1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/sched/mm.h>
39 #include <linux/shmem_fs.h>
40 #include <linux/hugetlb.h>
41 #include <linux/pagemap.h>
42 #include <linux/smp.h>
43 #include <linux/page-flags.h>
44 #include <linux/backing-dev.h>
45 #include <linux/bit_spinlock.h>
46 #include <linux/rcupdate.h>
47 #include <linux/limits.h>
48 #include <linux/export.h>
49 #include <linux/mutex.h>
50 #include <linux/rbtree.h>
51 #include <linux/slab.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/spinlock.h>
55 #include <linux/eventfd.h>
56 #include <linux/poll.h>
57 #include <linux/sort.h>
59 #include <linux/seq_file.h>
60 #include <linux/vmpressure.h>
61 #include <linux/mm_inline.h>
62 #include <linux/swap_cgroup.h>
63 #include <linux/cpu.h>
64 #include <linux/oom.h>
65 #include <linux/lockdep.h>
66 #include <linux/file.h>
67 #include <linux/tracehook.h>
73 #include <linux/uaccess.h>
75 #include <trace/events/vmscan.h>
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
78 EXPORT_SYMBOL(memory_cgrp_subsys);
80 struct mem_cgroup *root_mem_cgroup __read_mostly;
82 #define MEM_CGROUP_RECLAIM_RETRIES 5
84 /* Socket memory accounting disabled? */
85 static bool cgroup_memory_nosocket;
87 /* Kernel memory accounting disabled? */
88 static bool cgroup_memory_nokmem;
90 /* Whether the swap controller is active */
91 #ifdef CONFIG_MEMCG_SWAP
92 int do_swap_account __read_mostly;
94 #define do_swap_account 0
97 /* Whether legacy memory+swap accounting is active */
98 static bool do_memsw_account(void)
100 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
103 static const char *const mem_cgroup_lru_names[] = {
111 #define THRESHOLDS_EVENTS_TARGET 128
112 #define SOFTLIMIT_EVENTS_TARGET 1024
113 #define NUMAINFO_EVENTS_TARGET 1024
116 * Cgroups above their limits are maintained in a RB-Tree, independent of
117 * their hierarchy representation
120 struct mem_cgroup_tree_per_node {
121 struct rb_root rb_root;
122 struct rb_node *rb_rightmost;
126 struct mem_cgroup_tree {
127 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
130 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
133 struct mem_cgroup_eventfd_list {
134 struct list_head list;
135 struct eventfd_ctx *eventfd;
139 * cgroup_event represents events which userspace want to receive.
141 struct mem_cgroup_event {
143 * memcg which the event belongs to.
145 struct mem_cgroup *memcg;
147 * eventfd to signal userspace about the event.
149 struct eventfd_ctx *eventfd;
151 * Each of these stored in a list by the cgroup.
153 struct list_head list;
155 * register_event() callback will be used to add new userspace
156 * waiter for changes related to this event. Use eventfd_signal()
157 * on eventfd to send notification to userspace.
159 int (*register_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd, const char *args);
162 * unregister_event() callback will be called when userspace closes
163 * the eventfd or on cgroup removing. This callback must be set,
164 * if you want provide notification functionality.
166 void (*unregister_event)(struct mem_cgroup *memcg,
167 struct eventfd_ctx *eventfd);
169 * All fields below needed to unregister event when
170 * userspace closes eventfd.
173 wait_queue_head_t *wqh;
174 wait_queue_entry_t wait;
175 struct work_struct remove;
178 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
179 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
181 /* Stuffs for move charges at task migration. */
183 * Types of charges to be moved.
185 #define MOVE_ANON 0x1U
186 #define MOVE_FILE 0x2U
187 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
189 /* "mc" and its members are protected by cgroup_mutex */
190 static struct move_charge_struct {
191 spinlock_t lock; /* for from, to */
192 struct mm_struct *mm;
193 struct mem_cgroup *from;
194 struct mem_cgroup *to;
196 unsigned long precharge;
197 unsigned long moved_charge;
198 unsigned long moved_swap;
199 struct task_struct *moving_task; /* a task moving charges */
200 wait_queue_head_t waitq; /* a waitq for other context */
202 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
203 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
207 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
208 * limit reclaim to prevent infinite loops, if they ever occur.
210 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
211 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
214 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
215 MEM_CGROUP_CHARGE_TYPE_ANON,
216 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
217 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
221 /* for encoding cft->private value on file */
230 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
231 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
232 #define MEMFILE_ATTR(val) ((val) & 0xffff)
233 /* Used for OOM nofiier */
234 #define OOM_CONTROL (0)
236 /* Some nice accessors for the vmpressure. */
237 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
240 memcg = root_mem_cgroup;
241 return &memcg->vmpressure;
244 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
246 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
249 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
251 return (memcg == root_mem_cgroup);
256 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
257 * The main reason for not using cgroup id for this:
258 * this works better in sparse environments, where we have a lot of memcgs,
259 * but only a few kmem-limited. Or also, if we have, for instance, 200
260 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
261 * 200 entry array for that.
263 * The current size of the caches array is stored in memcg_nr_cache_ids. It
264 * will double each time we have to increase it.
266 static DEFINE_IDA(memcg_cache_ida);
267 int memcg_nr_cache_ids;
269 /* Protects memcg_nr_cache_ids */
270 static DECLARE_RWSEM(memcg_cache_ids_sem);
272 void memcg_get_cache_ids(void)
274 down_read(&memcg_cache_ids_sem);
277 void memcg_put_cache_ids(void)
279 up_read(&memcg_cache_ids_sem);
283 * MIN_SIZE is different than 1, because we would like to avoid going through
284 * the alloc/free process all the time. In a small machine, 4 kmem-limited
285 * cgroups is a reasonable guess. In the future, it could be a parameter or
286 * tunable, but that is strictly not necessary.
288 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
289 * this constant directly from cgroup, but it is understandable that this is
290 * better kept as an internal representation in cgroup.c. In any case, the
291 * cgrp_id space is not getting any smaller, and we don't have to necessarily
292 * increase ours as well if it increases.
294 #define MEMCG_CACHES_MIN_SIZE 4
295 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
298 * A lot of the calls to the cache allocation functions are expected to be
299 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
300 * conditional to this static branch, we'll have to allow modules that does
301 * kmem_cache_alloc and the such to see this symbol as well
303 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
304 EXPORT_SYMBOL(memcg_kmem_enabled_key);
306 struct workqueue_struct *memcg_kmem_cache_wq;
308 #endif /* !CONFIG_SLOB */
311 * mem_cgroup_css_from_page - css of the memcg associated with a page
312 * @page: page of interest
314 * If memcg is bound to the default hierarchy, css of the memcg associated
315 * with @page is returned. The returned css remains associated with @page
316 * until it is released.
318 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
321 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
323 struct mem_cgroup *memcg;
325 memcg = page->mem_cgroup;
327 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
328 memcg = root_mem_cgroup;
334 * page_cgroup_ino - return inode number of the memcg a page is charged to
337 * Look up the closest online ancestor of the memory cgroup @page is charged to
338 * and return its inode number or 0 if @page is not charged to any cgroup. It
339 * is safe to call this function without holding a reference to @page.
341 * Note, this function is inherently racy, because there is nothing to prevent
342 * the cgroup inode from getting torn down and potentially reallocated a moment
343 * after page_cgroup_ino() returns, so it only should be used by callers that
344 * do not care (such as procfs interfaces).
346 ino_t page_cgroup_ino(struct page *page)
348 struct mem_cgroup *memcg;
349 unsigned long ino = 0;
352 memcg = READ_ONCE(page->mem_cgroup);
353 while (memcg && !(memcg->css.flags & CSS_ONLINE))
354 memcg = parent_mem_cgroup(memcg);
356 ino = cgroup_ino(memcg->css.cgroup);
361 static struct mem_cgroup_per_node *
362 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
364 int nid = page_to_nid(page);
366 return memcg->nodeinfo[nid];
369 static struct mem_cgroup_tree_per_node *
370 soft_limit_tree_node(int nid)
372 return soft_limit_tree.rb_tree_per_node[nid];
375 static struct mem_cgroup_tree_per_node *
376 soft_limit_tree_from_page(struct page *page)
378 int nid = page_to_nid(page);
380 return soft_limit_tree.rb_tree_per_node[nid];
383 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
384 struct mem_cgroup_tree_per_node *mctz,
385 unsigned long new_usage_in_excess)
387 struct rb_node **p = &mctz->rb_root.rb_node;
388 struct rb_node *parent = NULL;
389 struct mem_cgroup_per_node *mz_node;
390 bool rightmost = true;
395 mz->usage_in_excess = new_usage_in_excess;
396 if (!mz->usage_in_excess)
400 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
402 if (mz->usage_in_excess < mz_node->usage_in_excess) {
408 * We can't avoid mem cgroups that are over their soft
409 * limit by the same amount
411 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
416 mctz->rb_rightmost = &mz->tree_node;
418 rb_link_node(&mz->tree_node, parent, p);
419 rb_insert_color(&mz->tree_node, &mctz->rb_root);
423 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
424 struct mem_cgroup_tree_per_node *mctz)
429 if (&mz->tree_node == mctz->rb_rightmost)
430 mctz->rb_rightmost = rb_prev(&mz->tree_node);
432 rb_erase(&mz->tree_node, &mctz->rb_root);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
437 struct mem_cgroup_tree_per_node *mctz)
441 spin_lock_irqsave(&mctz->lock, flags);
442 __mem_cgroup_remove_exceeded(mz, mctz);
443 spin_unlock_irqrestore(&mctz->lock, flags);
446 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
448 unsigned long nr_pages = page_counter_read(&memcg->memory);
449 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
450 unsigned long excess = 0;
452 if (nr_pages > soft_limit)
453 excess = nr_pages - soft_limit;
458 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
460 unsigned long excess;
461 struct mem_cgroup_per_node *mz;
462 struct mem_cgroup_tree_per_node *mctz;
464 mctz = soft_limit_tree_from_page(page);
468 * Necessary to update all ancestors when hierarchy is used.
469 * because their event counter is not touched.
471 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
472 mz = mem_cgroup_page_nodeinfo(memcg, page);
473 excess = soft_limit_excess(memcg);
475 * We have to update the tree if mz is on RB-tree or
476 * mem is over its softlimit.
478 if (excess || mz->on_tree) {
481 spin_lock_irqsave(&mctz->lock, flags);
482 /* if on-tree, remove it */
484 __mem_cgroup_remove_exceeded(mz, mctz);
486 * Insert again. mz->usage_in_excess will be updated.
487 * If excess is 0, no tree ops.
489 __mem_cgroup_insert_exceeded(mz, mctz, excess);
490 spin_unlock_irqrestore(&mctz->lock, flags);
495 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
497 struct mem_cgroup_tree_per_node *mctz;
498 struct mem_cgroup_per_node *mz;
502 mz = mem_cgroup_nodeinfo(memcg, nid);
503 mctz = soft_limit_tree_node(nid);
505 mem_cgroup_remove_exceeded(mz, mctz);
509 static struct mem_cgroup_per_node *
510 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
512 struct mem_cgroup_per_node *mz;
516 if (!mctz->rb_rightmost)
517 goto done; /* Nothing to reclaim from */
519 mz = rb_entry(mctz->rb_rightmost,
520 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);
545 static unsigned long memcg_sum_events(struct mem_cgroup *memcg,
548 return atomic_long_read(&memcg->events[event]);
551 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
553 bool compound, int nr_pages)
556 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
557 * counted as CACHE even if it's on ANON LRU.
560 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
562 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
563 if (PageSwapBacked(page))
564 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
568 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
569 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
572 /* pagein of a big page is an event. So, ignore page size */
574 __count_memcg_events(memcg, PGPGIN, 1);
576 __count_memcg_events(memcg, PGPGOUT, 1);
577 nr_pages = -nr_pages; /* for event */
580 __this_cpu_add(memcg->stat_cpu->nr_page_events, nr_pages);
583 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
584 int nid, unsigned int lru_mask)
586 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
587 unsigned long nr = 0;
590 VM_BUG_ON((unsigned)nid >= nr_node_ids);
593 if (!(BIT(lru) & lru_mask))
595 nr += mem_cgroup_get_lru_size(lruvec, lru);
600 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
601 unsigned int lru_mask)
603 unsigned long nr = 0;
606 for_each_node_state(nid, N_MEMORY)
607 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
611 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
612 enum mem_cgroup_events_target target)
614 unsigned long val, next;
616 val = __this_cpu_read(memcg->stat_cpu->nr_page_events);
617 next = __this_cpu_read(memcg->stat_cpu->targets[target]);
618 /* from time_after() in jiffies.h */
619 if ((long)(next - val) < 0) {
621 case MEM_CGROUP_TARGET_THRESH:
622 next = val + THRESHOLDS_EVENTS_TARGET;
624 case MEM_CGROUP_TARGET_SOFTLIMIT:
625 next = val + SOFTLIMIT_EVENTS_TARGET;
627 case MEM_CGROUP_TARGET_NUMAINFO:
628 next = val + NUMAINFO_EVENTS_TARGET;
633 __this_cpu_write(memcg->stat_cpu->targets[target], next);
640 * Check events in order.
643 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
645 /* threshold event is triggered in finer grain than soft limit */
646 if (unlikely(mem_cgroup_event_ratelimit(memcg,
647 MEM_CGROUP_TARGET_THRESH))) {
649 bool do_numainfo __maybe_unused;
651 do_softlimit = mem_cgroup_event_ratelimit(memcg,
652 MEM_CGROUP_TARGET_SOFTLIMIT);
654 do_numainfo = mem_cgroup_event_ratelimit(memcg,
655 MEM_CGROUP_TARGET_NUMAINFO);
657 mem_cgroup_threshold(memcg);
658 if (unlikely(do_softlimit))
659 mem_cgroup_update_tree(memcg, page);
661 if (unlikely(do_numainfo))
662 atomic_inc(&memcg->numainfo_events);
667 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
670 * mm_update_next_owner() may clear mm->owner to NULL
671 * if it races with swapoff, page migration, etc.
672 * So this can be called with p == NULL.
677 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
679 EXPORT_SYMBOL(mem_cgroup_from_task);
681 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
683 struct mem_cgroup *memcg = NULL;
688 * Page cache insertions can happen withou an
689 * actual mm context, e.g. during disk probing
690 * on boot, loopback IO, acct() writes etc.
693 memcg = root_mem_cgroup;
695 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
696 if (unlikely(!memcg))
697 memcg = root_mem_cgroup;
699 } while (!css_tryget(&memcg->css));
705 * mem_cgroup_iter - iterate over memory cgroup hierarchy
706 * @root: hierarchy root
707 * @prev: previously returned memcg, NULL on first invocation
708 * @reclaim: cookie for shared reclaim walks, NULL for full walks
710 * Returns references to children of the hierarchy below @root, or
711 * @root itself, or %NULL after a full round-trip.
713 * Caller must pass the return value in @prev on subsequent
714 * invocations for reference counting, or use mem_cgroup_iter_break()
715 * to cancel a hierarchy walk before the round-trip is complete.
717 * Reclaimers can specify a zone and a priority level in @reclaim to
718 * divide up the memcgs in the hierarchy among all concurrent
719 * reclaimers operating on the same zone and priority.
721 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
722 struct mem_cgroup *prev,
723 struct mem_cgroup_reclaim_cookie *reclaim)
725 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
726 struct cgroup_subsys_state *css = NULL;
727 struct mem_cgroup *memcg = NULL;
728 struct mem_cgroup *pos = NULL;
730 if (mem_cgroup_disabled())
734 root = root_mem_cgroup;
736 if (prev && !reclaim)
739 if (!root->use_hierarchy && root != root_mem_cgroup) {
748 struct mem_cgroup_per_node *mz;
750 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
751 iter = &mz->iter[reclaim->priority];
753 if (prev && reclaim->generation != iter->generation)
757 pos = READ_ONCE(iter->position);
758 if (!pos || css_tryget(&pos->css))
761 * css reference reached zero, so iter->position will
762 * be cleared by ->css_released. However, we should not
763 * rely on this happening soon, because ->css_released
764 * is called from a work queue, and by busy-waiting we
765 * might block it. So we clear iter->position right
768 (void)cmpxchg(&iter->position, pos, NULL);
776 css = css_next_descendant_pre(css, &root->css);
779 * Reclaimers share the hierarchy walk, and a
780 * new one might jump in right at the end of
781 * the hierarchy - make sure they see at least
782 * one group and restart from the beginning.
790 * Verify the css and acquire a reference. The root
791 * is provided by the caller, so we know it's alive
792 * and kicking, and don't take an extra reference.
794 memcg = mem_cgroup_from_css(css);
796 if (css == &root->css)
807 * The position could have already been updated by a competing
808 * thread, so check that the value hasn't changed since we read
809 * it to avoid reclaiming from the same cgroup twice.
