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);
427 memcg_shrinker_map_size = size;
428 mutex_unlock(&memcg_shrinker_map_mutex);
432 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
434 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
435 struct memcg_shrinker_map *map;
438 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
439 /* Pairs with smp mb in shrink_slab() */
440 smp_mb__before_atomic();
441 set_bit(shrinker_id, map->map);
446 #else /* CONFIG_MEMCG_KMEM */
447 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
451 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
452 #endif /* CONFIG_MEMCG_KMEM */
455 * mem_cgroup_css_from_page - css of the memcg associated with a page
456 * @page: page of interest
458 * If memcg is bound to the default hierarchy, css of the memcg associated
459 * with @page is returned. The returned css remains associated with @page
460 * until it is released.
462 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
465 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
467 struct mem_cgroup *memcg;
469 memcg = page->mem_cgroup;
471 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
472 memcg = root_mem_cgroup;
478 * page_cgroup_ino - return inode number of the memcg a page is charged to
481 * Look up the closest online ancestor of the memory cgroup @page is charged to
482 * and return its inode number or 0 if @page is not charged to any cgroup. It
483 * is safe to call this function without holding a reference to @page.
485 * Note, this function is inherently racy, because there is nothing to prevent
486 * the cgroup inode from getting torn down and potentially reallocated a moment
487 * after page_cgroup_ino() returns, so it only should be used by callers that
488 * do not care (such as procfs interfaces).
490 ino_t page_cgroup_ino(struct page *page)
492 struct mem_cgroup *memcg;
493 unsigned long ino = 0;
496 memcg = READ_ONCE(page->mem_cgroup);
497 while (memcg && !(memcg->css.flags & CSS_ONLINE))
498 memcg = parent_mem_cgroup(memcg);
500 ino = cgroup_ino(memcg->css.cgroup);
505 static struct mem_cgroup_per_node *
506 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
508 int nid = page_to_nid(page);
510 return memcg->nodeinfo[nid];
513 static struct mem_cgroup_tree_per_node *
514 soft_limit_tree_node(int nid)
516 return soft_limit_tree.rb_tree_per_node[nid];
519 static struct mem_cgroup_tree_per_node *
520 soft_limit_tree_from_page(struct page *page)
522 int nid = page_to_nid(page);
524 return soft_limit_tree.rb_tree_per_node[nid];
527 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
528 struct mem_cgroup_tree_per_node *mctz,
529 unsigned long new_usage_in_excess)
531 struct rb_node **p = &mctz->rb_root.rb_node;
532 struct rb_node *parent = NULL;
533 struct mem_cgroup_per_node *mz_node;
534 bool rightmost = true;
539 mz->usage_in_excess = new_usage_in_excess;
540 if (!mz->usage_in_excess)
544 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
546 if (mz->usage_in_excess < mz_node->usage_in_excess) {
552 * We can't avoid mem cgroups that are over their soft
553 * limit by the same amount
555 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
560 mctz->rb_rightmost = &mz->tree_node;
562 rb_link_node(&mz->tree_node, parent, p);
563 rb_insert_color(&mz->tree_node, &mctz->rb_root);
567 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
568 struct mem_cgroup_tree_per_node *mctz)
573 if (&mz->tree_node == mctz->rb_rightmost)
574 mctz->rb_rightmost = rb_prev(&mz->tree_node);
576 rb_erase(&mz->tree_node, &mctz->rb_root);
580 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
581 struct mem_cgroup_tree_per_node *mctz)
585 spin_lock_irqsave(&mctz->lock, flags);
586 __mem_cgroup_remove_exceeded(mz, mctz);
587 spin_unlock_irqrestore(&mctz->lock, flags);
590 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
592 unsigned long nr_pages = page_counter_read(&memcg->memory);
593 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
594 unsigned long excess = 0;
596 if (nr_pages > soft_limit)
597 excess = nr_pages - soft_limit;
602 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
604 unsigned long excess;
605 struct mem_cgroup_per_node *mz;
606 struct mem_cgroup_tree_per_node *mctz;
608 mctz = soft_limit_tree_from_page(page);
612 * Necessary to update all ancestors when hierarchy is used.
