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)
237 * Iteration constructs for visiting all cgroups (under a tree). If
238 * loops are exited prematurely (break), mem_cgroup_iter_break() must
239 * be used for reference counting.
241 #define for_each_mem_cgroup_tree(iter, root) \
242 for (iter = mem_cgroup_iter(root, NULL, NULL); \
244 iter = mem_cgroup_iter(root, iter, NULL))
246 #define for_each_mem_cgroup(iter) \
247 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
249 iter = mem_cgroup_iter(NULL, iter, NULL))
251 static inline bool should_force_charge(void)
253 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
254 (current->flags & PF_EXITING);
257 /* Some nice accessors for the vmpressure. */
258 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
261 memcg = root_mem_cgroup;
262 return &memcg->vmpressure;
265 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
267 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
270 #ifdef CONFIG_MEMCG_KMEM
272 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
273 * The main reason for not using cgroup id for this:
274 * this works better in sparse environments, where we have a lot of memcgs,
275 * but only a few kmem-limited. Or also, if we have, for instance, 200
276 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
277 * 200 entry array for that.
279 * The current size of the caches array is stored in memcg_nr_cache_ids. It
280 * will double each time we have to increase it.
282 static DEFINE_IDA(memcg_cache_ida);
283 int memcg_nr_cache_ids;
285 /* Protects memcg_nr_cache_ids */
286 static DECLARE_RWSEM(memcg_cache_ids_sem);
288 void memcg_get_cache_ids(void)
290 down_read(&memcg_cache_ids_sem);
293 void memcg_put_cache_ids(void)
295 up_read(&memcg_cache_ids_sem);
299 * MIN_SIZE is different than 1, because we would like to avoid going through
300 * the alloc/free process all the time. In a small machine, 4 kmem-limited
301 * cgroups is a reasonable guess. In the future, it could be a parameter or
302 * tunable, but that is strictly not necessary.
304 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
305 * this constant directly from cgroup, but it is understandable that this is
306 * better kept as an internal representation in cgroup.c. In any case, the
307 * cgrp_id space is not getting any smaller, and we don't have to necessarily
308 * increase ours as well if it increases.
310 #define MEMCG_CACHES_MIN_SIZE 4
311 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
314 * A lot of the calls to the cache allocation functions are expected to be
315 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
316 * conditional to this static branch, we'll have to allow modules that does
317 * kmem_cache_alloc and the such to see this symbol as well
319 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
320 EXPORT_SYMBOL(memcg_kmem_enabled_key);
322 struct workqueue_struct *memcg_kmem_cache_wq;
324 static int memcg_shrinker_map_size;
325 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
327 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
329 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
332 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
333 int size, int old_size)
335 struct memcg_shrinker_map *new, *old;
338 lockdep_assert_held(&memcg_shrinker_map_mutex);
341 old = rcu_dereference_protected(
342 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
343 /* Not yet online memcg */
347 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
351 /* Set all old bits, clear all new bits */
352 memset(new->map, (int)0xff, old_size);
353 memset((void *)new->map + old_size, 0, size - old_size);
355 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
356 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
362 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
364 struct mem_cgroup_per_node *pn;
365 struct memcg_shrinker_map *map;
368 if (mem_cgroup_is_root(memcg))
372 pn = mem_cgroup_nodeinfo(memcg, nid);
373 map = rcu_dereference_protected(pn->shrinker_map, true);
376 rcu_assign_pointer(pn->shrinker_map, NULL);
380 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
382 struct memcg_shrinker_map *map;
383 int nid, size, ret = 0;
385 if (mem_cgroup_is_root(memcg))
388 mutex_lock(&memcg_shrinker_map_mutex);
389 size = memcg_shrinker_map_size;
391 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
393 memcg_free_shrinker_maps(memcg);
397 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
399 mutex_unlock(&memcg_shrinker_map_mutex);
404 int memcg_expand_shrinker_maps(int new_id)
406 int size, old_size, ret = 0;
407 struct mem_cgroup *memcg;
409 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
410 old_size = memcg_shrinker_map_size;
411 if (size <= old_size)
414 mutex_lock(&memcg_shrinker_map_mutex);
415 if (!root_mem_cgroup)
418 for_each_mem_cgroup(memcg) {
419 if (mem_cgroup_is_root(memcg))
421 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
423 mem_cgroup_iter_break(NULL, memcg);
429 memcg_shrinker_map_size = size;
430 mutex_unlock(&memcg_shrinker_map_mutex);
434 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
436 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
437 struct memcg_shrinker_map *map;
440 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
441 /* Pairs with smp mb in shrink_slab() */
442 smp_mb__before_atomic();
443 set_bit(shrinker_id, map->map);
448 #else /* CONFIG_MEMCG_KMEM */
449 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
453 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
454 #endif /* CONFIG_MEMCG_KMEM */
457 * mem_cgroup_css_from_page - css of the memcg associated with a page
458 * @page: page of interest
460 * If memcg is bound to the default hierarchy, css of the memcg associated
461 * with @page is returned. The returned css remains associated with @page
462 * until it is released.
464 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
467 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
469 struct mem_cgroup *memcg;
471 memcg = page->mem_cgroup;
473 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
474 memcg = root_mem_cgroup;
480 * page_cgroup_ino - return inode number of the memcg a page is charged to
483 * Look up the closest online ancestor of the memory cgroup @page is charged to
484 * and return its inode number or 0 if @page is not charged to any cgroup. It
485 * is safe to call this function without holding a reference to @page.
487 * Note, this function is inherently racy, because there is nothing to prevent
488 * the cgroup inode from getting torn down and potentially reallocated a moment
489 * after page_cgroup_ino() returns, so it only should be used by callers that
490 * do not care (such as procfs interfaces).
492 ino_t page_cgroup_ino(struct page *page)
494 struct mem_cgroup *memcg;
495 unsigned long ino = 0;
498 memcg = READ_ONCE(page->mem_cgroup);
499 while (memcg && !(memcg->css.flags & CSS_ONLINE))
500 memcg = parent_mem_cgroup(memcg);
502 ino = cgroup_ino(memcg->css.cgroup);
507 static struct mem_cgroup_per_node *
508 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
510 int nid = page_to_nid(page);
512 return memcg->nodeinfo[nid];
515 static struct mem_cgroup_tree_per_node *
516 soft_limit_tree_node(int nid)
518 return soft_limit_tree.rb_tree_per_node[nid];
521 static struct mem_cgroup_tree_per_node *
522 soft_limit_tree_from_page(struct page *page)
524 int nid = page_to_nid(page);
526 return soft_limit_tree.rb_tree_per_node[nid];
529 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
530 struct mem_cgroup_tree_per_node *mctz,
531 unsigned long new_usage_in_excess)
533 struct rb_node **p = &mctz->rb_root.rb_node;
534 struct rb_node *parent = NULL;
535 struct mem_cgroup_per_node *mz_node;
536 bool rightmost = true;
541 mz->usage_in_excess = new_usage_in_excess;
542 if (!mz->usage_in_excess)
546 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
548 if (mz->usage_in_excess < mz_node->usage_in_excess) {
554 * We can't avoid mem cgroups that are over their soft
555 * limit by the same amount
557 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
562 mctz->rb_rightmost = &mz->tree_node;
564 rb_link_node(&mz->tree_node, parent, p);
565 rb_insert_color(&mz->tree_node, &mctz->rb_root);
569 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
570 struct mem_cgroup_tree_per_node *mctz)
575 if (&mz->tree_node == mctz->rb_rightmost)
576 mctz->rb_rightmost = rb_prev(&mz->tree_node);
578 rb_erase(&mz->tree_node, &mctz->rb_root);
582 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
583 struct mem_cgroup_tree_per_node *mctz)
587 spin_lock_irqsave(&mctz->lock, flags);
588 __mem_cgroup_remove_exceeded(mz, mctz);
589 spin_unlock_irqrestore(&mctz->lock, flags);
592 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
594 unsigned long nr_pages = page_counter_read(&memcg->memory);
595 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
596 unsigned long excess = 0;
598 if (nr_pages > soft_limit)
599 excess = nr_pages - soft_limit;
604 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
606 unsigned long excess;
607 struct mem_cgroup_per_node *mz;
608 struct mem_cgroup_tree_per_node *mctz;
610 mctz = soft_limit_tree_from_page(page);
614 * Necessary to update all ancestors when hierarchy is used.
615 * because their event counter is not touched.
617 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
618 mz = mem_cgroup_page_nodeinfo(memcg, page);
619 excess = soft_limit_excess(memcg);
621 * We have to update the tree if mz is on RB-tree or
622 * mem is over its softlimit.
624 if (excess || mz->on_tree) {
627 spin_lock_irqsave(&mctz->lock, flags);
628 /* if on-tree, remove it */
630 __mem_cgroup_remove_exceeded(mz, mctz);
632 * Insert again. mz->usage_in_excess will be updated.
633 * If excess is 0, no tree ops.
635 __mem_cgroup_insert_exceeded(mz, mctz, excess);
636 spin_unlock_irqrestore(&mctz->lock, flags);
641 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
643 struct mem_cgroup_tree_per_node *mctz;
644 struct mem_cgroup_per_node *mz;
648 mz = mem_cgroup_nodeinfo(memcg, nid);
649 mctz = soft_limit_tree_node(nid);
651 mem_cgroup_remove_exceeded(mz, mctz);
655 static struct mem_cgroup_per_node *
656 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
658 struct mem_cgroup_per_node *mz;
662 if (!mctz->rb_rightmost)
663 goto done; /* Nothing to reclaim from */
665 mz = rb_entry(mctz->rb_rightmost,
666 struct mem_cgroup_per_node, tree_node);
668 * Remove the node now but someone else can add it back,
669 * we will to add it back at the end of reclaim to its correct
670 * position in the tree.
672 __mem_cgroup_remove_exceeded(mz, mctz);
673 if (!soft_limit_excess(mz->memcg) ||
674 !css_tryget_online(&mz->memcg->css))
680 static struct mem_cgroup_per_node *
681 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
683 struct mem_cgroup_per_node *mz;
685 spin_lock_irq(&mctz->lock);
686 mz = __mem_cgroup_largest_soft_limit_node(mctz);
687 spin_unlock_irq(&mctz->lock);
691 static unsigned long memcg_sum_events(struct mem_cgroup *memcg,
694 return atomic_long_read(&memcg->events[event]);
697 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
699 bool compound, int nr_pages)
702 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
703 * counted as CACHE even if it's on ANON LRU.
706 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
708 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
709 if (PageSwapBacked(page))
710 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
714 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
715 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
718 /* pagein of a big page is an event. So, ignore page size */
720 __count_memcg_events(memcg, PGPGIN, 1);
722 __count_memcg_events(memcg, PGPGOUT, 1);
723 nr_pages = -nr_pages; /* for event */
726 __this_cpu_add(memcg->stat_cpu->nr_page_events, nr_pages);
729 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
730 int nid, unsigned int lru_mask)
732 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
733 unsigned long nr = 0;
736 VM_BUG_ON((unsigned)nid >= nr_node_ids);
739 if (!(BIT(lru) & lru_mask))
741 nr += mem_cgroup_get_lru_size(lruvec, lru);
746 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
747 unsigned int lru_mask)
749 unsigned long nr = 0;
752 for_each_node_state(nid, N_MEMORY)
753 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
757 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
758 enum mem_cgroup_events_target target)
760 unsigned long val, next;
762 val = __this_cpu_read(memcg->stat_cpu->nr_page_events);
763 next = __this_cpu_read(memcg->stat_cpu->targets[target]);
764 /* from time_after() in jiffies.h */
765 if ((long)(next - val) < 0) {
767 case MEM_CGROUP_TARGET_THRESH:
768 next = val + THRESHOLDS_EVENTS_TARGET;
770 case MEM_CGROUP_TARGET_SOFTLIMIT:
771 next = val + SOFTLIMIT_EVENTS_TARGET;
773 case MEM_CGROUP_TARGET_NUMAINFO:
774 next = val + NUMAINFO_EVENTS_TARGET;
779 __this_cpu_write(memcg->stat_cpu->targets[target], next);
786 * Check events in order.
789 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
791 /* threshold event is triggered in finer grain than soft limit */
792 if (unlikely(mem_cgroup_event_ratelimit(memcg,
793 MEM_CGROUP_TARGET_THRESH))) {
795 bool do_numainfo __maybe_unused;
797 do_softlimit = mem_cgroup_event_ratelimit(memcg,
798 MEM_CGROUP_TARGET_SOFTLIMIT);
800 do_numainfo = mem_cgroup_event_ratelimit(memcg,
801 MEM_CGROUP_TARGET_NUMAINFO);
803 mem_cgroup_threshold(memcg);
804 if (unlikely(do_softlimit))
805 mem_cgroup_update_tree(memcg, page);
807 if (unlikely(do_numainfo))
808 atomic_inc(&memcg->numainfo_events);
813 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
816 * mm_update_next_owner() may clear mm->owner to NULL
817 * if it races with swapoff, page migration, etc.
818 * So this can be called with p == NULL.
823 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
825 EXPORT_SYMBOL(mem_cgroup_from_task);
828 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
829 * @mm: mm from which memcg should be extracted. It can be NULL.
831 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
832 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
835 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
837 struct mem_cgroup *memcg;
839 if (mem_cgroup_disabled())
845 * Page cache insertions can happen withou an
846 * actual mm context, e.g. during disk probing
847 * on boot, loopback IO, acct() writes etc.
850 memcg = root_mem_cgroup;
852 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
853 if (unlikely(!memcg))
854 memcg = root_mem_cgroup;
856 } while (!css_tryget(&memcg->css));
860 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
863 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
864 * @page: page from which memcg should be extracted.
866 * Obtain a reference on page->memcg and returns it if successful. Otherwise
867 * root_mem_cgroup is returned.
869 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
871 struct mem_cgroup *memcg = page->mem_cgroup;
873 if (mem_cgroup_disabled())
877 if (!memcg || !css_tryget_online(&memcg->css))
878 memcg = root_mem_cgroup;
882 EXPORT_SYMBOL(get_mem_cgroup_from_page);
885 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
887 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
889 if (unlikely(current->active_memcg)) {
890 struct mem_cgroup *memcg = root_mem_cgroup;
893 if (css_tryget_online(¤t->active_memcg->css))
894 memcg = current->active_memcg;
898 return get_mem_cgroup_from_mm(current->mm);
902 * mem_cgroup_iter - iterate over memory cgroup hierarchy
903 * @root: hierarchy root
904 * @prev: previously returned memcg, NULL on first invocation
905 * @reclaim: cookie for shared reclaim walks, NULL for full walks
907 * Returns references to children of the hierarchy below @root, or
908 * @root itself, or %NULL after a full round-trip.
910 * Caller must pass the return value in @prev on subsequent
911 * invocations for reference counting, or use mem_cgroup_iter_break()
912 * to cancel a hierarchy walk before the round-trip is complete.
914 * Reclaimers can specify a node and a priority level in @reclaim to
915 * divide up the memcgs in the hierarchy among all concurrent
916 * reclaimers operating on the same node and priority.
918 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
919 struct mem_cgroup *prev,
920 struct mem_cgroup_reclaim_cookie *reclaim)
922 struct mem_cgroup_reclaim_iter *iter;
923 struct cgroup_subsys_state *css = NULL;
924 struct mem_cgroup *memcg = NULL;
925 struct mem_cgroup *pos = NULL;
927 if (mem_cgroup_disabled())
931 root = root_mem_cgroup;
933 if (prev && !reclaim)
936 if (!root->use_hierarchy && root != root_mem_cgroup) {
945 struct mem_cgroup_per_node *mz;
947 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
948 iter = &mz->iter[reclaim->priority];
950 if (prev && reclaim->generation != iter->generation)
954 pos = READ_ONCE(iter->position);
955 if (!pos || css_tryget(&pos->css))
958 * css reference reached zero, so iter->position will
959 * be cleared by ->css_released. However, we should not
960 * rely on this happening soon, because ->css_released
961 * is called from a work queue, and by busy-waiting we
962 * might block it. So we clear iter->position right
965 (void)cmpxchg(&iter->position, pos, NULL);
973 css = css_next_descendant_pre(css, &root->css);
976 * Reclaimers share the hierarchy walk, and a
977 * new one might jump in right at the end of
978 * the hierarchy - make sure they see at least
979 * one group and restart from the beginning.
987 * Verify the css and acquire a reference. The root
988 * is provided by the caller, so we know it's alive
989 * and kicking, and don't take an extra reference.
991 memcg = mem_cgroup_from_css(css);
993 if (css == &root->css)
1004 * The position could have already been updated by a competing
1005 * thread, so check that the value hasn't changed since we read
1006 * it to avoid reclaiming from the same cgroup twice.
1008 (void)cmpxchg(&iter->position, pos, memcg);
1016 reclaim->generation = iter->generation;
1022 if (prev && prev != root)
1023 css_put(&prev->css);
1029 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1030 * @root: hierarchy root
1031 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1033 void mem_cgroup_iter_break(struct mem_cgroup *root,
1034 struct mem_cgroup *prev)
1037 root = root_mem_cgroup;
1038 if (prev && prev != root)
1039 css_put(&prev->css);
1042 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1043 struct mem_cgroup *dead_memcg)
1045 struct mem_cgroup_reclaim_iter *iter;
1046 struct mem_cgroup_per_node *mz;
1050 for_each_node(nid) {
1051 mz = mem_cgroup_nodeinfo(from, nid);
1052 for (i = 0; i <= DEF_PRIORITY; i++) {
1053 iter = &mz->iter[i];
1054 cmpxchg(&iter->position,
1060 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1062 struct mem_cgroup *memcg = dead_memcg;
1063 struct mem_cgroup *last;
1066 __invalidate_reclaim_iterators(memcg, dead_memcg);
1068 } while ((memcg = parent_mem_cgroup(memcg)));
1071 * When cgruop1 non-hierarchy mode is used,
1072 * parent_mem_cgroup() does not walk all the way up to the
1073 * cgroup root (root_mem_cgroup). So we have to handle
1074 * dead_memcg from cgroup root separately.