811 (void)cmpxchg(&iter->position, pos, memcg);
819 reclaim->generation = iter->generation;
825 if (prev && prev != root)
832 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
833 * @root: hierarchy root
834 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
836 void mem_cgroup_iter_break(struct mem_cgroup *root,
837 struct mem_cgroup *prev)
840 root = root_mem_cgroup;
841 if (prev && prev != root)
845 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
846 struct mem_cgroup *dead_memcg)
848 struct mem_cgroup_reclaim_iter *iter;
849 struct mem_cgroup_per_node *mz;
854 mz = mem_cgroup_nodeinfo(from, nid);
855 for (i = 0; i <= DEF_PRIORITY; i++) {
857 cmpxchg(&iter->position,
863 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
865 struct mem_cgroup *memcg = dead_memcg;
866 struct mem_cgroup *last;
869 __invalidate_reclaim_iterators(memcg, dead_memcg);
871 } while ((memcg = parent_mem_cgroup(memcg)));
874 * When cgruop1 non-hierarchy mode is used,
875 * parent_mem_cgroup() does not walk all the way up to the
876 * cgroup root (root_mem_cgroup). So we have to handle
877 * dead_memcg from cgroup root separately.
879 if (last != root_mem_cgroup)
880 __invalidate_reclaim_iterators(root_mem_cgroup,
885 * Iteration constructs for visiting all cgroups (under a tree). If
886 * loops are exited prematurely (break), mem_cgroup_iter_break() must
887 * be used for reference counting.
889 #define for_each_mem_cgroup_tree(iter, root) \
890 for (iter = mem_cgroup_iter(root, NULL, NULL); \
892 iter = mem_cgroup_iter(root, iter, NULL))
894 #define for_each_mem_cgroup(iter) \
895 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
897 iter = mem_cgroup_iter(NULL, iter, NULL))
900 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
901 * @memcg: hierarchy root
902 * @fn: function to call for each task
903 * @arg: argument passed to @fn
905 * This function iterates over tasks attached to @memcg or to any of its
906 * descendants and calls @fn for each task. If @fn returns a non-zero
907 * value, the function breaks the iteration loop and returns the value.
908 * Otherwise, it will iterate over all tasks and return 0.
910 * This function must not be called for the root memory cgroup.
912 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
913 int (*fn)(struct task_struct *, void *), void *arg)
915 struct mem_cgroup *iter;
918 BUG_ON(memcg == root_mem_cgroup);
920 for_each_mem_cgroup_tree(iter, memcg) {
921 struct css_task_iter it;
922 struct task_struct *task;
924 css_task_iter_start(&iter->css, 0, &it);
925 while (!ret && (task = css_task_iter_next(&it)))
927 css_task_iter_end(&it);
929 mem_cgroup_iter_break(memcg, iter);
937 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
939 * @zone: zone of the page
941 * This function is only safe when following the LRU page isolation
942 * and putback protocol: the LRU lock must be held, and the page must
943 * either be PageLRU() or the caller must have isolated/allocated it.
945 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
947 struct mem_cgroup_per_node *mz;
948 struct mem_cgroup *memcg;
949 struct lruvec *lruvec;
951 if (mem_cgroup_disabled()) {
952 lruvec = &pgdat->lruvec;
956 memcg = page->mem_cgroup;
958 * Swapcache readahead pages are added to the LRU - and
959 * possibly migrated - before they are charged.
962 memcg = root_mem_cgroup;
964 mz = mem_cgroup_page_nodeinfo(memcg, page);
965 lruvec = &mz->lruvec;
968 * Since a node can be onlined after the mem_cgroup was created,
969 * we have to be prepared to initialize lruvec->zone here;
970 * and if offlined then reonlined, we need to reinitialize it.
972 if (unlikely(lruvec->pgdat != pgdat))
973 lruvec->pgdat = pgdat;
978 * mem_cgroup_update_lru_size - account for adding or removing an lru page
979 * @lruvec: mem_cgroup per zone lru vector
980 * @lru: index of lru list the page is sitting on
981 * @zid: zone id of the accounted pages
982 * @nr_pages: positive when adding or negative when removing
984 * This function must be called under lru_lock, just before a page is added
985 * to or just after a page is removed from an lru list (that ordering being
986 * so as to allow it to check that lru_size 0 is consistent with list_empty).
988 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
989 int zid, int nr_pages)
991 struct mem_cgroup_per_node *mz;
992 unsigned long *lru_size;
995 if (mem_cgroup_disabled())
998 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
999 lru_size = &mz->lru_zone_size[zid][lru];
1002 *lru_size += nr_pages;
1005 if (WARN_ONCE(size < 0,
1006 "%s(%p, %d, %d): lru_size %ld\n",
1007 __func__, lruvec, lru, nr_pages, size)) {
1013 *lru_size += nr_pages;
1016 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1018 struct mem_cgroup *task_memcg;
1019 struct task_struct *p;
1022 p = find_lock_task_mm(task);
1024 task_memcg = get_mem_cgroup_from_mm(p->mm);
1028 * All threads may have already detached their mm's, but the oom
1029 * killer still needs to detect if they have already been oom
1030 * killed to prevent needlessly killing additional tasks.
1033 task_memcg = mem_cgroup_from_task(task);
1034 css_get(&task_memcg->css);
1037 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1038 css_put(&task_memcg->css);
1043 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1044 * @memcg: the memory cgroup
1046 * Returns the maximum amount of memory @mem can be charged with, in
1049 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1051 unsigned long margin = 0;
1052 unsigned long count;
1053 unsigned long limit;
1055 count = page_counter_read(&memcg->memory);
1056 limit = READ_ONCE(memcg->memory.limit);
1058 margin = limit - count;
1060 if (do_memsw_account()) {
1061 count = page_counter_read(&memcg->memsw);
1062 limit = READ_ONCE(memcg->memsw.limit);
1064 margin = min(margin, limit - count);
1073 * A routine for checking "mem" is under move_account() or not.
1075 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1076 * moving cgroups. This is for waiting at high-memory pressure
1079 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1081 struct mem_cgroup *from;
1082 struct mem_cgroup *to;
1085 * Unlike task_move routines, we access mc.to, mc.from not under
1086 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1088 spin_lock(&mc.lock);
1094 ret = mem_cgroup_is_descendant(from, memcg) ||
1095 mem_cgroup_is_descendant(to, memcg);
1097 spin_unlock(&mc.lock);
1101 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1103 if (mc.moving_task && current != mc.moving_task) {
1104 if (mem_cgroup_under_move(memcg)) {
1106 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1107 /* moving charge context might have finished. */
1110 finish_wait(&mc.waitq, &wait);
1117 unsigned int memcg1_stats[] = {
1128 static const char *const memcg1_stat_names[] = {
1139 #define K(x) ((x) << (PAGE_SHIFT-10))
1141 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1142 * @memcg: The memory cgroup that went over limit
1143 * @p: Task that is going to be killed
1145 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1148 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1150 struct mem_cgroup *iter;
1156 pr_info("Task in ");
1157 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1158 pr_cont(" killed as a result of limit of ");
1160 pr_info("Memory limit reached of cgroup ");
1163 pr_cont_cgroup_path(memcg->css.cgroup);
1168 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1169 K((u64)page_counter_read(&memcg->memory)),
1170 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1171 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1172 K((u64)page_counter_read(&memcg->memsw)),
1173 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1174 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1175 K((u64)page_counter_read(&memcg->kmem)),
1176 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1178 for_each_mem_cgroup_tree(iter, memcg) {
1179 pr_info("Memory cgroup stats for ");
1180 pr_cont_cgroup_path(iter->css.cgroup);
1183 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1184 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1186 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1187 K(memcg_page_state(iter, memcg1_stats[i])));
1190 for (i = 0; i < NR_LRU_LISTS; i++)
1191 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1192 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1199 * This function returns the number of memcg under hierarchy tree. Returns
1200 * 1(self count) if no children.
1202 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1205 struct mem_cgroup *iter;
1207 for_each_mem_cgroup_tree(iter, memcg)
1213 * Return the memory (and swap, if configured) limit for a memcg.
1215 unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1217 unsigned long limit;
1219 limit = memcg->memory.limit;
1220 if (mem_cgroup_swappiness(memcg)) {
1221 unsigned long memsw_limit;
1222 unsigned long swap_limit;
1224 memsw_limit = memcg->memsw.limit;
1225 swap_limit = memcg->swap.limit;
1226 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1227 limit = min(limit + swap_limit, memsw_limit);
1232 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1235 struct oom_control oc = {
1239 .gfp_mask = gfp_mask,
1244 mutex_lock(&oom_lock);
1245 ret = out_of_memory(&oc);
1246 mutex_unlock(&oom_lock);
1250 #if MAX_NUMNODES > 1
1253 * test_mem_cgroup_node_reclaimable
1254 * @memcg: the target memcg
1255 * @nid: the node ID to be checked.
1256 * @noswap : specify true here if the user wants flle only information.
1258 * This function returns whether the specified memcg contains any
1259 * reclaimable pages on a node. Returns true if there are any reclaimable
1260 * pages in the node.
1262 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1263 int nid, bool noswap)
1265 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1267 if (noswap || !total_swap_pages)
1269 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1276 * Always updating the nodemask is not very good - even if we have an empty
1277 * list or the wrong list here, we can start from some node and traverse all
1278 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1281 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1285 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1286 * pagein/pageout changes since the last update.
1288 if (!atomic_read(&memcg->numainfo_events))
1290 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1293 /* make a nodemask where this memcg uses memory from */
1294 memcg->scan_nodes = node_states[N_MEMORY];
1296 for_each_node_mask(nid, node_states[N_MEMORY]) {
1298 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1299 node_clear(nid, memcg->scan_nodes);
1302 atomic_set(&memcg->numainfo_events, 0);
1303 atomic_set(&memcg->numainfo_updating, 0);
1307 * Selecting a node where we start reclaim from. Because what we need is just
1308 * reducing usage counter, start from anywhere is O,K. Considering
1309 * memory reclaim from current node, there are pros. and cons.
1311 * Freeing memory from current node means freeing memory from a node which
1312 * we'll use or we've used. So, it may make LRU bad. And if several threads
1313 * hit limits, it will see a contention on a node. But freeing from remote
1314 * node means more costs for memory reclaim because of memory latency.
1316 * Now, we use round-robin. Better algorithm is welcomed.
1318 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1322 mem_cgroup_may_update_nodemask(memcg);
1323 node = memcg->last_scanned_node;
1325 node = next_node_in(node, memcg->scan_nodes);
1327 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1328 * last time it really checked all the LRUs due to rate limiting.
1329 * Fallback to the current node in that case for simplicity.
1331 if (unlikely(node == MAX_NUMNODES))
1332 node = numa_node_id();
1334 memcg->last_scanned_node = node;
1338 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1344 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1347 unsigned long *total_scanned)
1349 struct mem_cgroup *victim = NULL;
1352 unsigned long excess;
1353 unsigned long nr_scanned;
1354 struct mem_cgroup_reclaim_cookie reclaim = {
1359 excess = soft_limit_excess(root_memcg);
1362 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1367 * If we have not been able to reclaim
1368 * anything, it might because there are
1369 * no reclaimable pages under this hierarchy
1374 * We want to do more targeted reclaim.
1375 * excess >> 2 is not to excessive so as to
1376 * reclaim too much, nor too less that we keep
1377 * coming back to reclaim from this cgroup
1379 if (total >= (excess >> 2) ||
1380 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1385 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1386 pgdat, &nr_scanned);
1387 *total_scanned += nr_scanned;
1388 if (!soft_limit_excess(root_memcg))
1391 mem_cgroup_iter_break(root_memcg, victim);
1395 #ifdef CONFIG_LOCKDEP
1396 static struct lockdep_map memcg_oom_lock_dep_map = {
1397 .name = "memcg_oom_lock",
1401 static DEFINE_SPINLOCK(memcg_oom_lock);
1404 * Check OOM-Killer is already running under our hierarchy.
1405 * If someone is running, return false.
1407 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1409 struct mem_cgroup *iter, *failed = NULL;
1411 spin_lock(&memcg_oom_lock);
1413 for_each_mem_cgroup_tree(iter, memcg) {
1414 if (iter->oom_lock) {
1416 * this subtree of our hierarchy is already locked
1417 * so we cannot give a lock.
1420 mem_cgroup_iter_break(memcg, iter);
1423 iter->oom_lock = true;
1428 * OK, we failed to lock the whole subtree so we have
1429 * to clean up what we set up to the failing subtree
1431 for_each_mem_cgroup_tree(iter, memcg) {
1432 if (iter == failed) {
1433 mem_cgroup_iter_break(memcg, iter);
1436 iter->oom_lock = false;
1439 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1441 spin_unlock(&memcg_oom_lock);
1446 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1448 struct mem_cgroup *iter;
1450 spin_lock(&memcg_oom_lock);
1451 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1452 for_each_mem_cgroup_tree(iter, memcg)
1453 iter->oom_lock = false;
1454 spin_unlock(&memcg_oom_lock);
1457 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1459 struct mem_cgroup *iter;
1461 spin_lock(&memcg_oom_lock);
1462 for_each_mem_cgroup_tree(iter, memcg)
1464 spin_unlock(&memcg_oom_lock);
1467 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1469 struct mem_cgroup *iter;
1472 * When a new child is created while the hierarchy is under oom,
1473 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1475 spin_lock(&memcg_oom_lock);
1476 for_each_mem_cgroup_tree(iter, memcg)
1477 if (iter->under_oom > 0)
1479 spin_unlock(&memcg_oom_lock);
1482 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1484 struct oom_wait_info {
1485 struct mem_cgroup *memcg;
1486 wait_queue_entry_t wait;
1489 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1490 unsigned mode, int sync, void *arg)
1492 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1493 struct mem_cgroup *oom_wait_memcg;
1494 struct oom_wait_info *oom_wait_info;
1496 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1497 oom_wait_memcg = oom_wait_info->memcg;
1499 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1500 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1502 return autoremove_wake_function(wait, mode, sync, arg);
1505 static void memcg_oom_recover(struct mem_cgroup *memcg)
1508 * For the following lockless ->under_oom test, the only required
1509 * guarantee is that it must see the state asserted by an OOM when
1510 * this function is called as a result of userland actions
1511 * triggered by the notification of the OOM. This is trivially
1512 * achieved by invoking mem_cgroup_mark_under_oom() before
1513 * triggering notification.
1515 if (memcg && memcg->under_oom)
1516 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1519 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1521 if (!current->memcg_may_oom)
1524 * We are in the middle of the charge context here, so we
1525 * don't want to block when potentially sitting on a callstack
1526 * that holds all kinds of filesystem and mm locks.
1528 * Also, the caller may handle a failed allocation gracefully
1529 * (like optional page cache readahead) and so an OOM killer
1530 * invocation might not even be necessary.
1532 * That's why we don't do anything here except remember the
1533 * OOM context and then deal with it at the end of the page
1534 * fault when the stack is unwound, the locks are released,
1535 * and when we know whether the fault was overall successful.
1537 css_get(&memcg->css);
1538 current->memcg_in_oom = memcg;
1539 current->memcg_oom_gfp_mask = mask;
1540 current->memcg_oom_order = order;
1544 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1545 * @handle: actually kill/wait or just clean up the OOM state
1547 * This has to be called at the end of a page fault if the memcg OOM
1548 * handler was enabled.
1550 * Memcg supports userspace OOM handling where failed allocations must
1551 * sleep on a waitqueue until the userspace task resolves the
1552 * situation. Sleeping directly in the charge context with all kinds
1553 * of locks held is not a good idea, instead we remember an OOM state
1554 * in the task and mem_cgroup_oom_synchronize() has to be called at
1555 * the end of the page fault to complete the OOM handling.
1557 * Returns %true if an ongoing memcg OOM situation was detected and
1558 * completed, %false otherwise.
1560 bool mem_cgroup_oom_synchronize(bool handle)
1562 struct mem_cgroup *memcg = current->memcg_in_oom;
1563 struct oom_wait_info owait;
1566 /* OOM is global, do not handle */
1573 owait.memcg = memcg;
1574 owait.wait.flags = 0;
1575 owait.wait.func = memcg_oom_wake_function;
1576 owait.wait.private = current;
1577 INIT_LIST_HEAD(&owait.wait.entry);
1579 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1580 mem_cgroup_mark_under_oom(memcg);
1582 locked = mem_cgroup_oom_trylock(memcg);
1585 mem_cgroup_oom_notify(memcg);
1587 if (locked && !memcg->oom_kill_disable) {
1588 mem_cgroup_unmark_under_oom(memcg);
1589 finish_wait(&memcg_oom_waitq, &owait.wait);
1590 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1591 current->memcg_oom_order);
1594 mem_cgroup_unmark_under_oom(memcg);
1595 finish_wait(&memcg_oom_waitq, &owait.wait);
1599 mem_cgroup_oom_unlock(memcg);
1601 * There is no guarantee that an OOM-lock contender
1602 * sees the wakeups triggered by the OOM kill
1603 * uncharges. Wake any sleepers explicitely.
1605 memcg_oom_recover(memcg);
1608 current->memcg_in_oom = NULL;
1609 css_put(&memcg->css);
1614 * lock_page_memcg - lock a page->mem_cgroup binding
1617 * This function protects unlocked LRU pages from being moved to
1620 * It ensures lifetime of the returned memcg. Caller is responsible
1621 * for the lifetime of the page; __unlock_page_memcg() is available
1622 * when @page might get freed inside the locked section.
1624 struct mem_cgroup *lock_page_memcg(struct page *page)
1626 struct mem_cgroup *memcg;
1627 unsigned long flags;
1630 * The RCU lock is held throughout the transaction. The fast
1631 * path can get away without acquiring the memcg->move_lock
1632 * because page moving starts with an RCU grace period.
