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 <linux/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 struct workqueue_struct *memcg_kmem_cache_wq;
322 #endif /* !CONFIG_SLOB */
325 * mem_cgroup_css_from_page - css of the memcg associated with a page
326 * @page: page of interest
328 * If memcg is bound to the default hierarchy, css of the memcg associated
329 * with @page is returned. The returned css remains associated with @page
330 * until it is released.
332 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
335 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
337 struct mem_cgroup *memcg;
339 memcg = page->mem_cgroup;
341 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
342 memcg = root_mem_cgroup;
348 * page_cgroup_ino - return inode number of the memcg a page is charged to
351 * Look up the closest online ancestor of the memory cgroup @page is charged to
352 * and return its inode number or 0 if @page is not charged to any cgroup. It
353 * is safe to call this function without holding a reference to @page.
355 * Note, this function is inherently racy, because there is nothing to prevent
356 * the cgroup inode from getting torn down and potentially reallocated a moment
357 * after page_cgroup_ino() returns, so it only should be used by callers that
358 * do not care (such as procfs interfaces).
360 ino_t page_cgroup_ino(struct page *page)
362 struct mem_cgroup *memcg;
363 unsigned long ino = 0;
366 memcg = READ_ONCE(page->mem_cgroup);
367 while (memcg && !(memcg->css.flags & CSS_ONLINE))
368 memcg = parent_mem_cgroup(memcg);
370 ino = cgroup_ino(memcg->css.cgroup);
375 static struct mem_cgroup_per_node *
376 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
378 int nid = page_to_nid(page);
380 return memcg->nodeinfo[nid];
383 static struct mem_cgroup_tree_per_node *
384 soft_limit_tree_node(int nid)
386 return soft_limit_tree.rb_tree_per_node[nid];
389 static struct mem_cgroup_tree_per_node *
390 soft_limit_tree_from_page(struct page *page)
392 int nid = page_to_nid(page);
394 return soft_limit_tree.rb_tree_per_node[nid];
397 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
398 struct mem_cgroup_tree_per_node *mctz,
399 unsigned long new_usage_in_excess)
401 struct rb_node **p = &mctz->rb_root.rb_node;
402 struct rb_node *parent = NULL;
403 struct mem_cgroup_per_node *mz_node;
408 mz->usage_in_excess = new_usage_in_excess;
409 if (!mz->usage_in_excess)
413 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
415 if (mz->usage_in_excess < mz_node->usage_in_excess)
418 * We can't avoid mem cgroups that are over their soft
419 * limit by the same amount
421 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
424 rb_link_node(&mz->tree_node, parent, p);
425 rb_insert_color(&mz->tree_node, &mctz->rb_root);
429 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
430 struct mem_cgroup_tree_per_node *mctz)
434 rb_erase(&mz->tree_node, &mctz->rb_root);
438 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
439 struct mem_cgroup_tree_per_node *mctz)
443 spin_lock_irqsave(&mctz->lock, flags);
444 __mem_cgroup_remove_exceeded(mz, mctz);
445 spin_unlock_irqrestore(&mctz->lock, flags);
448 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
450 unsigned long nr_pages = page_counter_read(&memcg->memory);
451 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
452 unsigned long excess = 0;
454 if (nr_pages > soft_limit)
455 excess = nr_pages - soft_limit;
460 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
462 unsigned long excess;
463 struct mem_cgroup_per_node *mz;
464 struct mem_cgroup_tree_per_node *mctz;
466 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);
504 mem_cgroup_remove_exceeded(mz, mctz);
508 static struct mem_cgroup_per_node *
509 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
511 struct rb_node *rightmost = NULL;
512 struct mem_cgroup_per_node *mz;
516 rightmost = rb_last(&mctz->rb_root);
518 goto done; /* Nothing to reclaim from */
520 mz = rb_entry(rightmost, struct mem_cgroup_per_node, tree_node);
522 * Remove the node now but someone else can add it back,
523 * we will to add it back at the end of reclaim to its correct
524 * position in the tree.
526 __mem_cgroup_remove_exceeded(mz, mctz);
527 if (!soft_limit_excess(mz->memcg) ||
528 !css_tryget_online(&mz->memcg->css))
534 static struct mem_cgroup_per_node *
535 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
537 struct mem_cgroup_per_node *mz;
539 spin_lock_irq(&mctz->lock);
540 mz = __mem_cgroup_largest_soft_limit_node(mctz);
541 spin_unlock_irq(&mctz->lock);
546 * Return page count for single (non recursive) @memcg.
548 * Implementation Note: reading percpu statistics for memcg.
550 * Both of vmstat[] and percpu_counter has threshold and do periodic
551 * synchronization to implement "quick" read. There are trade-off between
552 * reading cost and precision of value. Then, we may have a chance to implement
553 * a periodic synchronization of counter in memcg's counter.
555 * But this _read() function is used for user interface now. The user accounts
556 * memory usage by memory cgroup and he _always_ requires exact value because
557 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
558 * have to visit all online cpus and make sum. So, for now, unnecessary
559 * synchronization is not implemented. (just implemented for cpu hotplug)
561 * If there are kernel internal actions which can make use of some not-exact
562 * value, and reading all cpu value can be performance bottleneck in some
563 * common workload, threshold and synchronization as vmstat[] should be
567 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
572 /* Per-cpu values can be negative, use a signed accumulator */
573 for_each_possible_cpu(cpu)
574 val += per_cpu(memcg->stat->count[idx], cpu);
576 * Summing races with updates, so val may be negative. Avoid exposing
577 * transient negative values.
584 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
585 enum mem_cgroup_events_index idx)
587 unsigned long val = 0;
590 for_each_possible_cpu(cpu)
591 val += per_cpu(memcg->stat->events[idx], cpu);
595 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
597 bool compound, int nr_pages)
600 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
601 * counted as CACHE even if it's on ANON LRU.
604 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
607 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
611 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
612 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
616 /* pagein of a big page is an event. So, ignore page size */
618 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
620 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
621 nr_pages = -nr_pages; /* for event */
624 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
627 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
628 int nid, unsigned int lru_mask)
630 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
631 unsigned long nr = 0;
634 VM_BUG_ON((unsigned)nid >= nr_node_ids);
637 if (!(BIT(lru) & lru_mask))
639 nr += mem_cgroup_get_lru_size(lruvec, lru);
644 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
645 unsigned int lru_mask)
647 unsigned long nr = 0;
650 for_each_node_state(nid, N_MEMORY)
651 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
655 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
656 enum mem_cgroup_events_target target)
658 unsigned long val, next;
660 val = __this_cpu_read(memcg->stat->nr_page_events);
661 next = __this_cpu_read(memcg->stat->targets[target]);
662 /* from time_after() in jiffies.h */
663 if ((long)next - (long)val < 0) {
665 case MEM_CGROUP_TARGET_THRESH:
666 next = val + THRESHOLDS_EVENTS_TARGET;
668 case MEM_CGROUP_TARGET_SOFTLIMIT:
669 next = val + SOFTLIMIT_EVENTS_TARGET;
671 case MEM_CGROUP_TARGET_NUMAINFO:
672 next = val + NUMAINFO_EVENTS_TARGET;
677 __this_cpu_write(memcg->stat->targets[target], next);
684 * Check events in order.
687 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
689 /* threshold event is triggered in finer grain than soft limit */
690 if (unlikely(mem_cgroup_event_ratelimit(memcg,
691 MEM_CGROUP_TARGET_THRESH))) {
693 bool do_numainfo __maybe_unused;
695 do_softlimit = mem_cgroup_event_ratelimit(memcg,
696 MEM_CGROUP_TARGET_SOFTLIMIT);
698 do_numainfo = mem_cgroup_event_ratelimit(memcg,
699 MEM_CGROUP_TARGET_NUMAINFO);
701 mem_cgroup_threshold(memcg);
702 if (unlikely(do_softlimit))
703 mem_cgroup_update_tree(memcg, page);
705 if (unlikely(do_numainfo))
706 atomic_inc(&memcg->numainfo_events);
711 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
714 * mm_update_next_owner() may clear mm->owner to NULL
715 * if it races with swapoff, page migration, etc.
716 * So this can be called with p == NULL.
721 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
723 EXPORT_SYMBOL(mem_cgroup_from_task);
725 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
727 struct mem_cgroup *memcg = NULL;
732 * Page cache insertions can happen withou an
733 * actual mm context, e.g. during disk probing
734 * on boot, loopback IO, acct() writes etc.
737 memcg = root_mem_cgroup;
739 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
740 if (unlikely(!memcg))
741 memcg = root_mem_cgroup;
743 } while (!css_tryget_online(&memcg->css));
749 * mem_cgroup_iter - iterate over memory cgroup hierarchy
750 * @root: hierarchy root
751 * @prev: previously returned memcg, NULL on first invocation
752 * @reclaim: cookie for shared reclaim walks, NULL for full walks
754 * Returns references to children of the hierarchy below @root, or
755 * @root itself, or %NULL after a full round-trip.
757 * Caller must pass the return value in @prev on subsequent
758 * invocations for reference counting, or use mem_cgroup_iter_break()
759 * to cancel a hierarchy walk before the round-trip is complete.
761 * Reclaimers can specify a zone and a priority level in @reclaim to
762 * divide up the memcgs in the hierarchy among all concurrent
763 * reclaimers operating on the same zone and priority.
765 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
766 struct mem_cgroup *prev,
767 struct mem_cgroup_reclaim_cookie *reclaim)
769 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
770 struct cgroup_subsys_state *css = NULL;
771 struct mem_cgroup *memcg = NULL;
772 struct mem_cgroup *pos = NULL;
774 if (mem_cgroup_disabled())
778 root = root_mem_cgroup;
780 if (prev && !reclaim)
783 if (!root->use_hierarchy && root != root_mem_cgroup) {
792 struct mem_cgroup_per_node *mz;
794 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
795 iter = &mz->iter[reclaim->priority];
797 if (prev && reclaim->generation != iter->generation)
801 pos = READ_ONCE(iter->position);
802 if (!pos || css_tryget(&pos->css))
805 * css reference reached zero, so iter->position will
806 * be cleared by ->css_released. However, we should not
807 * rely on this happening soon, because ->css_released
808 * is called from a work queue, and by busy-waiting we
809 * might block it. So we clear iter->position right
812 (void)cmpxchg(&iter->position, pos, NULL);
820 css = css_next_descendant_pre(css, &root->css);
823 * Reclaimers share the hierarchy walk, and a
824 * new one might jump in right at the end of
825 * the hierarchy - make sure they see at least
826 * one group and restart from the beginning.
834 * Verify the css and acquire a reference. The root
835 * is provided by the caller, so we know it's alive
836 * and kicking, and don't take an extra reference.
838 memcg = mem_cgroup_from_css(css);
840 if (css == &root->css)
851 * The position could have already been updated by a competing
852 * thread, so check that the value hasn't changed since we read
853 * it to avoid reclaiming from the same cgroup twice.
855 (void)cmpxchg(&iter->position, pos, memcg);
863 reclaim->generation = iter->generation;
869 if (prev && prev != root)
876 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
877 * @root: hierarchy root
878 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
880 void mem_cgroup_iter_break(struct mem_cgroup *root,
881 struct mem_cgroup *prev)
884 root = root_mem_cgroup;
885 if (prev && prev != root)
889 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
891 struct mem_cgroup *memcg = dead_memcg;
892 struct mem_cgroup_reclaim_iter *iter;
893 struct mem_cgroup_per_node *mz;
897 while ((memcg = parent_mem_cgroup(memcg))) {
899 mz = mem_cgroup_nodeinfo(memcg, nid);
900 for (i = 0; i <= DEF_PRIORITY; i++) {
902 cmpxchg(&iter->position,
910 * Iteration constructs for visiting all cgroups (under a tree). If
911 * loops are exited prematurely (break), mem_cgroup_iter_break() must
912 * be used for reference counting.
914 #define for_each_mem_cgroup_tree(iter, root) \
915 for (iter = mem_cgroup_iter(root, NULL, NULL); \
917 iter = mem_cgroup_iter(root, iter, NULL))
919 #define for_each_mem_cgroup(iter) \
920 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
922 iter = mem_cgroup_iter(NULL, iter, NULL))
925 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
926 * @memcg: hierarchy root
927 * @fn: function to call for each task
928 * @arg: argument passed to @fn
930 * This function iterates over tasks attached to @memcg or to any of its
931 * descendants and calls @fn for each task. If @fn returns a non-zero
932 * value, the function breaks the iteration loop and returns the value.
933 * Otherwise, it will iterate over all tasks and return 0.
935 * This function must not be called for the root memory cgroup.
937 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
938 int (*fn)(struct task_struct *, void *), void *arg)
940 struct mem_cgroup *iter;
943 BUG_ON(memcg == root_mem_cgroup);
945 for_each_mem_cgroup_tree(iter, memcg) {
946 struct css_task_iter it;
947 struct task_struct *task;
949 css_task_iter_start(&iter->css, &it);
950 while (!ret && (task = css_task_iter_next(&it)))
952 css_task_iter_end(&it);
954 mem_cgroup_iter_break(memcg, iter);
962 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
964 * @zone: zone of the page
966 * This function is only safe when following the LRU page isolation
967 * and putback protocol: the LRU lock must be held, and the page must
968 * either be PageLRU() or the caller must have isolated/allocated it.
970 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
972 struct mem_cgroup_per_node *mz;
973 struct mem_cgroup *memcg;
974 struct lruvec *lruvec;
976 if (mem_cgroup_disabled()) {
977 lruvec = &pgdat->lruvec;
981 memcg = page->mem_cgroup;
983 * Swapcache readahead pages are added to the LRU - and
984 * possibly migrated - before they are charged.
987 memcg = root_mem_cgroup;
989 mz = mem_cgroup_page_nodeinfo(memcg, page);
990 lruvec = &mz->lruvec;
993 * Since a node can be onlined after the mem_cgroup was created,
994 * we have to be prepared to initialize lruvec->zone here;
995 * and if offlined then reonlined, we need to reinitialize it.
997 if (unlikely(lruvec->pgdat != pgdat))
998 lruvec->pgdat = pgdat;
1003 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1004 * @lruvec: mem_cgroup per zone lru vector
1005 * @lru: index of lru list the page is sitting on
1006 * @zid: zone id of the accounted pages
1007 * @nr_pages: positive when adding or negative when removing
1009 * This function must be called under lru_lock, just before a page is added
1010 * to or just after a page is removed from an lru list (that ordering being
1011 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1013 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1014 int zid, int nr_pages)
1016 struct mem_cgroup_per_node *mz;
1017 unsigned long *lru_size;
1020 if (mem_cgroup_disabled())
1023 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1024 lru_size = &mz->lru_zone_size[zid][lru];
1027 *lru_size += nr_pages;
1030 if (WARN_ONCE(size < 0,
1031 "%s(%p, %d, %d): lru_size %ld\n",
1032 __func__, lruvec, lru, nr_pages, size)) {
1038 *lru_size += nr_pages;
1041 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1043 struct mem_cgroup *task_memcg;
1044 struct task_struct *p;
1047 p = find_lock_task_mm(task);
1049 task_memcg = get_mem_cgroup_from_mm(p->mm);
1053 * All threads may have already detached their mm's, but the oom
1054 * killer still needs to detect if they have already been oom
1055 * killed to prevent needlessly killing additional tasks.