1076 if (last != root_mem_cgroup)
1077 __invalidate_reclaim_iterators(root_mem_cgroup,
1082 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1083 * @memcg: hierarchy root
1084 * @fn: function to call for each task
1085 * @arg: argument passed to @fn
1087 * This function iterates over tasks attached to @memcg or to any of its
1088 * descendants and calls @fn for each task. If @fn returns a non-zero
1089 * value, the function breaks the iteration loop and returns the value.
1090 * Otherwise, it will iterate over all tasks and return 0.
1092 * This function must not be called for the root memory cgroup.
1094 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1095 int (*fn)(struct task_struct *, void *), void *arg)
1097 struct mem_cgroup *iter;
1100 BUG_ON(memcg == root_mem_cgroup);
1102 for_each_mem_cgroup_tree(iter, memcg) {
1103 struct css_task_iter it;
1104 struct task_struct *task;
1106 css_task_iter_start(&iter->css, 0, &it);
1107 while (!ret && (task = css_task_iter_next(&it)))
1108 ret = fn(task, arg);
1109 css_task_iter_end(&it);
1111 mem_cgroup_iter_break(memcg, iter);
1119 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1121 * @pgdat: pgdat of the page
1123 * This function is only safe when following the LRU page isolation
1124 * and putback protocol: the LRU lock must be held, and the page must
1125 * either be PageLRU() or the caller must have isolated/allocated it.
1127 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1129 struct mem_cgroup_per_node *mz;
1130 struct mem_cgroup *memcg;
1131 struct lruvec *lruvec;
1133 if (mem_cgroup_disabled()) {
1134 lruvec = &pgdat->lruvec;
1138 memcg = page->mem_cgroup;
1140 * Swapcache readahead pages are added to the LRU - and
1141 * possibly migrated - before they are charged.
1144 memcg = root_mem_cgroup;
1146 mz = mem_cgroup_page_nodeinfo(memcg, page);
1147 lruvec = &mz->lruvec;
1150 * Since a node can be onlined after the mem_cgroup was created,
1151 * we have to be prepared to initialize lruvec->zone here;
1152 * and if offlined then reonlined, we need to reinitialize it.
1154 if (unlikely(lruvec->pgdat != pgdat))
1155 lruvec->pgdat = pgdat;
1160 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1161 * @lruvec: mem_cgroup per zone lru vector
1162 * @lru: index of lru list the page is sitting on
1163 * @zid: zone id of the accounted pages
1164 * @nr_pages: positive when adding or negative when removing
1166 * This function must be called under lru_lock, just before a page is added
1167 * to or just after a page is removed from an lru list (that ordering being
1168 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1170 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1171 int zid, int nr_pages)
1173 struct mem_cgroup_per_node *mz;
1174 unsigned long *lru_size;
1177 if (mem_cgroup_disabled())
1180 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1181 lru_size = &mz->lru_zone_size[zid][lru];
1184 *lru_size += nr_pages;
1187 if (WARN_ONCE(size < 0,
1188 "%s(%p, %d, %d): lru_size %ld\n",
1189 __func__, lruvec, lru, nr_pages, size)) {
1195 *lru_size += nr_pages;
1198 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1200 struct mem_cgroup *task_memcg;
1201 struct task_struct *p;
1204 p = find_lock_task_mm(task);
1206 task_memcg = get_mem_cgroup_from_mm(p->mm);
1210 * All threads may have already detached their mm's, but the oom
1211 * killer still needs to detect if they have already been oom
1212 * killed to prevent needlessly killing additional tasks.
1215 task_memcg = mem_cgroup_from_task(task);
1216 css_get(&task_memcg->css);
1219 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1220 css_put(&task_memcg->css);
1225 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1226 * @memcg: the memory cgroup
1228 * Returns the maximum amount of memory @mem can be charged with, in
1231 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1233 unsigned long margin = 0;
1234 unsigned long count;
1235 unsigned long limit;
1237 count = page_counter_read(&memcg->memory);
1238 limit = READ_ONCE(memcg->memory.max);
1240 margin = limit - count;
1242 if (do_memsw_account()) {
1243 count = page_counter_read(&memcg->memsw);
1244 limit = READ_ONCE(memcg->memsw.max);
1246 margin = min(margin, limit - count);
1255 * A routine for checking "mem" is under move_account() or not.
1257 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1258 * moving cgroups. This is for waiting at high-memory pressure
1261 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1263 struct mem_cgroup *from;
1264 struct mem_cgroup *to;
1267 * Unlike task_move routines, we access mc.to, mc.from not under
1268 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1270 spin_lock(&mc.lock);
1276 ret = mem_cgroup_is_descendant(from, memcg) ||
1277 mem_cgroup_is_descendant(to, memcg);
1279 spin_unlock(&mc.lock);
1283 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1285 if (mc.moving_task && current != mc.moving_task) {
1286 if (mem_cgroup_under_move(memcg)) {
1288 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1289 /* moving charge context might have finished. */
1292 finish_wait(&mc.waitq, &wait);
1299 static const unsigned int memcg1_stats[] = {
1310 static const char *const memcg1_stat_names[] = {
1321 #define K(x) ((x) << (PAGE_SHIFT-10))
1323 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1324 * @memcg: The memory cgroup that went over limit
1325 * @p: Task that is going to be killed
1327 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1330 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1332 struct mem_cgroup *iter;
1338 pr_info("Task in ");
1339 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1340 pr_cont(" killed as a result of limit of ");
1342 pr_info("Memory limit reached of cgroup ");
1345 pr_cont_cgroup_path(memcg->css.cgroup);
1350 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1351 K((u64)page_counter_read(&memcg->memory)),
1352 K((u64)memcg->memory.max), memcg->memory.failcnt);
1353 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1354 K((u64)page_counter_read(&memcg->memsw)),
1355 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1356 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1357 K((u64)page_counter_read(&memcg->kmem)),
1358 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1360 for_each_mem_cgroup_tree(iter, memcg) {
1361 pr_info("Memory cgroup stats for ");
1362 pr_cont_cgroup_path(iter->css.cgroup);
1365 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1366 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1368 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1369 K(memcg_page_state(iter, memcg1_stats[i])));
1372 for (i = 0; i < NR_LRU_LISTS; i++)
1373 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1374 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1381 * Return the memory (and swap, if configured) limit for a memcg.
1383 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1387 max = memcg->memory.max;
1388 if (mem_cgroup_swappiness(memcg)) {
1389 unsigned long memsw_max;
1390 unsigned long swap_max;
1392 memsw_max = memcg->memsw.max;
1393 swap_max = memcg->swap.max;
1394 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1395 max = min(max + swap_max, memsw_max);
1400 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1403 struct oom_control oc = {
1407 .gfp_mask = gfp_mask,
1412 if (mutex_lock_killable(&oom_lock))
1415 * A few threads which were not waiting at mutex_lock_killable() can
1416 * fail to bail out. Therefore, check again after holding oom_lock.
1418 ret = should_force_charge() || out_of_memory(&oc);
1419 mutex_unlock(&oom_lock);
1423 #if MAX_NUMNODES > 1
1426 * test_mem_cgroup_node_reclaimable
1427 * @memcg: the target memcg
1428 * @nid: the node ID to be checked.
1429 * @noswap : specify true here if the user wants flle only information.
1431 * This function returns whether the specified memcg contains any
1432 * reclaimable pages on a node. Returns true if there are any reclaimable
1433 * pages in the node.
1435 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1436 int nid, bool noswap)
1438 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1440 if (noswap || !total_swap_pages)
1442 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1449 * Always updating the nodemask is not very good - even if we have an empty
1450 * list or the wrong list here, we can start from some node and traverse all
1451 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1454 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1458 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1459 * pagein/pageout changes since the last update.
1461 if (!atomic_read(&memcg->numainfo_events))
1463 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1466 /* make a nodemask where this memcg uses memory from */
1467 memcg->scan_nodes = node_states[N_MEMORY];
1469 for_each_node_mask(nid, node_states[N_MEMORY]) {
1471 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1472 node_clear(nid, memcg->scan_nodes);
1475 atomic_set(&memcg->numainfo_events, 0);
1476 atomic_set(&memcg->numainfo_updating, 0);
1480 * Selecting a node where we start reclaim from. Because what we need is just
1481 * reducing usage counter, start from anywhere is O,K. Considering
1482 * memory reclaim from current node, there are pros. and cons.
1484 * Freeing memory from current node means freeing memory from a node which
1485 * we'll use or we've used. So, it may make LRU bad. And if several threads
1486 * hit limits, it will see a contention on a node. But freeing from remote
1487 * node means more costs for memory reclaim because of memory latency.
1489 * Now, we use round-robin. Better algorithm is welcomed.
1491 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1495 mem_cgroup_may_update_nodemask(memcg);
1496 node = memcg->last_scanned_node;
1498 node = next_node_in(node, memcg->scan_nodes);
1500 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1501 * last time it really checked all the LRUs due to rate limiting.
1502 * Fallback to the current node in that case for simplicity.
1504 if (unlikely(node == MAX_NUMNODES))
1505 node = numa_node_id();
1507 memcg->last_scanned_node = node;
1511 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1517 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1520 unsigned long *total_scanned)
1522 struct mem_cgroup *victim = NULL;
1525 unsigned long excess;
1526 unsigned long nr_scanned;
1527 struct mem_cgroup_reclaim_cookie reclaim = {
1532 excess = soft_limit_excess(root_memcg);
1535 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1540 * If we have not been able to reclaim
1541 * anything, it might because there are
1542 * no reclaimable pages under this hierarchy
1547 * We want to do more targeted reclaim.
1548 * excess >> 2 is not to excessive so as to
1549 * reclaim too much, nor too less that we keep
1550 * coming back to reclaim from this cgroup
1552 if (total >= (excess >> 2) ||
1553 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1558 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1559 pgdat, &nr_scanned);
1560 *total_scanned += nr_scanned;
1561 if (!soft_limit_excess(root_memcg))
1564 mem_cgroup_iter_break(root_memcg, victim);
1568 #ifdef CONFIG_LOCKDEP
1569 static struct lockdep_map memcg_oom_lock_dep_map = {
1570 .name = "memcg_oom_lock",
1574 static DEFINE_SPINLOCK(memcg_oom_lock);
1577 * Check OOM-Killer is already running under our hierarchy.
1578 * If someone is running, return false.
1580 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1582 struct mem_cgroup *iter, *failed = NULL;
1584 spin_lock(&memcg_oom_lock);
1586 for_each_mem_cgroup_tree(iter, memcg) {
1587 if (iter->oom_lock) {
1589 * this subtree of our hierarchy is already locked
1590 * so we cannot give a lock.
1593 mem_cgroup_iter_break(memcg, iter);
1596 iter->oom_lock = true;
1601 * OK, we failed to lock the whole subtree so we have
1602 * to clean up what we set up to the failing subtree
1604 for_each_mem_cgroup_tree(iter, memcg) {
1605 if (iter == failed) {
1606 mem_cgroup_iter_break(memcg, iter);
1609 iter->oom_lock = false;
1612 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1614 spin_unlock(&memcg_oom_lock);
1619 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1621 struct mem_cgroup *iter;
1623 spin_lock(&memcg_oom_lock);
1624 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1625 for_each_mem_cgroup_tree(iter, memcg)
1626 iter->oom_lock = false;
1627 spin_unlock(&memcg_oom_lock);
1630 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1632 struct mem_cgroup *iter;
1634 spin_lock(&memcg_oom_lock);
1635 for_each_mem_cgroup_tree(iter, memcg)
1637 spin_unlock(&memcg_oom_lock);
1640 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1642 struct mem_cgroup *iter;
1645 * When a new child is created while the hierarchy is under oom,
1646 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1648 spin_lock(&memcg_oom_lock);
1649 for_each_mem_cgroup_tree(iter, memcg)
1650 if (iter->under_oom > 0)
1652 spin_unlock(&memcg_oom_lock);
1655 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1657 struct oom_wait_info {
1658 struct mem_cgroup *memcg;
1659 wait_queue_entry_t wait;
1662 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1663 unsigned mode, int sync, void *arg)
1665 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1666 struct mem_cgroup *oom_wait_memcg;
1667 struct oom_wait_info *oom_wait_info;
1669 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1670 oom_wait_memcg = oom_wait_info->memcg;
1672 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1673 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1675 return autoremove_wake_function(wait, mode, sync, arg);
1678 static void memcg_oom_recover(struct mem_cgroup *memcg)
1681 * For the following lockless ->under_oom test, the only required
1682 * guarantee is that it must see the state asserted by an OOM when
1683 * this function is called as a result of userland actions
1684 * triggered by the notification of the OOM. This is trivially
1685 * achieved by invoking mem_cgroup_mark_under_oom() before
1686 * triggering notification.
1688 if (memcg && memcg->under_oom)
1689 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1699 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1701 enum oom_status ret;
1704 if (order > PAGE_ALLOC_COSTLY_ORDER)
1708 * We are in the middle of the charge context here, so we
1709 * don't want to block when potentially sitting on a callstack
1710 * that holds all kinds of filesystem and mm locks.
1712 * cgroup1 allows disabling the OOM killer and waiting for outside
1713 * handling until the charge can succeed; remember the context and put
1714 * the task to sleep at the end of the page fault when all locks are
1717 * On the other hand, in-kernel OOM killer allows for an async victim
1718 * memory reclaim (oom_reaper) and that means that we are not solely
1719 * relying on the oom victim to make a forward progress and we can
1720 * invoke the oom killer here.
1722 * Please note that mem_cgroup_out_of_memory might fail to find a
1723 * victim and then we have to bail out from the charge path.
1725 if (memcg->oom_kill_disable) {
1726 if (!current->in_user_fault)
1728 css_get(&memcg->css);
1729 current->memcg_in_oom = memcg;
1730 current->memcg_oom_gfp_mask = mask;
1731 current->memcg_oom_order = order;
1736 mem_cgroup_mark_under_oom(memcg);
1738 locked = mem_cgroup_oom_trylock(memcg);
1741 mem_cgroup_oom_notify(memcg);
1743 mem_cgroup_unmark_under_oom(memcg);
1744 if (mem_cgroup_out_of_memory(memcg, mask, order))
1750 mem_cgroup_oom_unlock(memcg);
1756 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1757 * @handle: actually kill/wait or just clean up the OOM state
1759 * This has to be called at the end of a page fault if the memcg OOM
1760 * handler was enabled.
1762 * Memcg supports userspace OOM handling where failed allocations must
1763 * sleep on a waitqueue until the userspace task resolves the
1764 * situation. Sleeping directly in the charge context with all kinds
1765 * of locks held is not a good idea, instead we remember an OOM state
1766 * in the task and mem_cgroup_oom_synchronize() has to be called at
1767 * the end of the page fault to complete the OOM handling.
1769 * Returns %true if an ongoing memcg OOM situation was detected and
1770 * completed, %false otherwise.
1772 bool mem_cgroup_oom_synchronize(bool handle)
1774 struct mem_cgroup *memcg = current->memcg_in_oom;
1775 struct oom_wait_info owait;
1778 /* OOM is global, do not handle */
1785 owait.memcg = memcg;
1786 owait.wait.flags = 0;
1787 owait.wait.func = memcg_oom_wake_function;
1788 owait.wait.private = current;
1789 INIT_LIST_HEAD(&owait.wait.entry);
1791 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1792 mem_cgroup_mark_under_oom(memcg);
1794 locked = mem_cgroup_oom_trylock(memcg);
1797 mem_cgroup_oom_notify(memcg);
1799 if (locked && !memcg->oom_kill_disable) {
1800 mem_cgroup_unmark_under_oom(memcg);
1801 finish_wait(&memcg_oom_waitq, &owait.wait);
1802 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1803 current->memcg_oom_order);
1806 mem_cgroup_unmark_under_oom(memcg);
1807 finish_wait(&memcg_oom_waitq, &owait.wait);
1811 mem_cgroup_oom_unlock(memcg);
1813 * There is no guarantee that an OOM-lock contender
1814 * sees the wakeups triggered by the OOM kill
1815 * uncharges. Wake any sleepers explicitely.
1817 memcg_oom_recover(memcg);
1820 current->memcg_in_oom = NULL;
1821 css_put(&memcg->css);
1826 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1827 * @victim: task to be killed by the OOM killer
1828 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1830 * Returns a pointer to a memory cgroup, which has to be cleaned up
1831 * by killing all belonging OOM-killable tasks.
1833 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1835 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1836 struct mem_cgroup *oom_domain)
1838 struct mem_cgroup *oom_group = NULL;
1839 struct mem_cgroup *memcg;
1841 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1845 oom_domain = root_mem_cgroup;
1849 memcg = mem_cgroup_from_task(victim);
1850 if (memcg == root_mem_cgroup)
1854 * Traverse the memory cgroup hierarchy from the victim task's
1855 * cgroup up to the OOMing cgroup (or root) to find the
1856 * highest-level memory cgroup with oom.group set.
1858 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1859 if (memcg->oom_group)
1862 if (memcg == oom_domain)
1867 css_get(&oom_group->css);
1874 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1876 pr_info("Tasks in ");
1877 pr_cont_cgroup_path(memcg->css.cgroup);
1878 pr_cont(" are going to be killed due to memory.oom.group set\n");
1882 * lock_page_memcg - lock a page->mem_cgroup binding
1885 * This function protects unlocked LRU pages from being moved to
1888 * It ensures lifetime of the returned memcg. Caller is responsible
1889 * for the lifetime of the page; __unlock_page_memcg() is available
1890 * when @page might get freed inside the locked section.
1892 struct mem_cgroup *lock_page_memcg(struct page *page)
1894 struct mem_cgroup *memcg;
1895 unsigned long flags;
1898 * The RCU lock is held throughout the transaction. The fast
1899 * path can get away without acquiring the memcg->move_lock
1900 * because page moving starts with an RCU grace period.
1902 * The RCU lock also protects the memcg from being freed when
1903 * the page state that is going to change is the only thing
1904 * preventing the page itself from being freed. E.g. writeback
1905 * doesn't hold a page reference and relies on PG_writeback to
1906 * keep off truncation, migration and so forth.