1634 * The RCU lock also protects the memcg from being freed when
1635 * the page state that is going to change is the only thing
1636 * preventing the page itself from being freed. E.g. writeback
1637 * doesn't hold a page reference and relies on PG_writeback to
1638 * keep off truncation, migration and so forth.
1642 if (mem_cgroup_disabled())
1645 memcg = page->mem_cgroup;
1646 if (unlikely(!memcg))
1649 if (atomic_read(&memcg->moving_account) <= 0)
1652 spin_lock_irqsave(&memcg->move_lock, flags);
1653 if (memcg != page->mem_cgroup) {
1654 spin_unlock_irqrestore(&memcg->move_lock, flags);
1659 * When charge migration first begins, we can have locked and
1660 * unlocked page stat updates happening concurrently. Track
1661 * the task who has the lock for unlock_page_memcg().
1663 memcg->move_lock_task = current;
1664 memcg->move_lock_flags = flags;
1668 EXPORT_SYMBOL(lock_page_memcg);
1671 * __unlock_page_memcg - unlock and unpin a memcg
1674 * Unlock and unpin a memcg returned by lock_page_memcg().
1676 void __unlock_page_memcg(struct mem_cgroup *memcg)
1678 if (memcg && memcg->move_lock_task == current) {
1679 unsigned long flags = memcg->move_lock_flags;
1681 memcg->move_lock_task = NULL;
1682 memcg->move_lock_flags = 0;
1684 spin_unlock_irqrestore(&memcg->move_lock, flags);
1691 * unlock_page_memcg - unlock a page->mem_cgroup binding
1694 void unlock_page_memcg(struct page *page)
1696 __unlock_page_memcg(page->mem_cgroup);
1698 EXPORT_SYMBOL(unlock_page_memcg);
1700 struct memcg_stock_pcp {
1701 struct mem_cgroup *cached; /* this never be root cgroup */
1702 unsigned int nr_pages;
1703 struct work_struct work;
1704 unsigned long flags;
1705 #define FLUSHING_CACHED_CHARGE 0
1707 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1708 static DEFINE_MUTEX(percpu_charge_mutex);
1711 * consume_stock: Try to consume stocked charge on this cpu.
1712 * @memcg: memcg to consume from.
1713 * @nr_pages: how many pages to charge.
1715 * The charges will only happen if @memcg matches the current cpu's memcg
1716 * stock, and at least @nr_pages are available in that stock. Failure to
1717 * service an allocation will refill the stock.
1719 * returns true if successful, false otherwise.
1721 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1723 struct memcg_stock_pcp *stock;
1724 unsigned long flags;
1727 if (nr_pages > MEMCG_CHARGE_BATCH)
1730 local_irq_save(flags);
1732 stock = this_cpu_ptr(&memcg_stock);
1733 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1734 stock->nr_pages -= nr_pages;
1738 local_irq_restore(flags);
1744 * Returns stocks cached in percpu and reset cached information.
1746 static void drain_stock(struct memcg_stock_pcp *stock)
1748 struct mem_cgroup *old = stock->cached;
1750 if (stock->nr_pages) {
1751 page_counter_uncharge(&old->memory, stock->nr_pages);
1752 if (do_memsw_account())
1753 page_counter_uncharge(&old->memsw, stock->nr_pages);
1754 css_put_many(&old->css, stock->nr_pages);
1755 stock->nr_pages = 0;
1757 stock->cached = NULL;
1760 static void drain_local_stock(struct work_struct *dummy)
1762 struct memcg_stock_pcp *stock;
1763 unsigned long flags;
1766 * The only protection from memory hotplug vs. drain_stock races is
1767 * that we always operate on local CPU stock here with IRQ disabled
1769 local_irq_save(flags);
1771 stock = this_cpu_ptr(&memcg_stock);
1773 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1775 local_irq_restore(flags);
1779 * Cache charges(val) to local per_cpu area.
1780 * This will be consumed by consume_stock() function, later.
1782 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1784 struct memcg_stock_pcp *stock;
1785 unsigned long flags;
1787 local_irq_save(flags);
1789 stock = this_cpu_ptr(&memcg_stock);
1790 if (stock->cached != memcg) { /* reset if necessary */
1792 stock->cached = memcg;
1794 stock->nr_pages += nr_pages;
1796 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
1799 local_irq_restore(flags);
1803 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1804 * of the hierarchy under it.
1806 static void drain_all_stock(struct mem_cgroup *root_memcg)
1810 /* If someone's already draining, avoid adding running more workers. */
1811 if (!mutex_trylock(&percpu_charge_mutex))
1814 * Notify other cpus that system-wide "drain" is running
1815 * We do not care about races with the cpu hotplug because cpu down
1816 * as well as workers from this path always operate on the local
1817 * per-cpu data. CPU up doesn't touch memcg_stock at all.
1820 for_each_online_cpu(cpu) {
1821 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1822 struct mem_cgroup *memcg;
1824 memcg = stock->cached;
1825 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
1827 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
1828 css_put(&memcg->css);
1831 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1833 drain_local_stock(&stock->work);
1835 schedule_work_on(cpu, &stock->work);
1837 css_put(&memcg->css);
1840 mutex_unlock(&percpu_charge_mutex);
1843 static int memcg_hotplug_cpu_dead(unsigned int cpu)
1845 struct memcg_stock_pcp *stock;
1846 struct mem_cgroup *memcg;
1848 stock = &per_cpu(memcg_stock, cpu);
1851 for_each_mem_cgroup(memcg) {
1854 for (i = 0; i < MEMCG_NR_STAT; i++) {
1858 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
1860 atomic_long_add(x, &memcg->stat[i]);
1862 if (i >= NR_VM_NODE_STAT_ITEMS)
1865 for_each_node(nid) {
1866 struct mem_cgroup_per_node *pn;
1868 pn = mem_cgroup_nodeinfo(memcg, nid);
1869 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
1871 atomic_long_add(x, &pn->lruvec_stat[i]);
1875 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
1878 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
1880 atomic_long_add(x, &memcg->events[i]);
1887 static void reclaim_high(struct mem_cgroup *memcg,
1888 unsigned int nr_pages,
1892 if (page_counter_read(&memcg->memory) <= memcg->high)
1894 memcg_memory_event(memcg, MEMCG_HIGH);
1895 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1896 } while ((memcg = parent_mem_cgroup(memcg)));
1899 static void high_work_func(struct work_struct *work)
1901 struct mem_cgroup *memcg;
1903 memcg = container_of(work, struct mem_cgroup, high_work);
1904 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
1908 * Scheduled by try_charge() to be executed from the userland return path
1909 * and reclaims memory over the high limit.
1911 void mem_cgroup_handle_over_high(void)
1913 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1914 struct mem_cgroup *memcg;
1916 if (likely(!nr_pages))
1919 memcg = get_mem_cgroup_from_mm(current->mm);
1920 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1921 css_put(&memcg->css);
1922 current->memcg_nr_pages_over_high = 0;
1925 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1926 unsigned int nr_pages)
1928 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
1929 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1930 struct mem_cgroup *mem_over_limit;
1931 struct page_counter *counter;
1932 unsigned long nr_reclaimed;
1933 bool may_swap = true;
1934 bool drained = false;
1936 if (mem_cgroup_is_root(memcg))
1939 if (consume_stock(memcg, nr_pages))
1942 if (!do_memsw_account() ||
1943 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1944 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1946 if (do_memsw_account())
1947 page_counter_uncharge(&memcg->memsw, batch);
1948 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1950 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1954 if (batch > nr_pages) {
1960 * Unlike in global OOM situations, memcg is not in a physical
1961 * memory shortage. Allow dying and OOM-killed tasks to
1962 * bypass the last charges so that they can exit quickly and
1963 * free their memory.
1965 if (unlikely(tsk_is_oom_victim(current) ||
1966 fatal_signal_pending(current) ||
1967 current->flags & PF_EXITING))
1971 * Prevent unbounded recursion when reclaim operations need to
1972 * allocate memory. This might exceed the limits temporarily,
1973 * but we prefer facilitating memory reclaim and getting back
1974 * under the limit over triggering OOM kills in these cases.
1976 if (unlikely(current->flags & PF_MEMALLOC))
1979 if (unlikely(task_in_memcg_oom(current)))
1982 if (!gfpflags_allow_blocking(gfp_mask))
1985 memcg_memory_event(mem_over_limit, MEMCG_MAX);
1987 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1988 gfp_mask, may_swap);
1990 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1994 drain_all_stock(mem_over_limit);
1999 if (gfp_mask & __GFP_NORETRY)
2002 * Even though the limit is exceeded at this point, reclaim
2003 * may have been able to free some pages. Retry the charge
2004 * before killing the task.
2006 * Only for regular pages, though: huge pages are rather
2007 * unlikely to succeed so close to the limit, and we fall back
2008 * to regular pages anyway in case of failure.
2010 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2013 * At task move, charge accounts can be doubly counted. So, it's
2014 * better to wait until the end of task_move if something is going on.
2016 if (mem_cgroup_wait_acct_move(mem_over_limit))
2022 if (gfp_mask & __GFP_NOFAIL)
2025 if (fatal_signal_pending(current))
2028 memcg_memory_event(mem_over_limit, MEMCG_OOM);
2030 mem_cgroup_oom(mem_over_limit, gfp_mask,
2031 get_order(nr_pages * PAGE_SIZE));
2033 if (!(gfp_mask & __GFP_NOFAIL))
2037 * The allocation either can't fail or will lead to more memory
2038 * being freed very soon. Allow memory usage go over the limit
2039 * temporarily by force charging it.
2041 page_counter_charge(&memcg->memory, nr_pages);
2042 if (do_memsw_account())
2043 page_counter_charge(&memcg->memsw, nr_pages);
2044 css_get_many(&memcg->css, nr_pages);
2049 css_get_many(&memcg->css, batch);
2050 if (batch > nr_pages)
2051 refill_stock(memcg, batch - nr_pages);
2054 * If the hierarchy is above the normal consumption range, schedule
2055 * reclaim on returning to userland. We can perform reclaim here
2056 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2057 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2058 * not recorded as it most likely matches current's and won't
2059 * change in the meantime. As high limit is checked again before
2060 * reclaim, the cost of mismatch is negligible.
2063 if (page_counter_read(&memcg->memory) > memcg->high) {
2064 /* Don't bother a random interrupted task */
2065 if (in_interrupt()) {
2066 schedule_work(&memcg->high_work);
2069 current->memcg_nr_pages_over_high += batch;
2070 set_notify_resume(current);
2073 } while ((memcg = parent_mem_cgroup(memcg)));
2078 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2080 if (mem_cgroup_is_root(memcg))
2083 page_counter_uncharge(&memcg->memory, nr_pages);
2084 if (do_memsw_account())
2085 page_counter_uncharge(&memcg->memsw, nr_pages);
2087 css_put_many(&memcg->css, nr_pages);
2090 static void lock_page_lru(struct page *page, int *isolated)
2092 struct zone *zone = page_zone(page);
2094 spin_lock_irq(zone_lru_lock(zone));
2095 if (PageLRU(page)) {
2096 struct lruvec *lruvec;
2098 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2100 del_page_from_lru_list(page, lruvec, page_lru(page));
2106 static void unlock_page_lru(struct page *page, int isolated)
2108 struct zone *zone = page_zone(page);
2111 struct lruvec *lruvec;
2113 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2114 VM_BUG_ON_PAGE(PageLRU(page), page);
2116 add_page_to_lru_list(page, lruvec, page_lru(page));
2118 spin_unlock_irq(zone_lru_lock(zone));
2121 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2126 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2129 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2130 * may already be on some other mem_cgroup's LRU. Take care of it.
2133 lock_page_lru(page, &isolated);
2136 * Nobody should be changing or seriously looking at
2137 * page->mem_cgroup at this point:
2139 * - the page is uncharged
2141 * - the page is off-LRU
2143 * - an anonymous fault has exclusive page access, except for
2144 * a locked page table
2146 * - a page cache insertion, a swapin fault, or a migration
2147 * have the page locked
2149 page->mem_cgroup = memcg;
2152 unlock_page_lru(page, isolated);
2156 static int memcg_alloc_cache_id(void)
2161 id = ida_simple_get(&memcg_cache_ida,
2162 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2166 if (id < memcg_nr_cache_ids)
2170 * There's no space for the new id in memcg_caches arrays,
2171 * so we have to grow them.
2173 down_write(&memcg_cache_ids_sem);
2175 size = 2 * (id + 1);
2176 if (size < MEMCG_CACHES_MIN_SIZE)
2177 size = MEMCG_CACHES_MIN_SIZE;
2178 else if (size > MEMCG_CACHES_MAX_SIZE)
2179 size = MEMCG_CACHES_MAX_SIZE;
2181 err = memcg_update_all_caches(size);
2183 err = memcg_update_all_list_lrus(size);
2185 memcg_nr_cache_ids = size;
2187 up_write(&memcg_cache_ids_sem);
2190 ida_simple_remove(&memcg_cache_ida, id);
2196 static void memcg_free_cache_id(int id)
2198 ida_simple_remove(&memcg_cache_ida, id);
2201 struct memcg_kmem_cache_create_work {
2202 struct mem_cgroup *memcg;
2203 struct kmem_cache *cachep;
2204 struct work_struct work;
2207 static void memcg_kmem_cache_create_func(struct work_struct *w)
2209 struct memcg_kmem_cache_create_work *cw =
2210 container_of(w, struct memcg_kmem_cache_create_work, work);
2211 struct mem_cgroup *memcg = cw->memcg;
2212 struct kmem_cache *cachep = cw->cachep;
2214 memcg_create_kmem_cache(memcg, cachep);
2216 css_put(&memcg->css);
2221 * Enqueue the creation of a per-memcg kmem_cache.
2223 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2224 struct kmem_cache *cachep)
2226 struct memcg_kmem_cache_create_work *cw;
2228 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2232 css_get(&memcg->css);
2235 cw->cachep = cachep;
2236 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2238 queue_work(memcg_kmem_cache_wq, &cw->work);
2241 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2242 struct kmem_cache *cachep)
2245 * We need to stop accounting when we kmalloc, because if the
2246 * corresponding kmalloc cache is not yet created, the first allocation
2247 * in __memcg_schedule_kmem_cache_create will recurse.
2249 * However, it is better to enclose the whole function. Depending on
2250 * the debugging options enabled, INIT_WORK(), for instance, can
2251 * trigger an allocation. This too, will make us recurse. Because at
2252 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2253 * the safest choice is to do it like this, wrapping the whole function.
2255 current->memcg_kmem_skip_account = 1;
2256 __memcg_schedule_kmem_cache_create(memcg, cachep);
2257 current->memcg_kmem_skip_account = 0;
2260 static inline bool memcg_kmem_bypass(void)
2262 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2268 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2269 * @cachep: the original global kmem cache
2271 * Return the kmem_cache we're supposed to use for a slab allocation.
2272 * We try to use the current memcg's version of the cache.
2274 * If the cache does not exist yet, if we are the first user of it, we
2275 * create it asynchronously in a workqueue and let the current allocation
2276 * go through with the original cache.
2278 * This function takes a reference to the cache it returns to assure it
2279 * won't get destroyed while we are working with it. Once the caller is
2280 * done with it, memcg_kmem_put_cache() must be called to release the
2283 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2285 struct mem_cgroup *memcg;
2286 struct kmem_cache *memcg_cachep;
2289 VM_BUG_ON(!is_root_cache(cachep));
2291 if (memcg_kmem_bypass())
2294 if (current->memcg_kmem_skip_account)
2297 memcg = get_mem_cgroup_from_mm(current->mm);
2298 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2302 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2303 if (likely(memcg_cachep))
2304 return memcg_cachep;
2307 * If we are in a safe context (can wait, and not in interrupt
2308 * context), we could be be predictable and return right away.
2309 * This would guarantee that the allocation being performed
2310 * already belongs in the new cache.
2312 * However, there are some clashes that can arrive from locking.
2313 * For instance, because we acquire the slab_mutex while doing
2314 * memcg_create_kmem_cache, this means no further allocation
2315 * could happen with the slab_mutex held. So it's better to
2318 memcg_schedule_kmem_cache_create(memcg, cachep);
2320 css_put(&memcg->css);
2325 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2326 * @cachep: the cache returned by memcg_kmem_get_cache
2328 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2330 if (!is_root_cache(cachep))
2331 css_put(&cachep->memcg_params.memcg->css);
2335 * memcg_kmem_charge: charge a kmem page
2336 * @page: page to charge
2337 * @gfp: reclaim mode
2338 * @order: allocation order
2339 * @memcg: memory cgroup to charge
2341 * Returns 0 on success, an error code on failure.
2343 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2344 struct mem_cgroup *memcg)
2346 unsigned int nr_pages = 1 << order;
2347 struct page_counter *counter;
2350 ret = try_charge(memcg, gfp, nr_pages);
2354 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2355 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2358 * Enforce __GFP_NOFAIL allocation because callers are not
2359 * prepared to see failures and likely do not have any failure
2362 if (gfp & __GFP_NOFAIL) {
2363 page_counter_charge(&memcg->kmem, nr_pages);
2366 cancel_charge(memcg, nr_pages);
2370 page->mem_cgroup = memcg;
2376 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2377 * @page: page to charge
2378 * @gfp: reclaim mode
2379 * @order: allocation order
2381 * Returns 0 on success, an error code on failure.
