_pages = nr_anon + nr_file + nr_kmem;
5515 unsigned long flags;
5517 if (!mem_cgroup_is_root(memcg)) {
5518 page_counter_uncharge(&memcg->memory, nr_pages);
5519 if (do_memsw_account())
5520 page_counter_uncharge(&memcg->memsw, nr_pages);
5521 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && nr_kmem)
5522 page_counter_uncharge(&memcg->kmem, nr_kmem);
5523 memcg_oom_recover(memcg);
5526 local_irq_save(flags);
5527 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5528 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5529 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5530 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5531 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5532 memcg_check_events(memcg, dummy_page);
5533 local_irq_restore(flags);
5535 if (!mem_cgroup_is_root(memcg))
5536 css_put_many(&memcg->css, nr_pages);
5539 static void uncharge_list(struct list_head *page_list)
5541 struct mem_cgroup *memcg = NULL;
5542 unsigned long nr_anon = 0;
5543 unsigned long nr_file = 0;
5544 unsigned long nr_huge = 0;
5545 unsigned long nr_kmem = 0;
5546 unsigned long pgpgout = 0;
5547 struct list_head *next;
5551 * Note that the list can be a single page->lru; hence the
5552 * do-while loop instead of a simple list_for_each_entry().
5554 next = page_list->next;
5556 page = list_entry(next, struct page, lru);
5557 next = page->lru.next;
5559 VM_BUG_ON_PAGE(PageLRU(page), page);
5560 VM_BUG_ON_PAGE(!PageHWPoison(page) && page_count(page), page);
5562 if (!page->mem_cgroup)
5566 * Nobody should be changing or seriously looking at
5567 * page->mem_cgroup at this point, we have fully
5568 * exclusive access to the page.
5571 if (memcg != page->mem_cgroup) {
5573 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5574 nr_huge, nr_kmem, page);
5575 pgpgout = nr_anon = nr_file =
5576 nr_huge = nr_kmem = 0;
5578 memcg = page->mem_cgroup;
5581 if (!PageKmemcg(page)) {
5582 unsigned int nr_pages = 1;
5584 if (PageTransHuge(page)) {
5585 nr_pages <<= compound_order(page);
5586 nr_huge += nr_pages;
5589 nr_anon += nr_pages;
5591 nr_file += nr_pages;
5594 nr_kmem += 1 << compound_order(page);
5595 __ClearPageKmemcg(page);
5598 page->mem_cgroup = NULL;
5599 } while (next != page_list);
5602 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5603 nr_huge, nr_kmem, page);
5607 * mem_cgroup_uncharge - uncharge a page
5608 * @page: page to uncharge
5610 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5611 * mem_cgroup_commit_charge().
5613 void mem_cgroup_uncharge(struct page *page)
5615 if (mem_cgroup_disabled())
5618 /* Don't touch page->lru of any random page, pre-check: */
5619 if (!page->mem_cgroup)
5622 INIT_LIST_HEAD(&page->lru);
5623 uncharge_list(&page->lru);
5627 * mem_cgroup_uncharge_list - uncharge a list of page
5628 * @page_list: list of pages to uncharge
5630 * Uncharge a list of pages previously charged with
5631 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5633 void mem_cgroup_uncharge_list(struct list_head *page_list)
5635 if (mem_cgroup_disabled())
5638 if (!list_empty(page_list))
5639 uncharge_list(page_list);
5643 * mem_cgroup_migrate - charge a page's replacement
5644 * @oldpage: currently circulating page
5645 * @newpage: replacement page
5647 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5648 * be uncharged upon free.
5650 * Both pages must be locked, @newpage->mapping must be set up.
5652 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5654 struct mem_cgroup *memcg;
5655 unsigned int nr_pages;
5657 unsigned long flags;
5659 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5660 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5661 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5662 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5665 if (mem_cgroup_disabled())
5668 /* Page cache replacement: new page already charged? */
5669 if (newpage->mem_cgroup)
5672 /* Swapcache readahead pages can get replaced before being charged */
5673 memcg = oldpage->mem_cgroup;
5677 /* Force-charge the new page. The old one will be freed soon */
5678 compound = PageTransHuge(newpage);
5679 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5681 page_counter_charge(&memcg->memory, nr_pages);
5682 if (do_memsw_account())
5683 page_counter_charge(&memcg->memsw, nr_pages);
5684 css_get_many(&memcg->css, nr_pages);
5686 commit_charge(newpage, memcg, false);
5688 local_irq_save(flags);
5689 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5690 memcg_check_events(memcg, newpage);
5691 local_irq_restore(flags);
5694 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5695 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5697 void mem_cgroup_sk_alloc(struct sock *sk)
5699 struct mem_cgroup *memcg;
5701 if (!mem_cgroup_sockets_enabled)
5705 * Socket cloning can throw us here with sk_memcg already
5706 * filled. It won't however, necessarily happen from
5707 * process context. So the test for root memcg given
5708 * the current task's memcg won't help us in this case.