1058 task_memcg = mem_cgroup_from_task(task);
1059 css_get(&task_memcg->css);
1062 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1063 css_put(&task_memcg->css);
1068 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1069 * @memcg: the memory cgroup
1071 * Returns the maximum amount of memory @mem can be charged with, in
1074 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1076 unsigned long margin = 0;
1077 unsigned long count;
1078 unsigned long limit;
1080 count = page_counter_read(&memcg->memory);
1081 limit = READ_ONCE(memcg->memory.limit);
1083 margin = limit - count;
1085 if (do_memsw_account()) {
1086 count = page_counter_read(&memcg->memsw);
1087 limit = READ_ONCE(memcg->memsw.limit);
1089 margin = min(margin, limit - count);
1098 * A routine for checking "mem" is under move_account() or not.
1100 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1101 * moving cgroups. This is for waiting at high-memory pressure
1104 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1106 struct mem_cgroup *from;
1107 struct mem_cgroup *to;
1110 * Unlike task_move routines, we access mc.to, mc.from not under
1111 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1113 spin_lock(&mc.lock);
1119 ret = mem_cgroup_is_descendant(from, memcg) ||
1120 mem_cgroup_is_descendant(to, memcg);
1122 spin_unlock(&mc.lock);
1126 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1128 if (mc.moving_task && current != mc.moving_task) {
1129 if (mem_cgroup_under_move(memcg)) {
1131 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1132 /* moving charge context might have finished. */
1135 finish_wait(&mc.waitq, &wait);
1142 #define K(x) ((x) << (PAGE_SHIFT-10))
1144 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1145 * @memcg: The memory cgroup that went over limit
1146 * @p: Task that is going to be killed
1148 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1151 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1153 struct mem_cgroup *iter;
1159 pr_info("Task in ");
1160 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1161 pr_cont(" killed as a result of limit of ");
1163 pr_info("Memory limit reached of cgroup ");
1166 pr_cont_cgroup_path(memcg->css.cgroup);
1171 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1172 K((u64)page_counter_read(&memcg->memory)),
1173 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1174 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1175 K((u64)page_counter_read(&memcg->memsw)),
1176 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1177 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1178 K((u64)page_counter_read(&memcg->kmem)),
1179 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1181 for_each_mem_cgroup_tree(iter, memcg) {
1182 pr_info("Memory cgroup stats for ");
1183 pr_cont_cgroup_path(iter->css.cgroup);
1186 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1187 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1189 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1190 K(mem_cgroup_read_stat(iter, i)));
1193 for (i = 0; i < NR_LRU_LISTS; i++)
1194 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1195 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1202 * This function returns the number of memcg under hierarchy tree. Returns
1203 * 1(self count) if no children.
1205 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1208 struct mem_cgroup *iter;
1210 for_each_mem_cgroup_tree(iter, memcg)
1216 * Return the memory (and swap, if configured) limit for a memcg.
1218 unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1220 unsigned long limit;
1222 limit = memcg->memory.limit;
1223 if (mem_cgroup_swappiness(memcg)) {
1224 unsigned long memsw_limit;
1225 unsigned long swap_limit;
1227 memsw_limit = memcg->memsw.limit;
1228 swap_limit = memcg->swap.limit;
1229 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1230 limit = min(limit + swap_limit, memsw_limit);
1235 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1238 struct oom_control oc = {
1242 .gfp_mask = gfp_mask,
1247 mutex_lock(&oom_lock);
1248 ret = out_of_memory(&oc);
1249 mutex_unlock(&oom_lock);
1253 #if MAX_NUMNODES > 1
1256 * test_mem_cgroup_node_reclaimable
1257 * @memcg: the target memcg
1258 * @nid: the node ID to be checked.
1259 * @noswap : specify true here if the user wants flle only information.
1261 * This function returns whether the specified memcg contains any
1262 * reclaimable pages on a node. Returns true if there are any reclaimable
1263 * pages in the node.
1265 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1266 int nid, bool noswap)
1268 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1270 if (noswap || !total_swap_pages)
1272 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1279 * Always updating the nodemask is not very good - even if we have an empty
1280 * list or the wrong list here, we can start from some node and traverse all
1281 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1284 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1288 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1289 * pagein/pageout changes since the last update.
1291 if (!atomic_read(&memcg->numainfo_events))
1293 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1296 /* make a nodemask where this memcg uses memory from */
1297 memcg->scan_nodes = node_states[N_MEMORY];
1299 for_each_node_mask(nid, node_states[N_MEMORY]) {
1301 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1302 node_clear(nid, memcg->scan_nodes);
1305 atomic_set(&memcg->numainfo_events, 0);
1306 atomic_set(&memcg->numainfo_updating, 0);
1310 * Selecting a node where we start reclaim from. Because what we need is just
1311 * reducing usage counter, start from anywhere is O,K. Considering
1312 * memory reclaim from current node, there are pros. and cons.
1314 * Freeing memory from current node means freeing memory from a node which
1315 * we'll use or we've used. So, it may make LRU bad. And if several threads
1316 * hit limits, it will see a contention on a node. But freeing from remote
1317 * node means more costs for memory reclaim because of memory latency.
1319 * Now, we use round-robin. Better algorithm is welcomed.
1321 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1325 mem_cgroup_may_update_nodemask(memcg);
1326 node = memcg->last_scanned_node;
1328 node = next_node_in(node, memcg->scan_nodes);
1330 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1331 * last time it really checked all the LRUs due to rate limiting.
1332 * Fallback to the current node in that case for simplicity.
1334 if (unlikely(node == MAX_NUMNODES))
1335 node = numa_node_id();
1337 memcg->last_scanned_node = node;
1341 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1347 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1350 unsigned long *total_scanned)
1352 struct mem_cgroup *victim = NULL;
1355 unsigned long excess;
1356 unsigned long nr_scanned;
1357 struct mem_cgroup_reclaim_cookie reclaim = {
1362 excess = soft_limit_excess(root_memcg);
1365 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1370 * If we have not been able to reclaim
1371 * anything, it might because there are
1372 * no reclaimable pages under this hierarchy
1377 * We want to do more targeted reclaim.
1378 * excess >> 2 is not to excessive so as to
1379 * reclaim too much, nor too less that we keep
1380 * coming back to reclaim from this cgroup
1382 if (total >= (excess >> 2) ||
1383 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1388 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1389 pgdat, &nr_scanned);
1390 *total_scanned += nr_scanned;
1391 if (!soft_limit_excess(root_memcg))
1394 mem_cgroup_iter_break(root_memcg, victim);
1398 #ifdef CONFIG_LOCKDEP
1399 static struct lockdep_map memcg_oom_lock_dep_map = {
1400 .name = "memcg_oom_lock",
1404 static DEFINE_SPINLOCK(memcg_oom_lock);
1407 * Check OOM-Killer is already running under our hierarchy.
1408 * If someone is running, return false.
1410 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1412 struct mem_cgroup *iter, *failed = NULL;
1414 spin_lock(&memcg_oom_lock);
1416 for_each_mem_cgroup_tree(iter, memcg) {
1417 if (iter->oom_lock) {
1419 * this subtree of our hierarchy is already locked
1420 * so we cannot give a lock.
1423 mem_cgroup_iter_break(memcg, iter);
1426 iter->oom_lock = true;
1431 * OK, we failed to lock the whole subtree so we have
1432 * to clean up what we set up to the failing subtree
1434 for_each_mem_cgroup_tree(iter, memcg) {
1435 if (iter == failed) {
1436 mem_cgroup_iter_break(memcg, iter);
1439 iter->oom_lock = false;
1442 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1444 spin_unlock(&memcg_oom_lock);
1449 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1451 struct mem_cgroup *iter;
1453 spin_lock(&memcg_oom_lock);
1454 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1455 for_each_mem_cgroup_tree(iter, memcg)
1456 iter->oom_lock = false;
1457 spin_unlock(&memcg_oom_lock);
1460 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1462 struct mem_cgroup *iter;
1464 spin_lock(&memcg_oom_lock);
1465 for_each_mem_cgroup_tree(iter, memcg)
1467 spin_unlock(&memcg_oom_lock);
1470 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1472 struct mem_cgroup *iter;
1475 * When a new child is created while the hierarchy is under oom,
1476 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1478 spin_lock(&memcg_oom_lock);
1479 for_each_mem_cgroup_tree(iter, memcg)
1480 if (iter->under_oom > 0)
1482 spin_unlock(&memcg_oom_lock);
1485 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1487 struct oom_wait_info {
1488 struct mem_cgroup *memcg;
1492 static int memcg_oom_wake_function(wait_queue_t *wait,
1493 unsigned mode, int sync, void *arg)
1495 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1496 struct mem_cgroup *oom_wait_memcg;
1497 struct oom_wait_info *oom_wait_info;
1499 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1500 oom_wait_memcg = oom_wait_info->memcg;
1502 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1503 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1505 return autoremove_wake_function(wait, mode, sync, arg);
1508 static void memcg_oom_recover(struct mem_cgroup *memcg)
1511 * For the following lockless ->under_oom test, the only required
1512 * guarantee is that it must see the state asserted by an OOM when
1513 * this function is called as a result of userland actions
1514 * triggered by the notification of the OOM. This is trivially
1515 * achieved by invoking mem_cgroup_mark_under_oom() before
1516 * triggering notification.
1518 if (memcg && memcg->under_oom)
1519 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1522 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1524 if (!current->memcg_may_oom)
1527 * We are in the middle of the charge context here, so we
1528 * don't want to block when potentially sitting on a callstack
1529 * that holds all kinds of filesystem and mm locks.
1531 * Also, the caller may handle a failed allocation gracefully
1532 * (like optional page cache readahead) and so an OOM killer
1533 * invocation might not even be necessary.
1535 * That's why we don't do anything here except remember the
1536 * OOM context and then deal with it at the end of the page
1537 * fault when the stack is unwound, the locks are released,
1538 * and when we know whether the fault was overall successful.
1540 css_get(&memcg->css);
1541 current->memcg_in_oom = memcg;
1542 current->memcg_oom_gfp_mask = mask;
1543 current->memcg_oom_order = order;
1547 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1548 * @handle: actually kill/wait or just clean up the OOM state
1550 * This has to be called at the end of a page fault if the memcg OOM
1551 * handler was enabled.
1553 * Memcg supports userspace OOM handling where failed allocations must
1554 * sleep on a waitqueue until the userspace task resolves the
1555 * situation. Sleeping directly in the charge context with all kinds
1556 * of locks held is not a good idea, instead we remember an OOM state
1557 * in the task and mem_cgroup_oom_synchronize() has to be called at
1558 * the end of the page fault to complete the OOM handling.
1560 * Returns %true if an ongoing memcg OOM situation was detected and
1561 * completed, %false otherwise.
1563 bool mem_cgroup_oom_synchronize(bool handle)
1565 struct mem_cgroup *memcg = current->memcg_in_oom;
1566 struct oom_wait_info owait;
1569 /* OOM is global, do not handle */
1576 owait.memcg = memcg;
1577 owait.wait.flags = 0;
1578 owait.wait.func = memcg_oom_wake_function;
1579 owait.wait.private = current;
1580 INIT_LIST_HEAD(&owait.wait.task_list);
1582 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1583 mem_cgroup_mark_under_oom(memcg);
1585 locked = mem_cgroup_oom_trylock(memcg);
1588 mem_cgroup_oom_notify(memcg);
1590 if (locked && !memcg->oom_kill_disable) {
1591 mem_cgroup_unmark_under_oom(memcg);
1592 finish_wait(&memcg_oom_waitq, &owait.wait);
1593 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1594 current->memcg_oom_order);
1597 mem_cgroup_unmark_under_oom(memcg);
1598 finish_wait(&memcg_oom_waitq, &owait.wait);
1602 mem_cgroup_oom_unlock(memcg);
1604 * There is no guarantee that an OOM-lock contender
1605 * sees the wakeups triggered by the OOM kill
1606 * uncharges. Wake any sleepers explicitely.
1608 memcg_oom_recover(memcg);
1611 current->memcg_in_oom = NULL;
1612 css_put(&memcg->css);
1617 * lock_page_memcg - lock a page->mem_cgroup binding
1620 * This function protects unlocked LRU pages from being moved to
1621 * another cgroup and stabilizes their page->mem_cgroup binding.
1623 void lock_page_memcg(struct page *page)
1625 struct mem_cgroup *memcg;
1626 unsigned long flags;
1629 * The RCU lock is held throughout the transaction. The fast
1630 * path can get away without acquiring the memcg->move_lock
1631 * because page moving starts with an RCU grace period.
1635 if (mem_cgroup_disabled())
1638 memcg = page->mem_cgroup;
1639 if (unlikely(!memcg))
1642 if (atomic_read(&memcg->moving_account) <= 0)
1645 spin_lock_irqsave(&memcg->move_lock, flags);
1646 if (memcg != page->mem_cgroup) {
1647 spin_unlock_irqrestore(&memcg->move_lock, flags);
1652 * When charge migration first begins, we can have locked and
1653 * unlocked page stat updates happening concurrently. Track
1654 * the task who has the lock for unlock_page_memcg().
1656 memcg->move_lock_task = current;
1657 memcg->move_lock_flags = flags;
1661 EXPORT_SYMBOL(lock_page_memcg);
1664 * unlock_page_memcg - unlock a page->mem_cgroup binding
1667 void unlock_page_memcg(struct page *page)
1669 struct mem_cgroup *memcg = page->mem_cgroup;
1671 if (memcg && memcg->move_lock_task == current) {
1672 unsigned long flags = memcg->move_lock_flags;
1674 memcg->move_lock_task = NULL;
1675 memcg->move_lock_flags = 0;
1677 spin_unlock_irqrestore(&memcg->move_lock, flags);
1682 EXPORT_SYMBOL(unlock_page_memcg);
1685 * size of first charge trial. "32" comes from vmscan.c's magic value.
1686 * TODO: maybe necessary to use big numbers in big irons.
1688 #define CHARGE_BATCH 32U
1689 struct memcg_stock_pcp {
1690 struct mem_cgroup *cached; /* this never be root cgroup */
1691 unsigned int nr_pages;
1692 struct work_struct work;
1693 unsigned long flags;
1694 #define FLUSHING_CACHED_CHARGE 0
1696 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1697 static DEFINE_MUTEX(percpu_charge_mutex);
1700 * consume_stock: Try to consume stocked charge on this cpu.
1701 * @memcg: memcg to consume from.
1702 * @nr_pages: how many pages to charge.
1704 * The charges will only happen if @memcg matches the current cpu's memcg
1705 * stock, and at least @nr_pages are available in that stock. Failure to
1706 * service an allocation will refill the stock.
1708 * returns true if successful, false otherwise.
1710 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1712 struct memcg_stock_pcp *stock;
1713 unsigned long flags;
1716 if (nr_pages > CHARGE_BATCH)
1719 local_irq_save(flags);
1721 stock = this_cpu_ptr(&memcg_stock);
1722 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1723 stock->nr_pages -= nr_pages;
1727 local_irq_restore(flags);
1733 * Returns stocks cached in percpu and reset cached information.
1735 static void drain_stock(struct memcg_stock_pcp *stock)
1737 struct mem_cgroup *old = stock->cached;
1739 if (stock->nr_pages) {
1740 page_counter_uncharge(&old->memory, stock->nr_pages);
1741 if (do_memsw_account())
1742 page_counter_uncharge(&old->memsw, stock->nr_pages);
1743 css_put_many(&old->css, stock->nr_pages);
1744 stock->nr_pages = 0;
1746 stock->cached = NULL;
1749 static void drain_local_stock(struct work_struct *dummy)
1751 struct memcg_stock_pcp *stock;
1752 unsigned long flags;
1754 local_irq_save(flags);
1756 stock = this_cpu_ptr(&memcg_stock);
1758 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1760 local_irq_restore(flags);
1764 * Cache charges(val) to local per_cpu area.
1765 * This will be consumed by consume_stock() function, later.