2383 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2385 struct mem_cgroup *memcg;
2388 if (memcg_kmem_bypass())
2391 memcg = get_mem_cgroup_from_mm(current->mm);
2392 if (!mem_cgroup_is_root(memcg)) {
2393 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2395 __SetPageKmemcg(page);
2397 css_put(&memcg->css);
2401 * memcg_kmem_uncharge: uncharge a kmem page
2402 * @page: page to uncharge
2403 * @order: allocation order
2405 void memcg_kmem_uncharge(struct page *page, int order)
2407 struct mem_cgroup *memcg = page->mem_cgroup;
2408 unsigned int nr_pages = 1 << order;
2413 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2415 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2416 page_counter_uncharge(&memcg->kmem, nr_pages);
2418 page_counter_uncharge(&memcg->memory, nr_pages);
2419 if (do_memsw_account())
2420 page_counter_uncharge(&memcg->memsw, nr_pages);
2422 page->mem_cgroup = NULL;
2424 /* slab pages do not have PageKmemcg flag set */
2425 if (PageKmemcg(page))
2426 __ClearPageKmemcg(page);
2428 css_put_many(&memcg->css, nr_pages);
2430 #endif /* !CONFIG_SLOB */
2432 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2435 * Because tail pages are not marked as "used", set it. We're under
2436 * zone_lru_lock and migration entries setup in all page mappings.
2438 void mem_cgroup_split_huge_fixup(struct page *head)
2442 if (mem_cgroup_disabled())
2445 for (i = 1; i < HPAGE_PMD_NR; i++)
2446 head[i].mem_cgroup = head->mem_cgroup;
2448 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2450 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2452 #ifdef CONFIG_MEMCG_SWAP
2454 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2455 * @entry: swap entry to be moved
2456 * @from: mem_cgroup which the entry is moved from
2457 * @to: mem_cgroup which the entry is moved to
2459 * It succeeds only when the swap_cgroup's record for this entry is the same
2460 * as the mem_cgroup's id of @from.
2462 * Returns 0 on success, -EINVAL on failure.
2464 * The caller must have charged to @to, IOW, called page_counter_charge() about
2465 * both res and memsw, and called css_get().
2467 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2468 struct mem_cgroup *from, struct mem_cgroup *to)
2470 unsigned short old_id, new_id;
2472 old_id = mem_cgroup_id(from);
2473 new_id = mem_cgroup_id(to);
2475 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2476 mod_memcg_state(from, MEMCG_SWAP, -1);
2477 mod_memcg_state(to, MEMCG_SWAP, 1);
2483 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2484 struct mem_cgroup *from, struct mem_cgroup *to)
2490 static DEFINE_MUTEX(memcg_limit_mutex);
2492 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2493 unsigned long limit)
2495 unsigned long curusage;
2496 unsigned long oldusage;
2497 bool enlarge = false;
2502 * For keeping hierarchical_reclaim simple, how long we should retry
2503 * is depends on callers. We set our retry-count to be function
2504 * of # of children which we should visit in this loop.
2506 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2507 mem_cgroup_count_children(memcg);
2509 oldusage = page_counter_read(&memcg->memory);
2512 if (signal_pending(current)) {
2517 mutex_lock(&memcg_limit_mutex);
2518 if (limit > memcg->memsw.limit) {
2519 mutex_unlock(&memcg_limit_mutex);
2523 if (limit > memcg->memory.limit)
2525 ret = page_counter_limit(&memcg->memory, limit);
2526 mutex_unlock(&memcg_limit_mutex);
2531 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2533 curusage = page_counter_read(&memcg->memory);
2534 /* Usage is reduced ? */
2535 if (curusage >= oldusage)
2538 oldusage = curusage;
2539 } while (retry_count);
2541 if (!ret && enlarge)
2542 memcg_oom_recover(memcg);
2547 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2548 unsigned long limit)
2550 unsigned long curusage;
2551 unsigned long oldusage;
2552 bool enlarge = false;
2556 /* see mem_cgroup_resize_res_limit */
2557 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2558 mem_cgroup_count_children(memcg);
2560 oldusage = page_counter_read(&memcg->memsw);
2563 if (signal_pending(current)) {
2568 mutex_lock(&memcg_limit_mutex);
2569 if (limit < memcg->memory.limit) {
2570 mutex_unlock(&memcg_limit_mutex);
2574 if (limit > memcg->memsw.limit)
2576 ret = page_counter_limit(&memcg->memsw, limit);
2577 mutex_unlock(&memcg_limit_mutex);
2582 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2584 curusage = page_counter_read(&memcg->memsw);
2585 /* Usage is reduced ? */
2586 if (curusage >= oldusage)
2589 oldusage = curusage;
2590 } while (retry_count);
2592 if (!ret && enlarge)
2593 memcg_oom_recover(memcg);
2598 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2600 unsigned long *total_scanned)
2602 unsigned long nr_reclaimed = 0;
2603 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2604 unsigned long reclaimed;
2606 struct mem_cgroup_tree_per_node *mctz;
2607 unsigned long excess;
2608 unsigned long nr_scanned;
2613 mctz = soft_limit_tree_node(pgdat->node_id);
2616 * Do not even bother to check the largest node if the root
2617 * is empty. Do it lockless to prevent lock bouncing. Races
2618 * are acceptable as soft limit is best effort anyway.
2620 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2624 * This loop can run a while, specially if mem_cgroup's continuously
2625 * keep exceeding their soft limit and putting the system under
2632 mz = mem_cgroup_largest_soft_limit_node(mctz);
2637 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2638 gfp_mask, &nr_scanned);
2639 nr_reclaimed += reclaimed;
2640 *total_scanned += nr_scanned;
2641 spin_lock_irq(&mctz->lock);
2642 __mem_cgroup_remove_exceeded(mz, mctz);
2645 * If we failed to reclaim anything from this memory cgroup
2646 * it is time to move on to the next cgroup
2650 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2652 excess = soft_limit_excess(mz->memcg);
2654 * One school of thought says that we should not add
2655 * back the node to the tree if reclaim returns 0.
2656 * But our reclaim could return 0, simply because due
2657 * to priority we are exposing a smaller subset of
2658 * memory to reclaim from. Consider this as a longer
2661 /* If excess == 0, no tree ops */
2662 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2663 spin_unlock_irq(&mctz->lock);
2664 css_put(&mz->memcg->css);
2667 * Could not reclaim anything and there are no more
2668 * mem cgroups to try or we seem to be looping without
2669 * reclaiming anything.
2671 if (!nr_reclaimed &&
2673 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2675 } while (!nr_reclaimed);
2677 css_put(&next_mz->memcg->css);
2678 return nr_reclaimed;
2682 * Test whether @memcg has children, dead or alive. Note that this
2683 * function doesn't care whether @memcg has use_hierarchy enabled and
2684 * returns %true if there are child csses according to the cgroup
2685 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2687 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2692 ret = css_next_child(NULL, &memcg->css);
2698 * Reclaims as many pages from the given memcg as possible.
2700 * Caller is responsible for holding css reference for memcg.
2702 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2704 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2706 /* we call try-to-free pages for make this cgroup empty */
2707 lru_add_drain_all();
2708 /* try to free all pages in this cgroup */
2709 while (nr_retries && page_counter_read(&memcg->memory)) {
2712 if (signal_pending(current))
2715 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2719 /* maybe some writeback is necessary */
2720 congestion_wait(BLK_RW_ASYNC, HZ/10);
2728 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2729 char *buf, size_t nbytes,
2732 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2734 if (mem_cgroup_is_root(memcg))
2736 return mem_cgroup_force_empty(memcg) ?: nbytes;
2739 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2742 return mem_cgroup_from_css(css)->use_hierarchy;
2745 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2746 struct cftype *cft, u64 val)
2749 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2750 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2752 if (memcg->use_hierarchy == val)
2756 * If parent's use_hierarchy is set, we can't make any modifications
2757 * in the child subtrees. If it is unset, then the change can
2758 * occur, provided the current cgroup has no children.
2760 * For the root cgroup, parent_mem is NULL, we allow value to be
2761 * set if there are no children.
2763 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2764 (val == 1 || val == 0)) {
2765 if (!memcg_has_children(memcg))
2766 memcg->use_hierarchy = val;
2775 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2777 struct mem_cgroup *iter;
2780 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2782 for_each_mem_cgroup_tree(iter, memcg) {
2783 for (i = 0; i < MEMCG_NR_STAT; i++)
2784 stat[i] += memcg_page_state(iter, i);
2788 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2790 struct mem_cgroup *iter;
2793 memset(events, 0, sizeof(*events) * NR_VM_EVENT_ITEMS);
2795 for_each_mem_cgroup_tree(iter, memcg) {
2796 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
2797 events[i] += memcg_sum_events(iter, i);
2801 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2803 unsigned long val = 0;
2805 if (mem_cgroup_is_root(memcg)) {
2806 struct mem_cgroup *iter;
2808 for_each_mem_cgroup_tree(iter, memcg) {
2809 val += memcg_page_state(iter, MEMCG_CACHE);
2810 val += memcg_page_state(iter, MEMCG_RSS);
2812 val += memcg_page_state(iter, MEMCG_SWAP);
2816 val = page_counter_read(&memcg->memory);
2818 val = page_counter_read(&memcg->memsw);
2831 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2834 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2835 struct page_counter *counter;
2837 switch (MEMFILE_TYPE(cft->private)) {
2839 counter = &memcg->memory;
2842 counter = &memcg->memsw;
2845 counter = &memcg->kmem;
2848 counter = &memcg->tcpmem;
2854 switch (MEMFILE_ATTR(cft->private)) {
2856 if (counter == &memcg->memory)
2857 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2858 if (counter == &memcg->memsw)
2859 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2860 return (u64)page_counter_read(counter) * PAGE_SIZE;
2862 return (u64)counter->limit * PAGE_SIZE;
2864 return (u64)counter->watermark * PAGE_SIZE;
2866 return counter->failcnt;
2867 case RES_SOFT_LIMIT:
2868 return (u64)memcg->soft_limit * PAGE_SIZE;
2875 static int memcg_online_kmem(struct mem_cgroup *memcg)
2879 if (cgroup_memory_nokmem)
2882 BUG_ON(memcg->kmemcg_id >= 0);
2883 BUG_ON(memcg->kmem_state);
2885 memcg_id = memcg_alloc_cache_id();
2889 static_branch_inc(&memcg_kmem_enabled_key);
2891 * A memory cgroup is considered kmem-online as soon as it gets
2892 * kmemcg_id. Setting the id after enabling static branching will
2893 * guarantee no one starts accounting before all call sites are
2896 memcg->kmemcg_id = memcg_id;
2897 memcg->kmem_state = KMEM_ONLINE;
2898 INIT_LIST_HEAD(&memcg->kmem_caches);
2903 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2905 struct cgroup_subsys_state *css;
2906 struct mem_cgroup *parent, *child;
2909 if (memcg->kmem_state != KMEM_ONLINE)
2912 * Clear the online state before clearing memcg_caches array
2913 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2914 * guarantees that no cache will be created for this cgroup
2915 * after we are done (see memcg_create_kmem_cache()).
2917 memcg->kmem_state = KMEM_ALLOCATED;
2919 memcg_deactivate_kmem_caches(memcg);
2921 kmemcg_id = memcg->kmemcg_id;
2922 BUG_ON(kmemcg_id < 0);
2924 parent = parent_mem_cgroup(memcg);
2926 parent = root_mem_cgroup;
2929 * Change kmemcg_id of this cgroup and all its descendants to the
2930 * parent's id, and then move all entries from this cgroup's list_lrus
2931 * to ones of the parent. After we have finished, all list_lrus
2932 * corresponding to this cgroup are guaranteed to remain empty. The
2933 * ordering is imposed by list_lru_node->lock taken by
2934 * memcg_drain_all_list_lrus().
2936 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2937 css_for_each_descendant_pre(css, &memcg->css) {
2938 child = mem_cgroup_from_css(css);
2939 BUG_ON(child->kmemcg_id != kmemcg_id);
2940 child->kmemcg_id = parent->kmemcg_id;
2941 if (!memcg->use_hierarchy)
2946 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2948 memcg_free_cache_id(kmemcg_id);
2951 static void memcg_free_kmem(struct mem_cgroup *memcg)
2953 /* css_alloc() failed, offlining didn't happen */
2954 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2955 memcg_offline_kmem(memcg);
2957 if (memcg->kmem_state == KMEM_ALLOCATED) {
2958 memcg_destroy_kmem_caches(memcg);
2959 static_branch_dec(&memcg_kmem_enabled_key);
2960 WARN_ON(page_counter_read(&memcg->kmem));
2964 static int memcg_online_kmem(struct mem_cgroup *memcg)
2968 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2971 static void memcg_free_kmem(struct mem_cgroup *memcg)
2974 #endif /* !CONFIG_SLOB */
2976 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2977 unsigned long limit)
2981 mutex_lock(&memcg_limit_mutex);
2982 ret = page_counter_limit(&memcg->kmem, limit);
2983 mutex_unlock(&memcg_limit_mutex);
2987 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2991 mutex_lock(&memcg_limit_mutex);
2993 ret = page_counter_limit(&memcg->tcpmem, limit);
2997 if (!memcg->tcpmem_active) {
2999 * The active flag needs to be written after the static_key
3000 * update. This is what guarantees that the socket activation
3001 * function is the last one to run. See mem_cgroup_sk_alloc()
3002 * for details, and note that we don't mark any socket as
3003 * belonging to this memcg until that flag is up.
3005 * We need to do this, because static_keys will span multiple
3006 * sites, but we can't control their order. If we mark a socket
3007 * as accounted, but the accounting functions are not patched in
3008 * yet, we'll lose accounting.
3010 * We never race with the readers in mem_cgroup_sk_alloc(),
3011 * because when this value change, the code to process it is not
3014 static_branch_inc(&memcg_sockets_enabled_key);
3015 memcg->tcpmem_active = true;
3018 mutex_unlock(&memcg_limit_mutex);
3023 * The user of this function is...
3026 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3027 char *buf, size_t nbytes, loff_t off)
3029 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3030 unsigned long nr_pages;
3033 buf = strstrip(buf);
3034 ret = page_counter_memparse(buf, "-1", &nr_pages);
3038 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3040 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3044 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3046 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3049 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3052 ret = memcg_update_kmem_limit(memcg, nr_pages);
3055 ret = memcg_update_tcp_limit(memcg, nr_pages);
3059 case RES_SOFT_LIMIT:
3060 memcg->soft_limit = nr_pages;
3064 return ret ?: nbytes;
3067 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3068 size_t nbytes, loff_t off)
3070 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3071 struct page_counter *counter;
3073 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3075 counter = &memcg->memory;
3078 counter = &memcg->memsw;
3081 counter = &memcg->kmem;
3084 counter = &memcg->tcpmem;
3090 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3092 page_counter_reset_watermark(counter);
3095 counter->failcnt = 0;
3104 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3107 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3111 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3112 struct cftype *cft, u64 val)
3114 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3116 if (val & ~MOVE_MASK)
3120 * No kind of locking is needed in here, because ->can_attach() will
3121 * check this value once in the beginning of the process, and then carry
3122 * on with stale data. This means that changes to this value will only
3123 * affect task migrations starting after the change.
3125 memcg->move_charge_at_immigrate = val;
3129 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3130 struct cftype *cft, u64 val)
3137 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3141 unsigned int lru_mask;
3144 static const struct numa_stat stats[] = {
3145 { "total", LRU_ALL },
3146 { "file", LRU_ALL_FILE },
3147 { "anon", LRU_ALL_ANON },
3148 { "unevictable", BIT(LRU_UNEVICTABLE) },
3150 const struct numa_stat *stat;
3153 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3155 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3156 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3157 seq_printf(m, "%s=%lu", stat->name, nr);
3158 for_each_node_state(nid, N_MEMORY) {
3159 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3161 seq_printf(m, " N%d=%lu", nid, nr);
3166 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3167 struct mem_cgroup *iter;
3170 for_each_mem_cgroup_tree(iter, memcg)
3171 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3172 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3173 for_each_node_state(nid, N_MEMORY) {
3175 for_each_mem_cgroup_tree(iter, memcg)
3176 nr += mem_cgroup_node_nr_lru_pages(
3177 iter, nid, stat->lru_mask);
3178 seq_printf(m, " N%d=%lu", nid, nr);
3185 #endif /* CONFIG_NUMA */
3187 /* Universal VM events cgroup1 shows, original sort order */
3188 unsigned int memcg1_events[] = {
3195 static const char *const memcg1_event_names[] = {
3202 static int memcg_stat_show(struct seq_file *m, void *v)
3204 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3205 unsigned long memory, memsw;
3206 struct mem_cgroup *mi;
3209 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3210 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3212 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3213 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3215 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3216 memcg_page_state(memcg, memcg1_stats[i]) *
3220 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3221 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3222 memcg_sum_events(memcg, memcg1_events[i]));
3224 for (i = 0; i < NR_LRU_LISTS; i++)
3225 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3226 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3228 /* Hierarchical information */
3229 memory = memsw = PAGE_COUNTER_MAX;
3230 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3231 memory = min(memory, mi->memory.limit);
3232 memsw = min(memsw, mi->memsw.limit);
3234 seq_printf(m, "hierarchical_memory_limit %llu\n",
3235 (u64)memory * PAGE_SIZE);
3236 if (do_memsw_account())
3237 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3238 (u64)memsw * PAGE_SIZE);
3240 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3241 unsigned long long val = 0;
3243 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3245 for_each_mem_cgroup_tree(mi, memcg)
3246 val += memcg_page_state(mi, memcg1_stats[i]) *
3248 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i], val);
3251 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++) {
3252 unsigned long long val = 0;
3254 for_each_mem_cgroup_tree(mi, memcg)
3255 val += memcg_sum_events(mi, memcg1_events[i]);
3256 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i], val);
3259 for (i = 0; i < NR_LRU_LISTS; i++) {
3260 unsigned long long val = 0;
3262 for_each_mem_cgroup_tree(mi, memcg)
3263 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3264 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3267 #ifdef CONFIG_DEBUG_VM
3270 struct mem_cgroup_per_node *mz;
3271 struct zone_reclaim_stat *rstat;
3272 unsigned long recent_rotated[2] = {0, 0};
3273 unsigned long recent_scanned[2] = {0, 0};
3275 for_each_online_pgdat(pgdat) {
3276 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3277 rstat = &mz->lruvec.reclaim_stat;
3279 recent_rotated[0] += rstat->recent_rotated[0];
3280 recent_rotated[1] += rstat->recent_rotated[1];
3281 recent_scanned[0] += rstat->recent_scanned[0];
3282 recent_scanned[1] += rstat->recent_scanned[1];
3284 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3285 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3286 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3287 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3294 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3297 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3299 return mem_cgroup_swappiness(memcg);
3302 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3303 struct cftype *cft, u64 val)
3305 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3311 memcg->swappiness = val;
3313 vm_swappiness = val;
3318 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3320 struct mem_cgroup_threshold_ary *t;
3321 unsigned long usage;
3326 t = rcu_dereference(memcg->thresholds.primary);
3328 t = rcu_dereference(memcg->memsw_thresholds.primary);
3333 usage = mem_cgroup_usage(memcg, swap);
3336 * current_threshold points to threshold just below or equal to usage.