1910 if (mem_cgroup_disabled())
1913 memcg = page->mem_cgroup;
1914 if (unlikely(!memcg))
1917 if (atomic_read(&memcg->moving_account) <= 0)
1920 spin_lock_irqsave(&memcg->move_lock, flags);
1921 if (memcg != page->mem_cgroup) {
1922 spin_unlock_irqrestore(&memcg->move_lock, flags);
1927 * When charge migration first begins, we can have locked and
1928 * unlocked page stat updates happening concurrently. Track
1929 * the task who has the lock for unlock_page_memcg().
1931 memcg->move_lock_task = current;
1932 memcg->move_lock_flags = flags;
1936 EXPORT_SYMBOL(lock_page_memcg);
1939 * __unlock_page_memcg - unlock and unpin a memcg
1942 * Unlock and unpin a memcg returned by lock_page_memcg().
1944 void __unlock_page_memcg(struct mem_cgroup *memcg)
1946 if (memcg && memcg->move_lock_task == current) {
1947 unsigned long flags = memcg->move_lock_flags;
1949 memcg->move_lock_task = NULL;
1950 memcg->move_lock_flags = 0;
1952 spin_unlock_irqrestore(&memcg->move_lock, flags);
1959 * unlock_page_memcg - unlock a page->mem_cgroup binding
1962 void unlock_page_memcg(struct page *page)
1964 __unlock_page_memcg(page->mem_cgroup);
1966 EXPORT_SYMBOL(unlock_page_memcg);
1968 struct memcg_stock_pcp {
1969 struct mem_cgroup *cached; /* this never be root cgroup */
1970 unsigned int nr_pages;
1971 struct work_struct work;
1972 unsigned long flags;
1973 #define FLUSHING_CACHED_CHARGE 0
1975 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1976 static DEFINE_MUTEX(percpu_charge_mutex);
1979 * consume_stock: Try to consume stocked charge on this cpu.
1980 * @memcg: memcg to consume from.
1981 * @nr_pages: how many pages to charge.
1983 * The charges will only happen if @memcg matches the current cpu's memcg
1984 * stock, and at least @nr_pages are available in that stock. Failure to
1985 * service an allocation will refill the stock.
1987 * returns true if successful, false otherwise.
1989 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1991 struct memcg_stock_pcp *stock;
1992 unsigned long flags;
1995 if (nr_pages > MEMCG_CHARGE_BATCH)
1998 local_irq_save(flags);
2000 stock = this_cpu_ptr(&memcg_stock);
2001 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2002 stock->nr_pages -= nr_pages;
2006 local_irq_restore(flags);
2012 * Returns stocks cached in percpu and reset cached information.
2014 static void drain_stock(struct memcg_stock_pcp *stock)
2016 struct mem_cgroup *old = stock->cached;
2021 if (stock->nr_pages) {
2022 page_counter_uncharge(&old->memory, stock->nr_pages);
2023 if (do_memsw_account())
2024 page_counter_uncharge(&old->memsw, stock->nr_pages);
2025 css_put_many(&old->css, stock->nr_pages);
2026 stock->nr_pages = 0;
2030 stock->cached = NULL;
2033 static void drain_local_stock(struct work_struct *dummy)
2035 struct memcg_stock_pcp *stock;
2036 unsigned long flags;
2039 * The only protection from memory hotplug vs. drain_stock races is
2040 * that we always operate on local CPU stock here with IRQ disabled
2042 local_irq_save(flags);
2044 stock = this_cpu_ptr(&memcg_stock);
2046 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2048 local_irq_restore(flags);
2052 * Cache charges(val) to local per_cpu area.
2053 * This will be consumed by consume_stock() function, later.
2055 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2057 struct memcg_stock_pcp *stock;
2058 unsigned long flags;
2060 local_irq_save(flags);
2062 stock = this_cpu_ptr(&memcg_stock);
2063 if (stock->cached != memcg) { /* reset if necessary */
2065 css_get(&memcg->css);
2066 stock->cached = memcg;
2068 stock->nr_pages += nr_pages;
2070 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2073 local_irq_restore(flags);
2077 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2078 * of the hierarchy under it.
2080 static void drain_all_stock(struct mem_cgroup *root_memcg)
2084 /* If someone's already draining, avoid adding running more workers. */
2085 if (!mutex_trylock(&percpu_charge_mutex))
2088 * Notify other cpus that system-wide "drain" is running
2089 * We do not care about races with the cpu hotplug because cpu down
2090 * as well as workers from this path always operate on the local
2091 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2094 for_each_online_cpu(cpu) {
2095 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2096 struct mem_cgroup *memcg;
2100 memcg = stock->cached;
2101 if (memcg && stock->nr_pages &&
2102 mem_cgroup_is_descendant(memcg, root_memcg))
2107 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2109 drain_local_stock(&stock->work);
2111 schedule_work_on(cpu, &stock->work);
2115 mutex_unlock(&percpu_charge_mutex);
2118 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2120 struct memcg_stock_pcp *stock;
2121 struct mem_cgroup *memcg;
2123 stock = &per_cpu(memcg_stock, cpu);
2126 for_each_mem_cgroup(memcg) {
2129 for (i = 0; i < MEMCG_NR_STAT; i++) {
2133 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
2135 atomic_long_add(x, &memcg->stat[i]);
2137 if (i >= NR_VM_NODE_STAT_ITEMS)
2140 for_each_node(nid) {
2141 struct mem_cgroup_per_node *pn;
2143 pn = mem_cgroup_nodeinfo(memcg, nid);
2144 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2146 atomic_long_add(x, &pn->lruvec_stat[i]);
2150 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2153 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
2155 atomic_long_add(x, &memcg->events[i]);
2162 static void reclaim_high(struct mem_cgroup *memcg,
2163 unsigned int nr_pages,
2167 if (page_counter_read(&memcg->memory) <= memcg->high)
2169 memcg_memory_event(memcg, MEMCG_HIGH);
2170 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2171 } while ((memcg = parent_mem_cgroup(memcg)));
2174 static void high_work_func(struct work_struct *work)
2176 struct mem_cgroup *memcg;
2178 memcg = container_of(work, struct mem_cgroup, high_work);
2179 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2183 * Scheduled by try_charge() to be executed from the userland return path
2184 * and reclaims memory over the high limit.
2186 void mem_cgroup_handle_over_high(void)
2188 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2189 struct mem_cgroup *memcg;
2191 if (likely(!nr_pages))
2194 memcg = get_mem_cgroup_from_mm(current->mm);
2195 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2196 css_put(&memcg->css);
2197 current->memcg_nr_pages_over_high = 0;
2200 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2201 unsigned int nr_pages)
2203 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2204 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2205 struct mem_cgroup *mem_over_limit;
2206 struct page_counter *counter;
2207 unsigned long nr_reclaimed;
2208 bool may_swap = true;
2209 bool drained = false;
2211 enum oom_status oom_status;
2213 if (mem_cgroup_is_root(memcg))
2216 if (consume_stock(memcg, nr_pages))
2219 if (!do_memsw_account() ||
2220 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2221 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2223 if (do_memsw_account())
2224 page_counter_uncharge(&memcg->memsw, batch);
2225 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2227 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2231 if (batch > nr_pages) {
2237 * Memcg doesn't have a dedicated reserve for atomic
2238 * allocations. But like the global atomic pool, we need to
2239 * put the burden of reclaim on regular allocation requests
2240 * and let these go through as privileged allocations.
2242 if (gfp_mask & __GFP_ATOMIC)
2246 * Unlike in global OOM situations, memcg is not in a physical
2247 * memory shortage. Allow dying and OOM-killed tasks to
2248 * bypass the last charges so that they can exit quickly and
2249 * free their memory.
2251 if (unlikely(should_force_charge()))
2255 * Prevent unbounded recursion when reclaim operations need to
2256 * allocate memory. This might exceed the limits temporarily,
2257 * but we prefer facilitating memory reclaim and getting back
2258 * under the limit over triggering OOM kills in these cases.
2260 if (unlikely(current->flags & PF_MEMALLOC))
2263 if (unlikely(task_in_memcg_oom(current)))
2266 if (!gfpflags_allow_blocking(gfp_mask))
2269 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2271 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2272 gfp_mask, may_swap);
2274 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2278 drain_all_stock(mem_over_limit);
2283 if (gfp_mask & __GFP_NORETRY)
2286 * Even though the limit is exceeded at this point, reclaim
2287 * may have been able to free some pages. Retry the charge
2288 * before killing the task.
2290 * Only for regular pages, though: huge pages are rather
2291 * unlikely to succeed so close to the limit, and we fall back
2292 * to regular pages anyway in case of failure.
2294 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2297 * At task move, charge accounts can be doubly counted. So, it's
2298 * better to wait until the end of task_move if something is going on.
2300 if (mem_cgroup_wait_acct_move(mem_over_limit))
2306 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2309 if (gfp_mask & __GFP_NOFAIL)
2312 if (fatal_signal_pending(current))
2315 memcg_memory_event(mem_over_limit, MEMCG_OOM);
2318 * keep retrying as long as the memcg oom killer is able to make
2319 * a forward progress or bypass the charge if the oom killer
2320 * couldn't make any progress.
2322 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2323 get_order(nr_pages * PAGE_SIZE));
2324 switch (oom_status) {
2326 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2335 if (!(gfp_mask & __GFP_NOFAIL))
2339 * The allocation either can't fail or will lead to more memory
2340 * being freed very soon. Allow memory usage go over the limit
2341 * temporarily by force charging it.
2343 page_counter_charge(&memcg->memory, nr_pages);
2344 if (do_memsw_account())
2345 page_counter_charge(&memcg->memsw, nr_pages);
2346 css_get_many(&memcg->css, nr_pages);
2351 css_get_many(&memcg->css, batch);
2352 if (batch > nr_pages)
2353 refill_stock(memcg, batch - nr_pages);
2356 * If the hierarchy is above the normal consumption range, schedule
2357 * reclaim on returning to userland. We can perform reclaim here
2358 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2359 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2360 * not recorded as it most likely matches current's and won't
2361 * change in the meantime. As high limit is checked again before
2362 * reclaim, the cost of mismatch is negligible.
2365 if (page_counter_read(&memcg->memory) > memcg->high) {
2366 /* Don't bother a random interrupted task */
2367 if (in_interrupt()) {
2368 schedule_work(&memcg->high_work);
2371 current->memcg_nr_pages_over_high += batch;
2372 set_notify_resume(current);
2375 } while ((memcg = parent_mem_cgroup(memcg)));
2380 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2382 if (mem_cgroup_is_root(memcg))
2385 page_counter_uncharge(&memcg->memory, nr_pages);
2386 if (do_memsw_account())
2387 page_counter_uncharge(&memcg->memsw, nr_pages);
2389 css_put_many(&memcg->css, nr_pages);
2392 static void lock_page_lru(struct page *page, int *isolated)
2394 struct zone *zone = page_zone(page);
2396 spin_lock_irq(zone_lru_lock(zone));
2397 if (PageLRU(page)) {
2398 struct lruvec *lruvec;
2400 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2402 del_page_from_lru_list(page, lruvec, page_lru(page));
2408 static void unlock_page_lru(struct page *page, int isolated)
2410 struct zone *zone = page_zone(page);
2413 struct lruvec *lruvec;
2415 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2416 VM_BUG_ON_PAGE(PageLRU(page), page);
2418 add_page_to_lru_list(page, lruvec, page_lru(page));
2420 spin_unlock_irq(zone_lru_lock(zone));
2423 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2428 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2431 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2432 * may already be on some other mem_cgroup's LRU. Take care of it.
2435 lock_page_lru(page, &isolated);
2438 * Nobody should be changing or seriously looking at
2439 * page->mem_cgroup at this point:
2441 * - the page is uncharged
2443 * - the page is off-LRU
2445 * - an anonymous fault has exclusive page access, except for
2446 * a locked page table
2448 * - a page cache insertion, a swapin fault, or a migration
2449 * have the page locked
2451 page->mem_cgroup = memcg;
2454 unlock_page_lru(page, isolated);
2457 #ifdef CONFIG_MEMCG_KMEM
2458 static int memcg_alloc_cache_id(void)
2463 id = ida_simple_get(&memcg_cache_ida,
2464 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2468 if (id < memcg_nr_cache_ids)
2472 * There's no space for the new id in memcg_caches arrays,
2473 * so we have to grow them.
2475 down_write(&memcg_cache_ids_sem);
2477 size = 2 * (id + 1);
2478 if (size < MEMCG_CACHES_MIN_SIZE)
2479 size = MEMCG_CACHES_MIN_SIZE;
2480 else if (size > MEMCG_CACHES_MAX_SIZE)
2481 size = MEMCG_CACHES_MAX_SIZE;
2483 err = memcg_update_all_caches(size);
2485 err = memcg_update_all_list_lrus(size);
2487 memcg_nr_cache_ids = size;
2489 up_write(&memcg_cache_ids_sem);
2492 ida_simple_remove(&memcg_cache_ida, id);
2498 static void memcg_free_cache_id(int id)
2500 ida_simple_remove(&memcg_cache_ida, id);
2503 struct memcg_kmem_cache_create_work {
2504 struct mem_cgroup *memcg;
2505 struct kmem_cache *cachep;
2506 struct work_struct work;
2509 static void memcg_kmem_cache_create_func(struct work_struct *w)
2511 struct memcg_kmem_cache_create_work *cw =
2512 container_of(w, struct memcg_kmem_cache_create_work, work);
2513 struct mem_cgroup *memcg = cw->memcg;
2514 struct kmem_cache *cachep = cw->cachep;
2516 memcg_create_kmem_cache(memcg, cachep);
2518 css_put(&memcg->css);
2523 * Enqueue the creation of a per-memcg kmem_cache.
2525 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2526 struct kmem_cache *cachep)
2528 struct memcg_kmem_cache_create_work *cw;
2530 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2534 css_get(&memcg->css);
2537 cw->cachep = cachep;
2538 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2540 queue_work(memcg_kmem_cache_wq, &cw->work);
2543 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2544 struct kmem_cache *cachep)
2547 * We need to stop accounting when we kmalloc, because if the
2548 * corresponding kmalloc cache is not yet created, the first allocation
2549 * in __memcg_schedule_kmem_cache_create will recurse.
2551 * However, it is better to enclose the whole function. Depending on
2552 * the debugging options enabled, INIT_WORK(), for instance, can
2553 * trigger an allocation. This too, will make us recurse. Because at
2554 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2555 * the safest choice is to do it like this, wrapping the whole function.
2557 current->memcg_kmem_skip_account = 1;
2558 __memcg_schedule_kmem_cache_create(memcg, cachep);
2559 current->memcg_kmem_skip_account = 0;
2562 static inline bool memcg_kmem_bypass(void)
2564 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2570 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2571 * @cachep: the original global kmem cache
2573 * Return the kmem_cache we're supposed to use for a slab allocation.
2574 * We try to use the current memcg's version of the cache.
2576 * If the cache does not exist yet, if we are the first user of it, we
2577 * create it asynchronously in a workqueue and let the current allocation
2578 * go through with the original cache.
2580 * This function takes a reference to the cache it returns to assure it
2581 * won't get destroyed while we are working with it. Once the caller is
2582 * done with it, memcg_kmem_put_cache() must be called to release the
2585 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2587 struct mem_cgroup *memcg;
2588 struct kmem_cache *memcg_cachep;
2591 VM_BUG_ON(!is_root_cache(cachep));
2593 if (memcg_kmem_bypass())
2596 if (current->memcg_kmem_skip_account)
2599 memcg = get_mem_cgroup_from_current();
2600 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2604 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2605 if (likely(memcg_cachep))
2606 return memcg_cachep;
2609 * If we are in a safe context (can wait, and not in interrupt
2610 * context), we could be be predictable and return right away.
2611 * This would guarantee that the allocation being performed
2612 * already belongs in the new cache.
2614 * However, there are some clashes that can arrive from locking.
2615 * For instance, because we acquire the slab_mutex while doing
2616 * memcg_create_kmem_cache, this means no further allocation
2617 * could happen with the slab_mutex held. So it's better to
2620 memcg_schedule_kmem_cache_create(memcg, cachep);
2622 css_put(&memcg->css);
2627 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2628 * @cachep: the cache returned by memcg_kmem_get_cache
2630 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2632 if (!is_root_cache(cachep))
2633 css_put(&cachep->memcg_params.memcg->css);
2637 * memcg_kmem_charge_memcg: charge a kmem page
2638 * @page: page to charge
2639 * @gfp: reclaim mode
2640 * @order: allocation order
2641 * @memcg: memory cgroup to charge
2643 * Returns 0 on success, an error code on failure.
2645 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2646 struct mem_cgroup *memcg)
2648 unsigned int nr_pages = 1 << order;
2649 struct page_counter *counter;
2652 ret = try_charge(memcg, gfp, nr_pages);
2656 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2657 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2660 * Enforce __GFP_NOFAIL allocation because callers are not
2661 * prepared to see failures and likely do not have any failure
2664 if (gfp & __GFP_NOFAIL) {
2665 page_counter_charge(&memcg->kmem, nr_pages);
2668 cancel_charge(memcg, nr_pages);
2672 page->mem_cgroup = memcg;
2678 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2679 * @page: page to charge
2680 * @gfp: reclaim mode
2681 * @order: allocation order
2683 * Returns 0 on success, an error code on failure.
2685 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2687 struct mem_cgroup *memcg;
2690 if (mem_cgroup_disabled() || memcg_kmem_bypass())
2693 memcg = get_mem_cgroup_from_current();
2694 if (!mem_cgroup_is_root(memcg)) {
2695 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2697 __SetPageKmemcg(page);
2699 css_put(&memcg->css);
2703 * memcg_kmem_uncharge: uncharge a kmem page
2704 * @page: page to uncharge
2705 * @order: allocation order
2707 void memcg_kmem_uncharge(struct page *page, int order)
2709 struct mem_cgroup *memcg = page->mem_cgroup;
2710 unsigned int nr_pages = 1 << order;
2715 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2717 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2718 page_counter_uncharge(&memcg->kmem, nr_pages);
2720 page_counter_uncharge(&memcg->memory, nr_pages);
2721 if (do_memsw_account())
2722 page_counter_uncharge(&memcg->memsw, nr_pages);
2724 page->mem_cgroup = NULL;
2726 /* slab pages do not have PageKmemcg flag set */
2727 if (PageKmemcg(page))
2728 __ClearPageKmemcg(page);
2730 css_put_many(&memcg->css, nr_pages);
2732 #endif /* CONFIG_MEMCG_KMEM */
2734 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2737 * Because tail pages are not marked as "used", set it. We're under
2738 * zone_lru_lock and migration entries setup in all page mappings.