613 * because their event counter is not touched.
615 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
616 mz = mem_cgroup_page_nodeinfo(memcg, page);
617 excess = soft_limit_excess(memcg);
619 * We have to update the tree if mz is on RB-tree or
620 * mem is over its softlimit.
622 if (excess || mz->on_tree) {
625 spin_lock_irqsave(&mctz->lock, flags);
626 /* if on-tree, remove it */
628 __mem_cgroup_remove_exceeded(mz, mctz);
630 * Insert again. mz->usage_in_excess will be updated.
631 * If excess is 0, no tree ops.
633 __mem_cgroup_insert_exceeded(mz, mctz, excess);
634 spin_unlock_irqrestore(&mctz->lock, flags);
639 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
641 struct mem_cgroup_tree_per_node *mctz;
642 struct mem_cgroup_per_node *mz;
646 mz = mem_cgroup_nodeinfo(memcg, nid);
647 mctz = soft_limit_tree_node(nid);
649 mem_cgroup_remove_exceeded(mz, mctz);
653 static struct mem_cgroup_per_node *
654 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
656 struct mem_cgroup_per_node *mz;
660 if (!mctz->rb_rightmost)
661 goto done; /* Nothing to reclaim from */
663 mz = rb_entry(mctz->rb_rightmost,
664 struct mem_cgroup_per_node, tree_node);
666 * Remove the node now but someone else can add it back,
667 * we will to add it back at the end of reclaim to its correct
668 * position in the tree.
670 __mem_cgroup_remove_exceeded(mz, mctz);
671 if (!soft_limit_excess(mz->memcg) ||
672 !css_tryget_online(&mz->memcg->css))
678 static struct mem_cgroup_per_node *
679 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
681 struct mem_cgroup_per_node *mz;
683 spin_lock_irq(&mctz->lock);
684 mz = __mem_cgroup_largest_soft_limit_node(mctz);
685 spin_unlock_irq(&mctz->lock);
689 static unsigned long memcg_sum_events(struct mem_cgroup *memcg,
692 return atomic_long_read(&memcg->events[event]);
695 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
697 bool compound, int nr_pages)
700 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
701 * counted as CACHE even if it's on ANON LRU.
704 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
706 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
707 if (PageSwapBacked(page))
708 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
712 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
713 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
716 /* pagein of a big page is an event. So, ignore page size */
718 __count_memcg_events(memcg, PGPGIN, 1);
720 __count_memcg_events(memcg, PGPGOUT, 1);
721 nr_pages = -nr_pages; /* for event */
724 __this_cpu_add(memcg->stat_cpu->nr_page_events, nr_pages);
727 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
728 int nid, unsigned int lru_mask)
730 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
731 unsigned long nr = 0;
734 VM_BUG_ON((unsigned)nid >= nr_node_ids);
737 if (!(BIT(lru) & lru_mask))
739 nr += mem_cgroup_get_lru_size(lruvec, lru);
744 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
745 unsigned int lru_mask)
747 unsigned long nr = 0;
750 for_each_node_state(nid, N_MEMORY)
751 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
755 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
756 enum mem_cgroup_events_target target)
758 unsigned long val, next;
760 val = __this_cpu_read(memcg->stat_cpu->nr_page_events);
761 next = __this_cpu_read(memcg->stat_cpu->targets[target]);
762 /* from time_after() in jiffies.h */
763 if ((long)(next - val) < 0) {
765 case MEM_CGROUP_TARGET_THRESH:
766 next = val + THRESHOLDS_EVENTS_TARGET;
768 case MEM_CGROUP_TARGET_SOFTLIMIT:
769 next = val + SOFTLIMIT_EVENTS_TARGET;
771 case MEM_CGROUP_TARGET_NUMAINFO:
772 next = val + NUMAINFO_EVENTS_TARGET;
777 __this_cpu_write(memcg->stat_cpu->targets[target], next);
784 * Check events in order.