3337 * If it's not true, a threshold was crossed after last
3338 * call of __mem_cgroup_threshold().
3340 i = t->current_threshold;
3343 * Iterate backward over array of thresholds starting from
3344 * current_threshold and check if a threshold is crossed.
3345 * If none of thresholds below usage is crossed, we read
3346 * only one element of the array here.
3348 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3349 eventfd_signal(t->entries[i].eventfd, 1);
3351 /* i = current_threshold + 1 */
3355 * Iterate forward over array of thresholds starting from
3356 * current_threshold+1 and check if a threshold is crossed.
3357 * If none of thresholds above usage is crossed, we read
3358 * only one element of the array here.
3360 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3361 eventfd_signal(t->entries[i].eventfd, 1);
3363 /* Update current_threshold */
3364 t->current_threshold = i - 1;
3369 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3372 __mem_cgroup_threshold(memcg, false);
3373 if (do_memsw_account())
3374 __mem_cgroup_threshold(memcg, true);
3376 memcg = parent_mem_cgroup(memcg);
3380 static int compare_thresholds(const void *a, const void *b)
3382 const struct mem_cgroup_threshold *_a = a;
3383 const struct mem_cgroup_threshold *_b = b;
3385 if (_a->threshold > _b->threshold)
3388 if (_a->threshold < _b->threshold)
3394 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3396 struct mem_cgroup_eventfd_list *ev;
3398 spin_lock(&memcg_oom_lock);
3400 list_for_each_entry(ev, &memcg->oom_notify, list)
3401 eventfd_signal(ev->eventfd, 1);
3403 spin_unlock(&memcg_oom_lock);
3407 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3409 struct mem_cgroup *iter;
3411 for_each_mem_cgroup_tree(iter, memcg)
3412 mem_cgroup_oom_notify_cb(iter);
3415 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3416 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3418 struct mem_cgroup_thresholds *thresholds;
3419 struct mem_cgroup_threshold_ary *new;
3420 unsigned long threshold;
3421 unsigned long usage;
3424 ret = page_counter_memparse(args, "-1", &threshold);
3428 mutex_lock(&memcg->thresholds_lock);
3431 thresholds = &memcg->thresholds;
3432 usage = mem_cgroup_usage(memcg, false);
3433 } else if (type == _MEMSWAP) {
3434 thresholds = &memcg->memsw_thresholds;
3435 usage = mem_cgroup_usage(memcg, true);
3439 /* Check if a threshold crossed before adding a new one */
3440 if (thresholds->primary)
3441 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3443 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3445 /* Allocate memory for new array of thresholds */
3446 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3454 /* Copy thresholds (if any) to new array */
3455 if (thresholds->primary) {
3456 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3457 sizeof(struct mem_cgroup_threshold));
3460 /* Add new threshold */
3461 new->entries[size - 1].eventfd = eventfd;
3462 new->entries[size - 1].threshold = threshold;
3464 /* Sort thresholds. Registering of new threshold isn't time-critical */
3465 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3466 compare_thresholds, NULL);
3468 /* Find current threshold */
3469 new->current_threshold = -1;
3470 for (i = 0; i < size; i++) {
3471 if (new->entries[i].threshold <= usage) {
3473 * new->current_threshold will not be used until
3474 * rcu_assign_pointer(), so it's safe to increment
3477 ++new->current_threshold;
3482 /* Free old spare buffer and save old primary buffer as spare */
3483 kfree(thresholds->spare);
3484 thresholds->spare = thresholds->primary;
3486 rcu_assign_pointer(thresholds->primary, new);
3488 /* To be sure that nobody uses thresholds */
3492 mutex_unlock(&memcg->thresholds_lock);
3497 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3498 struct eventfd_ctx *eventfd, const char *args)
3500 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3503 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3504 struct eventfd_ctx *eventfd, const char *args)
3506 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3509 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3510 struct eventfd_ctx *eventfd, enum res_type type)
3512 struct mem_cgroup_thresholds *thresholds;
3513 struct mem_cgroup_threshold_ary *new;
3514 unsigned long usage;
3515 int i, j, size, entries;
3517 mutex_lock(&memcg->thresholds_lock);
3520 thresholds = &memcg->thresholds;
3521 usage = mem_cgroup_usage(memcg, false);
3522 } else if (type == _MEMSWAP) {
3523 thresholds = &memcg->memsw_thresholds;
3524 usage = mem_cgroup_usage(memcg, true);
3528 if (!thresholds->primary)
3531 /* Check if a threshold crossed before removing */
3532 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3534 /* Calculate new number of threshold */
3536 for (i = 0; i < thresholds->primary->size; i++) {
3537 if (thresholds->primary->entries[i].eventfd != eventfd)
3543 new = thresholds->spare;
3545 /* If no items related to eventfd have been cleared, nothing to do */
3549 /* Set thresholds array to NULL if we don't have thresholds */
3558 /* Copy thresholds and find current threshold */
3559 new->current_threshold = -1;
3560 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3561 if (thresholds->primary->entries[i].eventfd == eventfd)
3564 new->entries[j] = thresholds->primary->entries[i];
3565 if (new->entries[j].threshold <= usage) {
3567 * new->current_threshold will not be used
3568 * until rcu_assign_pointer(), so it's safe to increment
3571 ++new->current_threshold;
3577 /* Swap primary and spare array */
3578 thresholds->spare = thresholds->primary;
3580 rcu_assign_pointer(thresholds->primary, new);
3582 /* To be sure that nobody uses thresholds */
3585 /* If all events are unregistered, free the spare array */
3587 kfree(thresholds->spare);
3588 thresholds->spare = NULL;
3591 mutex_unlock(&memcg->thresholds_lock);
3594 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3595 struct eventfd_ctx *eventfd)
3597 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3600 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3601 struct eventfd_ctx *eventfd)
3603 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3606 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3607 struct eventfd_ctx *eventfd, const char *args)
3609 struct mem_cgroup_eventfd_list *event;
3611 event = kmalloc(sizeof(*event), GFP_KERNEL);
3615 spin_lock(&memcg_oom_lock);
3617 event->eventfd = eventfd;
3618 list_add(&event->list, &memcg->oom_notify);
3620 /* already in OOM ? */
3621 if (memcg->under_oom)
3622 eventfd_signal(eventfd, 1);
3623 spin_unlock(&memcg_oom_lock);
3628 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3629 struct eventfd_ctx *eventfd)
3631 struct mem_cgroup_eventfd_list *ev, *tmp;
3633 spin_lock(&memcg_oom_lock);
3635 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3636 if (ev->eventfd == eventfd) {
3637 list_del(&ev->list);
3642 spin_unlock(&memcg_oom_lock);
3645 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3647 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3649 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3650 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3651 seq_printf(sf, "oom_kill %lu\n",
3652 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3656 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3657 struct cftype *cft, u64 val)
3659 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3661 /* cannot set to root cgroup and only 0 and 1 are allowed */
3662 if (!css->parent || !((val == 0) || (val == 1)))
3665 memcg->oom_kill_disable = val;
3667 memcg_oom_recover(memcg);
3672 #ifdef CONFIG_CGROUP_WRITEBACK
3674 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3676 return &memcg->cgwb_list;
3679 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3681 return wb_domain_init(&memcg->cgwb_domain, gfp);
3684 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3686 wb_domain_exit(&memcg->cgwb_domain);
3689 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3691 wb_domain_size_changed(&memcg->cgwb_domain);
3694 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3696 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3698 if (!memcg->css.parent)
3701 return &memcg->cgwb_domain;
3705 * idx can be of type enum memcg_stat_item or node_stat_item.
3706 * Keep in sync with memcg_exact_page().
3708 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
3710 long x = atomic_long_read(&memcg->stat[idx]);
3713 for_each_online_cpu(cpu)
3714 x += per_cpu_ptr(memcg->stat_cpu, cpu)->count[idx];
3721 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3722 * @wb: bdi_writeback in question
3723 * @pfilepages: out parameter for number of file pages
3724 * @pheadroom: out parameter for number of allocatable pages according to memcg
3725 * @pdirty: out parameter for number of dirty pages
3726 * @pwriteback: out parameter for number of pages under writeback
3728 * Determine the numbers of file, headroom, dirty, and writeback pages in
3729 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3730 * is a bit more involved.
3732 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3733 * headroom is calculated as the lowest headroom of itself and the
3734 * ancestors. Note that this doesn't consider the actual amount of
3735 * available memory in the system. The caller should further cap
3736 * *@pheadroom accordingly.
3738 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3739 unsigned long *pheadroom, unsigned long *pdirty,
3740 unsigned long *pwriteback)
3742 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3743 struct mem_cgroup *parent;
3745 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
3747 /* this should eventually include NR_UNSTABLE_NFS */
3748 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
3749 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3750 (1 << LRU_ACTIVE_FILE));
3751 *pheadroom = PAGE_COUNTER_MAX;
3753 while ((parent = parent_mem_cgroup(memcg))) {
3754 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3755 unsigned long used = page_counter_read(&memcg->memory);
3757 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3762 #else /* CONFIG_CGROUP_WRITEBACK */
3764 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3769 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3773 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3777 #endif /* CONFIG_CGROUP_WRITEBACK */
3780 * DO NOT USE IN NEW FILES.
3782 * "cgroup.event_control" implementation.
3784 * This is way over-engineered. It tries to support fully configurable
3785 * events for each user. Such level of flexibility is completely
3786 * unnecessary especially in the light of the planned unified hierarchy.
3788 * Please deprecate this and replace with something simpler if at all
3793 * Unregister event and free resources.
3795 * Gets called from workqueue.
3797 static void memcg_event_remove(struct work_struct *work)
3799 struct mem_cgroup_event *event =
3800 container_of(work, struct mem_cgroup_event, remove);
3801 struct mem_cgroup *memcg = event->memcg;
3803 remove_wait_queue(event->wqh, &event->wait);
3805 event->unregister_event(memcg, event->eventfd);
3807 /* Notify userspace the event is going away. */
3808 eventfd_signal(event->eventfd, 1);
3810 eventfd_ctx_put(event->eventfd);
3812 css_put(&memcg->css);
3816 * Gets called on POLLHUP on eventfd when user closes it.
3818 * Called with wqh->lock held and interrupts disabled.
3820 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
3821 int sync, void *key)
3823 struct mem_cgroup_event *event =
3824 container_of(wait, struct mem_cgroup_event, wait);
3825 struct mem_cgroup *memcg = event->memcg;
3826 unsigned long flags = (unsigned long)key;
3828 if (flags & POLLHUP) {
3830 * If the event has been detached at cgroup removal, we
3831 * can simply return knowing the other side will cleanup
3834 * We can't race against event freeing since the other
3835 * side will require wqh->lock via remove_wait_queue(),
3838 spin_lock(&memcg->event_list_lock);
3839 if (!list_empty(&event->list)) {
3840 list_del_init(&event->list);
3842 * We are in atomic context, but cgroup_event_remove()
3843 * may sleep, so we have to call it in workqueue.
3845 schedule_work(&event->remove);
3847 spin_unlock(&memcg->event_list_lock);
3853 static void memcg_event_ptable_queue_proc(struct file *file,
3854 wait_queue_head_t *wqh, poll_table *pt)
3856 struct mem_cgroup_event *event =
3857 container_of(pt, struct mem_cgroup_event, pt);
3860 add_wait_queue(wqh, &event->wait);
3864 * DO NOT USE IN NEW FILES.
3866 * Parse input and register new cgroup event handler.
3868 * Input must be in format '<event_fd> <control_fd> <args>'.
3869 * Interpretation of args is defined by control file implementation.
3871 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3872 char *buf, size_t nbytes, loff_t off)
3874 struct cgroup_subsys_state *css = of_css(of);
3875 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3876 struct mem_cgroup_event *event;
3877 struct cgroup_subsys_state *cfile_css;
3878 unsigned int efd, cfd;
3881 struct dentry *cdentry;
3886 buf = strstrip(buf);
3888 efd = simple_strtoul(buf, &endp, 10);
3893 cfd = simple_strtoul(buf, &endp, 10);
3894 if ((*endp != ' ') && (*endp != '\0'))
3898 event = kzalloc(sizeof(*event), GFP_KERNEL);
3902 event->memcg = memcg;
3903 INIT_LIST_HEAD(&event->list);
3904 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3905 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3906 INIT_WORK(&event->remove, memcg_event_remove);
3914 event->eventfd = eventfd_ctx_fileget(efile.file);
3915 if (IS_ERR(event->eventfd)) {
3916 ret = PTR_ERR(event->eventfd);
3923 goto out_put_eventfd;
3926 /* the process need read permission on control file */
3927 /* AV: shouldn't we check that it's been opened for read instead? */
3928 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3933 * The control file must be a regular cgroup1 file. As a regular cgroup
3934 * file can't be renamed, it's safe to access its name afterwards.
3936 cdentry = cfile.file->f_path.dentry;
3937 if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
3943 * Determine the event callbacks and set them in @event. This used
3944 * to be done via struct cftype but cgroup core no longer knows
3945 * about these events. The following is crude but the whole thing
3946 * is for compatibility anyway.
3948 * DO NOT ADD NEW FILES.
3950 name = cdentry->d_name.name;
3952 if (!strcmp(name, "memory.usage_in_bytes")) {
3953 event->register_event = mem_cgroup_usage_register_event;
3954 event->unregister_event = mem_cgroup_usage_unregister_event;
3955 } else if (!strcmp(name, "memory.oom_control")) {
3956 event->register_event = mem_cgroup_oom_register_event;
3957 event->unregister_event = mem_cgroup_oom_unregister_event;
3958 } else if (!strcmp(name, "memory.pressure_level")) {
3959 event->register_event = vmpressure_register_event;
3960 event->unregister_event = vmpressure_unregister_event;
3961 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3962 event->register_event = memsw_cgroup_usage_register_event;
3963 event->unregister_event = memsw_cgroup_usage_unregister_event;
3970 * Verify @cfile should belong to @css. Also, remaining events are
3971 * automatically removed on cgroup destruction but the removal is
3972 * asynchronous, so take an extra ref on @css.