2740 void mem_cgroup_split_huge_fixup(struct page *head)
2744 if (mem_cgroup_disabled())
2747 for (i = 1; i < HPAGE_PMD_NR; i++)
2748 head[i].mem_cgroup = head->mem_cgroup;
2750 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2752 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2754 #ifdef CONFIG_MEMCG_SWAP
2756 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2757 * @entry: swap entry to be moved
2758 * @from: mem_cgroup which the entry is moved from
2759 * @to: mem_cgroup which the entry is moved to
2761 * It succeeds only when the swap_cgroup's record for this entry is the same
2762 * as the mem_cgroup's id of @from.
2764 * Returns 0 on success, -EINVAL on failure.
2766 * The caller must have charged to @to, IOW, called page_counter_charge() about
2767 * both res and memsw, and called css_get().
2769 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2770 struct mem_cgroup *from, struct mem_cgroup *to)
2772 unsigned short old_id, new_id;
2774 old_id = mem_cgroup_id(from);
2775 new_id = mem_cgroup_id(to);
2777 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2778 mod_memcg_state(from, MEMCG_SWAP, -1);
2779 mod_memcg_state(to, MEMCG_SWAP, 1);
2785 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2786 struct mem_cgroup *from, struct mem_cgroup *to)
2792 static DEFINE_MUTEX(memcg_max_mutex);
2794 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2795 unsigned long max, bool memsw)
2797 bool enlarge = false;
2798 bool drained = false;
2800 bool limits_invariant;
2801 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2804 if (signal_pending(current)) {
2809 mutex_lock(&memcg_max_mutex);
2811 * Make sure that the new limit (memsw or memory limit) doesn't
2812 * break our basic invariant rule memory.max <= memsw.max.
2814 limits_invariant = memsw ? max >= memcg->memory.max :
2815 max <= memcg->memsw.max;
2816 if (!limits_invariant) {
2817 mutex_unlock(&memcg_max_mutex);
2821 if (max > counter->max)
2823 ret = page_counter_set_max(counter, max);
2824 mutex_unlock(&memcg_max_mutex);
2830 drain_all_stock(memcg);
2835 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2836 GFP_KERNEL, !memsw)) {
2842 if (!ret && enlarge)
2843 memcg_oom_recover(memcg);
2848 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2850 unsigned long *total_scanned)
2852 unsigned long nr_reclaimed = 0;
2853 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2854 unsigned long reclaimed;
2856 struct mem_cgroup_tree_per_node *mctz;
2857 unsigned long excess;
2858 unsigned long nr_scanned;
2863 mctz = soft_limit_tree_node(pgdat->node_id);
2866 * Do not even bother to check the largest node if the root
2867 * is empty. Do it lockless to prevent lock bouncing. Races
2868 * are acceptable as soft limit is best effort anyway.
2870 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2874 * This loop can run a while, specially if mem_cgroup's continuously
2875 * keep exceeding their soft limit and putting the system under
2882 mz = mem_cgroup_largest_soft_limit_node(mctz);
2887 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2888 gfp_mask, &nr_scanned);
2889 nr_reclaimed += reclaimed;
2890 *total_scanned += nr_scanned;
2891 spin_lock_irq(&mctz->lock);
2892 __mem_cgroup_remove_exceeded(mz, mctz);
2895 * If we failed to reclaim anything from this memory cgroup
2896 * it is time to move on to the next cgroup
2900 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2902 excess = soft_limit_excess(mz->memcg);
2904 * One school of thought says that we should not add
2905 * back the node to the tree if reclaim returns 0.
2906 * But our reclaim could return 0, simply because due
2907 * to priority we are exposing a smaller subset of
2908 * memory to reclaim from. Consider this as a longer
2911 /* If excess == 0, no tree ops */
2912 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2913 spin_unlock_irq(&mctz->lock);
2914 css_put(&mz->memcg->css);
2917 * Could not reclaim anything and there are no more
2918 * mem cgroups to try or we seem to be looping without
2919 * reclaiming anything.
2921 if (!nr_reclaimed &&
2923 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2925 } while (!nr_reclaimed);
2927 css_put(&next_mz->memcg->css);
2928 return nr_reclaimed;
2932 * Test whether @memcg has children, dead or alive. Note that this
2933 * function doesn't care whether @memcg has use_hierarchy enabled and
2934 * returns %true if there are child csses according to the cgroup
2935 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2937 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2942 ret = css_next_child(NULL, &memcg->css);
2948 * Reclaims as many pages from the given memcg as possible.
2950 * Caller is responsible for holding css reference for memcg.
2952 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2954 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2956 /* we call try-to-free pages for make this cgroup empty */
2957 lru_add_drain_all();
2959 drain_all_stock(memcg);
2961 /* try to free all pages in this cgroup */
2962 while (nr_retries && page_counter_read(&memcg->memory)) {
2965 if (signal_pending(current))
2968 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2972 /* maybe some writeback is necessary */
2973 congestion_wait(BLK_RW_ASYNC, HZ/10);
2981 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2982 char *buf, size_t nbytes,
2985 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2987 if (mem_cgroup_is_root(memcg))
2989 return mem_cgroup_force_empty(memcg) ?: nbytes;
2992 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2995 return mem_cgroup_from_css(css)->use_hierarchy;
2998 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2999 struct cftype *cft, u64 val)
3002 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3003 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3005 if (memcg->use_hierarchy == val)
3009 * If parent's use_hierarchy is set, we can't make any modifications
3010 * in the child subtrees. If it is unset, then the change can
3011 * occur, provided the current cgroup has no children.
3013 * For the root cgroup, parent_mem is NULL, we allow value to be
3014 * set if there are no children.
3016 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3017 (val == 1 || val == 0)) {
3018 if (!memcg_has_children(memcg))
3019 memcg->use_hierarchy = val;
3028 struct accumulated_stats {
3029 unsigned long stat[MEMCG_NR_STAT];
3030 unsigned long events[NR_VM_EVENT_ITEMS];
3031 unsigned long lru_pages[NR_LRU_LISTS];
3032 const unsigned int *stats_array;
3033 const unsigned int *events_array;
3038 static void accumulate_memcg_tree(struct mem_cgroup *memcg,
3039 struct accumulated_stats *acc)
3041 struct mem_cgroup *mi;
3044 for_each_mem_cgroup_tree(mi, memcg) {
3045 for (i = 0; i < acc->stats_size; i++)
3046 acc->stat[i] += memcg_page_state(mi,
3047 acc->stats_array ? acc->stats_array[i] : i);
3049 for (i = 0; i < acc->events_size; i++)
3050 acc->events[i] += memcg_sum_events(mi,
3051 acc->events_array ? acc->events_array[i] : i);
3053 for (i = 0; i < NR_LRU_LISTS; i++)
3054 acc->lru_pages[i] +=
3055 mem_cgroup_nr_lru_pages(mi, BIT(i));
3059 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3061 unsigned long val = 0;
3063 if (mem_cgroup_is_root(memcg)) {
3064 struct mem_cgroup *iter;
3066 for_each_mem_cgroup_tree(iter, memcg) {
3067 val += memcg_page_state(iter, MEMCG_CACHE);
3068 val += memcg_page_state(iter, MEMCG_RSS);
3070 val += memcg_page_state(iter, MEMCG_SWAP);
3074 val = page_counter_read(&memcg->memory);
3076 val = page_counter_read(&memcg->memsw);
3089 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3092 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3093 struct page_counter *counter;
3095 switch (MEMFILE_TYPE(cft->private)) {
3097 counter = &memcg->memory;
3100 counter = &memcg->memsw;
3103 counter = &memcg->kmem;
3106 counter = &memcg->tcpmem;
3112 switch (MEMFILE_ATTR(cft->private)) {
3114 if (counter == &memcg->memory)
3115 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3116 if (counter == &memcg->memsw)
3117 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3118 return (u64)page_counter_read(counter) * PAGE_SIZE;
3120 return (u64)counter->max * PAGE_SIZE;
3122 return (u64)counter->watermark * PAGE_SIZE;
3124 return counter->failcnt;
3125 case RES_SOFT_LIMIT:
3126 return (u64)memcg->soft_limit * PAGE_SIZE;
3132 #ifdef CONFIG_MEMCG_KMEM
3133 static int memcg_online_kmem(struct mem_cgroup *memcg)
3137 if (cgroup_memory_nokmem)
3140 BUG_ON(memcg->kmemcg_id >= 0);
3141 BUG_ON(memcg->kmem_state);
3143 memcg_id = memcg_alloc_cache_id();
3147 static_branch_inc(&memcg_kmem_enabled_key);
3149 * A memory cgroup is considered kmem-online as soon as it gets
3150 * kmemcg_id. Setting the id after enabling static branching will
3151 * guarantee no one starts accounting before all call sites are
3154 memcg->kmemcg_id = memcg_id;
3155 memcg->kmem_state = KMEM_ONLINE;
3156 INIT_LIST_HEAD(&memcg->kmem_caches);
3161 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3163 struct cgroup_subsys_state *css;
3164 struct mem_cgroup *parent, *child;
3167 if (memcg->kmem_state != KMEM_ONLINE)
3170 * Clear the online state before clearing memcg_caches array
3171 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3172 * guarantees that no cache will be created for this cgroup
3173 * after we are done (see memcg_create_kmem_cache()).
3175 memcg->kmem_state = KMEM_ALLOCATED;
3177 memcg_deactivate_kmem_caches(memcg);
3179 kmemcg_id = memcg->kmemcg_id;
3180 BUG_ON(kmemcg_id < 0);
3182 parent = parent_mem_cgroup(memcg);
3184 parent = root_mem_cgroup;
3187 * Change kmemcg_id of this cgroup and all its descendants to the
3188 * parent's id, and then move all entries from this cgroup's list_lrus
3189 * to ones of the parent. After we have finished, all list_lrus
3190 * corresponding to this cgroup are guaranteed to remain empty. The
3191 * ordering is imposed by list_lru_node->lock taken by
3192 * memcg_drain_all_list_lrus().
3194 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3195 css_for_each_descendant_pre(css, &memcg->css) {
3196 child = mem_cgroup_from_css(css);
3197 BUG_ON(child->kmemcg_id != kmemcg_id);
3198 child->kmemcg_id = parent->kmemcg_id;
3199 if (!memcg->use_hierarchy)
3204 memcg_drain_all_list_lrus(kmemcg_id, parent);
3206 memcg_free_cache_id(kmemcg_id);
3209 static void memcg_free_kmem(struct mem_cgroup *memcg)
3211 /* css_alloc() failed, offlining didn't happen */
3212 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3213 memcg_offline_kmem(memcg);
3215 if (memcg->kmem_state == KMEM_ALLOCATED) {
3216 memcg_destroy_kmem_caches(memcg);
3217 static_branch_dec(&memcg_kmem_enabled_key);
3218 WARN_ON(page_counter_read(&memcg->kmem));
3222 static int memcg_online_kmem(struct mem_cgroup *memcg)
3226 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3229 static void memcg_free_kmem(struct mem_cgroup *memcg)
3232 #endif /* CONFIG_MEMCG_KMEM */
3234 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3239 mutex_lock(&memcg_max_mutex);
3240 ret = page_counter_set_max(&memcg->kmem, max);
3241 mutex_unlock(&memcg_max_mutex);
3245 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3249 mutex_lock(&memcg_max_mutex);
3251 ret = page_counter_set_max(&memcg->tcpmem, max);
3255 if (!memcg->tcpmem_active) {
3257 * The active flag needs to be written after the static_key
3258 * update. This is what guarantees that the socket activation
3259 * function is the last one to run. See mem_cgroup_sk_alloc()
3260 * for details, and note that we don't mark any socket as
3261 * belonging to this memcg until that flag is up.
3263 * We need to do this, because static_keys will span multiple
3264 * sites, but we can't control their order. If we mark a socket
3265 * as accounted, but the accounting functions are not patched in
3266 * yet, we'll lose accounting.
3268 * We never race with the readers in mem_cgroup_sk_alloc(),
3269 * because when this value change, the code to process it is not
3272 static_branch_inc(&memcg_sockets_enabled_key);
3273 memcg->tcpmem_active = true;
3276 mutex_unlock(&memcg_max_mutex);
3281 * The user of this function is...
3284 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3285 char *buf, size_t nbytes, loff_t off)
3287 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3288 unsigned long nr_pages;
3291 buf = strstrip(buf);
3292 ret = page_counter_memparse(buf, "-1", &nr_pages);
3296 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3298 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3302 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3304 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3307 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3310 ret = memcg_update_kmem_max(memcg, nr_pages);
3313 ret = memcg_update_tcp_max(memcg, nr_pages);
3317 case RES_SOFT_LIMIT:
3318 memcg->soft_limit = nr_pages;
3322 return ret ?: nbytes;
3325 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3326 size_t nbytes, loff_t off)
3328 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3329 struct page_counter *counter;
3331 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3333 counter = &memcg->memory;
3336 counter = &memcg->memsw;
3339 counter = &memcg->kmem;
3342 counter = &memcg->tcpmem;
3348 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3350 page_counter_reset_watermark(counter);
3353 counter->failcnt = 0;
3362 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3365 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3369 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3370 struct cftype *cft, u64 val)
3372 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3374 if (val & ~MOVE_MASK)
3378 * No kind of locking is needed in here, because ->can_attach() will
3379 * check this value once in the beginning of the process, and then carry
3380 * on with stale data. This means that changes to this value will only
3381 * affect task migrations starting after the change.
3383 memcg->move_charge_at_immigrate = val;
3387 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3388 struct cftype *cft, u64 val)
3395 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3399 unsigned int lru_mask;
3402 static const struct numa_stat stats[] = {
3403 { "total", LRU_ALL },
3404 { "file", LRU_ALL_FILE },
3405 { "anon", LRU_ALL_ANON },
3406 { "unevictable", BIT(LRU_UNEVICTABLE) },
3408 const struct numa_stat *stat;
3411 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3413 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3414 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3415 seq_printf(m, "%s=%lu", stat->name, nr);
3416 for_each_node_state(nid, N_MEMORY) {
3417 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3419 seq_printf(m, " N%d=%lu", nid, nr);
3424 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3425 struct mem_cgroup *iter;
3428 for_each_mem_cgroup_tree(iter, memcg)
3429 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3430 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3431 for_each_node_state(nid, N_MEMORY) {
3433 for_each_mem_cgroup_tree(iter, memcg)
3434 nr += mem_cgroup_node_nr_lru_pages(
3435 iter, nid, stat->lru_mask);
3436 seq_printf(m, " N%d=%lu", nid, nr);
3443 #endif /* CONFIG_NUMA */
3445 /* Universal VM events cgroup1 shows, original sort order */
3446 static const unsigned int memcg1_events[] = {
3453 static const char *const memcg1_event_names[] = {
3460 static int memcg_stat_show(struct seq_file *m, void *v)
3462 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3463 unsigned long memory, memsw;
3464 struct mem_cgroup *mi;
3466 struct accumulated_stats acc;
3468 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3469 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3471 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3472 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3474 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3475 memcg_page_state(memcg, memcg1_stats[i]) *
3479 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3480 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3481 memcg_sum_events(memcg, memcg1_events[i]));
3483 for (i = 0; i < NR_LRU_LISTS; i++)
3484 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3485 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3487 /* Hierarchical information */
3488 memory = memsw = PAGE_COUNTER_MAX;
3489 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3490 memory = min(memory, mi->memory.max);
3491 memsw = min(memsw, mi->memsw.max);
3493 seq_printf(m, "hierarchical_memory_limit %llu\n",
3494 (u64)memory * PAGE_SIZE);
3495 if (do_memsw_account())
3496 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3497 (u64)memsw * PAGE_SIZE);
3499 memset(&acc, 0, sizeof(acc));
3500 acc.stats_size = ARRAY_SIZE(memcg1_stats);
3501 acc.stats_array = memcg1_stats;
3502 acc.events_size = ARRAY_SIZE(memcg1_events);
3503 acc.events_array = memcg1_events;
3504 accumulate_memcg_tree(memcg, &acc);
3506 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3507 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3509 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3510 (u64)acc.stat[i] * PAGE_SIZE);
3513 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3514 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3515 (u64)acc.events[i]);
3517 for (i = 0; i < NR_LRU_LISTS; i++)
3518 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3519 (u64)acc.lru_pages[i] * PAGE_SIZE);
3521 #ifdef CONFIG_DEBUG_VM
3524 struct mem_cgroup_per_node *mz;
3525 struct zone_reclaim_stat *rstat;
3526 unsigned long recent_rotated[2] = {0, 0};
3527 unsigned long recent_scanned[2] = {0, 0};
3529 for_each_online_pgdat(pgdat) {
3530 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3531 rstat = &mz->lruvec.reclaim_stat;
3533 recent_rotated[0] += rstat->recent_rotated[0];
3534 recent_rotated[1] += rstat->recent_rotated[1];
3535 recent_scanned[0] += rstat->recent_scanned[0];
3536 recent_scanned[1] += rstat->recent_scanned[1];
3538 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3539 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3540 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3541 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3548 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3551 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3553 return mem_cgroup_swappiness(memcg);
3556 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3557 struct cftype *cft, u64 val)
3559 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3565 memcg->swappiness = val;
3567 vm_swappiness = val;
3572 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3574 struct mem_cgroup_threshold_ary *t;
3575 unsigned long usage;
3580 t = rcu_dereference(memcg->thresholds.primary);
3582 t = rcu_dereference(memcg->memsw_thresholds.primary);
3587 usage = mem_cgroup_usage(memcg, swap);
3590 * current_threshold points to threshold just below or equal to usage.
3591 * If it's not true, a threshold was crossed after last
3592 * call of __mem_cgroup_threshold().
3594 i = t->current_threshold;
3597 * Iterate backward over array of thresholds starting from
3598 * current_threshold and check if a threshold is crossed.
3599 * If none of thresholds below usage is crossed, we read
3600 * only one element of the array here.