787 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
789 /* threshold event is triggered in finer grain than soft limit */
790 if (unlikely(mem_cgroup_event_ratelimit(memcg,
791 MEM_CGROUP_TARGET_THRESH))) {
793 bool do_numainfo __maybe_unused;
795 do_softlimit = mem_cgroup_event_ratelimit(memcg,
796 MEM_CGROUP_TARGET_SOFTLIMIT);
798 do_numainfo = mem_cgroup_event_ratelimit(memcg,
799 MEM_CGROUP_TARGET_NUMAINFO);
801 mem_cgroup_threshold(memcg);
802 if (unlikely(do_softlimit))
803 mem_cgroup_update_tree(memcg, page);
805 if (unlikely(do_numainfo))
806 atomic_inc(&memcg->numainfo_events);
811 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
814 * mm_update_next_owner() may clear mm->owner to NULL
815 * if it races with swapoff, page migration, etc.
816 * So this can be called with p == NULL.
821 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
823 EXPORT_SYMBOL(mem_cgroup_from_task);
826 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
827 * @mm: mm from which memcg should be extracted. It can be NULL.
829 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
830 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
833 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
835 struct mem_cgroup *memcg;
837 if (mem_cgroup_disabled())
843 * Page cache insertions can happen withou an
844 * actual mm context, e.g. during disk probing
845 * on boot, loopback IO, acct() writes etc.
848 memcg = root_mem_cgroup;
850 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
851 if (unlikely(!memcg))
852 memcg = root_mem_cgroup;
854 } while (!css_tryget(&memcg->css));
858 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
861 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
862 * @page: page from which memcg should be extracted.
864 * Obtain a reference on page->memcg and returns it if successful. Otherwise
865 * root_mem_cgroup is returned.
867 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
869 struct mem_cgroup *memcg = page->mem_cgroup;
871 if (mem_cgroup_disabled())
875 if (!memcg || !css_tryget_online(&memcg->css))
876 memcg = root_mem_cgroup;
880 EXPORT_SYMBOL(get_mem_cgroup_from_page);
883 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
885 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
887 if (unlikely(current->active_memcg)) {
888 struct mem_cgroup *memcg = root_mem_cgroup;
891 if (css_tryget_online(¤t->active_memcg->css))
892 memcg = current->active_memcg;
896 return get_mem_cgroup_from_mm(current->mm);
900 * mem_cgroup_iter - iterate over memory cgroup hierarchy
901 * @root: hierarchy root
902 * @prev: previously returned memcg, NULL on first invocation
903 * @reclaim: cookie for shared reclaim walks, NULL for full walks
905 * Returns references to children of the hierarchy below @root, or
906 * @root itself, or %NULL after a full round-trip.
908 * Caller must pass the return value in @prev on subsequent
909 * invocations for reference counting, or use mem_cgroup_iter_break()
910 * to cancel a hierarchy walk before the round-trip is complete.
912 * Reclaimers can specify a node and a priority level in @reclaim to
913 * divide up the memcgs in the hierarchy among all concurrent
914 * reclaimers operating on the same node and priority.