5710 * Respecting the original socket's memcg is a better
5711 * decision in this case.
5714 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5715 css_get(&sk->sk_memcg->css);
5720 memcg = mem_cgroup_from_task(current);
5721 if (memcg == root_mem_cgroup)
5723 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5725 if (css_tryget_online(&memcg->css))
5726 sk->sk_memcg = memcg;
5731 void mem_cgroup_sk_free(struct sock *sk)
5734 css_put(&sk->sk_memcg->css);
5738 * mem_cgroup_charge_skmem - charge socket memory
5739 * @memcg: memcg to charge
5740 * @nr_pages: number of pages to charge
5742 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5743 * @memcg's configured limit, %false if the charge had to be forced.
5745 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5747 gfp_t gfp_mask = GFP_KERNEL;
5749 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5750 struct page_counter *fail;
5752 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5753 memcg->tcpmem_pressure = 0;
5756 page_counter_charge(&memcg->tcpmem, nr_pages);
5757 memcg->tcpmem_pressure = 1;
5761 /* Don't block in the packet receive path */
5763 gfp_mask = GFP_NOWAIT;
5765 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5767 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5770 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5775 * mem_cgroup_uncharge_skmem - uncharge socket memory
5776 * @memcg - memcg to uncharge
5777 * @nr_pages - number of pages to uncharge
5779 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5781 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5782 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5786 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5788 page_counter_uncharge(&memcg->memory, nr_pages);
5789 css_put_many(&memcg->css, nr_pages);
5792 static int __init cgroup_memory(char *s)
5796 while ((token = strsep(&s, ",")) != NULL) {
5799 if (!strcmp(token, "nosocket"))
5800 cgroup_memory_nosocket = true;
5801 if (!strcmp(token, "nokmem"))
5802 cgroup_memory_nokmem = true;
5806 __setup("cgroup.memory=", cgroup_memory);
5809 * subsys_initcall() for memory controller.
5811 * Some parts like hotcpu_notifier() have to be initialized from this context
5812 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5813 * everything that doesn't depend on a specific mem_cgroup structure should
5814 * be initialized from here.
5816 static int __init mem_cgroup_init(void)
5822 * Kmem cache creation is mostly done with the slab_mutex held,
5823 * so use a special workqueue to avoid stalling all worker
5824 * threads in case lots of cgroups are created simultaneously.
5826 memcg_kmem_cache_create_wq =
5827 alloc_ordered_workqueue("memcg_kmem_cache_create", 0);
5828 BUG_ON(!memcg_kmem_cache_create_wq);
5831 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5833 for_each_possible_cpu(cpu)
5834 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5837 for_each_node(node) {
5838 struct mem_cgroup_tree_per_node *rtpn;
5840 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5841 node_online(node) ? node : NUMA_NO_NODE);
5843 rtpn->rb_root = RB_ROOT;
5844 spin_lock_init(&rtpn->lock);
5845 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5850 subsys_initcall(mem_cgroup_init);
5852 #ifdef CONFIG_MEMCG_SWAP
5853 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
5855 while (!atomic_inc_not_zero(&memcg->id.ref)) {
5857 * The root cgroup cannot be destroyed, so it's refcount must
5860 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
5864 memcg = parent_mem_cgroup(memcg);
5866 memcg = root_mem_cgroup;
5872 * mem_cgroup_swapout - transfer a memsw charge to swap
5873 * @page: page whose memsw charge to transfer
5874 * @entry: swap entry to move the charge to
5876 * Transfer the memsw charge of @page to @entry.
5878 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5880 struct mem_cgroup *memcg, *swap_memcg;
5881 unsigned short oldid;
5883 VM_BUG_ON_PAGE(PageLRU(page), page);
5884 VM_BUG_ON_PAGE(page_count(page), page);
5886 if (!do_memsw_account())
5889 memcg = page->mem_cgroup;
5891 /* Readahead page, never charged */
5896 * In case the memcg owning these pages has been offlined and doesn't
5897 * have an ID allocated to it anymore, charge the closest online
5898 * ancestor for the swap instead and transfer the memory+swap charge.
5900 swap_memcg = mem_cgroup_id_get_online(memcg);
5901 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg));
5902 VM_BUG_ON_PAGE(oldid, page);
5903 mem_cgroup_swap_statistics(swap_memcg, true);
5905 page->mem_cgroup = NULL;
5907 if (!mem_cgroup_is_root(memcg))
5908 page_counter_uncharge(&memcg->memory, 1);
5910 if (memcg != swap_memcg) {
5911 if (!mem_cgroup_is_root(swap_memcg))
5912 page_counter_charge(&swap_memcg->memsw, 1);
5913 page_counter_uncharge(&memcg->memsw, 1);
5917 * Interrupts should be disabled here because the caller holds the
5918 * mapping->tree_lock lock which is taken with interrupts-off. It is
5919 * important here to have the interrupts disabled because it is the
5920 * only synchronisation we have for udpating the per-CPU variables.