3602 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3603 eventfd_signal(t->entries[i].eventfd, 1);
3605 /* i = current_threshold + 1 */
3609 * Iterate forward over array of thresholds starting from
3610 * current_threshold+1 and check if a threshold is crossed.
3611 * If none of thresholds above usage is crossed, we read
3612 * only one element of the array here.
3614 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3615 eventfd_signal(t->entries[i].eventfd, 1);
3617 /* Update current_threshold */
3618 t->current_threshold = i - 1;
3623 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3626 __mem_cgroup_threshold(memcg, false);
3627 if (do_memsw_account())
3628 __mem_cgroup_threshold(memcg, true);
3630 memcg = parent_mem_cgroup(memcg);
3634 static int compare_thresholds(const void *a, const void *b)
3636 const struct mem_cgroup_threshold *_a = a;
3637 const struct mem_cgroup_threshold *_b = b;
3639 if (_a->threshold > _b->threshold)
3642 if (_a->threshold < _b->threshold)
3648 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3650 struct mem_cgroup_eventfd_list *ev;
3652 spin_lock(&memcg_oom_lock);
3654 list_for_each_entry(ev, &memcg->oom_notify, list)
3655 eventfd_signal(ev->eventfd, 1);
3657 spin_unlock(&memcg_oom_lock);
3661 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3663 struct mem_cgroup *iter;
3665 for_each_mem_cgroup_tree(iter, memcg)
3666 mem_cgroup_oom_notify_cb(iter);
3669 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3670 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3672 struct mem_cgroup_thresholds *thresholds;
3673 struct mem_cgroup_threshold_ary *new;
3674 unsigned long threshold;
3675 unsigned long usage;
3678 ret = page_counter_memparse(args, "-1", &threshold);
3682 mutex_lock(&memcg->thresholds_lock);
3685 thresholds = &memcg->thresholds;
3686 usage = mem_cgroup_usage(memcg, false);
3687 } else if (type == _MEMSWAP) {
3688 thresholds = &memcg->memsw_thresholds;
3689 usage = mem_cgroup_usage(memcg, true);
3693 /* Check if a threshold crossed before adding a new one */
3694 if (thresholds->primary)
3695 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3697 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3699 /* Allocate memory for new array of thresholds */
3700 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3708 /* Copy thresholds (if any) to new array */
3709 if (thresholds->primary) {
3710 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3711 sizeof(struct mem_cgroup_threshold));
3714 /* Add new threshold */
3715 new->entries[size - 1].eventfd = eventfd;
3716 new->entries[size - 1].threshold = threshold;
3718 /* Sort thresholds. Registering of new threshold isn't time-critical */
3719 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3720 compare_thresholds, NULL);
3722 /* Find current threshold */
3723 new->current_threshold = -1;
3724 for (i = 0; i < size; i++) {
3725 if (new->entries[i].threshold <= usage) {
3727 * new->current_threshold will not be used until
3728 * rcu_assign_pointer(), so it's safe to increment
3731 ++new->current_threshold;
3736 /* Free old spare buffer and save old primary buffer as spare */
3737 kfree(thresholds->spare);
3738 thresholds->spare = thresholds->primary;
3740 rcu_assign_pointer(thresholds->primary, new);
3742 /* To be sure that nobody uses thresholds */
3746 mutex_unlock(&memcg->thresholds_lock);
3751 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3752 struct eventfd_ctx *eventfd, const char *args)
3754 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3757 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3758 struct eventfd_ctx *eventfd, const char *args)
3760 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3763 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3764 struct eventfd_ctx *eventfd, enum res_type type)
3766 struct mem_cgroup_thresholds *thresholds;
3767 struct mem_cgroup_threshold_ary *new;
3768 unsigned long usage;
3769 int i, j, size, entries;
3771 mutex_lock(&memcg->thresholds_lock);
3774 thresholds = &memcg->thresholds;
3775 usage = mem_cgroup_usage(memcg, false);
3776 } else if (type == _MEMSWAP) {
3777 thresholds = &memcg->memsw_thresholds;
3778 usage = mem_cgroup_usage(memcg, true);
3782 if (!thresholds->primary)
3785 /* Check if a threshold crossed before removing */
3786 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3788 /* Calculate new number of threshold */
3790 for (i = 0; i < thresholds->primary->size; i++) {
3791 if (thresholds->primary->entries[i].eventfd != eventfd)
3797 new = thresholds->spare;
3799 /* If no items related to eventfd have been cleared, nothing to do */
3803 /* Set thresholds array to NULL if we don't have thresholds */
3812 /* Copy thresholds and find current threshold */
3813 new->current_threshold = -1;
3814 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3815 if (thresholds->primary->entries[i].eventfd == eventfd)
3818 new->entries[j] = thresholds->primary->entries[i];
3819 if (new->entries[j].threshold <= usage) {
3821 * new->current_threshold will not be used
3822 * until rcu_assign_pointer(), so it's safe to increment
3825 ++new->current_threshold;
3831 /* Swap primary and spare array */
3832 thresholds->spare = thresholds->primary;
3834 rcu_assign_pointer(thresholds->primary, new);
3836 /* To be sure that nobody uses thresholds */
3839 /* If all events are unregistered, free the spare array */
3841 kfree(thresholds->spare);
3842 thresholds->spare = NULL;
3845 mutex_unlock(&memcg->thresholds_lock);
3848 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3849 struct eventfd_ctx *eventfd)
3851 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3854 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3855 struct eventfd_ctx *eventfd)
3857 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3860 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3861 struct eventfd_ctx *eventfd, const char *args)
3863 struct mem_cgroup_eventfd_list *event;
3865 event = kmalloc(sizeof(*event), GFP_KERNEL);
3869 spin_lock(&memcg_oom_lock);
3871 event->eventfd = eventfd;
3872 list_add(&event->list, &memcg->oom_notify);
3874 /* already in OOM ? */
3875 if (memcg->under_oom)
3876 eventfd_signal(eventfd, 1);
3877 spin_unlock(&memcg_oom_lock);
3882 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3883 struct eventfd_ctx *eventfd)
3885 struct mem_cgroup_eventfd_list *ev, *tmp;
3887 spin_lock(&memcg_oom_lock);
3889 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3890 if (ev->eventfd == eventfd) {
3891 list_del(&ev->list);
3896 spin_unlock(&memcg_oom_lock);
3899 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3901 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3903 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3904 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3905 seq_printf(sf, "oom_kill %lu\n",
3906 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3910 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3911 struct cftype *cft, u64 val)
3913 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3915 /* cannot set to root cgroup and only 0 and 1 are allowed */
3916 if (!css->parent || !((val == 0) || (val == 1)))
3919 memcg->oom_kill_disable = val;
3921 memcg_oom_recover(memcg);
3926 #ifdef CONFIG_CGROUP_WRITEBACK
3928 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3930 return wb_domain_init(&memcg->cgwb_domain, gfp);
3933 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3935 wb_domain_exit(&memcg->cgwb_domain);
3938 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3940 wb_domain_size_changed(&memcg->cgwb_domain);
3943 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3945 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3947 if (!memcg->css.parent)
3950 return &memcg->cgwb_domain;
3954 * idx can be of type enum memcg_stat_item or node_stat_item.
3955 * Keep in sync with memcg_exact_page().
3957 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
3959 long x = atomic_long_read(&memcg->stat[idx]);
3962 for_each_online_cpu(cpu)
3963 x += per_cpu_ptr(memcg->stat_cpu, cpu)->count[idx];
3970 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3971 * @wb: bdi_writeback in question
3972 * @pfilepages: out parameter for number of file pages
3973 * @pheadroom: out parameter for number of allocatable pages according to memcg
3974 * @pdirty: out parameter for number of dirty pages
3975 * @pwriteback: out parameter for number of pages under writeback
3977 * Determine the numbers of file, headroom, dirty, and writeback pages in
3978 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3979 * is a bit more involved.
3981 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3982 * headroom is calculated as the lowest headroom of itself and the
3983 * ancestors. Note that this doesn't consider the actual amount of
3984 * available memory in the system. The caller should further cap
3985 * *@pheadroom accordingly.
3987 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3988 unsigned long *pheadroom, unsigned long *pdirty,
3989 unsigned long *pwriteback)
3991 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3992 struct mem_cgroup *parent;
3994 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
3996 /* this should eventually include NR_UNSTABLE_NFS */
3997 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
3998 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3999 (1 << LRU_ACTIVE_FILE));
4000 *pheadroom = PAGE_COUNTER_MAX;
4002 while ((parent = parent_mem_cgroup(memcg))) {
4003 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4004 unsigned long used = page_counter_read(&memcg->memory);
4006 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4011 #else /* CONFIG_CGROUP_WRITEBACK */
4013 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4018 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4022 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4026 #endif /* CONFIG_CGROUP_WRITEBACK */
4029 * DO NOT USE IN NEW FILES.
4031 * "cgroup.event_control" implementation.
4033 * This is way over-engineered. It tries to support fully configurable
4034 * events for each user. Such level of flexibility is completely
4035 * unnecessary especially in the light of the planned unified hierarchy.
4037 * Please deprecate this and replace with something simpler if at all
4042 * Unregister event and free resources.
4044 * Gets called from workqueue.
4046 static void memcg_event_remove(struct work_struct *work)
4048 struct mem_cgroup_event *event =
4049 container_of(work, struct mem_cgroup_event, remove);
4050 struct mem_cgroup *memcg = event->memcg;
4052 remove_wait_queue(event->wqh, &event->wait);
4054 event->unregister_event(memcg, event->eventfd);
4056 /* Notify userspace the event is going away. */
4057 eventfd_signal(event->eventfd, 1);
4059 eventfd_ctx_put(event->eventfd);
4061 css_put(&memcg->css);
4065 * Gets called on EPOLLHUP on eventfd when user closes it.
4067 * Called with wqh->lock held and interrupts disabled.
4069 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4070 int sync, void *key)
4072 struct mem_cgroup_event *event =
4073 container_of(wait, struct mem_cgroup_event, wait);
4074 struct mem_cgroup *memcg = event->memcg;
4075 __poll_t flags = key_to_poll(key);
4077 if (flags & EPOLLHUP) {
4079 * If the event has been detached at cgroup removal, we
4080 * can simply return knowing the other side will cleanup
4083 * We can't race against event freeing since the other
4084 * side will require wqh->lock via remove_wait_queue(),
4087 spin_lock(&memcg->event_list_lock);
4088 if (!list_empty(&event->list)) {
4089 list_del_init(&event->list);
4091 * We are in atomic context, but cgroup_event_remove()
4092 * may sleep, so we have to call it in workqueue.
4094 schedule_work(&event->remove);
4096 spin_unlock(&memcg->event_list_lock);
4102 static void memcg_event_ptable_queue_proc(struct file *file,
4103 wait_queue_head_t *wqh, poll_table *pt)
4105 struct mem_cgroup_event *event =
4106 container_of(pt, struct mem_cgroup_event, pt);
4109 add_wait_queue(wqh, &event->wait);
4113 * DO NOT USE IN NEW FILES.
4115 * Parse input and register new cgroup event handler.
4117 * Input must be in format '<event_fd> <control_fd> <args>'.
4118 * Interpretation of args is defined by control file implementation.
4120 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4121 char *buf, size_t nbytes, loff_t off)
4123 struct cgroup_subsys_state *css = of_css(of);
4124 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4125 struct mem_cgroup_event *event;
4126 struct cgroup_subsys_state *cfile_css;
4127 unsigned int efd, cfd;
4130 struct dentry *cdentry;
4135 buf = strstrip(buf);
4137 efd = simple_strtoul(buf, &endp, 10);
4142 cfd = simple_strtoul(buf, &endp, 10);
4143 if ((*endp != ' ') && (*endp != '\0'))
4147 event = kzalloc(sizeof(*event), GFP_KERNEL);
4151 event->memcg = memcg;
4152 INIT_LIST_HEAD(&event->list);
4153 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4154 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4155 INIT_WORK(&event->remove, memcg_event_remove);
4163 event->eventfd = eventfd_ctx_fileget(efile.file);
4164 if (IS_ERR(event->eventfd)) {
4165 ret = PTR_ERR(event->eventfd);
4172 goto out_put_eventfd;
4175 /* the process need read permission on control file */
4176 /* AV: shouldn't we check that it's been opened for read instead? */
4177 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4182 * The control file must be a regular cgroup1 file. As a regular cgroup
4183 * file can't be renamed, it's safe to access its name afterwards.
4185 cdentry = cfile.file->f_path.dentry;
4186 if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
4192 * Determine the event callbacks and set them in @event. This used
4193 * to be done via struct cftype but cgroup core no longer knows
4194 * about these events. The following is crude but the whole thing
4195 * is for compatibility anyway.
4197 * DO NOT ADD NEW FILES.
4199 name = cdentry->d_name.name;
4201 if (!strcmp(name, "memory.usage_in_bytes")) {
4202 event->register_event = mem_cgroup_usage_register_event;
4203 event->unregister_event = mem_cgroup_usage_unregister_event;
4204 } else if (!strcmp(name, "memory.oom_control")) {
4205 event->register_event = mem_cgroup_oom_register_event;
4206 event->unregister_event = mem_cgroup_oom_unregister_event;
4207 } else if (!strcmp(name, "memory.pressure_level")) {
4208 event->register_event = vmpressure_register_event;
4209 event->unregister_event = vmpressure_unregister_event;
4210 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4211 event->register_event = memsw_cgroup_usage_register_event;
4212 event->unregister_event = memsw_cgroup_usage_unregister_event;
4219 * Verify @cfile should belong to @css. Also, remaining events are
4220 * automatically removed on cgroup destruction but the removal is
4221 * asynchronous, so take an extra ref on @css.
4223 cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
4224 &memory_cgrp_subsys);
4226 if (IS_ERR(cfile_css))
4228 if (cfile_css != css) {
4233 ret = event->register_event(memcg, event->eventfd, buf);
4237 vfs_poll(efile.file, &event->pt);
4239 spin_lock(&memcg->event_list_lock);
4240 list_add(&event->list, &memcg->event_list);
4241 spin_unlock(&memcg->event_list_lock);
4253 eventfd_ctx_put(event->eventfd);
4262 static struct cftype mem_cgroup_legacy_files[] = {
4264 .name = "usage_in_bytes",
4265 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4266 .read_u64 = mem_cgroup_read_u64,
4269 .name = "max_usage_in_bytes",
4270 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4271 .write = mem_cgroup_reset,
4272 .read_u64 = mem_cgroup_read_u64,
4275 .name = "limit_in_bytes",
4276 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4277 .write = mem_cgroup_write,
4278 .read_u64 = mem_cgroup_read_u64,
4281 .name = "soft_limit_in_bytes",
4282 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4283 .write = mem_cgroup_write,
4284 .read_u64 = mem_cgroup_read_u64,
4288 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4289 .write = mem_cgroup_reset,
4290 .read_u64 = mem_cgroup_read_u64,
4294 .seq_show = memcg_stat_show,
4297 .name = "force_empty",
4298 .write = mem_cgroup_force_empty_write,
4301 .name = "use_hierarchy",
4302 .write_u64 = mem_cgroup_hierarchy_write,
4303 .read_u64 = mem_cgroup_hierarchy_read,
4306 .name = "cgroup.event_control", /* XXX: for compat */
4307 .write = memcg_write_event_control,
4308 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4311 .name = "swappiness",
4312 .read_u64 = mem_cgroup_swappiness_read,
4313 .write_u64 = mem_cgroup_swappiness_write,
4316 .name = "move_charge_at_immigrate",
4317 .read_u64 = mem_cgroup_move_charge_read,
4318 .write_u64 = mem_cgroup_move_charge_write,
4321 .name = "oom_control",
4322 .seq_show = mem_cgroup_oom_control_read,
4323 .write_u64 = mem_cgroup_oom_control_write,
4324 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4327 .name = "pressure_level",
4331 .name = "numa_stat",
4332 .seq_show = memcg_numa_stat_show,
4336 .name = "kmem.limit_in_bytes",
4337 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4338 .write = mem_cgroup_write,
4339 .read_u64 = mem_cgroup_read_u64,
4342 .name = "kmem.usage_in_bytes",
4343 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4344 .read_u64 = mem_cgroup_read_u64,
4347 .name = "kmem.failcnt",
4348 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4349 .write = mem_cgroup_reset,
4350 .read_u64 = mem_cgroup_read_u64,
4353 .name = "kmem.max_usage_in_bytes",
4354 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4355 .write = mem_cgroup_reset,
4356 .read_u64 = mem_cgroup_read_u64,
4358 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4360 .name = "kmem.slabinfo",
4361 .seq_start = memcg_slab_start,
4362 .seq_next = memcg_slab_next,
4363 .seq_stop = memcg_slab_stop,
4364 .seq_show = memcg_slab_show,
4368 .name = "kmem.tcp.limit_in_bytes",
4369 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4370 .write = mem_cgroup_write,
4371 .read_u64 = mem_cgroup_read_u64,
4374 .name = "kmem.tcp.usage_in_bytes",
4375 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4376 .read_u64 = mem_cgroup_read_u64,
4379 .name = "kmem.tcp.failcnt",
4380 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4381 .write = mem_cgroup_reset,
4382 .read_u64 = mem_cgroup_read_u64,
4385 .name = "kmem.tcp.max_usage_in_bytes",
4386 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4387 .write = mem_cgroup_reset,
4388 .read_u64 = mem_cgroup_read_u64,
4390 { }, /* terminate */
4394 * Private memory cgroup IDR
4396 * Swap-out records and page cache shadow entries need to store memcg
4397 * references in constrained space, so we maintain an ID space that is
4398 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4399 * memory-controlled cgroups to 64k.
4401 * However, there usually are many references to the oflline CSS after
4402 * the cgroup has been destroyed, such as page cache or reclaimable
4403 * slab objects, that don't need to hang on to the ID. We want to keep
4404 * those dead CSS from occupying IDs, or we might quickly exhaust the
4405 * relatively small ID space and prevent the creation of new cgroups
4406 * even when there are much fewer than 64k cgroups - possibly none.
4408 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4409 * be freed and recycled when it's no longer needed, which is usually
4410 * when the CSS is offlined.