1767 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1769 struct memcg_stock_pcp *stock;
1770 unsigned long flags;
1772 local_irq_save(flags);
1774 stock = this_cpu_ptr(&memcg_stock);
1775 if (stock->cached != memcg) { /* reset if necessary */
1777 stock->cached = memcg;
1779 stock->nr_pages += nr_pages;
1781 local_irq_restore(flags);
1785 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1786 * of the hierarchy under it.
1788 static void drain_all_stock(struct mem_cgroup *root_memcg)
1792 /* If someone's already draining, avoid adding running more workers. */
1793 if (!mutex_trylock(&percpu_charge_mutex))
1795 /* Notify other cpus that system-wide "drain" is running */
1798 for_each_online_cpu(cpu) {
1799 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1800 struct mem_cgroup *memcg;
1802 memcg = stock->cached;
1803 if (!memcg || !stock->nr_pages)
1805 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1807 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1809 drain_local_stock(&stock->work);
1811 schedule_work_on(cpu, &stock->work);
1816 mutex_unlock(&percpu_charge_mutex);
1819 static int memcg_hotplug_cpu_dead(unsigned int cpu)
1821 struct memcg_stock_pcp *stock;
1823 stock = &per_cpu(memcg_stock, cpu);
1828 static void reclaim_high(struct mem_cgroup *memcg,
1829 unsigned int nr_pages,
1833 if (page_counter_read(&memcg->memory) <= memcg->high)
1835 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1836 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1837 } while ((memcg = parent_mem_cgroup(memcg)));
1840 static void high_work_func(struct work_struct *work)
1842 struct mem_cgroup *memcg;
1844 memcg = container_of(work, struct mem_cgroup, high_work);
1845 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1849 * Scheduled by try_charge() to be executed from the userland return path
1850 * and reclaims memory over the high limit.
1852 void mem_cgroup_handle_over_high(void)
1854 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1855 struct mem_cgroup *memcg;
1857 if (likely(!nr_pages))
1860 memcg = get_mem_cgroup_from_mm(current->mm);
1861 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1862 css_put(&memcg->css);
1863 current->memcg_nr_pages_over_high = 0;
1866 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1867 unsigned int nr_pages)
1869 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1870 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1871 struct mem_cgroup *mem_over_limit;
1872 struct page_counter *counter;
1873 unsigned long nr_reclaimed;
1874 bool may_swap = true;
1875 bool drained = false;
1877 if (mem_cgroup_is_root(memcg))
1880 if (consume_stock(memcg, nr_pages))
1883 if (!do_memsw_account() ||
1884 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1885 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1887 if (do_memsw_account())
1888 page_counter_uncharge(&memcg->memsw, batch);
1889 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1891 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1895 if (batch > nr_pages) {
1901 * Unlike in global OOM situations, memcg is not in a physical
1902 * memory shortage. Allow dying and OOM-killed tasks to
1903 * bypass the last charges so that they can exit quickly and
1904 * free their memory.
1906 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1907 fatal_signal_pending(current) ||
1908 current->flags & PF_EXITING))
1912 * Prevent unbounded recursion when reclaim operations need to
1913 * allocate memory. This might exceed the limits temporarily,
1914 * but we prefer facilitating memory reclaim and getting back
1915 * under the limit over triggering OOM kills in these cases.
1917 if (unlikely(current->flags & PF_MEMALLOC))
1920 if (unlikely(task_in_memcg_oom(current)))
1923 if (!gfpflags_allow_blocking(gfp_mask))
1926 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1928 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1929 gfp_mask, may_swap);
1931 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1935 drain_all_stock(mem_over_limit);
1940 if (gfp_mask & __GFP_NORETRY)
1943 * Even though the limit is exceeded at this point, reclaim
1944 * may have been able to free some pages. Retry the charge
1945 * before killing the task.
1947 * Only for regular pages, though: huge pages are rather
1948 * unlikely to succeed so close to the limit, and we fall back
1949 * to regular pages anyway in case of failure.
1951 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
1954 * At task move, charge accounts can be doubly counted. So, it's
1955 * better to wait until the end of task_move if something is going on.
1957 if (mem_cgroup_wait_acct_move(mem_over_limit))
1963 if (gfp_mask & __GFP_NOFAIL)
1966 if (fatal_signal_pending(current))
1969 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
1971 mem_cgroup_oom(mem_over_limit, gfp_mask,
1972 get_order(nr_pages * PAGE_SIZE));
1974 if (!(gfp_mask & __GFP_NOFAIL))
1978 * The allocation either can't fail or will lead to more memory
1979 * being freed very soon. Allow memory usage go over the limit
1980 * temporarily by force charging it.
1982 page_counter_charge(&memcg->memory, nr_pages);
1983 if (do_memsw_account())
1984 page_counter_charge(&memcg->memsw, nr_pages);
1985 css_get_many(&memcg->css, nr_pages);
1990 css_get_many(&memcg->css, batch);
1991 if (batch > nr_pages)
1992 refill_stock(memcg, batch - nr_pages);
1995 * If the hierarchy is above the normal consumption range, schedule
1996 * reclaim on returning to userland. We can perform reclaim here
1997 * if __GFP_RECLAIM but let's always punt for simplicity and so that
1998 * GFP_KERNEL can consistently be used during reclaim. @memcg is
1999 * not recorded as it most likely matches current's and won't
2000 * change in the meantime. As high limit is checked again before
2001 * reclaim, the cost of mismatch is negligible.
2004 if (page_counter_read(&memcg->memory) > memcg->high) {
2005 /* Don't bother a random interrupted task */
2006 if (in_interrupt()) {
2007 schedule_work(&memcg->high_work);
2010 current->memcg_nr_pages_over_high += batch;
2011 set_notify_resume(current);
2014 } while ((memcg = parent_mem_cgroup(memcg)));
2019 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2021 if (mem_cgroup_is_root(memcg))
2024 page_counter_uncharge(&memcg->memory, nr_pages);
2025 if (do_memsw_account())
2026 page_counter_uncharge(&memcg->memsw, nr_pages);
2028 css_put_many(&memcg->css, nr_pages);
2031 static void lock_page_lru(struct page *page, int *isolated)
2033 struct zone *zone = page_zone(page);
2035 spin_lock_irq(zone_lru_lock(zone));
2036 if (PageLRU(page)) {
2037 struct lruvec *lruvec;
2039 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2041 del_page_from_lru_list(page, lruvec, page_lru(page));
2047 static void unlock_page_lru(struct page *page, int isolated)
2049 struct zone *zone = page_zone(page);
2052 struct lruvec *lruvec;
2054 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2055 VM_BUG_ON_PAGE(PageLRU(page), page);
2057 add_page_to_lru_list(page, lruvec, page_lru(page));
2059 spin_unlock_irq(zone_lru_lock(zone));
2062 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2067 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2070 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2071 * may already be on some other mem_cgroup's LRU. Take care of it.
2074 lock_page_lru(page, &isolated);
2077 * Nobody should be changing or seriously looking at
2078 * page->mem_cgroup at this point:
2080 * - the page is uncharged
2082 * - the page is off-LRU
2084 * - an anonymous fault has exclusive page access, except for
2085 * a locked page table
2087 * - a page cache insertion, a swapin fault, or a migration
2088 * have the page locked
2090 page->mem_cgroup = memcg;
2093 unlock_page_lru(page, isolated);
2097 static int memcg_alloc_cache_id(void)
2102 id = ida_simple_get(&memcg_cache_ida,
2103 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2107 if (id < memcg_nr_cache_ids)
2111 * There's no space for the new id in memcg_caches arrays,
2112 * so we have to grow them.
2114 down_write(&memcg_cache_ids_sem);
2116 size = 2 * (id + 1);
2117 if (size < MEMCG_CACHES_MIN_SIZE)
2118 size = MEMCG_CACHES_MIN_SIZE;
2119 else if (size > MEMCG_CACHES_MAX_SIZE)
2120 size = MEMCG_CACHES_MAX_SIZE;
2122 err = memcg_update_all_caches(size);
2124 err = memcg_update_all_list_lrus(size);
2126 memcg_nr_cache_ids = size;
2128 up_write(&memcg_cache_ids_sem);
2131 ida_simple_remove(&memcg_cache_ida, id);
2137 static void memcg_free_cache_id(int id)
2139 ida_simple_remove(&memcg_cache_ida, id);
2142 struct memcg_kmem_cache_create_work {
2143 struct mem_cgroup *memcg;
2144 struct kmem_cache *cachep;
2145 struct work_struct work;
2148 static void memcg_kmem_cache_create_func(struct work_struct *w)
2150 struct memcg_kmem_cache_create_work *cw =
2151 container_of(w, struct memcg_kmem_cache_create_work, work);
2152 struct mem_cgroup *memcg = cw->memcg;
2153 struct kmem_cache *cachep = cw->cachep;
2155 memcg_create_kmem_cache(memcg, cachep);
2157 css_put(&memcg->css);
2162 * Enqueue the creation of a per-memcg kmem_cache.
2164 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2165 struct kmem_cache *cachep)
2167 struct memcg_kmem_cache_create_work *cw;
2169 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2173 css_get(&memcg->css);
2176 cw->cachep = cachep;
2177 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2179 queue_work(memcg_kmem_cache_wq, &cw->work);
2182 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2183 struct kmem_cache *cachep)
2186 * We need to stop accounting when we kmalloc, because if the
2187 * corresponding kmalloc cache is not yet created, the first allocation
2188 * in __memcg_schedule_kmem_cache_create will recurse.
2190 * However, it is better to enclose the whole function. Depending on
2191 * the debugging options enabled, INIT_WORK(), for instance, can
2192 * trigger an allocation. This too, will make us recurse. Because at
2193 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2194 * the safest choice is to do it like this, wrapping the whole function.
2196 current->memcg_kmem_skip_account = 1;
2197 __memcg_schedule_kmem_cache_create(memcg, cachep);
2198 current->memcg_kmem_skip_account = 0;
2201 static inline bool memcg_kmem_bypass(void)
2203 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2209 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2210 * @cachep: the original global kmem cache
2212 * Return the kmem_cache we're supposed to use for a slab allocation.
2213 * We try to use the current memcg's version of the cache.
2215 * If the cache does not exist yet, if we are the first user of it, we
2216 * create it asynchronously in a workqueue and let the current allocation
2217 * go through with the original cache.
2219 * This function takes a reference to the cache it returns to assure it
2220 * won't get destroyed while we are working with it. Once the caller is
2221 * done with it, memcg_kmem_put_cache() must be called to release the
2224 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2226 struct mem_cgroup *memcg;
2227 struct kmem_cache *memcg_cachep;
2230 VM_BUG_ON(!is_root_cache(cachep));
2232 if (memcg_kmem_bypass())
2235 if (current->memcg_kmem_skip_account)
2238 memcg = get_mem_cgroup_from_mm(current->mm);
2239 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2243 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2244 if (likely(memcg_cachep))
2245 return memcg_cachep;
2248 * If we are in a safe context (can wait, and not in interrupt
2249 * context), we could be be predictable and return right away.
2250 * This would guarantee that the allocation being performed
2251 * already belongs in the new cache.
2253 * However, there are some clashes that can arrive from locking.
2254 * For instance, because we acquire the slab_mutex while doing
2255 * memcg_create_kmem_cache, this means no further allocation
2256 * could happen with the slab_mutex held. So it's better to
2259 memcg_schedule_kmem_cache_create(memcg, cachep);
2261 css_put(&memcg->css);
2266 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2267 * @cachep: the cache returned by memcg_kmem_get_cache
2269 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2271 if (!is_root_cache(cachep))
2272 css_put(&cachep->memcg_params.memcg->css);
2276 * memcg_kmem_charge: charge a kmem page
2277 * @page: page to charge
2278 * @gfp: reclaim mode
2279 * @order: allocation order
2280 * @memcg: memory cgroup to charge
2282 * Returns 0 on success, an error code on failure.
2284 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2285 struct mem_cgroup *memcg)
2287 unsigned int nr_pages = 1 << order;
2288 struct page_counter *counter;
2291 ret = try_charge(memcg, gfp, nr_pages);
2295 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2296 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2297 cancel_charge(memcg, nr_pages);
2301 page->mem_cgroup = memcg;
2307 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2308 * @page: page to charge
2309 * @gfp: reclaim mode
2310 * @order: allocation order
2312 * Returns 0 on success, an error code on failure.
2314 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2316 struct mem_cgroup *memcg;
2319 if (memcg_kmem_bypass())
2322 memcg = get_mem_cgroup_from_mm(current->mm);
2323 if (!mem_cgroup_is_root(memcg)) {
2324 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2326 __SetPageKmemcg(page);
2328 css_put(&memcg->css);
2332 * memcg_kmem_uncharge: uncharge a kmem page
2333 * @page: page to uncharge
2334 * @order: allocation order
2336 void memcg_kmem_uncharge(struct page *page, int order)
2338 struct mem_cgroup *memcg = page->mem_cgroup;
2339 unsigned int nr_pages = 1 << order;
2344 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2346 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2347 page_counter_uncharge(&memcg->kmem, nr_pages);
2349 page_counter_uncharge(&memcg->memory, nr_pages);
2350 if (do_memsw_account())
2351 page_counter_uncharge(&memcg->memsw, nr_pages);
2353 page->mem_cgroup = NULL;
2355 /* slab pages do not have PageKmemcg flag set */
2356 if (PageKmemcg(page))
2357 __ClearPageKmemcg(page);
2359 css_put_many(&memcg->css, nr_pages);
2361 #endif /* !CONFIG_SLOB */
2363 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2366 * Because tail pages are not marked as "used", set it. We're under
2367 * zone_lru_lock and migration entries setup in all page mappings.
2369 void mem_cgroup_split_huge_fixup(struct page *head)
2373 if (mem_cgroup_disabled())
2376 for (i = 1; i < HPAGE_PMD_NR; i++)
2377 head[i].mem_cgroup = head->mem_cgroup;
2379 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2382 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2384 #ifdef CONFIG_MEMCG_SWAP
2385 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2388 int val = (charge) ? 1 : -1;
2389 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2393 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2394 * @entry: swap entry to be moved
2395 * @from: mem_cgroup which the entry is moved from
2396 * @to: mem_cgroup which the entry is moved to
2398 * It succeeds only when the swap_cgroup's record for this entry is the same
2399 * as the mem_cgroup's id of @from.
2401 * Returns 0 on success, -EINVAL on failure.
2403 * The caller must have charged to @to, IOW, called page_counter_charge() about
2404 * both res and memsw, and called css_get().
2406 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2407 struct mem_cgroup *from, struct mem_cgroup *to)
2409 unsigned short old_id, new_id;
2411 old_id = mem_cgroup_id(from);
2412 new_id = mem_cgroup_id(to);
2414 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2415 mem_cgroup_swap_statistics(from, false);
2416 mem_cgroup_swap_statistics(to, true);
2422 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2423 struct mem_cgroup *from, struct mem_cgroup *to)
2429 static DEFINE_MUTEX(memcg_limit_mutex);
2431 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2432 unsigned long limit)
2434 unsigned long curusage;
2435 unsigned long oldusage;
2436 bool enlarge = false;
2441 * For keeping hierarchical_reclaim simple, how long we should retry
2442 * is depends on callers. We set our retry-count to be function
2443 * of # of children which we should visit in this loop.