3974 cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
3975 &memory_cgrp_subsys);
3977 if (IS_ERR(cfile_css))
3979 if (cfile_css != css) {
3984 ret = event->register_event(memcg, event->eventfd, buf);
3988 efile.file->f_op->poll(efile.file, &event->pt);
3990 spin_lock(&memcg->event_list_lock);
3991 list_add(&event->list, &memcg->event_list);
3992 spin_unlock(&memcg->event_list_lock);
4004 eventfd_ctx_put(event->eventfd);
4013 static struct cftype mem_cgroup_legacy_files[] = {
4015 .name = "usage_in_bytes",
4016 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4017 .read_u64 = mem_cgroup_read_u64,
4020 .name = "max_usage_in_bytes",
4021 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4022 .write = mem_cgroup_reset,
4023 .read_u64 = mem_cgroup_read_u64,
4026 .name = "limit_in_bytes",
4027 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4028 .write = mem_cgroup_write,
4029 .read_u64 = mem_cgroup_read_u64,
4032 .name = "soft_limit_in_bytes",
4033 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4034 .write = mem_cgroup_write,
4035 .read_u64 = mem_cgroup_read_u64,
4039 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4040 .write = mem_cgroup_reset,
4041 .read_u64 = mem_cgroup_read_u64,
4045 .seq_show = memcg_stat_show,
4048 .name = "force_empty",
4049 .write = mem_cgroup_force_empty_write,
4052 .name = "use_hierarchy",
4053 .write_u64 = mem_cgroup_hierarchy_write,
4054 .read_u64 = mem_cgroup_hierarchy_read,
4057 .name = "cgroup.event_control", /* XXX: for compat */
4058 .write = memcg_write_event_control,
4059 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4062 .name = "swappiness",
4063 .read_u64 = mem_cgroup_swappiness_read,
4064 .write_u64 = mem_cgroup_swappiness_write,
4067 .name = "move_charge_at_immigrate",
4068 .read_u64 = mem_cgroup_move_charge_read,
4069 .write_u64 = mem_cgroup_move_charge_write,
4072 .name = "oom_control",
4073 .seq_show = mem_cgroup_oom_control_read,
4074 .write_u64 = mem_cgroup_oom_control_write,
4075 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4078 .name = "pressure_level",
4082 .name = "numa_stat",
4083 .seq_show = memcg_numa_stat_show,
4087 .name = "kmem.limit_in_bytes",
4088 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4089 .write = mem_cgroup_write,
4090 .read_u64 = mem_cgroup_read_u64,
4093 .name = "kmem.usage_in_bytes",
4094 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4095 .read_u64 = mem_cgroup_read_u64,
4098 .name = "kmem.failcnt",
4099 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4100 .write = mem_cgroup_reset,
4101 .read_u64 = mem_cgroup_read_u64,
4104 .name = "kmem.max_usage_in_bytes",
4105 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4106 .write = mem_cgroup_reset,
4107 .read_u64 = mem_cgroup_read_u64,
4109 #ifdef CONFIG_SLABINFO
4111 .name = "kmem.slabinfo",
4112 .seq_start = memcg_slab_start,
4113 .seq_next = memcg_slab_next,
4114 .seq_stop = memcg_slab_stop,
4115 .seq_show = memcg_slab_show,
4119 .name = "kmem.tcp.limit_in_bytes",
4120 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4121 .write = mem_cgroup_write,
4122 .read_u64 = mem_cgroup_read_u64,
4125 .name = "kmem.tcp.usage_in_bytes",
4126 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4127 .read_u64 = mem_cgroup_read_u64,
4130 .name = "kmem.tcp.failcnt",
4131 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4132 .write = mem_cgroup_reset,
4133 .read_u64 = mem_cgroup_read_u64,
4136 .name = "kmem.tcp.max_usage_in_bytes",
4137 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4138 .write = mem_cgroup_reset,
4139 .read_u64 = mem_cgroup_read_u64,
4141 { }, /* terminate */
4145 * Private memory cgroup IDR
4147 * Swap-out records and page cache shadow entries need to store memcg
4148 * references in constrained space, so we maintain an ID space that is
4149 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4150 * memory-controlled cgroups to 64k.
4152 * However, there usually are many references to the oflline CSS after
4153 * the cgroup has been destroyed, such as page cache or reclaimable
4154 * slab objects, that don't need to hang on to the ID. We want to keep
4155 * those dead CSS from occupying IDs, or we might quickly exhaust the
4156 * relatively small ID space and prevent the creation of new cgroups
4157 * even when there are much fewer than 64k cgroups - possibly none.
4159 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4160 * be freed and recycled when it's no longer needed, which is usually
4161 * when the CSS is offlined.
4163 * The only exception to that are records of swapped out tmpfs/shmem
4164 * pages that need to be attributed to live ancestors on swapin. But
4165 * those references are manageable from userspace.
4168 static DEFINE_IDR(mem_cgroup_idr);
4170 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4172 if (memcg->id.id > 0) {
4173 idr_remove(&mem_cgroup_idr, memcg->id.id);
4178 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4180 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4181 atomic_add(n, &memcg->id.ref);
4184 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4186 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4187 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4188 mem_cgroup_id_remove(memcg);
4190 /* Memcg ID pins CSS */
4191 css_put(&memcg->css);
4195 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4197 mem_cgroup_id_get_many(memcg, 1);
4200 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4202 mem_cgroup_id_put_many(memcg, 1);
4206 * mem_cgroup_from_id - look up a memcg from a memcg id
4207 * @id: the memcg id to look up
4209 * Caller must hold rcu_read_lock().
4211 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4213 WARN_ON_ONCE(!rcu_read_lock_held());
4214 return idr_find(&mem_cgroup_idr, id);
4217 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4219 struct mem_cgroup_per_node *pn;
4222 * This routine is called against possible nodes.
4223 * But it's BUG to call kmalloc() against offline node.
4225 * TODO: this routine can waste much memory for nodes which will
4226 * never be onlined. It's better to use memory hotplug callback
4229 if (!node_state(node, N_NORMAL_MEMORY))
4231 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4235 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4236 if (!pn->lruvec_stat_cpu) {
4241 lruvec_init(&pn->lruvec);
4242 pn->usage_in_excess = 0;
4243 pn->on_tree = false;
4246 memcg->nodeinfo[node] = pn;
4250 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4252 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4257 free_percpu(pn->lruvec_stat_cpu);
4261 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4266 free_mem_cgroup_per_node_info(memcg, node);
4267 free_percpu(memcg->stat_cpu);
4271 static void mem_cgroup_free(struct mem_cgroup *memcg)
4273 memcg_wb_domain_exit(memcg);
4274 __mem_cgroup_free(memcg);
4277 static struct mem_cgroup *mem_cgroup_alloc(void)
4279 struct mem_cgroup *memcg;
4283 size = sizeof(struct mem_cgroup);
4284 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4286 memcg = kzalloc(size, GFP_KERNEL);
4290 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4291 1, MEM_CGROUP_ID_MAX,
4293 if (memcg->id.id < 0)
4296 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4297 if (!memcg->stat_cpu)
4301 if (alloc_mem_cgroup_per_node_info(memcg, node))
4304 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4307 INIT_WORK(&memcg->high_work, high_work_func);
4308 memcg->last_scanned_node = MAX_NUMNODES;
4309 INIT_LIST_HEAD(&memcg->oom_notify);
4310 mutex_init(&memcg->thresholds_lock);
4311 spin_lock_init(&memcg->move_lock);
4312 vmpressure_init(&memcg->vmpressure);
4313 INIT_LIST_HEAD(&memcg->event_list);
4314 spin_lock_init(&memcg->event_list_lock);
4315 memcg->socket_pressure = jiffies;
4317 memcg->kmemcg_id = -1;
4319 #ifdef CONFIG_CGROUP_WRITEBACK
4320 INIT_LIST_HEAD(&memcg->cgwb_list);
4322 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4325 mem_cgroup_id_remove(memcg);
4326 __mem_cgroup_free(memcg);
4330 static struct cgroup_subsys_state * __ref
4331 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4333 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4334 struct mem_cgroup *memcg;
4335 long error = -ENOMEM;
4337 memcg = mem_cgroup_alloc();
4339 return ERR_PTR(error);
4341 memcg->high = PAGE_COUNTER_MAX;
4342 memcg->soft_limit = PAGE_COUNTER_MAX;
4344 memcg->swappiness = mem_cgroup_swappiness(parent);
4345 memcg->oom_kill_disable = parent->oom_kill_disable;
4347 if (parent && parent->use_hierarchy) {
4348 memcg->use_hierarchy = true;
4349 page_counter_init(&memcg->memory, &parent->memory);
4350 page_counter_init(&memcg->swap, &parent->swap);
4351 page_counter_init(&memcg->memsw, &parent->memsw);
4352 page_counter_init(&memcg->kmem, &parent->kmem);
4353 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4355 page_counter_init(&memcg->memory, NULL);
4356 page_counter_init(&memcg->swap, NULL);
4357 page_counter_init(&memcg->memsw, NULL);
4358 page_counter_init(&memcg->kmem, NULL);
4359 page_counter_init(&memcg->tcpmem, NULL);
4361 * Deeper hierachy with use_hierarchy == false doesn't make
4362 * much sense so let cgroup subsystem know about this
4363 * unfortunate state in our controller.
4365 if (parent != root_mem_cgroup)
4366 memory_cgrp_subsys.broken_hierarchy = true;
4369 /* The following stuff does not apply to the root */
4371 root_mem_cgroup = memcg;
4375 error = memcg_online_kmem(memcg);
4379 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4380 static_branch_inc(&memcg_sockets_enabled_key);
4384 mem_cgroup_id_remove(memcg);
4385 mem_cgroup_free(memcg);
4386 return ERR_PTR(-ENOMEM);
4389 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4391 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4393 /* Online state pins memcg ID, memcg ID pins CSS */
4394 atomic_set(&memcg->id.ref, 1);
4399 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4401 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4402 struct mem_cgroup_event *event, *tmp;
4405 * Unregister events and notify userspace.
4406 * Notify userspace about cgroup removing only after rmdir of cgroup
4407 * directory to avoid race between userspace and kernelspace.
4409 spin_lock(&memcg->event_list_lock);
4410 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4411 list_del_init(&event->list);
4412 schedule_work(&event->remove);
4414 spin_unlock(&memcg->event_list_lock);
4418 memcg_offline_kmem(memcg);
4419 wb_memcg_offline(memcg);
4421 mem_cgroup_id_put(memcg);
4424 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4426 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4428 invalidate_reclaim_iterators(memcg);
4431 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4433 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4435 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4436 static_branch_dec(&memcg_sockets_enabled_key);
4438 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4439 static_branch_dec(&memcg_sockets_enabled_key);
4441 vmpressure_cleanup(&memcg->vmpressure);
4442 cancel_work_sync(&memcg->high_work);
4443 mem_cgroup_remove_from_trees(memcg);
4444 memcg_free_kmem(memcg);
4445 mem_cgroup_free(memcg);
4449 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4450 * @css: the target css
4452 * Reset the states of the mem_cgroup associated with @css. This is
4453 * invoked when the userland requests disabling on the default hierarchy
4454 * but the memcg is pinned through dependency. The memcg should stop
4455 * applying policies and should revert to the vanilla state as it may be
4456 * made visible again.
4458 * The current implementation only resets the essential configurations.
4459 * This needs to be expanded to cover all the visible parts.
4461 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4463 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4465 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4466 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4467 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4468 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4469 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4471 memcg->high = PAGE_COUNTER_MAX;
4472 memcg->soft_limit = PAGE_COUNTER_MAX;
4473 memcg_wb_domain_size_changed(memcg);
4477 /* Handlers for move charge at task migration. */
4478 static int mem_cgroup_do_precharge(unsigned long count)
4482 /* Try a single bulk charge without reclaim first, kswapd may wake */
4483 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4485 mc.precharge += count;
4489 /* Try charges one by one with reclaim, but do not retry */
4491 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4505 enum mc_target_type {
4512 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4513 unsigned long addr, pte_t ptent)
4515 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4517 if (!page || !page_mapped(page))
4519 if (PageAnon(page)) {
4520 if (!(mc.flags & MOVE_ANON))
4523 if (!(mc.flags & MOVE_FILE))
4526 if (!get_page_unless_zero(page))
4532 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4533 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4534 pte_t ptent, swp_entry_t *entry)
4536 struct page *page = NULL;
4537 swp_entry_t ent = pte_to_swp_entry(ptent);
4539 if (!(mc.flags & MOVE_ANON))
4543 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4544 * a device and because they are not accessible by CPU they are store
4545 * as special swap entry in the CPU page table.
4547 if (is_device_private_entry(ent)) {
4548 page = device_private_entry_to_page(ent);
4550 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4551 * a refcount of 1 when free (unlike normal page)
4553 if (!page_ref_add_unless(page, 1, 1))
4558 if (non_swap_entry(ent))
4562 * Because lookup_swap_cache() updates some statistics counter,
4563 * we call find_get_page() with swapper_space directly.
4565 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4566 if (do_memsw_account())
4567 entry->val = ent.val;
4572 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4573 pte_t ptent, swp_entry_t *entry)
4579 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4580 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4582 struct page *page = NULL;
4583 struct address_space *mapping;
4586 if (!vma->vm_file) /* anonymous vma */
4588 if (!(mc.flags & MOVE_FILE))
4591 mapping = vma->vm_file->f_mapping;
4592 pgoff = linear_page_index(vma, addr);
4594 /* page is moved even if it's not RSS of this task(page-faulted). */
4596 /* shmem/tmpfs may report page out on swap: account for that too. */
4597 if (shmem_mapping(mapping)) {
4598 page = find_get_entry(mapping, pgoff);
4599 if (radix_tree_exceptional_entry(page)) {
4600 swp_entry_t swp = radix_to_swp_entry(page);
4601 if (do_memsw_account())
4603 page = find_get_page(swap_address_space(swp),
4607 page = find_get_page(mapping, pgoff);
4609 page = find_get_page(mapping, pgoff);
4615 * mem_cgroup_move_account - move account of the page
4617 * @compound: charge the page as compound or small page
4618 * @from: mem_cgroup which the page is moved from.
4619 * @to: mem_cgroup which the page is moved to. @from != @to.
4621 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4623 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4626 static int mem_cgroup_move_account(struct page *page,
4628 struct mem_cgroup *from,
4629 struct mem_cgroup *to)
4631 unsigned long flags;
4632 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4636 VM_BUG_ON(from == to);
4637 VM_BUG_ON_PAGE(PageLRU(page), page);
4638 VM_BUG_ON(compound && !PageTransHuge(page));
4641 * Prevent mem_cgroup_migrate() from looking at
4642 * page->mem_cgroup of its source page while we change it.
4645 if (!trylock_page(page))
4649 if (page->mem_cgroup != from)
4652 anon = PageAnon(page);
4654 spin_lock_irqsave(&from->move_lock, flags);
4656 if (!anon && page_mapped(page)) {
4657 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4658 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4662 * move_lock grabbed above and caller set from->moving_account, so
4663 * mod_memcg_page_state will serialize updates to PageDirty.
4664 * So mapping should be stable for dirty pages.
4666 if (!anon && PageDirty(page)) {
4667 struct address_space *mapping = page_mapping(page);
4669 if (mapping_cap_account_dirty(mapping)) {
4670 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4671 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4675 if (PageWriteback(page)) {
4676 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4677 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4681 * It is safe to change page->mem_cgroup here because the page
4682 * is referenced, charged, and isolated - we can't race with
4683 * uncharging, charging, migration, or LRU putback.
4686 /* caller should have done css_get */
4687 page->mem_cgroup = to;
4688 spin_unlock_irqrestore(&from->move_lock, flags);
4692 local_irq_disable();
4693 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4694 memcg_check_events(to, page);
4695 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4696 memcg_check_events(from, page);
4705 * get_mctgt_type - get target type of moving charge
4706 * @vma: the vma the pte to be checked belongs
4707 * @addr: the address corresponding to the pte to be checked
4708 * @ptent: the pte to be checked
4709 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4712 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4713 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4714 * move charge. if @target is not NULL, the page is stored in target->page
4715 * with extra refcnt got(Callers should handle it).
4716 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4717 * target for charge migration. if @target is not NULL, the entry is stored
4719 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4720 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4721 * For now we such page is charge like a regular page would be as for all
4722 * intent and purposes it is just special memory taking the place of a
4725 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4727 * Called with pte lock held.
4730 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4731 unsigned long addr, pte_t ptent, union mc_target *target)
4733 struct page *page = NULL;
4734 enum mc_target_type ret = MC_TARGET_NONE;
4735 swp_entry_t ent = { .val = 0 };
4737 if (pte_present(ptent))
4738 page = mc_handle_present_pte(vma, addr, ptent);
4739 else if (is_swap_pte(ptent))
4740 page = mc_handle_swap_pte(vma, ptent, &ent);
4741 else if (pte_none(ptent))
4742 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4744 if (!page && !ent.val)
4748 * Do only loose check w/o serialization.
4749 * mem_cgroup_move_account() checks the page is valid or
4750 * not under LRU exclusion.
4752 if (page->mem_cgroup == mc.from) {
4753 ret = MC_TARGET_PAGE;
4754 if (is_device_private_page(page) ||
4755 is_device_public_page(page))
4756 ret = MC_TARGET_DEVICE;
4758 target->page = page;
4760 if (!ret || !target)
4764 * There is a swap entry and a page doesn't exist or isn't charged.
4765 * But we cannot move a tail-page in a THP.
4767 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
4768 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4769 ret = MC_TARGET_SWAP;
4776 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4778 * We don't consider PMD mapped swapping or file mapped pages because THP does
4779 * not support them for now.
4780 * Caller should make sure that pmd_trans_huge(pmd) is true.
4782 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4783 unsigned long addr, pmd_t pmd, union mc_target *target)
4785 struct page *page = NULL;
4786 enum mc_target_type ret = MC_TARGET_NONE;
4788 if (unlikely(is_swap_pmd(pmd))) {
4789 VM_BUG_ON(thp_migration_supported() &&
4790 !is_pmd_migration_entry(pmd));
4793 page = pmd_page(pmd);
4794 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4795 if (!(mc.flags & MOVE_ANON))
4797 if (page->mem_cgroup == mc.from) {
4798 ret = MC_TARGET_PAGE;
4801 target->page = page;
4807 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4808 unsigned long addr, pmd_t pmd, union mc_target *target)
4810 return MC_TARGET_NONE;
4814 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4815 unsigned long addr, unsigned long end,
4816 struct mm_walk *walk)
4818 struct vm_area_struct *vma = walk->vma;
4822 ptl = pmd_trans_huge_lock(pmd, vma);
4825 * Note their can not be MC_TARGET_DEVICE for now as we do not
4826 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
4827 * MEMORY_DEVICE_PRIVATE but this might change.