4412 * The only exception to that are records of swapped out tmpfs/shmem
4413 * pages that need to be attributed to live ancestors on swapin. But
4414 * those references are manageable from userspace.
4417 static DEFINE_IDR(mem_cgroup_idr);
4419 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4421 if (memcg->id.id > 0) {
4422 idr_remove(&mem_cgroup_idr, memcg->id.id);
4427 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4429 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4430 atomic_add(n, &memcg->id.ref);
4433 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4435 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4436 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4437 mem_cgroup_id_remove(memcg);
4439 /* Memcg ID pins CSS */
4440 css_put(&memcg->css);
4444 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4446 mem_cgroup_id_get_many(memcg, 1);
4449 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4451 mem_cgroup_id_put_many(memcg, 1);
4455 * mem_cgroup_from_id - look up a memcg from a memcg id
4456 * @id: the memcg id to look up
4458 * Caller must hold rcu_read_lock().
4460 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4462 WARN_ON_ONCE(!rcu_read_lock_held());
4463 return idr_find(&mem_cgroup_idr, id);
4466 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4468 struct mem_cgroup_per_node *pn;
4471 * This routine is called against possible nodes.
4472 * But it's BUG to call kmalloc() against offline node.
4474 * TODO: this routine can waste much memory for nodes which will
4475 * never be onlined. It's better to use memory hotplug callback
4478 if (!node_state(node, N_NORMAL_MEMORY))
4480 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4484 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4485 if (!pn->lruvec_stat_cpu) {
4490 lruvec_init(&pn->lruvec);
4491 pn->usage_in_excess = 0;
4492 pn->on_tree = false;
4495 memcg->nodeinfo[node] = pn;
4499 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4501 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4506 free_percpu(pn->lruvec_stat_cpu);
4510 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4515 free_mem_cgroup_per_node_info(memcg, node);
4516 free_percpu(memcg->stat_cpu);
4520 static void mem_cgroup_free(struct mem_cgroup *memcg)
4522 memcg_wb_domain_exit(memcg);
4523 __mem_cgroup_free(memcg);
4526 static struct mem_cgroup *mem_cgroup_alloc(void)
4528 struct mem_cgroup *memcg;
4532 size = sizeof(struct mem_cgroup);
4533 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4535 memcg = kzalloc(size, GFP_KERNEL);
4539 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4540 1, MEM_CGROUP_ID_MAX,
4542 if (memcg->id.id < 0)
4545 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4546 if (!memcg->stat_cpu)
4550 if (alloc_mem_cgroup_per_node_info(memcg, node))
4553 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4556 INIT_WORK(&memcg->high_work, high_work_func);
4557 memcg->last_scanned_node = MAX_NUMNODES;
4558 INIT_LIST_HEAD(&memcg->oom_notify);
4559 mutex_init(&memcg->thresholds_lock);
4560 spin_lock_init(&memcg->move_lock);
4561 vmpressure_init(&memcg->vmpressure);
4562 INIT_LIST_HEAD(&memcg->event_list);
4563 spin_lock_init(&memcg->event_list_lock);
4564 memcg->socket_pressure = jiffies;
4565 #ifdef CONFIG_MEMCG_KMEM
4566 memcg->kmemcg_id = -1;
4568 #ifdef CONFIG_CGROUP_WRITEBACK
4569 INIT_LIST_HEAD(&memcg->cgwb_list);
4571 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4574 mem_cgroup_id_remove(memcg);
4575 __mem_cgroup_free(memcg);
4579 static struct cgroup_subsys_state * __ref
4580 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4582 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4583 struct mem_cgroup *memcg;
4584 long error = -ENOMEM;
4586 memcg = mem_cgroup_alloc();
4588 return ERR_PTR(error);
4590 memcg->high = PAGE_COUNTER_MAX;
4591 memcg->soft_limit = PAGE_COUNTER_MAX;
4593 memcg->swappiness = mem_cgroup_swappiness(parent);
4594 memcg->oom_kill_disable = parent->oom_kill_disable;
4596 if (parent && parent->use_hierarchy) {
4597 memcg->use_hierarchy = true;
4598 page_counter_init(&memcg->memory, &parent->memory);
4599 page_counter_init(&memcg->swap, &parent->swap);
4600 page_counter_init(&memcg->memsw, &parent->memsw);
4601 page_counter_init(&memcg->kmem, &parent->kmem);
4602 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4604 page_counter_init(&memcg->memory, NULL);
4605 page_counter_init(&memcg->swap, NULL);
4606 page_counter_init(&memcg->memsw, NULL);
4607 page_counter_init(&memcg->kmem, NULL);
4608 page_counter_init(&memcg->tcpmem, NULL);
4610 * Deeper hierachy with use_hierarchy == false doesn't make
4611 * much sense so let cgroup subsystem know about this
4612 * unfortunate state in our controller.
4614 if (parent != root_mem_cgroup)
4615 memory_cgrp_subsys.broken_hierarchy = true;
4618 /* The following stuff does not apply to the root */
4620 root_mem_cgroup = memcg;
4624 error = memcg_online_kmem(memcg);
4628 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4629 static_branch_inc(&memcg_sockets_enabled_key);
4633 mem_cgroup_id_remove(memcg);
4634 mem_cgroup_free(memcg);
4635 return ERR_PTR(-ENOMEM);
4638 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4640 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4643 * A memcg must be visible for memcg_expand_shrinker_maps()
4644 * by the time the maps are allocated. So, we allocate maps
4645 * here, when for_each_mem_cgroup() can't skip it.
4647 if (memcg_alloc_shrinker_maps(memcg)) {
4648 mem_cgroup_id_remove(memcg);
4652 /* Online state pins memcg ID, memcg ID pins CSS */
4653 atomic_set(&memcg->id.ref, 1);
4658 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4660 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4661 struct mem_cgroup_event *event, *tmp;
4664 * Unregister events and notify userspace.
4665 * Notify userspace about cgroup removing only after rmdir of cgroup
4666 * directory to avoid race between userspace and kernelspace.
4668 spin_lock(&memcg->event_list_lock);
4669 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4670 list_del_init(&event->list);
4671 schedule_work(&event->remove);
4673 spin_unlock(&memcg->event_list_lock);
4675 page_counter_set_min(&memcg->memory, 0);
4676 page_counter_set_low(&memcg->memory, 0);
4678 memcg_offline_kmem(memcg);
4679 wb_memcg_offline(memcg);
4681 mem_cgroup_id_put(memcg);
4684 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4686 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4688 invalidate_reclaim_iterators(memcg);
4691 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4693 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4695 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4696 static_branch_dec(&memcg_sockets_enabled_key);
4698 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4699 static_branch_dec(&memcg_sockets_enabled_key);
4701 vmpressure_cleanup(&memcg->vmpressure);
4702 cancel_work_sync(&memcg->high_work);
4703 mem_cgroup_remove_from_trees(memcg);
4704 memcg_free_shrinker_maps(memcg);
4705 memcg_free_kmem(memcg);
4706 mem_cgroup_free(memcg);
4710 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4711 * @css: the target css
4713 * Reset the states of the mem_cgroup associated with @css. This is
4714 * invoked when the userland requests disabling on the default hierarchy
4715 * but the memcg is pinned through dependency. The memcg should stop
4716 * applying policies and should revert to the vanilla state as it may be
4717 * made visible again.
4719 * The current implementation only resets the essential configurations.
4720 * This needs to be expanded to cover all the visible parts.
4722 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4724 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4726 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4727 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4728 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4729 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4730 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4731 page_counter_set_min(&memcg->memory, 0);
4732 page_counter_set_low(&memcg->memory, 0);
4733 memcg->high = PAGE_COUNTER_MAX;
4734 memcg->soft_limit = PAGE_COUNTER_MAX;
4735 memcg_wb_domain_size_changed(memcg);
4739 /* Handlers for move charge at task migration. */
4740 static int mem_cgroup_do_precharge(unsigned long count)
4744 /* Try a single bulk charge without reclaim first, kswapd may wake */
4745 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4747 mc.precharge += count;
4751 /* Try charges one by one with reclaim, but do not retry */
4753 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4767 enum mc_target_type {
4774 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4775 unsigned long addr, pte_t ptent)
4777 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4779 if (!page || !page_mapped(page))
4781 if (PageAnon(page)) {
4782 if (!(mc.flags & MOVE_ANON))
4785 if (!(mc.flags & MOVE_FILE))
4788 if (!get_page_unless_zero(page))
4794 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4795 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4796 pte_t ptent, swp_entry_t *entry)
4798 struct page *page = NULL;
4799 swp_entry_t ent = pte_to_swp_entry(ptent);
4801 if (!(mc.flags & MOVE_ANON))
4805 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4806 * a device and because they are not accessible by CPU they are store
4807 * as special swap entry in the CPU page table.
4809 if (is_device_private_entry(ent)) {
4810 page = device_private_entry_to_page(ent);
4812 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4813 * a refcount of 1 when free (unlike normal page)
4815 if (!page_ref_add_unless(page, 1, 1))
4820 if (non_swap_entry(ent))
4824 * Because lookup_swap_cache() updates some statistics counter,
4825 * we call find_get_page() with swapper_space directly.
4827 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4828 if (do_memsw_account())
4829 entry->val = ent.val;
4834 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4835 pte_t ptent, swp_entry_t *entry)
4841 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4842 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4844 struct page *page = NULL;
4845 struct address_space *mapping;
4848 if (!vma->vm_file) /* anonymous vma */
4850 if (!(mc.flags & MOVE_FILE))
4853 mapping = vma->vm_file->f_mapping;
4854 pgoff = linear_page_index(vma, addr);
4856 /* page is moved even if it's not RSS of this task(page-faulted). */
4858 /* shmem/tmpfs may report page out on swap: account for that too. */
4859 if (shmem_mapping(mapping)) {
4860 page = find_get_entry(mapping, pgoff);
4861 if (radix_tree_exceptional_entry(page)) {
4862 swp_entry_t swp = radix_to_swp_entry(page);
4863 if (do_memsw_account())
4865 page = find_get_page(swap_address_space(swp),
4869 page = find_get_page(mapping, pgoff);
4871 page = find_get_page(mapping, pgoff);
4877 * mem_cgroup_move_account - move account of the page
4879 * @compound: charge the page as compound or small page
4880 * @from: mem_cgroup which the page is moved from.
4881 * @to: mem_cgroup which the page is moved to. @from != @to.
4883 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4885 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4888 static int mem_cgroup_move_account(struct page *page,
4890 struct mem_cgroup *from,
4891 struct mem_cgroup *to)
4893 unsigned long flags;
4894 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4898 VM_BUG_ON(from == to);
4899 VM_BUG_ON_PAGE(PageLRU(page), page);
4900 VM_BUG_ON(compound && !PageTransHuge(page));
4903 * Prevent mem_cgroup_migrate() from looking at
4904 * page->mem_cgroup of its source page while we change it.
4907 if (!trylock_page(page))
4911 if (page->mem_cgroup != from)
4914 anon = PageAnon(page);
4916 spin_lock_irqsave(&from->move_lock, flags);
4918 if (!anon && page_mapped(page)) {
4919 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4920 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4924 * move_lock grabbed above and caller set from->moving_account, so
4925 * mod_memcg_page_state will serialize updates to PageDirty.
4926 * So mapping should be stable for dirty pages.
4928 if (!anon && PageDirty(page)) {
4929 struct address_space *mapping = page_mapping(page);
4931 if (mapping_cap_account_dirty(mapping)) {
4932 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4933 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4937 if (PageWriteback(page)) {
4938 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4939 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4943 * It is safe to change page->mem_cgroup here because the page
4944 * is referenced, charged, and isolated - we can't race with
4945 * uncharging, charging, migration, or LRU putback.
4948 /* caller should have done css_get */
4949 page->mem_cgroup = to;
4950 spin_unlock_irqrestore(&from->move_lock, flags);
4954 local_irq_disable();
4955 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4956 memcg_check_events(to, page);
4957 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4958 memcg_check_events(from, page);
4967 * get_mctgt_type - get target type of moving charge
4968 * @vma: the vma the pte to be checked belongs
4969 * @addr: the address corresponding to the pte to be checked
4970 * @ptent: the pte to be checked
4971 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4974 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4975 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4976 * move charge. if @target is not NULL, the page is stored in target->page
4977 * with extra refcnt got(Callers should handle it).
4978 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4979 * target for charge migration. if @target is not NULL, the entry is stored
4981 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4982 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4983 * For now we such page is charge like a regular page would be as for all
4984 * intent and purposes it is just special memory taking the place of a
4987 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4989 * Called with pte lock held.
4992 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4993 unsigned long addr, pte_t ptent, union mc_target *target)
4995 struct page *page = NULL;
4996 enum mc_target_type ret = MC_TARGET_NONE;
4997 swp_entry_t ent = { .val = 0 };
4999 if (pte_present(ptent))
5000 page = mc_handle_present_pte(vma, addr, ptent);
5001 else if (is_swap_pte(ptent))
5002 page = mc_handle_swap_pte(vma, ptent, &ent);
5003 else if (pte_none(ptent))
5004 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5006 if (!page && !ent.val)
5010 * Do only loose check w/o serialization.
5011 * mem_cgroup_move_account() checks the page is valid or
5012 * not under LRU exclusion.
5014 if (page->mem_cgroup == mc.from) {
5015 ret = MC_TARGET_PAGE;
5016 if (is_device_private_page(page) ||
5017 is_device_public_page(page))
5018 ret = MC_TARGET_DEVICE;
5020 target->page = page;
5022 if (!ret || !target)
5026 * There is a swap entry and a page doesn't exist or isn't charged.
5027 * But we cannot move a tail-page in a THP.
5029 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5030 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5031 ret = MC_TARGET_SWAP;
5038 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5040 * We don't consider PMD mapped swapping or file mapped pages because THP does
5041 * not support them for now.
5042 * Caller should make sure that pmd_trans_huge(pmd) is true.
5044 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5045 unsigned long addr, pmd_t pmd, union mc_target *target)
5047 struct page *page = NULL;
5048 enum mc_target_type ret = MC_TARGET_NONE;
5050 if (unlikely(is_swap_pmd(pmd))) {
5051 VM_BUG_ON(thp_migration_supported() &&
5052 !is_pmd_migration_entry(pmd));
5055 page = pmd_page(pmd);
5056 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5057 if (!(mc.flags & MOVE_ANON))
5059 if (page->mem_cgroup == mc.from) {
5060 ret = MC_TARGET_PAGE;
5063 target->page = page;
5069 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5070 unsigned long addr, pmd_t pmd, union mc_target *target)
5072 return MC_TARGET_NONE;
5076 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5077 unsigned long addr, unsigned long end,
5078 struct mm_walk *walk)
5080 struct vm_area_struct *vma = walk->vma;
5084 ptl = pmd_trans_huge_lock(pmd, vma);
5087 * Note their can not be MC_TARGET_DEVICE for now as we do not
5088 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5089 * MEMORY_DEVICE_PRIVATE but this might change.
5091 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5092 mc.precharge += HPAGE_PMD_NR;
5097 if (pmd_trans_unstable(pmd))
5099 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5100 for (; addr != end; pte++, addr += PAGE_SIZE)
5101 if (get_mctgt_type(vma, addr, *pte, NULL))
5102 mc.precharge++; /* increment precharge temporarily */
5103 pte_unmap_unlock(pte - 1, ptl);
5109 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5111 unsigned long precharge;
5113 struct mm_walk mem_cgroup_count_precharge_walk = {
5114 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5117 down_read(&mm->mmap_sem);
5118 walk_page_range(0, mm->highest_vm_end,
5119 &mem_cgroup_count_precharge_walk);
5120 up_read(&mm->mmap_sem);
5122 precharge = mc.precharge;
5128 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5130 unsigned long precharge = mem_cgroup_count_precharge(mm);
5132 VM_BUG_ON(mc.moving_task);
5133 mc.moving_task = current;
5134 return mem_cgroup_do_precharge(precharge);
5137 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5138 static void __mem_cgroup_clear_mc(void)
5140 struct mem_cgroup *from = mc.from;
5141 struct mem_cgroup *to = mc.to;
5143 /* we must uncharge all the leftover precharges from mc.to */
5145 cancel_charge(mc.to, mc.precharge);
5149 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5150 * we must uncharge here.
5152 if (mc.moved_charge) {
5153 cancel_charge(mc.from, mc.moved_charge);
5154 mc.moved_charge = 0;
5156 /* we must fixup refcnts and charges */
5157 if (mc.moved_swap) {
5158 /* uncharge swap account from the old cgroup */
5159 if (!mem_cgroup_is_root(mc.from))
5160 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5162 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5165 * we charged both to->memory and to->memsw, so we
5166 * should uncharge to->memory.
5168 if (!mem_cgroup_is_root(mc.to))
5169 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5171 css_put_many(&mc.to->css, mc.moved_swap);
5175 memcg_oom_recover(from);
5176 memcg_oom_recover(to);
5177 wake_up_all(&mc.waitq);
5180 static void mem_cgroup_clear_mc(void)
5182 struct mm_struct *mm = mc.mm;
5185 * we must clear moving_task before waking up waiters at the end of
5188 mc.moving_task = NULL;
5189 __mem_cgroup_clear_mc();
5190 spin_lock(&mc.lock);
5194 spin_unlock(&mc.lock);
5199 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5201 struct cgroup_subsys_state *css;
5202 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5203 struct mem_cgroup *from;
5204 struct task_struct *leader, *p;
5205 struct mm_struct *mm;
5206 unsigned long move_flags;
5209 /* charge immigration isn't supported on the default hierarchy */
5210 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5214 * Multi-process migrations only happen on the default hierarchy
5215 * where charge immigration is not used. Perform charge
5216 * immigration if @tset contains a leader and whine if there are
5220 cgroup_taskset_for_each_leader(leader, css, tset) {
5223 memcg = mem_cgroup_from_css(css);
5229 * We are now commited to this value whatever it is. Changes in this
5230 * tunable will only affect upcoming migrations, not the current one.
5231 * So we need to save it, and keep it going.