2445 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2446 mem_cgroup_count_children(memcg);
2448 oldusage = page_counter_read(&memcg->memory);
2451 if (signal_pending(current)) {
2456 mutex_lock(&memcg_limit_mutex);
2457 if (limit > memcg->memsw.limit) {
2458 mutex_unlock(&memcg_limit_mutex);
2462 if (limit > memcg->memory.limit)
2464 ret = page_counter_limit(&memcg->memory, limit);
2465 mutex_unlock(&memcg_limit_mutex);
2470 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2472 curusage = page_counter_read(&memcg->memory);
2473 /* Usage is reduced ? */
2474 if (curusage >= oldusage)
2477 oldusage = curusage;
2478 } while (retry_count);
2480 if (!ret && enlarge)
2481 memcg_oom_recover(memcg);
2486 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2487 unsigned long limit)
2489 unsigned long curusage;
2490 unsigned long oldusage;
2491 bool enlarge = false;
2495 /* see mem_cgroup_resize_res_limit */
2496 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2497 mem_cgroup_count_children(memcg);
2499 oldusage = page_counter_read(&memcg->memsw);
2502 if (signal_pending(current)) {
2507 mutex_lock(&memcg_limit_mutex);
2508 if (limit < memcg->memory.limit) {
2509 mutex_unlock(&memcg_limit_mutex);
2513 if (limit > memcg->memsw.limit)
2515 ret = page_counter_limit(&memcg->memsw, limit);
2516 mutex_unlock(&memcg_limit_mutex);
2521 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2523 curusage = page_counter_read(&memcg->memsw);
2524 /* Usage is reduced ? */
2525 if (curusage >= oldusage)
2528 oldusage = curusage;
2529 } while (retry_count);
2531 if (!ret && enlarge)
2532 memcg_oom_recover(memcg);
2537 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2539 unsigned long *total_scanned)
2541 unsigned long nr_reclaimed = 0;
2542 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2543 unsigned long reclaimed;
2545 struct mem_cgroup_tree_per_node *mctz;
2546 unsigned long excess;
2547 unsigned long nr_scanned;
2552 mctz = soft_limit_tree_node(pgdat->node_id);
2555 * Do not even bother to check the largest node if the root
2556 * is empty. Do it lockless to prevent lock bouncing. Races
2557 * are acceptable as soft limit is best effort anyway.
2559 if (RB_EMPTY_ROOT(&mctz->rb_root))
2563 * This loop can run a while, specially if mem_cgroup's continuously
2564 * keep exceeding their soft limit and putting the system under
2571 mz = mem_cgroup_largest_soft_limit_node(mctz);
2576 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2577 gfp_mask, &nr_scanned);
2578 nr_reclaimed += reclaimed;
2579 *total_scanned += nr_scanned;
2580 spin_lock_irq(&mctz->lock);
2581 __mem_cgroup_remove_exceeded(mz, mctz);
2584 * If we failed to reclaim anything from this memory cgroup
2585 * it is time to move on to the next cgroup
2589 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2591 excess = soft_limit_excess(mz->memcg);
2593 * One school of thought says that we should not add
2594 * back the node to the tree if reclaim returns 0.
2595 * But our reclaim could return 0, simply because due
2596 * to priority we are exposing a smaller subset of
2597 * memory to reclaim from. Consider this as a longer
2600 /* If excess == 0, no tree ops */
2601 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2602 spin_unlock_irq(&mctz->lock);
2603 css_put(&mz->memcg->css);
2606 * Could not reclaim anything and there are no more
2607 * mem cgroups to try or we seem to be looping without
2608 * reclaiming anything.
2610 if (!nr_reclaimed &&
2612 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2614 } while (!nr_reclaimed);
2616 css_put(&next_mz->memcg->css);
2617 return nr_reclaimed;
2621 * Test whether @memcg has children, dead or alive. Note that this
2622 * function doesn't care whether @memcg has use_hierarchy enabled and
2623 * returns %true if there are child csses according to the cgroup
2624 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2626 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2631 ret = css_next_child(NULL, &memcg->css);
2637 * Reclaims as many pages from the given memcg as possible.
2639 * Caller is responsible for holding css reference for memcg.
2641 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2643 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2645 /* we call try-to-free pages for make this cgroup empty */
2646 lru_add_drain_all();
2647 /* try to free all pages in this cgroup */
2648 while (nr_retries && page_counter_read(&memcg->memory)) {
2651 if (signal_pending(current))
2654 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2658 /* maybe some writeback is necessary */
2659 congestion_wait(BLK_RW_ASYNC, HZ/10);
2667 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2668 char *buf, size_t nbytes,
2671 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2673 if (mem_cgroup_is_root(memcg))
2675 return mem_cgroup_force_empty(memcg) ?: nbytes;
2678 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2681 return mem_cgroup_from_css(css)->use_hierarchy;
2684 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2685 struct cftype *cft, u64 val)
2688 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2689 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2691 if (memcg->use_hierarchy == val)
2695 * If parent's use_hierarchy is set, we can't make any modifications
2696 * in the child subtrees. If it is unset, then the change can
2697 * occur, provided the current cgroup has no children.
2699 * For the root cgroup, parent_mem is NULL, we allow value to be
2700 * set if there are no children.
2702 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2703 (val == 1 || val == 0)) {
2704 if (!memcg_has_children(memcg))
2705 memcg->use_hierarchy = val;
2714 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2716 struct mem_cgroup *iter;
2719 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2721 for_each_mem_cgroup_tree(iter, memcg) {
2722 for (i = 0; i < MEMCG_NR_STAT; i++)
2723 stat[i] += mem_cgroup_read_stat(iter, i);
2727 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2729 struct mem_cgroup *iter;
2732 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2734 for_each_mem_cgroup_tree(iter, memcg) {
2735 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2736 events[i] += mem_cgroup_read_events(iter, i);
2740 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2742 unsigned long val = 0;
2744 if (mem_cgroup_is_root(memcg)) {
2745 struct mem_cgroup *iter;
2747 for_each_mem_cgroup_tree(iter, memcg) {
2748 val += mem_cgroup_read_stat(iter,
2749 MEM_CGROUP_STAT_CACHE);
2750 val += mem_cgroup_read_stat(iter,
2751 MEM_CGROUP_STAT_RSS);
2753 val += mem_cgroup_read_stat(iter,
2754 MEM_CGROUP_STAT_SWAP);
2758 val = page_counter_read(&memcg->memory);
2760 val = page_counter_read(&memcg->memsw);
2773 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2776 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2777 struct page_counter *counter;
2779 switch (MEMFILE_TYPE(cft->private)) {
2781 counter = &memcg->memory;
2784 counter = &memcg->memsw;
2787 counter = &memcg->kmem;
2790 counter = &memcg->tcpmem;
2796 switch (MEMFILE_ATTR(cft->private)) {
2798 if (counter == &memcg->memory)
2799 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2800 if (counter == &memcg->memsw)
2801 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2802 return (u64)page_counter_read(counter) * PAGE_SIZE;
2804 return (u64)counter->limit * PAGE_SIZE;
2806 return (u64)counter->watermark * PAGE_SIZE;
2808 return counter->failcnt;
2809 case RES_SOFT_LIMIT:
2810 return (u64)memcg->soft_limit * PAGE_SIZE;
2817 static int memcg_online_kmem(struct mem_cgroup *memcg)
2821 if (cgroup_memory_nokmem)
2824 BUG_ON(memcg->kmemcg_id >= 0);
2825 BUG_ON(memcg->kmem_state);
2827 memcg_id = memcg_alloc_cache_id();
2831 static_branch_inc(&memcg_kmem_enabled_key);
2833 * A memory cgroup is considered kmem-online as soon as it gets
2834 * kmemcg_id. Setting the id after enabling static branching will
2835 * guarantee no one starts accounting before all call sites are
2838 memcg->kmemcg_id = memcg_id;
2839 memcg->kmem_state = KMEM_ONLINE;
2840 INIT_LIST_HEAD(&memcg->kmem_caches);
2845 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2847 struct cgroup_subsys_state *css;
2848 struct mem_cgroup *parent, *child;
2851 if (memcg->kmem_state != KMEM_ONLINE)
2854 * Clear the online state before clearing memcg_caches array
2855 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2856 * guarantees that no cache will be created for this cgroup
2857 * after we are done (see memcg_create_kmem_cache()).
2859 memcg->kmem_state = KMEM_ALLOCATED;
2861 memcg_deactivate_kmem_caches(memcg);
2863 kmemcg_id = memcg->kmemcg_id;
2864 BUG_ON(kmemcg_id < 0);
2866 parent = parent_mem_cgroup(memcg);
2868 parent = root_mem_cgroup;
2871 * Change kmemcg_id of this cgroup and all its descendants to the
2872 * parent's id, and then move all entries from this cgroup's list_lrus
2873 * to ones of the parent. After we have finished, all list_lrus
2874 * corresponding to this cgroup are guaranteed to remain empty. The
2875 * ordering is imposed by list_lru_node->lock taken by
2876 * memcg_drain_all_list_lrus().
2878 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2879 css_for_each_descendant_pre(css, &memcg->css) {
2880 child = mem_cgroup_from_css(css);
2881 BUG_ON(child->kmemcg_id != kmemcg_id);
2882 child->kmemcg_id = parent->kmemcg_id;
2883 if (!memcg->use_hierarchy)
2888 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2890 memcg_free_cache_id(kmemcg_id);
2893 static void memcg_free_kmem(struct mem_cgroup *memcg)
2895 /* css_alloc() failed, offlining didn't happen */
2896 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2897 memcg_offline_kmem(memcg);
2899 if (memcg->kmem_state == KMEM_ALLOCATED) {
2900 memcg_destroy_kmem_caches(memcg);
2901 static_branch_dec(&memcg_kmem_enabled_key);
2902 WARN_ON(page_counter_read(&memcg->kmem));
2906 static int memcg_online_kmem(struct mem_cgroup *memcg)
2910 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2913 static void memcg_free_kmem(struct mem_cgroup *memcg)
2916 #endif /* !CONFIG_SLOB */
2918 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2919 unsigned long limit)
2923 mutex_lock(&memcg_limit_mutex);
2924 ret = page_counter_limit(&memcg->kmem, limit);
2925 mutex_unlock(&memcg_limit_mutex);
2929 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2933 mutex_lock(&memcg_limit_mutex);
2935 ret = page_counter_limit(&memcg->tcpmem, limit);
2939 if (!memcg->tcpmem_active) {
2941 * The active flag needs to be written after the static_key
2942 * update. This is what guarantees that the socket activation
2943 * function is the last one to run. See mem_cgroup_sk_alloc()
2944 * for details, and note that we don't mark any socket as
2945 * belonging to this memcg until that flag is up.
2947 * We need to do this, because static_keys will span multiple
2948 * sites, but we can't control their order. If we mark a socket
2949 * as accounted, but the accounting functions are not patched in
2950 * yet, we'll lose accounting.
2952 * We never race with the readers in mem_cgroup_sk_alloc(),
2953 * because when this value change, the code to process it is not
2956 static_branch_inc(&memcg_sockets_enabled_key);
2957 memcg->tcpmem_active = true;
2960 mutex_unlock(&memcg_limit_mutex);
2965 * The user of this function is...
2968 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2969 char *buf, size_t nbytes, loff_t off)
2971 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2972 unsigned long nr_pages;
2975 buf = strstrip(buf);
2976 ret = page_counter_memparse(buf, "-1", &nr_pages);
2980 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2982 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2986 switch (MEMFILE_TYPE(of_cft(of)->private)) {
2988 ret = mem_cgroup_resize_limit(memcg, nr_pages);
2991 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
2994 ret = memcg_update_kmem_limit(memcg, nr_pages);
2997 ret = memcg_update_tcp_limit(memcg, nr_pages);
3001 case RES_SOFT_LIMIT:
3002 memcg->soft_limit = nr_pages;
3006 return ret ?: nbytes;
3009 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3010 size_t nbytes, loff_t off)
3012 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3013 struct page_counter *counter;
3015 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3017 counter = &memcg->memory;
3020 counter = &memcg->memsw;
3023 counter = &memcg->kmem;
3026 counter = &memcg->tcpmem;
3032 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3034 page_counter_reset_watermark(counter);
3037 counter->failcnt = 0;
3046 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3049 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3053 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3054 struct cftype *cft, u64 val)
3056 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3058 if (val & ~MOVE_MASK)
3062 * No kind of locking is needed in here, because ->can_attach() will
3063 * check this value once in the beginning of the process, and then carry
3064 * on with stale data. This means that changes to this value will only
3065 * affect task migrations starting after the change.
3067 memcg->move_charge_at_immigrate = val;
3071 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3072 struct cftype *cft, u64 val)
3079 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3083 unsigned int lru_mask;
3086 static const struct numa_stat stats[] = {
3087 { "total", LRU_ALL },
3088 { "file", LRU_ALL_FILE },
3089 { "anon", LRU_ALL_ANON },
3090 { "unevictable", BIT(LRU_UNEVICTABLE) },
3092 const struct numa_stat *stat;
3095 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3097 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3098 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3099 seq_printf(m, "%s=%lu", stat->name, nr);
3100 for_each_node_state(nid, N_MEMORY) {
3101 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3103 seq_printf(m, " N%d=%lu", nid, nr);
3108 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3109 struct mem_cgroup *iter;
3112 for_each_mem_cgroup_tree(iter, memcg)
3113 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3114 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3115 for_each_node_state(nid, N_MEMORY) {
3117 for_each_mem_cgroup_tree(iter, memcg)
3118 nr += mem_cgroup_node_nr_lru_pages(
3119 iter, nid, stat->lru_mask);
3120 seq_printf(m, " N%d=%lu", nid, nr);
3127 #endif /* CONFIG_NUMA */
3129 static int memcg_stat_show(struct seq_file *m, void *v)
3131 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3132 unsigned long memory, memsw;
3133 struct mem_cgroup *mi;
3136 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3137 MEM_CGROUP_STAT_NSTATS);
3138 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3139 MEM_CGROUP_EVENTS_NSTATS);
3140 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3142 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3143 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3145 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3146 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3149 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3150 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3151 mem_cgroup_read_events(memcg, i));
3153 for (i = 0; i < NR_LRU_LISTS; i++)
3154 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3155 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3157 /* Hierarchical information */
3158 memory = memsw = PAGE_COUNTER_MAX;
3159 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3160 memory = min(memory, mi->memory.limit);
3161 memsw = min(memsw, mi->memsw.limit);
3163 seq_printf(m, "hierarchical_memory_limit %llu\n",
3164 (u64)memory * PAGE_SIZE);
3165 if (do_memsw_account())
3166 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3167 (u64)memsw * PAGE_SIZE);
3169 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3170 unsigned long long val = 0;
3172 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3174 for_each_mem_cgroup_tree(mi, memcg)
3175 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3176 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3179 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3180 unsigned long long val = 0;
3182 for_each_mem_cgroup_tree(mi, memcg)
3183 val += mem_cgroup_read_events(mi, i);
3184 seq_printf(m, "total_%s %llu\n",
3185 mem_cgroup_events_names[i], val);
3188 for (i = 0; i < NR_LRU_LISTS; i++) {
3189 unsigned long long val = 0;
3191 for_each_mem_cgroup_tree(mi, memcg)
3192 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3193 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3196 #ifdef CONFIG_DEBUG_VM
3199 struct mem_cgroup_per_node *mz;
3200 struct zone_reclaim_stat *rstat;
3201 unsigned long recent_rotated[2] = {0, 0};
3202 unsigned long recent_scanned[2] = {0, 0};
3204 for_each_online_pgdat(pgdat) {
3205 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3206 rstat = &mz->lruvec.reclaim_stat;
3208 recent_rotated[0] += rstat->recent_rotated[0];
3209 recent_rotated[1] += rstat->recent_rotated[1];
3210 recent_scanned[0] += rstat->recent_scanned[0];
3211 recent_scanned[1] += rstat->recent_scanned[1];
3213 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3214 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3215 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3216 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3223 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3226 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3228 return mem_cgroup_swappiness(memcg);
3231 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3232 struct cftype *cft, u64 val)
3234 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3240 memcg->swappiness = val;
3242 vm_swappiness = val;
3247 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3249 struct mem_cgroup_threshold_ary *t;
3250 unsigned long usage;
3255 t = rcu_dereference(memcg->thresholds.primary);
3257 t = rcu_dereference(memcg->memsw_thresholds.primary);
3262 usage = mem_cgroup_usage(memcg, swap);
3265 * current_threshold points to threshold just below or equal to usage.