5922 VM_BUG_ON(!irqs_disabled());
5923 mem_cgroup_charge_statistics(memcg, page, false, -1);
5924 memcg_check_events(memcg, page);
5926 if (!mem_cgroup_is_root(memcg))
5927 css_put(&memcg->css);
5931 * mem_cgroup_try_charge_swap - try charging a swap entry
5932 * @page: page being added to swap
5933 * @entry: swap entry to charge
5935 * Try to charge @entry to the memcg that @page belongs to.
5937 * Returns 0 on success, -ENOMEM on failure.
5939 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5941 struct mem_cgroup *memcg;
5942 struct page_counter *counter;
5943 unsigned short oldid;
5945 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5948 memcg = page->mem_cgroup;
5950 /* Readahead page, never charged */
5954 memcg = mem_cgroup_id_get_online(memcg);
5956 if (!mem_cgroup_is_root(memcg) &&
5957 !page_counter_try_charge(&memcg->swap, 1, &counter)) {
5958 mem_cgroup_id_put(memcg);
5962 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5963 VM_BUG_ON_PAGE(oldid, page);
5964 mem_cgroup_swap_statistics(memcg, true);
5970 * mem_cgroup_uncharge_swap - uncharge a swap entry
5971 * @entry: swap entry to uncharge
5973 * Drop the swap charge associated with @entry.
5975 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5977 struct mem_cgroup *memcg;
5980 if (!do_swap_account)
5983 id = swap_cgroup_record(entry, 0);
5985 memcg = mem_cgroup_from_id(id);
5987 if (!mem_cgroup_is_root(memcg)) {
5988 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5989 page_counter_uncharge(&memcg->swap, 1);
5991 page_counter_uncharge(&memcg->memsw, 1);
5993 mem_cgroup_swap_statistics(memcg, false);
5994 mem_cgroup_id_put(memcg);
5999 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6001 long nr_swap_pages = get_nr_swap_pages();
6003 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6004 return nr_swap_pages;
6005 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6006 nr_swap_pages = min_t(long, nr_swap_pages,
6007 READ_ONCE(memcg->swap.limit) -
6008 page_counter_read(&memcg->swap));
6009 return nr_swap_pages;
6012 bool mem_cgroup_swap_full(struct page *page)
6014 struct mem_cgroup *memcg;
6016 VM_BUG_ON_PAGE(!PageLocked(page), page);
6020 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6023 memcg = page->mem_cgroup;
6027 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6028 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
6034 /* for remember boot option*/
6035 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6036 static int really_do_swap_account __initdata = 1;
6038 static int really_do_swap_account __initdata;
6041 static int __init enable_swap_account(char *s)
6043 if (!strcmp(s, "1"))
6044 really_do_swap_account = 1;
6045 else if (!strcmp(s, "0"))
6046 really_do_swap_account = 0;
6049 __setup("swapaccount=", enable_swap_account);
6051 static u64 swap_current_read(struct cgroup_subsys_state *css,
6054 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6056 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6059 static int swap_max_show(struct seq_file *m, void *v)
6061 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6062 unsigned long max = READ_ONCE(memcg->swap.limit);
6064 if (max == PAGE_COUNTER_MAX)
6065 seq_puts(m, "max\n");
6067 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6072 static ssize_t swap_max_write(struct kernfs_open_file *of,
6073 char *buf, size_t nbytes, loff_t off)
6075 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6079 buf = strstrip(buf);
6080 err = page_counter_memparse(buf, "max", &max);
6084 mutex_lock(&memcg_limit_mutex);
6085 err = page_counter_limit(&memcg->swap, max);
6086 mutex_unlock(&memcg_limit_mutex);
6093 static struct cftype swap_files[] = {
6095 .name = "swap.current",
6096 .flags = CFTYPE_NOT_ON_ROOT,
6097 .read_u64 = swap_current_read,
6101 .flags = CFTYPE_NOT_ON_ROOT,
6102 .seq_show = swap_max_show,
6103 .write = swap_max_write,
6108 static struct cftype memsw_cgroup_files[] = {
6110 .name = "memsw.usage_in_bytes",
6111 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6112 .read_u64 = mem_cgroup_read_u64,
6115 .name = "memsw.max_usage_in_bytes",
6116 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6117 .write = mem_cgroup_reset,
6118 .read_u64 = mem_cgroup_read_u64,
6121 .name = "memsw.limit_in_bytes",
6122 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6123 .write = mem_cgroup_write,
6124 .read_u64 = mem_cgroup_read_u64,
6127 .name = "memsw.failcnt",
6128 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6129 .write = mem_cgroup_reset,
6130 .read_u64 = mem_cgroup_read_u64,
6132 { }, /* terminate */
6135 static int __init mem_cgroup_swap_init(void)
6137 if (!mem_cgroup_disabled() && really_do_swap_account) {
6138 do_swap_account = 1;
6139 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6141 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6142 memsw_cgroup_files));
6146 subsys_initcall(mem_cgroup_swap_init);
6148 #endif /* CONFIG_MEMCG_SWAP */