5233 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5237 from = mem_cgroup_from_task(p);
5239 VM_BUG_ON(from == memcg);
5241 mm = get_task_mm(p);
5244 /* We move charges only when we move a owner of the mm */
5245 if (mm->owner == p) {
5248 VM_BUG_ON(mc.precharge);
5249 VM_BUG_ON(mc.moved_charge);
5250 VM_BUG_ON(mc.moved_swap);
5252 spin_lock(&mc.lock);
5256 mc.flags = move_flags;
5257 spin_unlock(&mc.lock);
5258 /* We set mc.moving_task later */
5260 ret = mem_cgroup_precharge_mc(mm);
5262 mem_cgroup_clear_mc();
5269 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5272 mem_cgroup_clear_mc();
5275 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5276 unsigned long addr, unsigned long end,
5277 struct mm_walk *walk)
5280 struct vm_area_struct *vma = walk->vma;
5283 enum mc_target_type target_type;
5284 union mc_target target;
5287 ptl = pmd_trans_huge_lock(pmd, vma);
5289 if (mc.precharge < HPAGE_PMD_NR) {
5293 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5294 if (target_type == MC_TARGET_PAGE) {
5296 if (!isolate_lru_page(page)) {
5297 if (!mem_cgroup_move_account(page, true,
5299 mc.precharge -= HPAGE_PMD_NR;
5300 mc.moved_charge += HPAGE_PMD_NR;
5302 putback_lru_page(page);
5305 } else if (target_type == MC_TARGET_DEVICE) {
5307 if (!mem_cgroup_move_account(page, true,
5309 mc.precharge -= HPAGE_PMD_NR;
5310 mc.moved_charge += HPAGE_PMD_NR;
5318 if (pmd_trans_unstable(pmd))
5321 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5322 for (; addr != end; addr += PAGE_SIZE) {
5323 pte_t ptent = *(pte++);
5324 bool device = false;
5330 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5331 case MC_TARGET_DEVICE:
5334 case MC_TARGET_PAGE:
5337 * We can have a part of the split pmd here. Moving it
5338 * can be done but it would be too convoluted so simply
5339 * ignore such a partial THP and keep it in original
5340 * memcg. There should be somebody mapping the head.
5342 if (PageTransCompound(page))
5344 if (!device && isolate_lru_page(page))
5346 if (!mem_cgroup_move_account(page, false,
5349 /* we uncharge from mc.from later. */
5353 putback_lru_page(page);
5354 put: /* get_mctgt_type() gets the page */
5357 case MC_TARGET_SWAP:
5359 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5361 mem_cgroup_id_get_many(mc.to, 1);
5362 /* we fixup other refcnts and charges later. */
5370 pte_unmap_unlock(pte - 1, ptl);
5375 * We have consumed all precharges we got in can_attach().
5376 * We try charge one by one, but don't do any additional
5377 * charges to mc.to if we have failed in charge once in attach()
5380 ret = mem_cgroup_do_precharge(1);
5388 static void mem_cgroup_move_charge(void)
5390 struct mm_walk mem_cgroup_move_charge_walk = {
5391 .pmd_entry = mem_cgroup_move_charge_pte_range,
5395 lru_add_drain_all();
5397 * Signal lock_page_memcg() to take the memcg's move_lock
5398 * while we're moving its pages to another memcg. Then wait
5399 * for already started RCU-only updates to finish.
5401 atomic_inc(&mc.from->moving_account);
5404 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5406 * Someone who are holding the mmap_sem might be waiting in
5407 * waitq. So we cancel all extra charges, wake up all waiters,
5408 * and retry. Because we cancel precharges, we might not be able
5409 * to move enough charges, but moving charge is a best-effort
5410 * feature anyway, so it wouldn't be a big problem.
5412 __mem_cgroup_clear_mc();
5417 * When we have consumed all precharges and failed in doing
5418 * additional charge, the page walk just aborts.
5420 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5422 up_read(&mc.mm->mmap_sem);
5423 atomic_dec(&mc.from->moving_account);
5426 static void mem_cgroup_move_task(void)
5429 mem_cgroup_move_charge();
5430 mem_cgroup_clear_mc();
5433 #else /* !CONFIG_MMU */
5434 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5438 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5441 static void mem_cgroup_move_task(void)
5447 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5448 * to verify whether we're attached to the default hierarchy on each mount
5451 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5454 * use_hierarchy is forced on the default hierarchy. cgroup core
5455 * guarantees that @root doesn't have any children, so turning it
5456 * on for the root memcg is enough.
5458 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5459 root_mem_cgroup->use_hierarchy = true;
5461 root_mem_cgroup->use_hierarchy = false;
5464 static u64 memory_current_read(struct cgroup_subsys_state *css,
5467 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5469 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5472 static int memory_min_show(struct seq_file *m, void *v)
5474 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5475 unsigned long min = READ_ONCE(memcg->memory.min);
5477 if (min == PAGE_COUNTER_MAX)
5478 seq_puts(m, "max\n");
5480 seq_printf(m, "%llu\n", (u64)min * PAGE_SIZE);
5485 static ssize_t memory_min_write(struct kernfs_open_file *of,
5486 char *buf, size_t nbytes, loff_t off)
5488 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5492 buf = strstrip(buf);
5493 err = page_counter_memparse(buf, "max", &min);
5497 page_counter_set_min(&memcg->memory, min);
5502 static int memory_low_show(struct seq_file *m, void *v)
5504 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5505 unsigned long low = READ_ONCE(memcg->memory.low);
5507 if (low == PAGE_COUNTER_MAX)
5508 seq_puts(m, "max\n");
5510 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5515 static ssize_t memory_low_write(struct kernfs_open_file *of,
5516 char *buf, size_t nbytes, loff_t off)
5518 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5522 buf = strstrip(buf);
5523 err = page_counter_memparse(buf, "max", &low);
5527 page_counter_set_low(&memcg->memory, low);
5532 static int memory_high_show(struct seq_file *m, void *v)
5534 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5535 unsigned long high = READ_ONCE(memcg->high);
5537 if (high == PAGE_COUNTER_MAX)
5538 seq_puts(m, "max\n");
5540 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5545 static ssize_t memory_high_write(struct kernfs_open_file *of,
5546 char *buf, size_t nbytes, loff_t off)
5548 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5549 unsigned long nr_pages;
5553 buf = strstrip(buf);
5554 err = page_counter_memparse(buf, "max", &high);
5560 nr_pages = page_counter_read(&memcg->memory);
5561 if (nr_pages > high)
5562 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5565 memcg_wb_domain_size_changed(memcg);
5569 static int memory_max_show(struct seq_file *m, void *v)
5571 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5572 unsigned long max = READ_ONCE(memcg->memory.max);
5574 if (max == PAGE_COUNTER_MAX)
5575 seq_puts(m, "max\n");
5577 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5582 static ssize_t memory_max_write(struct kernfs_open_file *of,
5583 char *buf, size_t nbytes, loff_t off)
5585 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5586 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5587 bool drained = false;
5591 buf = strstrip(buf);
5592 err = page_counter_memparse(buf, "max", &max);
5596 xchg(&memcg->memory.max, max);
5599 unsigned long nr_pages = page_counter_read(&memcg->memory);
5601 if (nr_pages <= max)
5604 if (signal_pending(current)) {
5610 drain_all_stock(memcg);
5616 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5622 memcg_memory_event(memcg, MEMCG_OOM);
5623 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5627 memcg_wb_domain_size_changed(memcg);
5631 static int memory_events_show(struct seq_file *m, void *v)
5633 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5635 seq_printf(m, "low %lu\n",
5636 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5637 seq_printf(m, "high %lu\n",
5638 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5639 seq_printf(m, "max %lu\n",
5640 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5641 seq_printf(m, "oom %lu\n",
5642 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5643 seq_printf(m, "oom_kill %lu\n",
5644 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5649 static int memory_stat_show(struct seq_file *m, void *v)
5651 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5652 struct accumulated_stats acc;
5656 * Provide statistics on the state of the memory subsystem as
5657 * well as cumulative event counters that show past behavior.
5659 * This list is ordered following a combination of these gradients:
5660 * 1) generic big picture -> specifics and details
5661 * 2) reflecting userspace activity -> reflecting kernel heuristics
5663 * Current memory state:
5666 memset(&acc, 0, sizeof(acc));
5667 acc.stats_size = MEMCG_NR_STAT;
5668 acc.events_size = NR_VM_EVENT_ITEMS;
5669 accumulate_memcg_tree(memcg, &acc);
5671 seq_printf(m, "anon %llu\n",
5672 (u64)acc.stat[MEMCG_RSS] * PAGE_SIZE);
5673 seq_printf(m, "file %llu\n",
5674 (u64)acc.stat[MEMCG_CACHE] * PAGE_SIZE);
5675 seq_printf(m, "kernel_stack %llu\n",
5676 (u64)acc.stat[MEMCG_KERNEL_STACK_KB] * 1024);
5677 seq_printf(m, "slab %llu\n",
5678 (u64)(acc.stat[NR_SLAB_RECLAIMABLE] +
5679 acc.stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5680 seq_printf(m, "sock %llu\n",
5681 (u64)acc.stat[MEMCG_SOCK] * PAGE_SIZE);
5683 seq_printf(m, "shmem %llu\n",
5684 (u64)acc.stat[NR_SHMEM] * PAGE_SIZE);
5685 seq_printf(m, "file_mapped %llu\n",
5686 (u64)acc.stat[NR_FILE_MAPPED] * PAGE_SIZE);
5687 seq_printf(m, "file_dirty %llu\n",
5688 (u64)acc.stat[NR_FILE_DIRTY] * PAGE_SIZE);
5689 seq_printf(m, "file_writeback %llu\n",
5690 (u64)acc.stat[NR_WRITEBACK] * PAGE_SIZE);
5692 for (i = 0; i < NR_LRU_LISTS; i++)
5693 seq_printf(m, "%s %llu\n", mem_cgroup_lru_names[i],
5694 (u64)acc.lru_pages[i] * PAGE_SIZE);
5696 seq_printf(m, "slab_reclaimable %llu\n",
5697 (u64)acc.stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5698 seq_printf(m, "slab_unreclaimable %llu\n",
5699 (u64)acc.stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5701 /* Accumulated memory events */
5703 seq_printf(m, "pgfault %lu\n", acc.events[PGFAULT]);
5704 seq_printf(m, "pgmajfault %lu\n", acc.events[PGMAJFAULT]);
5706 seq_printf(m, "pgrefill %lu\n", acc.events[PGREFILL]);
5707 seq_printf(m, "pgscan %lu\n", acc.events[PGSCAN_KSWAPD] +
5708 acc.events[PGSCAN_DIRECT]);
5709 seq_printf(m, "pgsteal %lu\n", acc.events[PGSTEAL_KSWAPD] +
5710 acc.events[PGSTEAL_DIRECT]);
5711 seq_printf(m, "pgactivate %lu\n", acc.events[PGACTIVATE]);
5712 seq_printf(m, "pgdeactivate %lu\n", acc.events[PGDEACTIVATE]);
5713 seq_printf(m, "pglazyfree %lu\n", acc.events[PGLAZYFREE]);
5714 seq_printf(m, "pglazyfreed %lu\n", acc.events[PGLAZYFREED]);
5716 seq_printf(m, "workingset_refault %lu\n",
5717 acc.stat[WORKINGSET_REFAULT]);
5718 seq_printf(m, "workingset_activate %lu\n",
5719 acc.stat[WORKINGSET_ACTIVATE]);
5720 seq_printf(m, "workingset_nodereclaim %lu\n",
5721 acc.stat[WORKINGSET_NODERECLAIM]);
5726 static int memory_oom_group_show(struct seq_file *m, void *v)
5728 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5730 seq_printf(m, "%d\n", memcg->oom_group);
5735 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5736 char *buf, size_t nbytes, loff_t off)
5738 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5741 buf = strstrip(buf);
5745 ret = kstrtoint(buf, 0, &oom_group);
5749 if (oom_group != 0 && oom_group != 1)
5752 memcg->oom_group = oom_group;
5757 static struct cftype memory_files[] = {
5760 .flags = CFTYPE_NOT_ON_ROOT,
5761 .read_u64 = memory_current_read,
5765 .flags = CFTYPE_NOT_ON_ROOT,
5766 .seq_show = memory_min_show,
5767 .write = memory_min_write,
5771 .flags = CFTYPE_NOT_ON_ROOT,
5772 .seq_show = memory_low_show,
5773 .write = memory_low_write,
5777 .flags = CFTYPE_NOT_ON_ROOT,
5778 .seq_show = memory_high_show,
5779 .write = memory_high_write,
5783 .flags = CFTYPE_NOT_ON_ROOT,
5784 .seq_show = memory_max_show,
5785 .write = memory_max_write,
5789 .flags = CFTYPE_NOT_ON_ROOT,
5790 .file_offset = offsetof(struct mem_cgroup, events_file),
5791 .seq_show = memory_events_show,
5795 .flags = CFTYPE_NOT_ON_ROOT,
5796 .seq_show = memory_stat_show,
5799 .name = "oom.group",
5800 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5801 .seq_show = memory_oom_group_show,
5802 .write = memory_oom_group_write,
5807 struct cgroup_subsys memory_cgrp_subsys = {
5808 .css_alloc = mem_cgroup_css_alloc,
5809 .css_online = mem_cgroup_css_online,
5810 .css_offline = mem_cgroup_css_offline,
5811 .css_released = mem_cgroup_css_released,
5812 .css_free = mem_cgroup_css_free,
5813 .css_reset = mem_cgroup_css_reset,
5814 .can_attach = mem_cgroup_can_attach,
5815 .cancel_attach = mem_cgroup_cancel_attach,
5816 .post_attach = mem_cgroup_move_task,
5817 .bind = mem_cgroup_bind,
5818 .dfl_cftypes = memory_files,
5819 .legacy_cftypes = mem_cgroup_legacy_files,
5824 * mem_cgroup_protected - check if memory consumption is in the normal range
5825 * @root: the top ancestor of the sub-tree being checked
5826 * @memcg: the memory cgroup to check
5828 * WARNING: This function is not stateless! It can only be used as part
5829 * of a top-down tree iteration, not for isolated queries.
5831 * Returns one of the following:
5832 * MEMCG_PROT_NONE: cgroup memory is not protected
5833 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5834 * an unprotected supply of reclaimable memory from other cgroups.
5835 * MEMCG_PROT_MIN: cgroup memory is protected
5837 * @root is exclusive; it is never protected when looked at directly
5839 * To provide a proper hierarchical behavior, effective memory.min/low values
5840 * are used. Below is the description of how effective memory.low is calculated.
5841 * Effective memory.min values is calculated in the same way.
5843 * Effective memory.low is always equal or less than the original memory.low.
5844 * If there is no memory.low overcommittment (which is always true for
5845 * top-level memory cgroups), these two values are equal.
5846 * Otherwise, it's a part of parent's effective memory.low,
5847 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5848 * memory.low usages, where memory.low usage is the size of actually
5852 * elow = min( memory.low, parent->elow * ------------------ ),
5853 * siblings_low_usage
5855 * | memory.current, if memory.current < memory.low
5860 * Such definition of the effective memory.low provides the expected
5861 * hierarchical behavior: parent's memory.low value is limiting
5862 * children, unprotected memory is reclaimed first and cgroups,
5863 * which are not using their guarantee do not affect actual memory
5866 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5868 * A A/memory.low = 2G, A/memory.current = 6G
5870 * BC DE B/memory.low = 3G B/memory.current = 2G
5871 * C/memory.low = 1G C/memory.current = 2G
5872 * D/memory.low = 0 D/memory.current = 2G
5873 * E/memory.low = 10G E/memory.current = 0
5875 * and the memory pressure is applied, the following memory distribution
5876 * is expected (approximately):
5878 * A/memory.current = 2G
5880 * B/memory.current = 1.3G
5881 * C/memory.current = 0.6G
5882 * D/memory.current = 0
5883 * E/memory.current = 0
5885 * These calculations require constant tracking of the actual low usages
5886 * (see propagate_protected_usage()), as well as recursive calculation of
5887 * effective memory.low values. But as we do call mem_cgroup_protected()
5888 * path for each memory cgroup top-down from the reclaim,
5889 * it's possible to optimize this part, and save calculated elow
5890 * for next usage. This part is intentionally racy, but it's ok,
5891 * as memory.low is a best-effort mechanism.
5893 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5894 struct mem_cgroup *memcg)
5896 struct mem_cgroup *parent;
5897 unsigned long emin, parent_emin;
5898 unsigned long elow, parent_elow;
5899 unsigned long usage;
5901 if (mem_cgroup_disabled())
5902 return MEMCG_PROT_NONE;
5905 root = root_mem_cgroup;
5907 return MEMCG_PROT_NONE;
5909 usage = page_counter_read(&memcg->memory);
5911 return MEMCG_PROT_NONE;
5913 emin = memcg->memory.min;
5914 elow = memcg->memory.low;
5916 parent = parent_mem_cgroup(memcg);
5917 /* No parent means a non-hierarchical mode on v1 memcg */
5919 return MEMCG_PROT_NONE;
5924 parent_emin = READ_ONCE(parent->memory.emin);
5925 emin = min(emin, parent_emin);
5926 if (emin && parent_emin) {
5927 unsigned long min_usage, siblings_min_usage;
5929 min_usage = min(usage, memcg->memory.min);
5930 siblings_min_usage = atomic_long_read(
5931 &parent->memory.children_min_usage);
5933 if (min_usage && siblings_min_usage)
5934 emin = min(emin, parent_emin * min_usage /
5935 siblings_min_usage);
5938 parent_elow = READ_ONCE(parent->memory.elow);
5939 elow = min(elow, parent_elow);
5940 if (elow && parent_elow) {
5941 unsigned long low_usage, siblings_low_usage;
5943 low_usage = min(usage, memcg->memory.low);
5944 siblings_low_usage = atomic_long_read(
5945 &parent->memory.children_low_usage);
5947 if (low_usage && siblings_low_usage)
5948 elow = min(elow, parent_elow * low_usage /
5949 siblings_low_usage);
5953 memcg->memory.emin = emin;
5954 memcg->memory.elow = elow;
5957 return MEMCG_PROT_MIN;
5958 else if (usage <= elow)
5959 return MEMCG_PROT_LOW;
5961 return MEMCG_PROT_NONE;
5965 * mem_cgroup_try_charge - try charging a page
5966 * @page: page to charge
5967 * @mm: mm context of the victim
5968 * @gfp_mask: reclaim mode
5969 * @memcgp: charged memcg return
5970 * @compound: charge the page as compound or small page
5972 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5973 * pages according to @gfp_mask if necessary.