3266 * If it's not true, a threshold was crossed after last
3267 * call of __mem_cgroup_threshold().
3269 i = t->current_threshold;
3272 * Iterate backward over array of thresholds starting from
3273 * current_threshold and check if a threshold is crossed.
3274 * If none of thresholds below usage is crossed, we read
3275 * only one element of the array here.
3277 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3278 eventfd_signal(t->entries[i].eventfd, 1);
3280 /* i = current_threshold + 1 */
3284 * Iterate forward over array of thresholds starting from
3285 * current_threshold+1 and check if a threshold is crossed.
3286 * If none of thresholds above usage is crossed, we read
3287 * only one element of the array here.
3289 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3290 eventfd_signal(t->entries[i].eventfd, 1);
3292 /* Update current_threshold */
3293 t->current_threshold = i - 1;
3298 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3301 __mem_cgroup_threshold(memcg, false);
3302 if (do_memsw_account())
3303 __mem_cgroup_threshold(memcg, true);
3305 memcg = parent_mem_cgroup(memcg);
3309 static int compare_thresholds(const void *a, const void *b)
3311 const struct mem_cgroup_threshold *_a = a;
3312 const struct mem_cgroup_threshold *_b = b;
3314 if (_a->threshold > _b->threshold)
3317 if (_a->threshold < _b->threshold)
3323 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3325 struct mem_cgroup_eventfd_list *ev;
3327 spin_lock(&memcg_oom_lock);
3329 list_for_each_entry(ev, &memcg->oom_notify, list)
3330 eventfd_signal(ev->eventfd, 1);
3332 spin_unlock(&memcg_oom_lock);
3336 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3338 struct mem_cgroup *iter;
3340 for_each_mem_cgroup_tree(iter, memcg)
3341 mem_cgroup_oom_notify_cb(iter);
3344 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3345 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3347 struct mem_cgroup_thresholds *thresholds;
3348 struct mem_cgroup_threshold_ary *new;
3349 unsigned long threshold;
3350 unsigned long usage;
3353 ret = page_counter_memparse(args, "-1", &threshold);
3357 mutex_lock(&memcg->thresholds_lock);
3360 thresholds = &memcg->thresholds;
3361 usage = mem_cgroup_usage(memcg, false);
3362 } else if (type == _MEMSWAP) {
3363 thresholds = &memcg->memsw_thresholds;
3364 usage = mem_cgroup_usage(memcg, true);
3368 /* Check if a threshold crossed before adding a new one */
3369 if (thresholds->primary)
3370 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3372 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3374 /* Allocate memory for new array of thresholds */
3375 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3383 /* Copy thresholds (if any) to new array */
3384 if (thresholds->primary) {
3385 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3386 sizeof(struct mem_cgroup_threshold));
3389 /* Add new threshold */
3390 new->entries[size - 1].eventfd = eventfd;
3391 new->entries[size - 1].threshold = threshold;
3393 /* Sort thresholds. Registering of new threshold isn't time-critical */
3394 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3395 compare_thresholds, NULL);
3397 /* Find current threshold */
3398 new->current_threshold = -1;
3399 for (i = 0; i < size; i++) {
3400 if (new->entries[i].threshold <= usage) {
3402 * new->current_threshold will not be used until
3403 * rcu_assign_pointer(), so it's safe to increment
3406 ++new->current_threshold;
3411 /* Free old spare buffer and save old primary buffer as spare */
3412 kfree(thresholds->spare);
3413 thresholds->spare = thresholds->primary;
3415 rcu_assign_pointer(thresholds->primary, new);
3417 /* To be sure that nobody uses thresholds */
3421 mutex_unlock(&memcg->thresholds_lock);
3426 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3427 struct eventfd_ctx *eventfd, const char *args)
3429 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3432 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3433 struct eventfd_ctx *eventfd, const char *args)
3435 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3438 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3439 struct eventfd_ctx *eventfd, enum res_type type)
3441 struct mem_cgroup_thresholds *thresholds;
3442 struct mem_cgroup_threshold_ary *new;
3443 unsigned long usage;
3446 mutex_lock(&memcg->thresholds_lock);
3449 thresholds = &memcg->thresholds;
3450 usage = mem_cgroup_usage(memcg, false);
3451 } else if (type == _MEMSWAP) {
3452 thresholds = &memcg->memsw_thresholds;
3453 usage = mem_cgroup_usage(memcg, true);
3457 if (!thresholds->primary)
3460 /* Check if a threshold crossed before removing */
3461 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3463 /* Calculate new number of threshold */
3465 for (i = 0; i < thresholds->primary->size; i++) {
3466 if (thresholds->primary->entries[i].eventfd != eventfd)
3470 new = thresholds->spare;
3472 /* Set thresholds array to NULL if we don't have thresholds */
3481 /* Copy thresholds and find current threshold */
3482 new->current_threshold = -1;
3483 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3484 if (thresholds->primary->entries[i].eventfd == eventfd)
3487 new->entries[j] = thresholds->primary->entries[i];
3488 if (new->entries[j].threshold <= usage) {
3490 * new->current_threshold will not be used
3491 * until rcu_assign_pointer(), so it's safe to increment
3494 ++new->current_threshold;
3500 /* Swap primary and spare array */
3501 thresholds->spare = thresholds->primary;
3503 rcu_assign_pointer(thresholds->primary, new);
3505 /* To be sure that nobody uses thresholds */
3508 /* If all events are unregistered, free the spare array */
3510 kfree(thresholds->spare);
3511 thresholds->spare = NULL;
3514 mutex_unlock(&memcg->thresholds_lock);
3517 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3518 struct eventfd_ctx *eventfd)
3520 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3523 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3524 struct eventfd_ctx *eventfd)
3526 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3529 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3530 struct eventfd_ctx *eventfd, const char *args)
3532 struct mem_cgroup_eventfd_list *event;
3534 event = kmalloc(sizeof(*event), GFP_KERNEL);
3538 spin_lock(&memcg_oom_lock);
3540 event->eventfd = eventfd;
3541 list_add(&event->list, &memcg->oom_notify);
3543 /* already in OOM ? */
3544 if (memcg->under_oom)
3545 eventfd_signal(eventfd, 1);
3546 spin_unlock(&memcg_oom_lock);
3551 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3552 struct eventfd_ctx *eventfd)
3554 struct mem_cgroup_eventfd_list *ev, *tmp;
3556 spin_lock(&memcg_oom_lock);
3558 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3559 if (ev->eventfd == eventfd) {
3560 list_del(&ev->list);
3565 spin_unlock(&memcg_oom_lock);
3568 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3570 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3572 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3573 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3577 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3578 struct cftype *cft, u64 val)
3580 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3582 /* cannot set to root cgroup and only 0 and 1 are allowed */
3583 if (!css->parent || !((val == 0) || (val == 1)))
3586 memcg->oom_kill_disable = val;
3588 memcg_oom_recover(memcg);
3593 #ifdef CONFIG_CGROUP_WRITEBACK
3595 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3597 return &memcg->cgwb_list;
3600 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3602 return wb_domain_init(&memcg->cgwb_domain, gfp);
3605 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3607 wb_domain_exit(&memcg->cgwb_domain);
3610 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3612 wb_domain_size_changed(&memcg->cgwb_domain);
3615 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3617 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3619 if (!memcg->css.parent)
3622 return &memcg->cgwb_domain;
3626 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3627 * @wb: bdi_writeback in question
3628 * @pfilepages: out parameter for number of file pages
3629 * @pheadroom: out parameter for number of allocatable pages according to memcg
3630 * @pdirty: out parameter for number of dirty pages
3631 * @pwriteback: out parameter for number of pages under writeback
3633 * Determine the numbers of file, headroom, dirty, and writeback pages in
3634 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3635 * is a bit more involved.
3637 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3638 * headroom is calculated as the lowest headroom of itself and the
3639 * ancestors. Note that this doesn't consider the actual amount of
3640 * available memory in the system. The caller should further cap
3641 * *@pheadroom accordingly.
3643 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3644 unsigned long *pheadroom, unsigned long *pdirty,
3645 unsigned long *pwriteback)
3647 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3648 struct mem_cgroup *parent;
3650 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3652 /* this should eventually include NR_UNSTABLE_NFS */
3653 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3654 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3655 (1 << LRU_ACTIVE_FILE));
3656 *pheadroom = PAGE_COUNTER_MAX;
3658 while ((parent = parent_mem_cgroup(memcg))) {
3659 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3660 unsigned long used = page_counter_read(&memcg->memory);
3662 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3667 #else /* CONFIG_CGROUP_WRITEBACK */
3669 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3674 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3678 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3682 #endif /* CONFIG_CGROUP_WRITEBACK */
3685 * DO NOT USE IN NEW FILES.
3687 * "cgroup.event_control" implementation.
3689 * This is way over-engineered. It tries to support fully configurable
3690 * events for each user. Such level of flexibility is completely
3691 * unnecessary especially in the light of the planned unified hierarchy.
3693 * Please deprecate this and replace with something simpler if at all
3698 * Unregister event and free resources.
3700 * Gets called from workqueue.
3702 static void memcg_event_remove(struct work_struct *work)
3704 struct mem_cgroup_event *event =
3705 container_of(work, struct mem_cgroup_event, remove);
3706 struct mem_cgroup *memcg = event->memcg;
3708 remove_wait_queue(event->wqh, &event->wait);
3710 event->unregister_event(memcg, event->eventfd);
3712 /* Notify userspace the event is going away. */
3713 eventfd_signal(event->eventfd, 1);
3715 eventfd_ctx_put(event->eventfd);
3717 css_put(&memcg->css);
3721 * Gets called on POLLHUP on eventfd when user closes it.
3723 * Called with wqh->lock held and interrupts disabled.
3725 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3726 int sync, void *key)
3728 struct mem_cgroup_event *event =
3729 container_of(wait, struct mem_cgroup_event, wait);
3730 struct mem_cgroup *memcg = event->memcg;
3731 unsigned long flags = (unsigned long)key;
3733 if (flags & POLLHUP) {
3735 * If the event has been detached at cgroup removal, we
3736 * can simply return knowing the other side will cleanup
3739 * We can't race against event freeing since the other
3740 * side will require wqh->lock via remove_wait_queue(),
3743 spin_lock(&memcg->event_list_lock);
3744 if (!list_empty(&event->list)) {
3745 list_del_init(&event->list);
3747 * We are in atomic context, but cgroup_event_remove()
3748 * may sleep, so we have to call it in workqueue.
3750 schedule_work(&event->remove);
3752 spin_unlock(&memcg->event_list_lock);
3758 static void memcg_event_ptable_queue_proc(struct file *file,
3759 wait_queue_head_t *wqh, poll_table *pt)
3761 struct mem_cgroup_event *event =
3762 container_of(pt, struct mem_cgroup_event, pt);
3765 add_wait_queue(wqh, &event->wait);
3769 * DO NOT USE IN NEW FILES.
3771 * Parse input and register new cgroup event handler.
3773 * Input must be in format '<event_fd> <control_fd> <args>'.
3774 * Interpretation of args is defined by control file implementation.
3776 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3777 char *buf, size_t nbytes, loff_t off)
3779 struct cgroup_subsys_state *css = of_css(of);
3780 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3781 struct mem_cgroup_event *event;
3782 struct cgroup_subsys_state *cfile_css;
3783 unsigned int efd, cfd;
3790 buf = strstrip(buf);
3792 efd = simple_strtoul(buf, &endp, 10);
3797 cfd = simple_strtoul(buf, &endp, 10);
3798 if ((*endp != ' ') && (*endp != '\0'))
3802 event = kzalloc(sizeof(*event), GFP_KERNEL);
3806 event->memcg = memcg;
3807 INIT_LIST_HEAD(&event->list);
3808 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3809 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3810 INIT_WORK(&event->remove, memcg_event_remove);
3818 event->eventfd = eventfd_ctx_fileget(efile.file);
3819 if (IS_ERR(event->eventfd)) {
3820 ret = PTR_ERR(event->eventfd);
3827 goto out_put_eventfd;
3830 /* the process need read permission on control file */
3831 /* AV: shouldn't we check that it's been opened for read instead? */
3832 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3837 * Determine the event callbacks and set them in @event. This used
3838 * to be done via struct cftype but cgroup core no longer knows
3839 * about these events. The following is crude but the whole thing
3840 * is for compatibility anyway.
3842 * DO NOT ADD NEW FILES.
3844 name = cfile.file->f_path.dentry->d_name.name;
3846 if (!strcmp(name, "memory.usage_in_bytes")) {
3847 event->register_event = mem_cgroup_usage_register_event;
3848 event->unregister_event = mem_cgroup_usage_unregister_event;
3849 } else if (!strcmp(name, "memory.oom_control")) {
3850 event->register_event = mem_cgroup_oom_register_event;
3851 event->unregister_event = mem_cgroup_oom_unregister_event;
3852 } else if (!strcmp(name, "memory.pressure_level")) {
3853 event->register_event = vmpressure_register_event;
3854 event->unregister_event = vmpressure_unregister_event;
3855 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3856 event->register_event = memsw_cgroup_usage_register_event;
3857 event->unregister_event = memsw_cgroup_usage_unregister_event;
3864 * Verify @cfile should belong to @css. Also, remaining events are
3865 * automatically removed on cgroup destruction but the removal is
3866 * asynchronous, so take an extra ref on @css.
3868 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3869 &memory_cgrp_subsys);
3871 if (IS_ERR(cfile_css))
3873 if (cfile_css != css) {
3878 ret = event->register_event(memcg, event->eventfd, buf);
3882 efile.file->f_op->poll(efile.file, &event->pt);
3884 spin_lock(&memcg->event_list_lock);
3885 list_add(&event->list, &memcg->event_list);
3886 spin_unlock(&memcg->event_list_lock);
3898 eventfd_ctx_put(event->eventfd);
3907 static struct cftype mem_cgroup_legacy_files[] = {
3909 .name = "usage_in_bytes",
3910 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3911 .read_u64 = mem_cgroup_read_u64,
3914 .name = "max_usage_in_bytes",
3915 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3916 .write = mem_cgroup_reset,
3917 .read_u64 = mem_cgroup_read_u64,
3920 .name = "limit_in_bytes",
3921 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3922 .write = mem_cgroup_write,
3923 .read_u64 = mem_cgroup_read_u64,
3926 .name = "soft_limit_in_bytes",
3927 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3928 .write = mem_cgroup_write,
3929 .read_u64 = mem_cgroup_read_u64,
3933 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3934 .write = mem_cgroup_reset,
3935 .read_u64 = mem_cgroup_read_u64,
3939 .seq_show = memcg_stat_show,
3942 .name = "force_empty",
3943 .write = mem_cgroup_force_empty_write,
3946 .name = "use_hierarchy",
3947 .write_u64 = mem_cgroup_hierarchy_write,
3948 .read_u64 = mem_cgroup_hierarchy_read,
3951 .name = "cgroup.event_control", /* XXX: for compat */
3952 .write = memcg_write_event_control,
3953 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3956 .name = "swappiness",
3957 .read_u64 = mem_cgroup_swappiness_read,
3958 .write_u64 = mem_cgroup_swappiness_write,
3961 .name = "move_charge_at_immigrate",
3962 .read_u64 = mem_cgroup_move_charge_read,
3963 .write_u64 = mem_cgroup_move_charge_write,
3966 .name = "oom_control",
3967 .seq_show = mem_cgroup_oom_control_read,
3968 .write_u64 = mem_cgroup_oom_control_write,
3969 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3972 .name = "pressure_level",
3976 .name = "numa_stat",
3977 .seq_show = memcg_numa_stat_show,
3981 .name = "kmem.limit_in_bytes",
3982 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
3983 .write = mem_cgroup_write,
3984 .read_u64 = mem_cgroup_read_u64,
3987 .name = "kmem.usage_in_bytes",
3988 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
3989 .read_u64 = mem_cgroup_read_u64,
3992 .name = "kmem.failcnt",
3993 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
3994 .write = mem_cgroup_reset,
3995 .read_u64 = mem_cgroup_read_u64,
3998 .name = "kmem.max_usage_in_bytes",
3999 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4000 .write = mem_cgroup_reset,
4001 .read_u64 = mem_cgroup_read_u64,
4003 #ifdef CONFIG_SLABINFO
4005 .name = "kmem.slabinfo",
4006 .seq_start = memcg_slab_start,
4007 .seq_next = memcg_slab_next,
4008 .seq_stop = memcg_slab_stop,
4009 .seq_show = memcg_slab_show,
4013 .name = "kmem.tcp.limit_in_bytes",
4014 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4015 .write = mem_cgroup_write,
4016 .read_u64 = mem_cgroup_read_u64,
4019 .name = "kmem.tcp.usage_in_bytes",
4020 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4021 .read_u64 = mem_cgroup_read_u64,
4024 .name = "kmem.tcp.failcnt",
4025 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4026 .write = mem_cgroup_reset,
4027 .read_u64 = mem_cgroup_read_u64,
4030 .name = "kmem.tcp.max_usage_in_bytes",
4031 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4032 .write = mem_cgroup_reset,
4033 .read_u64 = mem_cgroup_read_u64,
4035 { }, /* terminate */
4039 * Private memory cgroup IDR
4041 * Swap-out records and page cache shadow entries need to store memcg
4042 * references in constrained space, so we maintain an ID space that is
4043 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4044 * memory-controlled cgroups to 64k.