4829 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4830 mc.precharge += HPAGE_PMD_NR;
4835 if (pmd_trans_unstable(pmd))
4837 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4838 for (; addr != end; pte++, addr += PAGE_SIZE)
4839 if (get_mctgt_type(vma, addr, *pte, NULL))
4840 mc.precharge++; /* increment precharge temporarily */
4841 pte_unmap_unlock(pte - 1, ptl);
4847 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4849 unsigned long precharge;
4851 struct mm_walk mem_cgroup_count_precharge_walk = {
4852 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4855 down_read(&mm->mmap_sem);
4856 walk_page_range(0, mm->highest_vm_end,
4857 &mem_cgroup_count_precharge_walk);
4858 up_read(&mm->mmap_sem);
4860 precharge = mc.precharge;
4866 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4868 unsigned long precharge = mem_cgroup_count_precharge(mm);
4870 VM_BUG_ON(mc.moving_task);
4871 mc.moving_task = current;
4872 return mem_cgroup_do_precharge(precharge);
4875 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4876 static void __mem_cgroup_clear_mc(void)
4878 struct mem_cgroup *from = mc.from;
4879 struct mem_cgroup *to = mc.to;
4881 /* we must uncharge all the leftover precharges from mc.to */
4883 cancel_charge(mc.to, mc.precharge);
4887 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4888 * we must uncharge here.
4890 if (mc.moved_charge) {
4891 cancel_charge(mc.from, mc.moved_charge);
4892 mc.moved_charge = 0;
4894 /* we must fixup refcnts and charges */
4895 if (mc.moved_swap) {
4896 /* uncharge swap account from the old cgroup */
4897 if (!mem_cgroup_is_root(mc.from))
4898 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4900 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4903 * we charged both to->memory and to->memsw, so we
4904 * should uncharge to->memory.
4906 if (!mem_cgroup_is_root(mc.to))
4907 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4909 css_put_many(&mc.to->css, mc.moved_swap);
4913 memcg_oom_recover(from);
4914 memcg_oom_recover(to);
4915 wake_up_all(&mc.waitq);
4918 static void mem_cgroup_clear_mc(void)
4920 struct mm_struct *mm = mc.mm;
4923 * we must clear moving_task before waking up waiters at the end of
4926 mc.moving_task = NULL;
4927 __mem_cgroup_clear_mc();
4928 spin_lock(&mc.lock);
4932 spin_unlock(&mc.lock);
4937 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4939 struct cgroup_subsys_state *css;
4940 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4941 struct mem_cgroup *from;
4942 struct task_struct *leader, *p;
4943 struct mm_struct *mm;
4944 unsigned long move_flags;
4947 /* charge immigration isn't supported on the default hierarchy */
4948 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4952 * Multi-process migrations only happen on the default hierarchy
4953 * where charge immigration is not used. Perform charge
4954 * immigration if @tset contains a leader and whine if there are
4958 cgroup_taskset_for_each_leader(leader, css, tset) {
4961 memcg = mem_cgroup_from_css(css);
4967 * We are now commited to this value whatever it is. Changes in this
4968 * tunable will only affect upcoming migrations, not the current one.
4969 * So we need to save it, and keep it going.
4971 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4975 from = mem_cgroup_from_task(p);
4977 VM_BUG_ON(from == memcg);
4979 mm = get_task_mm(p);
4982 /* We move charges only when we move a owner of the mm */
4983 if (mm->owner == p) {
4986 VM_BUG_ON(mc.precharge);
4987 VM_BUG_ON(mc.moved_charge);
4988 VM_BUG_ON(mc.moved_swap);
4990 spin_lock(&mc.lock);
4994 mc.flags = move_flags;
4995 spin_unlock(&mc.lock);
4996 /* We set mc.moving_task later */
4998 ret = mem_cgroup_precharge_mc(mm);
5000 mem_cgroup_clear_mc();
5007 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5010 mem_cgroup_clear_mc();
5013 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5014 unsigned long addr, unsigned long end,
5015 struct mm_walk *walk)
5018 struct vm_area_struct *vma = walk->vma;
5021 enum mc_target_type target_type;
5022 union mc_target target;
5025 ptl = pmd_trans_huge_lock(pmd, vma);
5027 if (mc.precharge < HPAGE_PMD_NR) {
5031 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5032 if (target_type == MC_TARGET_PAGE) {
5034 if (!isolate_lru_page(page)) {
5035 if (!mem_cgroup_move_account(page, true,
5037 mc.precharge -= HPAGE_PMD_NR;
5038 mc.moved_charge += HPAGE_PMD_NR;
5040 putback_lru_page(page);
5043 } else if (target_type == MC_TARGET_DEVICE) {
5045 if (!mem_cgroup_move_account(page, true,
5047 mc.precharge -= HPAGE_PMD_NR;
5048 mc.moved_charge += HPAGE_PMD_NR;
5056 if (pmd_trans_unstable(pmd))
5059 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5060 for (; addr != end; addr += PAGE_SIZE) {
5061 pte_t ptent = *(pte++);
5062 bool device = false;
5068 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5069 case MC_TARGET_DEVICE:
5072 case MC_TARGET_PAGE:
5075 * We can have a part of the split pmd here. Moving it
5076 * can be done but it would be too convoluted so simply
5077 * ignore such a partial THP and keep it in original
5078 * memcg. There should be somebody mapping the head.
5080 if (PageTransCompound(page))
5082 if (!device && isolate_lru_page(page))
5084 if (!mem_cgroup_move_account(page, false,
5087 /* we uncharge from mc.from later. */
5091 putback_lru_page(page);
5092 put: /* get_mctgt_type() gets the page */
5095 case MC_TARGET_SWAP:
5097 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5099 mem_cgroup_id_get_many(mc.to, 1);
5100 /* we fixup other refcnts and charges later. */
5108 pte_unmap_unlock(pte - 1, ptl);
5113 * We have consumed all precharges we got in can_attach().
5114 * We try charge one by one, but don't do any additional
5115 * charges to mc.to if we have failed in charge once in attach()
5118 ret = mem_cgroup_do_precharge(1);
5126 static void mem_cgroup_move_charge(void)
5128 struct mm_walk mem_cgroup_move_charge_walk = {
5129 .pmd_entry = mem_cgroup_move_charge_pte_range,
5133 lru_add_drain_all();
5135 * Signal lock_page_memcg() to take the memcg's move_lock
5136 * while we're moving its pages to another memcg. Then wait
5137 * for already started RCU-only updates to finish.
5139 atomic_inc(&mc.from->moving_account);
5142 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5144 * Someone who are holding the mmap_sem might be waiting in
5145 * waitq. So we cancel all extra charges, wake up all waiters,
5146 * and retry. Because we cancel precharges, we might not be able
5147 * to move enough charges, but moving charge is a best-effort
5148 * feature anyway, so it wouldn't be a big problem.
5150 __mem_cgroup_clear_mc();
5155 * When we have consumed all precharges and failed in doing
5156 * additional charge, the page walk just aborts.
5158 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5160 up_read(&mc.mm->mmap_sem);
5161 atomic_dec(&mc.from->moving_account);
5164 static void mem_cgroup_move_task(void)
5167 mem_cgroup_move_charge();
5168 mem_cgroup_clear_mc();
5171 #else /* !CONFIG_MMU */
5172 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5176 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5179 static void mem_cgroup_move_task(void)
5185 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5186 * to verify whether we're attached to the default hierarchy on each mount
5189 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5192 * use_hierarchy is forced on the default hierarchy. cgroup core
5193 * guarantees that @root doesn't have any children, so turning it
5194 * on for the root memcg is enough.
5196 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5197 root_mem_cgroup->use_hierarchy = true;
5199 root_mem_cgroup->use_hierarchy = false;
5202 static u64 memory_current_read(struct cgroup_subsys_state *css,
5205 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5207 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5210 static int memory_low_show(struct seq_file *m, void *v)
5212 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5213 unsigned long low = READ_ONCE(memcg->low);
5215 if (low == PAGE_COUNTER_MAX)
5216 seq_puts(m, "max\n");
5218 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5223 static ssize_t memory_low_write(struct kernfs_open_file *of,
5224 char *buf, size_t nbytes, loff_t off)
5226 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5230 buf = strstrip(buf);
5231 err = page_counter_memparse(buf, "max", &low);
5240 static int memory_high_show(struct seq_file *m, void *v)
5242 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5243 unsigned long high = READ_ONCE(memcg->high);
5245 if (high == PAGE_COUNTER_MAX)
5246 seq_puts(m, "max\n");
5248 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5253 static ssize_t memory_high_write(struct kernfs_open_file *of,
5254 char *buf, size_t nbytes, loff_t off)
5256 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5257 unsigned long nr_pages;
5261 buf = strstrip(buf);
5262 err = page_counter_memparse(buf, "max", &high);
5268 nr_pages = page_counter_read(&memcg->memory);
5269 if (nr_pages > high)
5270 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5273 memcg_wb_domain_size_changed(memcg);
5277 static int memory_max_show(struct seq_file *m, void *v)
5279 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5280 unsigned long max = READ_ONCE(memcg->memory.limit);
5282 if (max == PAGE_COUNTER_MAX)
5283 seq_puts(m, "max\n");
5285 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5290 static ssize_t memory_max_write(struct kernfs_open_file *of,
5291 char *buf, size_t nbytes, loff_t off)
5293 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5294 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5295 bool drained = false;
5299 buf = strstrip(buf);
5300 err = page_counter_memparse(buf, "max", &max);
5304 xchg(&memcg->memory.limit, max);
5307 unsigned long nr_pages = page_counter_read(&memcg->memory);
5309 if (nr_pages <= max)
5312 if (signal_pending(current)) {
5318 drain_all_stock(memcg);
5324 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5330 memcg_memory_event(memcg, MEMCG_OOM);
5331 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5335 memcg_wb_domain_size_changed(memcg);
5339 static int memory_events_show(struct seq_file *m, void *v)
5341 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5343 seq_printf(m, "low %lu\n",
5344 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5345 seq_printf(m, "high %lu\n",
5346 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5347 seq_printf(m, "max %lu\n",
5348 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5349 seq_printf(m, "oom %lu\n",
5350 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5351 seq_printf(m, "oom_kill %lu\n",
5352 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5357 static int memory_stat_show(struct seq_file *m, void *v)
5359 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5360 unsigned long stat[MEMCG_NR_STAT];
5361 unsigned long events[NR_VM_EVENT_ITEMS];
5365 * Provide statistics on the state of the memory subsystem as
5366 * well as cumulative event counters that show past behavior.
5368 * This list is ordered following a combination of these gradients:
5369 * 1) generic big picture -> specifics and details
5370 * 2) reflecting userspace activity -> reflecting kernel heuristics
5372 * Current memory state:
5375 tree_stat(memcg, stat);
5376 tree_events(memcg, events);
5378 seq_printf(m, "anon %llu\n",
5379 (u64)stat[MEMCG_RSS] * PAGE_SIZE);
5380 seq_printf(m, "file %llu\n",
5381 (u64)stat[MEMCG_CACHE] * PAGE_SIZE);
5382 seq_printf(m, "kernel_stack %llu\n",
5383 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5384 seq_printf(m, "slab %llu\n",
5385 (u64)(stat[NR_SLAB_RECLAIMABLE] +
5386 stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5387 seq_printf(m, "sock %llu\n",
5388 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5390 seq_printf(m, "shmem %llu\n",
5391 (u64)stat[NR_SHMEM] * PAGE_SIZE);
5392 seq_printf(m, "file_mapped %llu\n",
5393 (u64)stat[NR_FILE_MAPPED] * PAGE_SIZE);
5394 seq_printf(m, "file_dirty %llu\n",
5395 (u64)stat[NR_FILE_DIRTY] * PAGE_SIZE);
5396 seq_printf(m, "file_writeback %llu\n",
5397 (u64)stat[NR_WRITEBACK] * PAGE_SIZE);
5399 for (i = 0; i < NR_LRU_LISTS; i++) {
5400 struct mem_cgroup *mi;
5401 unsigned long val = 0;
5403 for_each_mem_cgroup_tree(mi, memcg)
5404 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5405 seq_printf(m, "%s %llu\n",
5406 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5409 seq_printf(m, "slab_reclaimable %llu\n",
5410 (u64)stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5411 seq_printf(m, "slab_unreclaimable %llu\n",
5412 (u64)stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5414 /* Accumulated memory events */
5416 seq_printf(m, "pgfault %lu\n", events[PGFAULT]);
5417 seq_printf(m, "pgmajfault %lu\n", events[PGMAJFAULT]);
5419 seq_printf(m, "pgrefill %lu\n", events[PGREFILL]);
5420 seq_printf(m, "pgscan %lu\n", events[PGSCAN_KSWAPD] +
5421 events[PGSCAN_DIRECT]);
5422 seq_printf(m, "pgsteal %lu\n", events[PGSTEAL_KSWAPD] +
5423 events[PGSTEAL_DIRECT]);
5424 seq_printf(m, "pgactivate %lu\n", events[PGACTIVATE]);
5425 seq_printf(m, "pgdeactivate %lu\n", events[PGDEACTIVATE]);
5426 seq_printf(m, "pglazyfree %lu\n", events[PGLAZYFREE]);
5427 seq_printf(m, "pglazyfreed %lu\n", events[PGLAZYFREED]);
5429 seq_printf(m, "workingset_refault %lu\n",
5430 stat[WORKINGSET_REFAULT]);
5431 seq_printf(m, "workingset_activate %lu\n",
5432 stat[WORKINGSET_ACTIVATE]);
5433 seq_printf(m, "workingset_nodereclaim %lu\n",
5434 stat[WORKINGSET_NODERECLAIM]);
5439 static struct cftype memory_files[] = {
5442 .flags = CFTYPE_NOT_ON_ROOT,
5443 .read_u64 = memory_current_read,
5447 .flags = CFTYPE_NOT_ON_ROOT,
5448 .seq_show = memory_low_show,
5449 .write = memory_low_write,
5453 .flags = CFTYPE_NOT_ON_ROOT,
5454 .seq_show = memory_high_show,
5455 .write = memory_high_write,
5459 .flags = CFTYPE_NOT_ON_ROOT,
5460 .seq_show = memory_max_show,
5461 .write = memory_max_write,
5465 .flags = CFTYPE_NOT_ON_ROOT,
5466 .file_offset = offsetof(struct mem_cgroup, events_file),
5467 .seq_show = memory_events_show,
5471 .flags = CFTYPE_NOT_ON_ROOT,
5472 .seq_show = memory_stat_show,
5477 struct cgroup_subsys memory_cgrp_subsys = {
5478 .css_alloc = mem_cgroup_css_alloc,
5479 .css_online = mem_cgroup_css_online,
5480 .css_offline = mem_cgroup_css_offline,
5481 .css_released = mem_cgroup_css_released,
5482 .css_free = mem_cgroup_css_free,
5483 .css_reset = mem_cgroup_css_reset,
5484 .can_attach = mem_cgroup_can_attach,
5485 .cancel_attach = mem_cgroup_cancel_attach,
5486 .post_attach = mem_cgroup_move_task,
5487 .bind = mem_cgroup_bind,
5488 .dfl_cftypes = memory_files,
5489 .legacy_cftypes = mem_cgroup_legacy_files,
5494 * mem_cgroup_low - check if memory consumption is below the normal range
5495 * @root: the top ancestor of the sub-tree being checked
5496 * @memcg: the memory cgroup to check
5498 * Returns %true if memory consumption of @memcg, and that of all
5499 * ancestors up to (but not including) @root, is below the normal range.
5501 * @root is exclusive; it is never low when looked at directly and isn't
5502 * checked when traversing the hierarchy.
5504 * Excluding @root enables using memory.low to prioritize memory usage
5505 * between cgroups within a subtree of the hierarchy that is limited by
5506 * memory.high or memory.max.
5508 * For example, given cgroup A with children B and C:
5516 * 1. A/memory.current > A/memory.high
5517 * 2. A/B/memory.current < A/B/memory.low
5518 * 3. A/C/memory.current >= A/C/memory.low
5520 * As 'A' is high, i.e. triggers reclaim from 'A', and 'B' is low, we
5521 * should reclaim from 'C' until 'A' is no longer high or until we can
5522 * no longer reclaim from 'C'. If 'A', i.e. @root, isn't excluded by
5523 * mem_cgroup_low when reclaming from 'A', then 'B' won't be considered
5524 * low and we will reclaim indiscriminately from both 'B' and 'C'.
5526 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5528 if (mem_cgroup_disabled())
5532 root = root_mem_cgroup;
5536 for (; memcg != root; memcg = parent_mem_cgroup(memcg)) {
5537 if (page_counter_read(&memcg->memory) >= memcg->low)
5545 * mem_cgroup_try_charge - try charging a page
5546 * @page: page to charge
5547 * @mm: mm context of the victim
5548 * @gfp_mask: reclaim mode
5549 * @memcgp: charged memcg return
5550 * @compound: charge the page as compound or small page
5552 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5553 * pages according to @gfp_mask if necessary.
5555 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5556 * Otherwise, an error code is returned.