5975 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5976 * Otherwise, an error code is returned.
5978 * After page->mapping has been set up, the caller must finalize the
5979 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5980 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5982 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5983 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5986 struct mem_cgroup *memcg = NULL;
5987 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5990 if (mem_cgroup_disabled())
5993 if (PageSwapCache(page)) {
5995 * Every swap fault against a single page tries to charge the
5996 * page, bail as early as possible. shmem_unuse() encounters
5997 * already charged pages, too. The USED bit is protected by
5998 * the page lock, which serializes swap cache removal, which
5999 * in turn serializes uncharging.
6001 VM_BUG_ON_PAGE(!PageLocked(page), page);
6002 if (compound_head(page)->mem_cgroup)
6005 if (do_swap_account) {
6006 swp_entry_t ent = { .val = page_private(page), };
6007 unsigned short id = lookup_swap_cgroup_id(ent);
6010 memcg = mem_cgroup_from_id(id);
6011 if (memcg && !css_tryget_online(&memcg->css))
6018 memcg = get_mem_cgroup_from_mm(mm);
6020 ret = try_charge(memcg, gfp_mask, nr_pages);
6022 css_put(&memcg->css);
6028 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6029 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6032 struct mem_cgroup *memcg;
6035 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6037 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6042 * mem_cgroup_commit_charge - commit a page charge
6043 * @page: page to charge
6044 * @memcg: memcg to charge the page to
6045 * @lrucare: page might be on LRU already
6046 * @compound: charge the page as compound or small page
6048 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6049 * after page->mapping has been set up. This must happen atomically
6050 * as part of the page instantiation, i.e. under the page table lock
6051 * for anonymous pages, under the page lock for page and swap cache.
6053 * In addition, the page must not be on the LRU during the commit, to
6054 * prevent racing with task migration. If it might be, use @lrucare.
6056 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6058 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6059 bool lrucare, bool compound)
6061 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6063 VM_BUG_ON_PAGE(!page->mapping, page);
6064 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6066 if (mem_cgroup_disabled())
6069 * Swap faults will attempt to charge the same page multiple
6070 * times. But reuse_swap_page() might have removed the page
6071 * from swapcache already, so we can't check PageSwapCache().
6076 commit_charge(page, memcg, lrucare);
6078 local_irq_disable();
6079 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6080 memcg_check_events(memcg, page);
6083 if (do_memsw_account() && PageSwapCache(page)) {
6084 swp_entry_t entry = { .val = page_private(page) };
6086 * The swap entry might not get freed for a long time,
6087 * let's not wait for it. The page already received a
6088 * memory+swap charge, drop the swap entry duplicate.
6090 mem_cgroup_uncharge_swap(entry, nr_pages);
6095 * mem_cgroup_cancel_charge - cancel a page charge
6096 * @page: page to charge
6097 * @memcg: memcg to charge the page to
6098 * @compound: charge the page as compound or small page
6100 * Cancel a charge transaction started by mem_cgroup_try_charge().
6102 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6105 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6107 if (mem_cgroup_disabled())
6110 * Swap faults will attempt to charge the same page multiple
6111 * times. But reuse_swap_page() might have removed the page
6112 * from swapcache already, so we can't check PageSwapCache().
6117 cancel_charge(memcg, nr_pages);
6120 struct uncharge_gather {
6121 struct mem_cgroup *memcg;
6122 unsigned long pgpgout;
6123 unsigned long nr_anon;
6124 unsigned long nr_file;
6125 unsigned long nr_kmem;
6126 unsigned long nr_huge;
6127 unsigned long nr_shmem;
6128 struct page *dummy_page;
6131 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6133 memset(ug, 0, sizeof(*ug));
6136 static void uncharge_batch(const struct uncharge_gather *ug)
6138 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6139 unsigned long flags;
6141 if (!mem_cgroup_is_root(ug->memcg)) {
6142 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6143 if (do_memsw_account())
6144 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6145 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6146 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6147 memcg_oom_recover(ug->memcg);
6150 local_irq_save(flags);
6151 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6152 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6153 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6154 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6155 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6156 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
6157 memcg_check_events(ug->memcg, ug->dummy_page);
6158 local_irq_restore(flags);
6160 if (!mem_cgroup_is_root(ug->memcg))
6161 css_put_many(&ug->memcg->css, nr_pages);
6164 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6166 VM_BUG_ON_PAGE(PageLRU(page), page);
6167 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6168 !PageHWPoison(page) , page);
6170 if (!page->mem_cgroup)
6174 * Nobody should be changing or seriously looking at
6175 * page->mem_cgroup at this point, we have fully
6176 * exclusive access to the page.
6179 if (ug->memcg != page->mem_cgroup) {
6182 uncharge_gather_clear(ug);
6184 ug->memcg = page->mem_cgroup;
6187 if (!PageKmemcg(page)) {
6188 unsigned int nr_pages = 1;
6190 if (PageTransHuge(page)) {
6191 nr_pages <<= compound_order(page);
6192 ug->nr_huge += nr_pages;
6195 ug->nr_anon += nr_pages;
6197 ug->nr_file += nr_pages;
6198 if (PageSwapBacked(page))
6199 ug->nr_shmem += nr_pages;
6203 ug->nr_kmem += 1 << compound_order(page);
6204 __ClearPageKmemcg(page);
6207 ug->dummy_page = page;
6208 page->mem_cgroup = NULL;
6211 static void uncharge_list(struct list_head *page_list)
6213 struct uncharge_gather ug;
6214 struct list_head *next;
6216 uncharge_gather_clear(&ug);
6219 * Note that the list can be a single page->lru; hence the
6220 * do-while loop instead of a simple list_for_each_entry().
6222 next = page_list->next;
6226 page = list_entry(next, struct page, lru);
6227 next = page->lru.next;
6229 uncharge_page(page, &ug);
6230 } while (next != page_list);
6233 uncharge_batch(&ug);
6237 * mem_cgroup_uncharge - uncharge a page
6238 * @page: page to uncharge
6240 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6241 * mem_cgroup_commit_charge().
6243 void mem_cgroup_uncharge(struct page *page)
6245 struct uncharge_gather ug;
6247 if (mem_cgroup_disabled())
6250 /* Don't touch page->lru of any random page, pre-check: */
6251 if (!page->mem_cgroup)
6254 uncharge_gather_clear(&ug);
6255 uncharge_page(page, &ug);
6256 uncharge_batch(&ug);
6260 * mem_cgroup_uncharge_list - uncharge a list of page
6261 * @page_list: list of pages to uncharge
6263 * Uncharge a list of pages previously charged with
6264 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6266 void mem_cgroup_uncharge_list(struct list_head *page_list)
6268 if (mem_cgroup_disabled())
6271 if (!list_empty(page_list))
6272 uncharge_list(page_list);
6276 * mem_cgroup_migrate - charge a page's replacement
6277 * @oldpage: currently circulating page
6278 * @newpage: replacement page
6280 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6281 * be uncharged upon free.
6283 * Both pages must be locked, @newpage->mapping must be set up.
6285 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6287 struct mem_cgroup *memcg;
6288 unsigned int nr_pages;
6290 unsigned long flags;
6292 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6293 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6294 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6295 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6298 if (mem_cgroup_disabled())
6301 /* Page cache replacement: new page already charged? */
6302 if (newpage->mem_cgroup)
6305 /* Swapcache readahead pages can get replaced before being charged */
6306 memcg = oldpage->mem_cgroup;
6310 /* Force-charge the new page. The old one will be freed soon */
6311 compound = PageTransHuge(newpage);
6312 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6314 page_counter_charge(&memcg->memory, nr_pages);
6315 if (do_memsw_account())
6316 page_counter_charge(&memcg->memsw, nr_pages);
6317 css_get_many(&memcg->css, nr_pages);
6319 commit_charge(newpage, memcg, false);
6321 local_irq_save(flags);
6322 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6323 memcg_check_events(memcg, newpage);
6324 local_irq_restore(flags);
6327 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6328 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6330 void mem_cgroup_sk_alloc(struct sock *sk)
6332 struct mem_cgroup *memcg;
6334 if (!mem_cgroup_sockets_enabled)
6337 /* Do not associate the sock with unrelated interrupted task's memcg. */
6342 memcg = mem_cgroup_from_task(current);
6343 if (memcg == root_mem_cgroup)
6345 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6347 if (css_tryget_online(&memcg->css))
6348 sk->sk_memcg = memcg;
6353 void mem_cgroup_sk_free(struct sock *sk)
6356 css_put(&sk->sk_memcg->css);
6360 * mem_cgroup_charge_skmem - charge socket memory
6361 * @memcg: memcg to charge
6362 * @nr_pages: number of pages to charge
6364 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6365 * @memcg's configured limit, %false if the charge had to be forced.
6367 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6369 gfp_t gfp_mask = GFP_KERNEL;
6371 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6372 struct page_counter *fail;
6374 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6375 memcg->tcpmem_pressure = 0;
6378 page_counter_charge(&memcg->tcpmem, nr_pages);
6379 memcg->tcpmem_pressure = 1;
6383 /* Don't block in the packet receive path */
6385 gfp_mask = GFP_NOWAIT;
6387 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6389 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6392 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6397 * mem_cgroup_uncharge_skmem - uncharge socket memory
6398 * @memcg: memcg to uncharge
6399 * @nr_pages: number of pages to uncharge
6401 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6403 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6404 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6408 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6410 refill_stock(memcg, nr_pages);
6413 static int __init cgroup_memory(char *s)
6417 while ((token = strsep(&s, ",")) != NULL) {
6420 if (!strcmp(token, "nosocket"))
6421 cgroup_memory_nosocket = true;
6422 if (!strcmp(token, "nokmem"))
6423 cgroup_memory_nokmem = true;
6427 __setup("cgroup.memory=", cgroup_memory);
6430 * subsys_initcall() for memory controller.
6432 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6433 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6434 * basically everything that doesn't depend on a specific mem_cgroup structure
6435 * should be initialized from here.
6437 static int __init mem_cgroup_init(void)
6441 #ifdef CONFIG_MEMCG_KMEM
6443 * Kmem cache creation is mostly done with the slab_mutex held,
6444 * so use a workqueue with limited concurrency to avoid stalling
6445 * all worker threads in case lots of cgroups are created and
6446 * destroyed simultaneously.
6448 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6449 BUG_ON(!memcg_kmem_cache_wq);
6452 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6453 memcg_hotplug_cpu_dead);
6455 for_each_possible_cpu(cpu)
6456 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6459 for_each_node(node) {
6460 struct mem_cgroup_tree_per_node *rtpn;
6462 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6463 node_online(node) ? node : NUMA_NO_NODE);
6465 rtpn->rb_root = RB_ROOT;
6466 rtpn->rb_rightmost = NULL;
6467 spin_lock_init(&rtpn->lock);
6468 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6473 subsys_initcall(mem_cgroup_init);
6475 #ifdef CONFIG_MEMCG_SWAP
6476 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6478 while (!atomic_inc_not_zero(&memcg->id.ref)) {
6480 * The root cgroup cannot be destroyed, so it's refcount must
6483 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6487 memcg = parent_mem_cgroup(memcg);
6489 memcg = root_mem_cgroup;
6495 * mem_cgroup_swapout - transfer a memsw charge to swap
6496 * @page: page whose memsw charge to transfer
6497 * @entry: swap entry to move the charge to
6499 * Transfer the memsw charge of @page to @entry.
6501 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6503 struct mem_cgroup *memcg, *swap_memcg;
6504 unsigned int nr_entries;
6505 unsigned short oldid;
6507 VM_BUG_ON_PAGE(PageLRU(page), page);
6508 VM_BUG_ON_PAGE(page_count(page), page);
6510 if (!do_memsw_account())
6513 memcg = page->mem_cgroup;
6515 /* Readahead page, never charged */
6520 * In case the memcg owning these pages has been offlined and doesn't
6521 * have an ID allocated to it anymore, charge the closest online
6522 * ancestor for the swap instead and transfer the memory+swap charge.
6524 swap_memcg = mem_cgroup_id_get_online(memcg);
6525 nr_entries = hpage_nr_pages(page);
6526 /* Get references for the tail pages, too */
6528 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6529 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6531 VM_BUG_ON_PAGE(oldid, page);
6532 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6534 page->mem_cgroup = NULL;
6536 if (!mem_cgroup_is_root(memcg))
6537 page_counter_uncharge(&memcg->memory, nr_entries);
6539 if (memcg != swap_memcg) {
6540 if (!mem_cgroup_is_root(swap_memcg))
6541 page_counter_charge(&swap_memcg->memsw, nr_entries);
6542 page_counter_uncharge(&memcg->memsw, nr_entries);
6546 * Interrupts should be disabled here because the caller holds the
6547 * i_pages lock which is taken with interrupts-off. It is
6548 * important here to have the interrupts disabled because it is the
6549 * only synchronisation we have for updating the per-CPU variables.
6551 VM_BUG_ON(!irqs_disabled());
6552 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6554 memcg_check_events(memcg, page);
6556 if (!mem_cgroup_is_root(memcg))
6557 css_put_many(&memcg->css, nr_entries);
6561 * mem_cgroup_try_charge_swap - try charging swap space for a page
6562 * @page: page being added to swap
6563 * @entry: swap entry to charge
6565 * Try to charge @page's memcg for the swap space at @entry.
6567 * Returns 0 on success, -ENOMEM on failure.
6569 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6571 unsigned int nr_pages = hpage_nr_pages(page);
6572 struct page_counter *counter;
6573 struct mem_cgroup *memcg;
6574 unsigned short oldid;
6576 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6579 memcg = page->mem_cgroup;
6581 /* Readahead page, never charged */
6586 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6590 memcg = mem_cgroup_id_get_online(memcg);
6592 if (!mem_cgroup_is_root(memcg) &&
6593 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6594 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6595 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6596 mem_cgroup_id_put(memcg);
6600 /* Get references for the tail pages, too */
6602 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6603 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6604 VM_BUG_ON_PAGE(oldid, page);
6605 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6611 * mem_cgroup_uncharge_swap - uncharge swap space
6612 * @entry: swap entry to uncharge
6613 * @nr_pages: the amount of swap space to uncharge
6615 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6617 struct mem_cgroup *memcg;
6620 if (!do_swap_account)
6623 id = swap_cgroup_record(entry, 0, nr_pages);
6625 memcg = mem_cgroup_from_id(id);
6627 if (!mem_cgroup_is_root(memcg)) {
6628 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6629 page_counter_uncharge(&memcg->swap, nr_pages);
6631 page_counter_uncharge(&memcg->memsw, nr_pages);
6633 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6634 mem_cgroup_id_put_many(memcg, nr_pages);
6639 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6641 long nr_swap_pages = get_nr_swap_pages();
6643 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6644 return nr_swap_pages;
6645 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6646 nr_swap_pages = min_t(long, nr_swap_pages,
6647 READ_ONCE(memcg->swap.max) -
6648 page_counter_read(&memcg->swap));
6649 return nr_swap_pages;
6652 bool mem_cgroup_swap_full(struct page *page)
6654 struct mem_cgroup *memcg;
6656 VM_BUG_ON_PAGE(!PageLocked(page), page);
6660 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6663 memcg = page->mem_cgroup;
6667 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6668 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6674 /* for remember boot option*/
6675 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6676 static int really_do_swap_account __initdata = 1;
6678 static int really_do_swap_account __initdata;
6681 static int __init enable_swap_account(char *s)
6683 if (!strcmp(s, "1"))
6684 really_do_swap_account = 1;
6685 else if (!strcmp(s, "0"))
6686 really_do_swap_account = 0;
6689 __setup("swapaccount=", enable_swap_account);
6691 static u64 swap_current_read(struct cgroup_subsys_state *css,
6694 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6696 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6699 static int swap_max_show(struct seq_file *m, void *v)
6701 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6702 unsigned long max = READ_ONCE(memcg->swap.max);
6704 if (max == PAGE_COUNTER_MAX)
6705 seq_puts(m, "max\n");
6707 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6712 static ssize_t swap_max_write(struct kernfs_open_file *of,
6713 char *buf, size_t nbytes, loff_t off)
6715 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6719 buf = strstrip(buf);
6720 err = page_counter_memparse(buf, "max", &max);
6724 xchg(&memcg->swap.max, max);
6729 static int swap_events_show(struct seq_file *m, void *v)
6731 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6733 seq_printf(m, "max %lu\n",
6734 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6735 seq_printf(m, "fail %lu\n",
6736 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6741 static struct cftype swap_files[] = {
6743 .name = "swap.current",
6744 .flags = CFTYPE_NOT_ON_ROOT,
6745 .read_u64 = swap_current_read,
6749 .flags = CFTYPE_NOT_ON_ROOT,
6750 .seq_show = swap_max_show,
6751 .write = swap_max_write,
6754 .name = "swap.events",
6755 .flags = CFTYPE_NOT_ON_ROOT,
6756 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6757 .seq_show = swap_events_show,
6762 static struct cftype memsw_cgroup_files[] = {
6764 .name = "memsw.usage_in_bytes",
6765 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6766 .read_u64 = mem_cgroup_read_u64,
6769 .name = "memsw.max_usage_in_bytes",
6770 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6771 .write = mem_cgroup_reset,
6772 .read_u64 = mem_cgroup_read_u64,
6775 .name = "memsw.limit_in_bytes",
6776 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6777 .write = mem_cgroup_write,
6778 .read_u64 = mem_cgroup_read_u64,
6781 .name = "memsw.failcnt",
6782 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6783 .write = mem_cgroup_reset,
6784 .read_u64 = mem_cgroup_read_u64,
6786 { }, /* terminate */
6789 static int __init mem_cgroup_swap_init(void)
6791 if (!mem_cgroup_disabled() && really_do_swap_account) {
6792 do_swap_account = 1;
6793 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6795 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6796 memsw_cgroup_files));
6800 subsys_initcall(mem_cgroup_swap_init);
6802 #endif /* CONFIG_MEMCG_SWAP */