4046 * However, there usually are many references to the oflline CSS after
4047 * the cgroup has been destroyed, such as page cache or reclaimable
4048 * slab objects, that don't need to hang on to the ID. We want to keep
4049 * those dead CSS from occupying IDs, or we might quickly exhaust the
4050 * relatively small ID space and prevent the creation of new cgroups
4051 * even when there are much fewer than 64k cgroups - possibly none.
4053 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4054 * be freed and recycled when it's no longer needed, which is usually
4055 * when the CSS is offlined.
4057 * The only exception to that are records of swapped out tmpfs/shmem
4058 * pages that need to be attributed to live ancestors on swapin. But
4059 * those references are manageable from userspace.
4062 static DEFINE_IDR(mem_cgroup_idr);
4064 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4066 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4067 atomic_add(n, &memcg->id.ref);
4070 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4072 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4073 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4074 idr_remove(&mem_cgroup_idr, memcg->id.id);
4077 /* Memcg ID pins CSS */
4078 css_put(&memcg->css);
4082 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4084 mem_cgroup_id_get_many(memcg, 1);
4087 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4089 mem_cgroup_id_put_many(memcg, 1);
4093 * mem_cgroup_from_id - look up a memcg from a memcg id
4094 * @id: the memcg id to look up
4096 * Caller must hold rcu_read_lock().
4098 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4100 WARN_ON_ONCE(!rcu_read_lock_held());
4101 return idr_find(&mem_cgroup_idr, id);
4104 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4106 struct mem_cgroup_per_node *pn;
4109 * This routine is called against possible nodes.
4110 * But it's BUG to call kmalloc() against offline node.
4112 * TODO: this routine can waste much memory for nodes which will
4113 * never be onlined. It's better to use memory hotplug callback
4116 if (!node_state(node, N_NORMAL_MEMORY))
4118 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4122 lruvec_init(&pn->lruvec);
4123 pn->usage_in_excess = 0;
4124 pn->on_tree = false;
4127 memcg->nodeinfo[node] = pn;
4131 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4133 kfree(memcg->nodeinfo[node]);
4136 static void mem_cgroup_free(struct mem_cgroup *memcg)
4140 memcg_wb_domain_exit(memcg);
4142 free_mem_cgroup_per_node_info(memcg, node);
4143 free_percpu(memcg->stat);
4147 static struct mem_cgroup *mem_cgroup_alloc(void)
4149 struct mem_cgroup *memcg;
4153 size = sizeof(struct mem_cgroup);
4154 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4156 memcg = kzalloc(size, GFP_KERNEL);
4160 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4161 1, MEM_CGROUP_ID_MAX,
4163 if (memcg->id.id < 0)
4166 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4171 if (alloc_mem_cgroup_per_node_info(memcg, node))
4174 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4177 INIT_WORK(&memcg->high_work, high_work_func);
4178 memcg->last_scanned_node = MAX_NUMNODES;
4179 INIT_LIST_HEAD(&memcg->oom_notify);
4180 mutex_init(&memcg->thresholds_lock);
4181 spin_lock_init(&memcg->move_lock);
4182 vmpressure_init(&memcg->vmpressure);
4183 INIT_LIST_HEAD(&memcg->event_list);
4184 spin_lock_init(&memcg->event_list_lock);
4185 memcg->socket_pressure = jiffies;
4187 memcg->kmemcg_id = -1;
4189 #ifdef CONFIG_CGROUP_WRITEBACK
4190 INIT_LIST_HEAD(&memcg->cgwb_list);
4192 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4195 if (memcg->id.id > 0)
4196 idr_remove(&mem_cgroup_idr, memcg->id.id);
4197 mem_cgroup_free(memcg);
4201 static struct cgroup_subsys_state * __ref
4202 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4204 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4205 struct mem_cgroup *memcg;
4206 long error = -ENOMEM;
4208 memcg = mem_cgroup_alloc();
4210 return ERR_PTR(error);
4212 memcg->high = PAGE_COUNTER_MAX;
4213 memcg->soft_limit = PAGE_COUNTER_MAX;
4215 memcg->swappiness = mem_cgroup_swappiness(parent);
4216 memcg->oom_kill_disable = parent->oom_kill_disable;
4218 if (parent && parent->use_hierarchy) {
4219 memcg->use_hierarchy = true;
4220 page_counter_init(&memcg->memory, &parent->memory);
4221 page_counter_init(&memcg->swap, &parent->swap);
4222 page_counter_init(&memcg->memsw, &parent->memsw);
4223 page_counter_init(&memcg->kmem, &parent->kmem);
4224 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4226 page_counter_init(&memcg->memory, NULL);
4227 page_counter_init(&memcg->swap, NULL);
4228 page_counter_init(&memcg->memsw, NULL);
4229 page_counter_init(&memcg->kmem, NULL);
4230 page_counter_init(&memcg->tcpmem, NULL);
4232 * Deeper hierachy with use_hierarchy == false doesn't make
4233 * much sense so let cgroup subsystem know about this
4234 * unfortunate state in our controller.
4236 if (parent != root_mem_cgroup)
4237 memory_cgrp_subsys.broken_hierarchy = true;
4240 /* The following stuff does not apply to the root */
4242 root_mem_cgroup = memcg;
4246 error = memcg_online_kmem(memcg);
4250 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4251 static_branch_inc(&memcg_sockets_enabled_key);
4255 mem_cgroup_free(memcg);
4256 return ERR_PTR(-ENOMEM);
4259 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4261 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4263 /* Online state pins memcg ID, memcg ID pins CSS */
4264 atomic_set(&memcg->id.ref, 1);
4269 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4271 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4272 struct mem_cgroup_event *event, *tmp;
4275 * Unregister events and notify userspace.
4276 * Notify userspace about cgroup removing only after rmdir of cgroup
4277 * directory to avoid race between userspace and kernelspace.
4279 spin_lock(&memcg->event_list_lock);
4280 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4281 list_del_init(&event->list);
4282 schedule_work(&event->remove);
4284 spin_unlock(&memcg->event_list_lock);
4286 memcg_offline_kmem(memcg);
4287 wb_memcg_offline(memcg);
4289 mem_cgroup_id_put(memcg);
4292 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4294 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4296 invalidate_reclaim_iterators(memcg);
4299 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4301 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4303 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4304 static_branch_dec(&memcg_sockets_enabled_key);
4306 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4307 static_branch_dec(&memcg_sockets_enabled_key);
4309 vmpressure_cleanup(&memcg->vmpressure);
4310 cancel_work_sync(&memcg->high_work);
4311 mem_cgroup_remove_from_trees(memcg);
4312 memcg_free_kmem(memcg);
4313 mem_cgroup_free(memcg);
4317 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4318 * @css: the target css
4320 * Reset the states of the mem_cgroup associated with @css. This is
4321 * invoked when the userland requests disabling on the default hierarchy
4322 * but the memcg is pinned through dependency. The memcg should stop
4323 * applying policies and should revert to the vanilla state as it may be
4324 * made visible again.
4326 * The current implementation only resets the essential configurations.
4327 * This needs to be expanded to cover all the visible parts.
4329 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4331 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4333 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4334 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4335 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4336 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4337 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4339 memcg->high = PAGE_COUNTER_MAX;
4340 memcg->soft_limit = PAGE_COUNTER_MAX;
4341 memcg_wb_domain_size_changed(memcg);
4345 /* Handlers for move charge at task migration. */
4346 static int mem_cgroup_do_precharge(unsigned long count)
4350 /* Try a single bulk charge without reclaim first, kswapd may wake */
4351 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4353 mc.precharge += count;
4357 /* Try charges one by one with reclaim, but do not retry */
4359 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4373 enum mc_target_type {
4379 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4380 unsigned long addr, pte_t ptent)
4382 struct page *page = vm_normal_page(vma, addr, ptent);
4384 if (!page || !page_mapped(page))
4386 if (PageAnon(page)) {
4387 if (!(mc.flags & MOVE_ANON))
4390 if (!(mc.flags & MOVE_FILE))
4393 if (!get_page_unless_zero(page))
4400 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4401 pte_t ptent, swp_entry_t *entry)
4403 struct page *page = NULL;
4404 swp_entry_t ent = pte_to_swp_entry(ptent);
4406 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4409 * Because lookup_swap_cache() updates some statistics counter,
4410 * we call find_get_page() with swapper_space directly.
4412 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4413 if (do_memsw_account())
4414 entry->val = ent.val;
4419 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4420 pte_t ptent, swp_entry_t *entry)
4426 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4427 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4429 struct page *page = NULL;
4430 struct address_space *mapping;
4433 if (!vma->vm_file) /* anonymous vma */
4435 if (!(mc.flags & MOVE_FILE))
4438 mapping = vma->vm_file->f_mapping;
4439 pgoff = linear_page_index(vma, addr);
4441 /* page is moved even if it's not RSS of this task(page-faulted). */
4443 /* shmem/tmpfs may report page out on swap: account for that too. */
4444 if (shmem_mapping(mapping)) {
4445 page = find_get_entry(mapping, pgoff);
4446 if (radix_tree_exceptional_entry(page)) {
4447 swp_entry_t swp = radix_to_swp_entry(page);
4448 if (do_memsw_account())
4450 page = find_get_page(swap_address_space(swp),
4454 page = find_get_page(mapping, pgoff);
4456 page = find_get_page(mapping, pgoff);
4462 * mem_cgroup_move_account - move account of the page
4464 * @compound: charge the page as compound or small page
4465 * @from: mem_cgroup which the page is moved from.
4466 * @to: mem_cgroup which the page is moved to. @from != @to.
4468 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4470 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4473 static int mem_cgroup_move_account(struct page *page,
4475 struct mem_cgroup *from,
4476 struct mem_cgroup *to)
4478 unsigned long flags;
4479 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4483 VM_BUG_ON(from == to);
4484 VM_BUG_ON_PAGE(PageLRU(page), page);
4485 VM_BUG_ON(compound && !PageTransHuge(page));
4488 * Prevent mem_cgroup_migrate() from looking at
4489 * page->mem_cgroup of its source page while we change it.
4492 if (!trylock_page(page))
4496 if (page->mem_cgroup != from)
4499 anon = PageAnon(page);
4501 spin_lock_irqsave(&from->move_lock, flags);
4503 if (!anon && page_mapped(page)) {
4504 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4506 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4511 * move_lock grabbed above and caller set from->moving_account, so
4512 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4513 * So mapping should be stable for dirty pages.
4515 if (!anon && PageDirty(page)) {
4516 struct address_space *mapping = page_mapping(page);
4518 if (mapping_cap_account_dirty(mapping)) {
4519 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4521 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4526 if (PageWriteback(page)) {
4527 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4529 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4534 * It is safe to change page->mem_cgroup here because the page
4535 * is referenced, charged, and isolated - we can't race with
4536 * uncharging, charging, migration, or LRU putback.
4539 /* caller should have done css_get */
4540 page->mem_cgroup = to;
4541 spin_unlock_irqrestore(&from->move_lock, flags);
4545 local_irq_disable();
4546 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4547 memcg_check_events(to, page);
4548 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4549 memcg_check_events(from, page);
4558 * get_mctgt_type - get target type of moving charge
4559 * @vma: the vma the pte to be checked belongs
4560 * @addr: the address corresponding to the pte to be checked
4561 * @ptent: the pte to be checked
4562 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4565 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4566 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4567 * move charge. if @target is not NULL, the page is stored in target->page
4568 * with extra refcnt got(Callers should handle it).
4569 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4570 * target for charge migration. if @target is not NULL, the entry is stored
4573 * Called with pte lock held.
4576 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4577 unsigned long addr, pte_t ptent, union mc_target *target)
4579 struct page *page = NULL;
4580 enum mc_target_type ret = MC_TARGET_NONE;
4581 swp_entry_t ent = { .val = 0 };
4583 if (pte_present(ptent))
4584 page = mc_handle_present_pte(vma, addr, ptent);
4585 else if (is_swap_pte(ptent))
4586 page = mc_handle_swap_pte(vma, ptent, &ent);
4587 else if (pte_none(ptent))
4588 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4590 if (!page && !ent.val)
4594 * Do only loose check w/o serialization.
4595 * mem_cgroup_move_account() checks the page is valid or
4596 * not under LRU exclusion.
4598 if (page->mem_cgroup == mc.from) {
4599 ret = MC_TARGET_PAGE;
4601 target->page = page;
4603 if (!ret || !target)
4606 /* There is a swap entry and a page doesn't exist or isn't charged */
4607 if (ent.val && !ret &&
4608 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4609 ret = MC_TARGET_SWAP;
4616 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4618 * We don't consider swapping or file mapped pages because THP does not
4619 * support them for now.
4620 * Caller should make sure that pmd_trans_huge(pmd) is true.
4622 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4623 unsigned long addr, pmd_t pmd, union mc_target *target)
4625 struct page *page = NULL;
4626 enum mc_target_type ret = MC_TARGET_NONE;
4628 page = pmd_page(pmd);
4629 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4630 if (!(mc.flags & MOVE_ANON))
4632 if (page->mem_cgroup == mc.from) {
4633 ret = MC_TARGET_PAGE;
4636 target->page = page;
4642 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4643 unsigned long addr, pmd_t pmd, union mc_target *target)
4645 return MC_TARGET_NONE;
4649 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4650 unsigned long addr, unsigned long end,
4651 struct mm_walk *walk)
4653 struct vm_area_struct *vma = walk->vma;
4657 ptl = pmd_trans_huge_lock(pmd, vma);
4659 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4660 mc.precharge += HPAGE_PMD_NR;
4665 if (pmd_trans_unstable(pmd))
4667 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4668 for (; addr != end; pte++, addr += PAGE_SIZE)
4669 if (get_mctgt_type(vma, addr, *pte, NULL))
4670 mc.precharge++; /* increment precharge temporarily */
4671 pte_unmap_unlock(pte - 1, ptl);
4677 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4679 unsigned long precharge;
4681 struct mm_walk mem_cgroup_count_precharge_walk = {
4682 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4685 down_read(&mm->mmap_sem);
4686 walk_page_range(0, mm->highest_vm_end,
4687 &mem_cgroup_count_precharge_walk);
4688 up_read(&mm->mmap_sem);
4690 precharge = mc.precharge;
4696 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4698 unsigned long precharge = mem_cgroup_count_precharge(mm);
4700 VM_BUG_ON(mc.moving_task);
4701 mc.moving_task = current;
4702 return mem_cgroup_do_precharge(precharge);
4705 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4706 static void __mem_cgroup_clear_mc(void)
4708 struct mem_cgroup *from = mc.from;
4709 struct mem_cgroup *to = mc.to;
4711 /* we must uncharge all the leftover precharges from mc.to */
4713 cancel_charge(mc.to, mc.precharge);
4717 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4718 * we must uncharge here.