5558 * After page->mapping has been set up, the caller must finalize the
5559 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5560 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5562 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5563 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5566 struct mem_cgroup *memcg = NULL;
5567 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5570 if (mem_cgroup_disabled())
5573 if (PageSwapCache(page)) {
5575 * Every swap fault against a single page tries to charge the
5576 * page, bail as early as possible. shmem_unuse() encounters
5577 * already charged pages, too. The USED bit is protected by
5578 * the page lock, which serializes swap cache removal, which
5579 * in turn serializes uncharging.
5581 VM_BUG_ON_PAGE(!PageLocked(page), page);
5582 if (compound_head(page)->mem_cgroup)
5585 if (do_swap_account) {
5586 swp_entry_t ent = { .val = page_private(page), };
5587 unsigned short id = lookup_swap_cgroup_id(ent);
5590 memcg = mem_cgroup_from_id(id);
5591 if (memcg && !css_tryget_online(&memcg->css))
5598 memcg = get_mem_cgroup_from_mm(mm);
5600 ret = try_charge(memcg, gfp_mask, nr_pages);
5602 css_put(&memcg->css);
5609 * mem_cgroup_commit_charge - commit a page charge
5610 * @page: page to charge
5611 * @memcg: memcg to charge the page to
5612 * @lrucare: page might be on LRU already
5613 * @compound: charge the page as compound or small page
5615 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5616 * after page->mapping has been set up. This must happen atomically
5617 * as part of the page instantiation, i.e. under the page table lock
5618 * for anonymous pages, under the page lock for page and swap cache.
5620 * In addition, the page must not be on the LRU during the commit, to
5621 * prevent racing with task migration. If it might be, use @lrucare.
5623 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5625 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5626 bool lrucare, bool compound)
5628 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5630 VM_BUG_ON_PAGE(!page->mapping, page);
5631 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5633 if (mem_cgroup_disabled())
5636 * Swap faults will attempt to charge the same page multiple
5637 * times. But reuse_swap_page() might have removed the page
5638 * from swapcache already, so we can't check PageSwapCache().
5643 commit_charge(page, memcg, lrucare);
5645 local_irq_disable();
5646 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5647 memcg_check_events(memcg, page);
5650 if (do_memsw_account() && PageSwapCache(page)) {
5651 swp_entry_t entry = { .val = page_private(page) };
5653 * The swap entry might not get freed for a long time,
5654 * let's not wait for it. The page already received a
5655 * memory+swap charge, drop the swap entry duplicate.
5657 mem_cgroup_uncharge_swap(entry, nr_pages);
5662 * mem_cgroup_cancel_charge - cancel a page charge
5663 * @page: page to charge
5664 * @memcg: memcg to charge the page to
5665 * @compound: charge the page as compound or small page
5667 * Cancel a charge transaction started by mem_cgroup_try_charge().
5669 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5672 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5674 if (mem_cgroup_disabled())
5677 * Swap faults will attempt to charge the same page multiple
5678 * times. But reuse_swap_page() might have removed the page
5679 * from swapcache already, so we can't check PageSwapCache().
5684 cancel_charge(memcg, nr_pages);
5687 struct uncharge_gather {
5688 struct mem_cgroup *memcg;
5689 unsigned long pgpgout;
5690 unsigned long nr_anon;
5691 unsigned long nr_file;
5692 unsigned long nr_kmem;
5693 unsigned long nr_huge;
5694 unsigned long nr_shmem;
5695 struct page *dummy_page;
5698 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
5700 memset(ug, 0, sizeof(*ug));
5703 static void uncharge_batch(const struct uncharge_gather *ug)
5705 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
5706 unsigned long flags;
5708 if (!mem_cgroup_is_root(ug->memcg)) {
5709 page_counter_uncharge(&ug->memcg->memory, nr_pages);
5710 if (do_memsw_account())
5711 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
5712 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
5713 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
5714 memcg_oom_recover(ug->memcg);
5717 local_irq_save(flags);
5718 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
5719 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
5720 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
5721 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
5722 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
5723 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
5724 memcg_check_events(ug->memcg, ug->dummy_page);
5725 local_irq_restore(flags);
5727 if (!mem_cgroup_is_root(ug->memcg))
5728 css_put_many(&ug->memcg->css, nr_pages);
5731 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
5733 VM_BUG_ON_PAGE(PageLRU(page), page);
5734 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
5735 !PageHWPoison(page) , page);
5737 if (!page->mem_cgroup)
5741 * Nobody should be changing or seriously looking at
5742 * page->mem_cgroup at this point, we have fully
5743 * exclusive access to the page.
5746 if (ug->memcg != page->mem_cgroup) {
5749 uncharge_gather_clear(ug);
5751 ug->memcg = page->mem_cgroup;
5754 if (!PageKmemcg(page)) {
5755 unsigned int nr_pages = 1;
5757 if (PageTransHuge(page)) {
5758 nr_pages <<= compound_order(page);
5759 ug->nr_huge += nr_pages;
5762 ug->nr_anon += nr_pages;
5764 ug->nr_file += nr_pages;
5765 if (PageSwapBacked(page))
5766 ug->nr_shmem += nr_pages;
5770 ug->nr_kmem += 1 << compound_order(page);
5771 __ClearPageKmemcg(page);
5774 ug->dummy_page = page;
5775 page->mem_cgroup = NULL;
5778 static void uncharge_list(struct list_head *page_list)
5780 struct uncharge_gather ug;
5781 struct list_head *next;
5783 uncharge_gather_clear(&ug);
5786 * Note that the list can be a single page->lru; hence the
5787 * do-while loop instead of a simple list_for_each_entry().
5789 next = page_list->next;
5793 page = list_entry(next, struct page, lru);
5794 next = page->lru.next;
5796 uncharge_page(page, &ug);
5797 } while (next != page_list);
5800 uncharge_batch(&ug);
5804 * mem_cgroup_uncharge - uncharge a page
5805 * @page: page to uncharge
5807 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5808 * mem_cgroup_commit_charge().
5810 void mem_cgroup_uncharge(struct page *page)
5812 struct uncharge_gather ug;
5814 if (mem_cgroup_disabled())
5817 /* Don't touch page->lru of any random page, pre-check: */
5818 if (!page->mem_cgroup)
5821 uncharge_gather_clear(&ug);
5822 uncharge_page(page, &ug);
5823 uncharge_batch(&ug);
5827 * mem_cgroup_uncharge_list - uncharge a list of page
5828 * @page_list: list of pages to uncharge
5830 * Uncharge a list of pages previously charged with
5831 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5833 void mem_cgroup_uncharge_list(struct list_head *page_list)
5835 if (mem_cgroup_disabled())
5838 if (!list_empty(page_list))
5839 uncharge_list(page_list);
5843 * mem_cgroup_migrate - charge a page's replacement
5844 * @oldpage: currently circulating page
5845 * @newpage: replacement page
5847 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5848 * be uncharged upon free.
5850 * Both pages must be locked, @newpage->mapping must be set up.
5852 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5854 struct mem_cgroup *memcg;
5855 unsigned int nr_pages;
5857 unsigned long flags;
5859 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5860 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5861 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5862 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5865 if (mem_cgroup_disabled())
5868 /* Page cache replacement: new page already charged? */
5869 if (newpage->mem_cgroup)
5872 /* Swapcache readahead pages can get replaced before being charged */
5873 memcg = oldpage->mem_cgroup;
5877 /* Force-charge the new page. The old one will be freed soon */
5878 compound = PageTransHuge(newpage);
5879 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5881 page_counter_charge(&memcg->memory, nr_pages);
5882 if (do_memsw_account())
5883 page_counter_charge(&memcg->memsw, nr_pages);
5884 css_get_many(&memcg->css, nr_pages);
5886 commit_charge(newpage, memcg, false);
5888 local_irq_save(flags);
5889 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5890 memcg_check_events(memcg, newpage);
5891 local_irq_restore(flags);
5894 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5895 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5897 void mem_cgroup_sk_alloc(struct sock *sk)
5899 struct mem_cgroup *memcg;
5901 if (!mem_cgroup_sockets_enabled)
5904 /* Do not associate the sock with unrelated interrupted task's memcg. */
5909 memcg = mem_cgroup_from_task(current);
5910 if (memcg == root_mem_cgroup)
5912 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5914 if (css_tryget_online(&memcg->css))
5915 sk->sk_memcg = memcg;
5920 void mem_cgroup_sk_free(struct sock *sk)
5923 css_put(&sk->sk_memcg->css);
5927 * mem_cgroup_charge_skmem - charge socket memory
5928 * @memcg: memcg to charge
5929 * @nr_pages: number of pages to charge
5931 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5932 * @memcg's configured limit, %false if the charge had to be forced.
5934 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5936 gfp_t gfp_mask = GFP_KERNEL;
5938 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5939 struct page_counter *fail;
5941 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5942 memcg->tcpmem_pressure = 0;
5945 page_counter_charge(&memcg->tcpmem, nr_pages);
5946 memcg->tcpmem_pressure = 1;
5950 /* Don't block in the packet receive path */
5952 gfp_mask = GFP_NOWAIT;
5954 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
5956 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5959 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5964 * mem_cgroup_uncharge_skmem - uncharge socket memory
5965 * @memcg - memcg to uncharge
5966 * @nr_pages - number of pages to uncharge
5968 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5970 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5971 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5975 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
5977 refill_stock(memcg, nr_pages);
5980 static int __init cgroup_memory(char *s)
5984 while ((token = strsep(&s, ",")) != NULL) {
5987 if (!strcmp(token, "nosocket"))
5988 cgroup_memory_nosocket = true;
5989 if (!strcmp(token, "nokmem"))
5990 cgroup_memory_nokmem = true;
5994 __setup("cgroup.memory=", cgroup_memory);
5997 * subsys_initcall() for memory controller.
5999 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6000 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6001 * basically everything that doesn't depend on a specific mem_cgroup structure
6002 * should be initialized from here.
6004 static int __init mem_cgroup_init(void)
6010 * Kmem cache creation is mostly done with the slab_mutex held,
6011 * so use a workqueue with limited concurrency to avoid stalling
6012 * all worker threads in case lots of cgroups are created and
6013 * destroyed simultaneously.
6015 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6016 BUG_ON(!memcg_kmem_cache_wq);
6019 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6020 memcg_hotplug_cpu_dead);
6022 for_each_possible_cpu(cpu)
6023 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6026 for_each_node(node) {
6027 struct mem_cgroup_tree_per_node *rtpn;
6029 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6030 node_online(node) ? node : NUMA_NO_NODE);
6032 rtpn->rb_root = RB_ROOT;
6033 rtpn->rb_rightmost = NULL;
6034 spin_lock_init(&rtpn->lock);
6035 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6040 subsys_initcall(mem_cgroup_init);
6042 #ifdef CONFIG_MEMCG_SWAP
6043 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6045 while (!atomic_inc_not_zero(&memcg->id.ref)) {
6047 * The root cgroup cannot be destroyed, so it's refcount must
6050 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6054 memcg = parent_mem_cgroup(memcg);
6056 memcg = root_mem_cgroup;
6062 * mem_cgroup_swapout - transfer a memsw charge to swap
6063 * @page: page whose memsw charge to transfer
6064 * @entry: swap entry to move the charge to
6066 * Transfer the memsw charge of @page to @entry.
6068 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6070 struct mem_cgroup *memcg, *swap_memcg;
6071 unsigned int nr_entries;
6072 unsigned short oldid;
6074 VM_BUG_ON_PAGE(PageLRU(page), page);
6075 VM_BUG_ON_PAGE(page_count(page), page);
6077 if (!do_memsw_account())
6080 memcg = page->mem_cgroup;
6082 /* Readahead page, never charged */
6087 * In case the memcg owning these pages has been offlined and doesn't
6088 * have an ID allocated to it anymore, charge the closest online
6089 * ancestor for the swap instead and transfer the memory+swap charge.
6091 swap_memcg = mem_cgroup_id_get_online(memcg);
6092 nr_entries = hpage_nr_pages(page);
6093 /* Get references for the tail pages, too */
6095 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6096 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6098 VM_BUG_ON_PAGE(oldid, page);
6099 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6101 page->mem_cgroup = NULL;
6103 if (!mem_cgroup_is_root(memcg))
6104 page_counter_uncharge(&memcg->memory, nr_entries);
6106 if (memcg != swap_memcg) {
6107 if (!mem_cgroup_is_root(swap_memcg))
6108 page_counter_charge(&swap_memcg->memsw, nr_entries);
6109 page_counter_uncharge(&memcg->memsw, nr_entries);
6113 * Interrupts should be disabled here because the caller holds the
6114 * mapping->tree_lock lock which is taken with interrupts-off. It is
6115 * important here to have the interrupts disabled because it is the
6116 * only synchronisation we have for udpating the per-CPU variables.
6118 VM_BUG_ON(!irqs_disabled());
6119 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6121 memcg_check_events(memcg, page);
6123 if (!mem_cgroup_is_root(memcg))
6124 css_put_many(&memcg->css, nr_entries);
6128 * mem_cgroup_try_charge_swap - try charging swap space for a page
6129 * @page: page being added to swap
6130 * @entry: swap entry to charge
6132 * Try to charge @page's memcg for the swap space at @entry.
6134 * Returns 0 on success, -ENOMEM on failure.
6136 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6138 unsigned int nr_pages = hpage_nr_pages(page);
6139 struct page_counter *counter;
6140 struct mem_cgroup *memcg;
6141 unsigned short oldid;
6143 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6146 memcg = page->mem_cgroup;
6148 /* Readahead page, never charged */
6152 memcg = mem_cgroup_id_get_online(memcg);
6154 if (!mem_cgroup_is_root(memcg) &&
6155 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6156 mem_cgroup_id_put(memcg);
6160 /* Get references for the tail pages, too */
6162 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6163 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6164 VM_BUG_ON_PAGE(oldid, page);
6165 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6171 * mem_cgroup_uncharge_swap - uncharge swap space
6172 * @entry: swap entry to uncharge
6173 * @nr_pages: the amount of swap space to uncharge
6175 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6177 struct mem_cgroup *memcg;
6180 if (!do_swap_account)
6183 id = swap_cgroup_record(entry, 0, nr_pages);
6185 memcg = mem_cgroup_from_id(id);
6187 if (!mem_cgroup_is_root(memcg)) {
6188 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6189 page_counter_uncharge(&memcg->swap, nr_pages);
6191 page_counter_uncharge(&memcg->memsw, nr_pages);
6193 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6194 mem_cgroup_id_put_many(memcg, nr_pages);
6199 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6201 long nr_swap_pages = get_nr_swap_pages();
6203 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6204 return nr_swap_pages;
6205 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6206 nr_swap_pages = min_t(long, nr_swap_pages,
6207 READ_ONCE(memcg->swap.limit) -
6208 page_counter_read(&memcg->swap));
6209 return nr_swap_pages;
6212 bool mem_cgroup_swap_full(struct page *page)
6214 struct mem_cgroup *memcg;
6216 VM_BUG_ON_PAGE(!PageLocked(page), page);
6220 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6223 memcg = page->mem_cgroup;
6227 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6228 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
6234 /* for remember boot option*/
6235 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6236 static int really_do_swap_account __initdata = 1;
6238 static int really_do_swap_account __initdata;
6241 static int __init enable_swap_account(char *s)
6243 if (!strcmp(s, "1"))
6244 really_do_swap_account = 1;
6245 else if (!strcmp(s, "0"))
6246 really_do_swap_account = 0;
6249 __setup("swapaccount=", enable_swap_account);
6251 static u64 swap_current_read(struct cgroup_subsys_state *css,
6254 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6256 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6259 static int swap_max_show(struct seq_file *m, void *v)
6261 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6262 unsigned long max = READ_ONCE(memcg->swap.limit);
6264 if (max == PAGE_COUNTER_MAX)
6265 seq_puts(m, "max\n");
6267 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6272 static ssize_t swap_max_write(struct kernfs_open_file *of,
6273 char *buf, size_t nbytes, loff_t off)
6275 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6279 buf = strstrip(buf);
6280 err = page_counter_memparse(buf, "max", &max);
6284 mutex_lock(&memcg_limit_mutex);
6285 err = page_counter_limit(&memcg->swap, max);
6286 mutex_unlock(&memcg_limit_mutex);
6293 static struct cftype swap_files[] = {
6295 .name = "swap.current",
6296 .flags = CFTYPE_NOT_ON_ROOT,
6297 .read_u64 = swap_current_read,
6301 .flags = CFTYPE_NOT_ON_ROOT,
6302 .seq_show = swap_max_show,
6303 .write = swap_max_write,
6308 static struct cftype memsw_cgroup_files[] = {
6310 .name = "memsw.usage_in_bytes",
6311 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6312 .read_u64 = mem_cgroup_read_u64,
6315 .name = "memsw.max_usage_in_bytes",
6316 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6317 .write = mem_cgroup_reset,
6318 .read_u64 = mem_cgroup_read_u64,
6321 .name = "memsw.limit_in_bytes",
6322 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6323 .write = mem_cgroup_write,
6324 .read_u64 = mem_cgroup_read_u64,
6327 .name = "memsw.failcnt",
6328 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6329 .write = mem_cgroup_reset,
6330 .read_u64 = mem_cgroup_read_u64,
6332 { }, /* terminate */
6335 static int __init mem_cgroup_swap_init(void)
6337 if (!mem_cgroup_disabled() && really_do_swap_account) {
6338 do_swap_account = 1;
6339 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6341 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6342 memsw_cgroup_files));
6346 subsys_initcall(mem_cgroup_swap_init);
6348 #endif /* CONFIG_MEMCG_SWAP */