4720 if (mc.moved_charge) {
4721 cancel_charge(mc.from, mc.moved_charge);
4722 mc.moved_charge = 0;
4724 /* we must fixup refcnts and charges */
4725 if (mc.moved_swap) {
4726 /* uncharge swap account from the old cgroup */
4727 if (!mem_cgroup_is_root(mc.from))
4728 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4730 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4733 * we charged both to->memory and to->memsw, so we
4734 * should uncharge to->memory.
4736 if (!mem_cgroup_is_root(mc.to))
4737 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4739 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4740 css_put_many(&mc.to->css, mc.moved_swap);
4744 memcg_oom_recover(from);
4745 memcg_oom_recover(to);
4746 wake_up_all(&mc.waitq);
4749 static void mem_cgroup_clear_mc(void)
4751 struct mm_struct *mm = mc.mm;
4754 * we must clear moving_task before waking up waiters at the end of
4757 mc.moving_task = NULL;
4758 __mem_cgroup_clear_mc();
4759 spin_lock(&mc.lock);
4763 spin_unlock(&mc.lock);
4768 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4770 struct cgroup_subsys_state *css;
4771 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4772 struct mem_cgroup *from;
4773 struct task_struct *leader, *p;
4774 struct mm_struct *mm;
4775 unsigned long move_flags;
4778 /* charge immigration isn't supported on the default hierarchy */
4779 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4783 * Multi-process migrations only happen on the default hierarchy
4784 * where charge immigration is not used. Perform charge
4785 * immigration if @tset contains a leader and whine if there are
4789 cgroup_taskset_for_each_leader(leader, css, tset) {
4792 memcg = mem_cgroup_from_css(css);
4798 * We are now commited to this value whatever it is. Changes in this
4799 * tunable will only affect upcoming migrations, not the current one.
4800 * So we need to save it, and keep it going.
4802 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4806 from = mem_cgroup_from_task(p);
4808 VM_BUG_ON(from == memcg);
4810 mm = get_task_mm(p);
4813 /* We move charges only when we move a owner of the mm */
4814 if (mm->owner == p) {
4817 VM_BUG_ON(mc.precharge);
4818 VM_BUG_ON(mc.moved_charge);
4819 VM_BUG_ON(mc.moved_swap);
4821 spin_lock(&mc.lock);
4825 mc.flags = move_flags;
4826 spin_unlock(&mc.lock);
4827 /* We set mc.moving_task later */
4829 ret = mem_cgroup_precharge_mc(mm);
4831 mem_cgroup_clear_mc();
4838 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4841 mem_cgroup_clear_mc();
4844 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4845 unsigned long addr, unsigned long end,
4846 struct mm_walk *walk)
4849 struct vm_area_struct *vma = walk->vma;
4852 enum mc_target_type target_type;
4853 union mc_target target;
4856 ptl = pmd_trans_huge_lock(pmd, vma);
4858 if (mc.precharge < HPAGE_PMD_NR) {
4862 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4863 if (target_type == MC_TARGET_PAGE) {
4865 if (!isolate_lru_page(page)) {
4866 if (!mem_cgroup_move_account(page, true,
4868 mc.precharge -= HPAGE_PMD_NR;
4869 mc.moved_charge += HPAGE_PMD_NR;
4871 putback_lru_page(page);
4879 if (pmd_trans_unstable(pmd))
4882 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4883 for (; addr != end; addr += PAGE_SIZE) {
4884 pte_t ptent = *(pte++);
4890 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4891 case MC_TARGET_PAGE:
4894 * We can have a part of the split pmd here. Moving it
4895 * can be done but it would be too convoluted so simply
4896 * ignore such a partial THP and keep it in original
4897 * memcg. There should be somebody mapping the head.
4899 if (PageTransCompound(page))
4901 if (isolate_lru_page(page))
4903 if (!mem_cgroup_move_account(page, false,
4906 /* we uncharge from mc.from later. */
4909 putback_lru_page(page);
4910 put: /* get_mctgt_type() gets the page */
4913 case MC_TARGET_SWAP:
4915 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4917 /* we fixup refcnts and charges later. */
4925 pte_unmap_unlock(pte - 1, ptl);
4930 * We have consumed all precharges we got in can_attach().
4931 * We try charge one by one, but don't do any additional
4932 * charges to mc.to if we have failed in charge once in attach()
4935 ret = mem_cgroup_do_precharge(1);
4943 static void mem_cgroup_move_charge(void)
4945 struct mm_walk mem_cgroup_move_charge_walk = {
4946 .pmd_entry = mem_cgroup_move_charge_pte_range,
4950 lru_add_drain_all();
4952 * Signal lock_page_memcg() to take the memcg's move_lock
4953 * while we're moving its pages to another memcg. Then wait
4954 * for already started RCU-only updates to finish.
4956 atomic_inc(&mc.from->moving_account);
4959 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
4961 * Someone who are holding the mmap_sem might be waiting in
4962 * waitq. So we cancel all extra charges, wake up all waiters,
4963 * and retry. Because we cancel precharges, we might not be able
4964 * to move enough charges, but moving charge is a best-effort
4965 * feature anyway, so it wouldn't be a big problem.
4967 __mem_cgroup_clear_mc();
4972 * When we have consumed all precharges and failed in doing
4973 * additional charge, the page walk just aborts.
4975 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
4977 up_read(&mc.mm->mmap_sem);
4978 atomic_dec(&mc.from->moving_account);
4981 static void mem_cgroup_move_task(void)
4984 mem_cgroup_move_charge();
4985 mem_cgroup_clear_mc();
4988 #else /* !CONFIG_MMU */
4989 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4993 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4996 static void mem_cgroup_move_task(void)
5002 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5003 * to verify whether we're attached to the default hierarchy on each mount
5006 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5009 * use_hierarchy is forced on the default hierarchy. cgroup core
5010 * guarantees that @root doesn't have any children, so turning it
5011 * on for the root memcg is enough.
5013 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5014 root_mem_cgroup->use_hierarchy = true;
5016 root_mem_cgroup->use_hierarchy = false;
5019 static u64 memory_current_read(struct cgroup_subsys_state *css,
5022 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5024 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5027 static int memory_low_show(struct seq_file *m, void *v)
5029 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5030 unsigned long low = READ_ONCE(memcg->low);
5032 if (low == PAGE_COUNTER_MAX)
5033 seq_puts(m, "max\n");
5035 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5040 static ssize_t memory_low_write(struct kernfs_open_file *of,
5041 char *buf, size_t nbytes, loff_t off)
5043 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5047 buf = strstrip(buf);
5048 err = page_counter_memparse(buf, "max", &low);
5057 static int memory_high_show(struct seq_file *m, void *v)
5059 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5060 unsigned long high = READ_ONCE(memcg->high);
5062 if (high == PAGE_COUNTER_MAX)
5063 seq_puts(m, "max\n");
5065 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5070 static ssize_t memory_high_write(struct kernfs_open_file *of,
5071 char *buf, size_t nbytes, loff_t off)
5073 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5074 unsigned long nr_pages;
5078 buf = strstrip(buf);
5079 err = page_counter_memparse(buf, "max", &high);
5085 nr_pages = page_counter_read(&memcg->memory);
5086 if (nr_pages > high)
5087 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5090 memcg_wb_domain_size_changed(memcg);
5094 static int memory_max_show(struct seq_file *m, void *v)
5096 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5097 unsigned long max = READ_ONCE(memcg->memory.limit);
5099 if (max == PAGE_COUNTER_MAX)
5100 seq_puts(m, "max\n");
5102 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5107 static ssize_t memory_max_write(struct kernfs_open_file *of,
5108 char *buf, size_t nbytes, loff_t off)
5110 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5111 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5112 bool drained = false;
5116 buf = strstrip(buf);
5117 err = page_counter_memparse(buf, "max", &max);
5121 xchg(&memcg->memory.limit, max);
5124 unsigned long nr_pages = page_counter_read(&memcg->memory);
5126 if (nr_pages <= max)
5129 if (signal_pending(current)) {
5135 drain_all_stock(memcg);
5141 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5147 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5148 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5152 memcg_wb_domain_size_changed(memcg);
5156 static int memory_events_show(struct seq_file *m, void *v)
5158 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5160 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5161 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5162 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5163 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5168 static int memory_stat_show(struct seq_file *m, void *v)
5170 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5171 unsigned long stat[MEMCG_NR_STAT];
5172 unsigned long events[MEMCG_NR_EVENTS];
5176 * Provide statistics on the state of the memory subsystem as
5177 * well as cumulative event counters that show past behavior.
5179 * This list is ordered following a combination of these gradients:
5180 * 1) generic big picture -> specifics and details
5181 * 2) reflecting userspace activity -> reflecting kernel heuristics
5183 * Current memory state:
5186 tree_stat(memcg, stat);
5187 tree_events(memcg, events);
5189 seq_printf(m, "anon %llu\n",
5190 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5191 seq_printf(m, "file %llu\n",
5192 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5193 seq_printf(m, "kernel_stack %llu\n",
5194 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5195 seq_printf(m, "slab %llu\n",
5196 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5197 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5198 seq_printf(m, "sock %llu\n",
5199 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5201 seq_printf(m, "file_mapped %llu\n",
5202 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5203 seq_printf(m, "file_dirty %llu\n",
5204 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5205 seq_printf(m, "file_writeback %llu\n",
5206 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5208 for (i = 0; i < NR_LRU_LISTS; i++) {
5209 struct mem_cgroup *mi;
5210 unsigned long val = 0;
5212 for_each_mem_cgroup_tree(mi, memcg)
5213 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5214 seq_printf(m, "%s %llu\n",
5215 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5218 seq_printf(m, "slab_reclaimable %llu\n",
5219 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5220 seq_printf(m, "slab_unreclaimable %llu\n",
5221 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5223 /* Accumulated memory events */
5225 seq_printf(m, "pgfault %lu\n",
5226 events[MEM_CGROUP_EVENTS_PGFAULT]);
5227 seq_printf(m, "pgmajfault %lu\n",
5228 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5233 static struct cftype memory_files[] = {
5236 .flags = CFTYPE_NOT_ON_ROOT,
5237 .read_u64 = memory_current_read,
5241 .flags = CFTYPE_NOT_ON_ROOT,
5242 .seq_show = memory_low_show,
5243 .write = memory_low_write,
5247 .flags = CFTYPE_NOT_ON_ROOT,
5248 .seq_show = memory_high_show,
5249 .write = memory_high_write,
5253 .flags = CFTYPE_NOT_ON_ROOT,
5254 .seq_show = memory_max_show,
5255 .write = memory_max_write,
5259 .flags = CFTYPE_NOT_ON_ROOT,
5260 .file_offset = offsetof(struct mem_cgroup, events_file),
5261 .seq_show = memory_events_show,
5265 .flags = CFTYPE_NOT_ON_ROOT,
5266 .seq_show = memory_stat_show,
5271 struct cgroup_subsys memory_cgrp_subsys = {
5272 .css_alloc = mem_cgroup_css_alloc,
5273 .css_online = mem_cgroup_css_online,
5274 .css_offline = mem_cgroup_css_offline,
5275 .css_released = mem_cgroup_css_released,
5276 .css_free = mem_cgroup_css_free,
5277 .css_reset = mem_cgroup_css_reset,
5278 .can_attach = mem_cgroup_can_attach,
5279 .cancel_attach = mem_cgroup_cancel_attach,
5280 .post_attach = mem_cgroup_move_task,
5281 .bind = mem_cgroup_bind,
5282 .dfl_cftypes = memory_files,
5283 .legacy_cftypes = mem_cgroup_legacy_files,
5288 * mem_cgroup_low - check if memory consumption is below the normal range
5289 * @root: the highest ancestor to consider
5290 * @memcg: the memory cgroup to check
5292 * Returns %true if memory consumption of @memcg, and that of all
5293 * configurable ancestors up to @root, is below the normal range.
5295 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5297 if (mem_cgroup_disabled())
5301 * The toplevel group doesn't have a configurable range, so
5302 * it's never low when looked at directly, and it is not
5303 * considered an ancestor when assessing the hierarchy.
5306 if (memcg == root_mem_cgroup)
5309 if (page_counter_read(&memcg->memory) >= memcg->low)
5312 while (memcg != root) {
5313 memcg = parent_mem_cgroup(memcg);
5315 if (memcg == root_mem_cgroup)
5318 if (page_counter_read(&memcg->memory) >= memcg->low)
5325 * mem_cgroup_try_charge - try charging a page
5326 * @page: page to charge
5327 * @mm: mm context of the victim
5328 * @gfp_mask: reclaim mode
5329 * @memcgp: charged memcg return
5330 * @compound: charge the page as compound or small page
5332 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5333 * pages according to @gfp_mask if necessary.
5335 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5336 * Otherwise, an error code is returned.
5338 * After page->mapping has been set up, the caller must finalize the
5339 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5340 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5342 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5343 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5346 struct mem_cgroup *memcg = NULL;
5347 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5350 if (mem_cgroup_disabled())
5353 if (PageSwapCache(page)) {
5355 * Every swap fault against a single page tries to charge the
5356 * page, bail as early as possible. shmem_unuse() encounters
5357 * already charged pages, too. The USED bit is protected by
5358 * the page lock, which serializes swap cache removal, which
5359 * in turn serializes uncharging.
5361 VM_BUG_ON_PAGE(!PageLocked(page), page);
5362 if (page->mem_cgroup)
5365 if (do_swap_account) {
5366 swp_entry_t ent = { .val = page_private(page), };
5367 unsigned short id = lookup_swap_cgroup_id(ent);
5370 memcg = mem_cgroup_from_id(id);
5371 if (memcg && !css_tryget_online(&memcg->css))
5378 memcg = get_mem_cgroup_from_mm(mm);
5380 ret = try_charge(memcg, gfp_mask, nr_pages);
5382 css_put(&memcg->css);
5389 * mem_cgroup_commit_charge - commit a page charge
5390 * @page: page to charge
5391 * @memcg: memcg to charge the page to
5392 * @lrucare: page might be on LRU already
5393 * @compound: charge the page as compound or small page
5395 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5396 * after page->mapping has been set up. This must happen atomically
5397 * as part of the page instantiation, i.e. under the page table lock
5398 * for anonymous pages, under the page lock for page and swap cache.
5400 * In addition, the page must not be on the LRU during the commit, to
5401 * prevent racing with task migration. If it might be, use @lrucare.
5403 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5405 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5406 bool lrucare, bool compound)
5408 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5410 VM_BUG_ON_PAGE(!page->mapping, page);
5411 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5413 if (mem_cgroup_disabled())
5416 * Swap faults will attempt to charge the same page multiple
5417 * times. But reuse_swap_page() might have removed the page
5418 * from swapcache already, so we can't check PageSwapCache().
5423 commit_charge(page, memcg, lrucare);
5425 local_irq_disable();
5426 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5427 memcg_check_events(memcg, page);
5430 if (do_memsw_account() && PageSwapCache(page)) {
5431 swp_entry_t entry = { .val = page_private(page) };
5433 * The swap entry might not get freed for a long time,
5434 * let's not wait for it. The page already received a
5435 * memory+swap charge, drop the swap entry duplicate.
5437 mem_cgroup_uncharge_swap(entry);
5442 * mem_cgroup_cancel_charge - cancel a page charge
5443 * @page: page to charge
5444 * @memcg: memcg to charge the page to
5445 * @compound: charge the page as compound or small page
5447 * Cancel a charge transaction started by mem_cgroup_try_charge().
5449 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5452 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5454 if (mem_cgroup_disabled())
5457 * Swap faults will attempt to charge the same page multiple
5458 * times. But reuse_swap_page() might have removed the page
5459 * from swapcache already, so we can't check PageSwapCache().
5464 cancel_charge(memcg, nr_pages);
5467 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5468 unsigned long nr_anon, unsigned long nr_file,
5469 unsigned long nr_huge, unsigned long nr_kmem,
5470 struct page *dummy_page)
5472 unsigned long nr_pages = nr_anon + nr_file + nr_kmem;
5473 unsigned long flags;
5475 if (!mem_cgroup_is_root(memcg)) {
5476 page_counter_uncharge(&memcg->memory, nr_pages);
5477 if (do_memsw_account())
5478 page_counter_uncharge(&memcg->memsw, nr_pages);
5479 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && nr_kmem)
5480 page_counter_uncharge(&memcg->kmem, nr_kmem);
5481 memcg_oom_recover(memcg);
5484 local_irq_save(flags);
5485 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5486 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5487 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5488 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5489 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5490 memcg_check_events(memcg, dummy_page);
5491 local_irq_restore(flags);
5493 if (!mem_cgroup_is_root(memcg))
5494 css_put_many(&memcg->css, nr_pages);
5497 static void uncharge_list(struct list_head *page_list)
5499 struct mem_cgroup *memcg = NULL;
5500 unsigned long nr_anon = 0;
5501 unsigned long nr_file = 0;
5502 unsigned long nr_huge = 0;
5503 unsigned long nr_kmem = 0;
5504 unsigned long pgpgout = 0;
5505 struct list_head *next;
5509 * Note that the list can be a single page->lru; hence the
5510 * do-while loop instead of a simple list_for_each_entry().
5512 next = page_list->next;
5514 page = list_entry(next, struct page, lru);
5515 next = page->lru.next;
5517 VM_BUG_ON_PAGE(PageLRU(page), page);
5518 VM_BUG_ON_PAGE(page_count(page), page);
5520 if (!page->mem_cgroup)
5524 * Nobody should be changing or seriously looking at
5525 * page->mem_cgroup at this point, we have fully
5526 * exclusive access to the page.
5529 if (memcg != page->mem_cgroup) {
5531 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5532 nr_huge, nr_kmem, page);
5533 pgpgout = nr_anon = nr_file =
5534 nr_huge = nr_kmem = 0;
5536 memcg = page->mem_cgroup;
5539 if (!PageKmemcg(page)) {
5540 unsigned int nr_pages = 1;
5542 if (PageTransHuge(page)) {
5543 nr_pages <<= compound_order(page);
5544 nr_huge += nr_pages;
5547 nr_anon += nr_pages;
5549 nr_file += nr_pages;
5552 nr_kmem += 1 << compound_order(page);
5553 __ClearPageKmemcg(page);
5556 page->mem_cgroup = NULL;
5557 } while (next != page_list);
5560 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5561 nr_huge, nr_kmem, page);
5565 * mem_cgroup_uncharge - uncharge a page
5566 * @page: page to uncharge
5568 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5569 * mem_cgroup_commit_charge().
5571 void mem_cgroup_uncharge(struct page *page)
5573 if (mem_cgroup_disabled())
5576 /* Don't touch page->lru of any random page, pre-check: */
5577 if (!page->mem_cgroup)
5580 INIT_LIST_HEAD(&page->lru);
5581 uncharge_list(&page->lru);
5585 * mem_cgroup_uncharge_list - uncharge a list of page
5586 * @page_list: list of pages to uncharge
5588 * Uncharge a list of pages previously charged with
5589 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5591 void mem_cgroup_uncharge_list(struct list_head *page_list)
5593 if (mem_cgroup_disabled())
5596 if (!list_empty(page_list))
5597 uncharge_list(page_list);
5601 * mem_cgroup_migrate - charge a page's replacement
5602 * @oldpage: currently circulating page
5603 * @newpage: replacement page
5605 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5606 * be uncharged upon free.
5608 * Both pages must be locked, @newpage->mapping must be set up.
5610 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5612 struct mem_cgroup *memcg;
5613 unsigned int nr_pages;
5615 unsigned long flags;
5617 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5618 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5619 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5620 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5623 if (mem_cgroup_disabled())
5626 /* Page cache replacement: new page already charged? */
5627 if (newpage->mem_cgroup)
5630 /* Swapcache readahead pages can get replaced before being charged */
5631 memcg = oldpage->mem_cgroup;
5635 /* Force-charge the new page. The old one will be freed soon */
5636 compound = PageTransHuge(newpage);
5637 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5639 page_counter_charge(&memcg->memory, nr_pages);
5640 if (do_memsw_account())
5641 page_counter_charge(&memcg->memsw, nr_pages);
5642 css_get_many(&memcg->css, nr_pages);
5644 commit_charge(newpage, memcg, false);
5646 local_irq_save(flags);
5647 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5648 memcg_check_events(memcg, newpage);
5649 local_irq_restore(flags);
5652 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5653 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5655 void mem_cgroup_sk_alloc(struct sock *sk)
5657 struct mem_cgroup *memcg;
5659 if (!mem_cgroup_sockets_enabled)
5663 * Socket cloning can throw us here with sk_memcg already
5664 * filled. It won't however, necessarily happen from
5665 * process context. So the test for root memcg given
5666 * the current task's memcg won't help us in this case.
5668 * Respecting the original socket's memcg is a better
5669 * decision in this case.
5672 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5673 css_get(&sk->sk_memcg->css);
5678 memcg = mem_cgroup_from_task(current);
5679 if (memcg == root_mem_cgroup)
5681 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5683 if (css_tryget_online(&memcg->css))
5684 sk->sk_memcg = memcg;
5689 void mem_cgroup_sk_free(struct sock *sk)
5692 css_put(&sk->sk_memcg->css);
5696 * mem_cgroup_charge_skmem - charge socket memory
5697 * @memcg: memcg to charge
5698 * @nr_pages: number of pages to charge
5700 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5701 * @memcg's configured limit, %false if the charge had to be forced.
5703 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5705 gfp_t gfp_mask = GFP_KERNEL;
5707 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5708 struct page_counter *fail;
5710 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5711 memcg->tcpmem_pressure = 0;
5714 page_counter_charge(&memcg->tcpmem, nr_pages);
5715 memcg->tcpmem_pressure = 1;
5719 /* Don't block in the packet receive path */
5721 gfp_mask = GFP_NOWAIT;
5723 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5725 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5728 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5733 * mem_cgroup_uncharge_skmem - uncharge socket memory
5734 * @memcg - memcg to uncharge
5735 * @nr_pages - number of pages to uncharge
5737 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5739 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5740 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5744 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5746 page_counter_uncharge(&memcg->memory, nr_pages);
5747 css_put_many(&memcg->css, nr_pages);
5750 static int __init cgroup_memory(char *s)
5754 while ((token = strsep(&s, ",")) != NULL) {
5757 if (!strcmp(token, "nosocket"))
5758 cgroup_memory_nosocket = true;
5759 if (!strcmp(token, "nokmem"))
5760 cgroup_memory_nokmem = true;
5764 __setup("cgroup.memory=", cgroup_memory);
5767 * subsys_initcall() for memory controller.
5769 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
5770 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
5771 * basically everything that doesn't depend on a specific mem_cgroup structure
5772 * should be initialized from here.
5774 static int __init mem_cgroup_init(void)
5780 * Kmem cache creation is mostly done with the slab_mutex held,
5781 * so use a workqueue with limited concurrency to avoid stalling
5782 * all worker threads in case lots of cgroups are created and
5783 * destroyed simultaneously.
5785 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
5786 BUG_ON(!memcg_kmem_cache_wq);
5789 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
5790 memcg_hotplug_cpu_dead);
5792 for_each_possible_cpu(cpu)
5793 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5796 for_each_node(node) {
5797 struct mem_cgroup_tree_per_node *rtpn;
5799 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5800 node_online(node) ? node : NUMA_NO_NODE);
5802 rtpn->rb_root = RB_ROOT;
5803 spin_lock_init(&rtpn->lock);
5804 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5809 subsys_initcall(mem_cgroup_init);
5811 #ifdef CONFIG_MEMCG_SWAP
5812 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
5814 while (!atomic_inc_not_zero(&memcg->id.ref)) {
5816 * The root cgroup cannot be destroyed, so it's refcount must
5819 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
5823 memcg = parent_mem_cgroup(memcg);
5825 memcg = root_mem_cgroup;
5831 * mem_cgroup_swapout - transfer a memsw charge to swap
5832 * @page: page whose memsw charge to transfer
5833 * @entry: swap entry to move the charge to
5835 * Transfer the memsw charge of @page to @entry.
5837 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5839 struct mem_cgroup *memcg, *swap_memcg;
5840 unsigned short oldid;
5842 VM_BUG_ON_PAGE(PageLRU(page), page);
5843 VM_BUG_ON_PAGE(page_count(page), page);
5845 if (!do_memsw_account())
5848 memcg = page->mem_cgroup;
5850 /* Readahead page, never charged */
5855 * In case the memcg owning these pages has been offlined and doesn't
5856 * have an ID allocated to it anymore, charge the closest online
5857 * ancestor for the swap instead and transfer the memory+swap charge.
5859 swap_memcg = mem_cgroup_id_get_online(memcg);
5860 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg));
5861 VM_BUG_ON_PAGE(oldid, page);
5862 mem_cgroup_swap_statistics(swap_memcg, true);
5864 page->mem_cgroup = NULL;
5866 if (!mem_cgroup_is_root(memcg))
5867 page_counter_uncharge(&memcg->memory, 1);
5869 if (memcg != swap_memcg) {
5870 if (!mem_cgroup_is_root(swap_memcg))
5871 page_counter_charge(&swap_memcg->memsw, 1);
5872 page_counter_uncharge(&memcg->memsw, 1);
5876 * Interrupts should be disabled here because the caller holds the
5877 * mapping->tree_lock lock which is taken with interrupts-off. It is
5878 * important here to have the interrupts disabled because it is the
5879 * only synchronisation we have for udpating the per-CPU variables.
5881 VM_BUG_ON(!irqs_disabled());
5882 mem_cgroup_charge_statistics(memcg, page, false, -1);
5883 memcg_check_events(memcg, page);
5885 if (!mem_cgroup_is_root(memcg))
5886 css_put(&memcg->css);
5890 * mem_cgroup_try_charge_swap - try charging a swap entry
5891 * @page: page being added to swap
5892 * @entry: swap entry to charge
5894 * Try to charge @entry to the memcg that @page belongs to.
5896 * Returns 0 on success, -ENOMEM on failure.
5898 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5900 struct mem_cgroup *memcg;
5901 struct page_counter *counter;
5902 unsigned short oldid;
5904 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5907 memcg = page->mem_cgroup;
5909 /* Readahead page, never charged */
5913 memcg = mem_cgroup_id_get_online(memcg);
5915 if (!mem_cgroup_is_root(memcg) &&
5916 !page_counter_try_charge(&memcg->swap, 1, &counter)) {
5917 mem_cgroup_id_put(memcg);
5921 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5922 VM_BUG_ON_PAGE(oldid, page);
5923 mem_cgroup_swap_statistics(memcg, true);
5929 * mem_cgroup_uncharge_swap - uncharge a swap entry
5930 * @entry: swap entry to uncharge
5932 * Drop the swap charge associated with @entry.
5934 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5936 struct mem_cgroup *memcg;
5939 if (!do_swap_account)
5942 id = swap_cgroup_record(entry, 0);
5944 memcg = mem_cgroup_from_id(id);
5946 if (!mem_cgroup_is_root(memcg)) {
5947 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5948 page_counter_uncharge(&memcg->swap, 1);
5950 page_counter_uncharge(&memcg->memsw, 1);
5952 mem_cgroup_swap_statistics(memcg, false);
5953 mem_cgroup_id_put(memcg);
5958 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
5960 long nr_swap_pages = get_nr_swap_pages();
5962 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5963 return nr_swap_pages;
5964 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5965 nr_swap_pages = min_t(long, nr_swap_pages,
5966 READ_ONCE(memcg->swap.limit) -
5967 page_counter_read(&memcg->swap));
5968 return nr_swap_pages;
5971 bool mem_cgroup_swap_full(struct page *page)
5973 struct mem_cgroup *memcg;
5975 VM_BUG_ON_PAGE(!PageLocked(page), page);
5979 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5982 memcg = page->mem_cgroup;
5986 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5987 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
5993 /* for remember boot option*/
5994 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5995 static int really_do_swap_account __initdata = 1;
5997 static int really_do_swap_account __initdata;
6000 static int __init enable_swap_account(char *s)
6002 if (!strcmp(s, "1"))
6003 really_do_swap_account = 1;
6004 else if (!strcmp(s, "0"))
6005 really_do_swap_account = 0;
6008 __setup("swapaccount=", enable_swap_account);
6010 static u64 swap_current_read(struct cgroup_subsys_state *css,
6013 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6015 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6018 static int swap_max_show(struct seq_file *m, void *v)
6020 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6021 unsigned long max = READ_ONCE(memcg->swap.limit);
6023 if (max == PAGE_COUNTER_MAX)
6024 seq_puts(m, "max\n");
6026 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6031 static ssize_t swap_max_write(struct kernfs_open_file *of,
6032 char *buf, size_t nbytes, loff_t off)
6034 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6038 buf = strstrip(buf);
6039 err = page_counter_memparse(buf, "max", &max);
6043 mutex_lock(&memcg_limit_mutex);
6044 err = page_counter_limit(&memcg->swap, max);
6045 mutex_unlock(&memcg_limit_mutex);
6052 static struct cftype swap_files[] = {
6054 .name = "swap.current",
6055 .flags = CFTYPE_NOT_ON_ROOT,
6056 .read_u64 = swap_current_read,
6060 .flags = CFTYPE_NOT_ON_ROOT,
6061 .seq_show = swap_max_show,
6062 .write = swap_max_write,
6067 static struct cftype memsw_cgroup_files[] = {
6069 .name = "memsw.usage_in_bytes",
6070 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6071 .read_u64 = mem_cgroup_read_u64,
6074 .name = "memsw.max_usage_in_bytes",
6075 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6076 .write = mem_cgroup_reset,
6077 .read_u64 = mem_cgroup_read_u64,
6080 .name = "memsw.limit_in_bytes",
6081 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6082 .write = mem_cgroup_write,
6083 .read_u64 = mem_cgroup_read_u64,
6086 .name = "memsw.failcnt",
6087 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6088 .write = mem_cgroup_reset,
6089 .read_u64 = mem_cgroup_read_u64,
6091 { }, /* terminate */
6094 static int __init mem_cgroup_swap_init(void)
6096 if (!mem_cgroup_disabled() && really_do_swap_account) {
6097 do_swap_account = 1;
6098 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6100 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6101 memsw_cgroup_files));
6105 subsys_initcall(mem_cgroup_swap_init);
6107 #endif /* CONFIG_MEMCG_SWAP */