916 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
917 struct mem_cgroup *prev,
918 struct mem_cgroup_reclaim_cookie *reclaim)
920 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
921 struct cgroup_subsys_state *css = NULL;
922 struct mem_cgroup *memcg = NULL;
923 struct mem_cgroup *pos = NULL;
925 if (mem_cgroup_disabled())
929 root = root_mem_cgroup;
931 if (prev && !reclaim)
934 if (!root->use_hierarchy && root != root_mem_cgroup) {
943 struct mem_cgroup_per_node *mz;
945 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
946 iter = &mz->iter[reclaim->priority];
948 if (prev && reclaim->generation != iter->generation)
952 pos = READ_ONCE(iter->position);
953 if (!pos || css_tryget(&pos->css))
956 * css reference reached zero, so iter->position will
957 * be cleared by ->css_released. However, we should not
958 * rely on this happening soon, because ->css_released
959 * is called from a work queue, and by busy-waiting we
960 * might block it. So we clear iter->position right
963 (void)cmpxchg(&iter->position, pos, NULL);
971 css = css_next_descendant_pre(css, &root->css);
974 * Reclaimers share the hierarchy walk, and a
975 * new one might jump in right at the end of
976 * the hierarchy - make sure they see at least
977 * one group and restart from the beginning.
985 * Verify the css and acquire a reference. The root
986 * is provided by the caller, so we know it's alive
987 * and kicking, and don't take an extra reference.
989 memcg = mem_cgroup_from_css(css);
991 if (css == &root->css)
1002 * The position could have already been updated by a competing
1003 * thread, so check that the value hasn't changed since we read
1004 * it to avoid reclaiming from the same cgroup twice.
1006 (void)cmpxchg(&iter->position, pos, memcg);
1014 reclaim->generation = iter->generation;
1020 if (prev && prev != root)
1021 css_put(&prev->css);
1027 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1028 * @root: hierarchy root
1029 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1031 void mem_cgroup_iter_break(struct mem_cgroup *root,
1032 struct mem_cgroup *prev)
1035 root = root_mem_cgroup;
1036 if (prev && prev != root)
1037 css_put(&prev->css);
1040 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1041 struct mem_cgroup *dead_memcg)
1043 struct mem_cgroup_reclaim_iter *iter;
1044 struct mem_cgroup_per_node *mz;
1048 for_each_node(nid) {
1049 mz = mem_cgroup_nodeinfo(from, nid);
1050 for (i = 0; i <= DEF_PRIORITY; i++) {
1051 iter = &mz->iter[i];
1052 cmpxchg(&iter->position,
1058 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1060 struct mem_cgroup *memcg = dead_memcg;
1061 struct mem_cgroup *last;
1064 __invalidate_reclaim_iterators(memcg, dead_memcg);
1066 } while ((memcg = parent_mem_cgroup(memcg)));
1069 * When cgruop1 non-hierarchy mode is used,
1070 * parent_mem_cgroup() does not walk all the way up to the
1071 * cgroup root (root_mem_cgroup). So we have to handle
1072 * dead_memcg from cgroup root separately.
1074 if (last != root_mem_cgroup)
1075 __invalidate_reclaim_iterators(root_mem_cgroup,
1080 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1081 * @memcg: hierarchy root
1082 * @fn: function to call for each task
1083 * @arg: argument passed to @fn
1085 * This function iterates over tasks attached to @memcg or to any of its
1086 * descendants and calls @fn for each task. If @fn returns a non-zero
1087 * value, the function breaks the iteration loop and returns the value.
1088 * Otherwise, it will iterate over all tasks and return 0.
1090 * This function must not be called for the root memory cgroup.
1092 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1093 int (*fn)(struct task_struct *, void *), void *arg)
1095 struct mem_cgroup *iter;
1098 BUG_ON(memcg == root_mem_cgroup);
1100 for_each_mem_cgroup_tree(iter, memcg) {
1101 struct css_task_iter it;
1102 struct task_struct *task;
1104 css_task_iter_start(&iter->css, 0, &it);
1105 while (!ret && (task = css_task_iter_next(&it)))
1106 ret = fn(task, arg);
1107 css_task_iter_end(&it);
1109 mem_cgroup_iter_break(memcg, iter);
1117 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1119 * @pgdat: pgdat of the page
1121 * This function is only safe when following the LRU page isolation
1122 * and putback protocol: the LRU lock must be held, and the page must
1123 * either be PageLRU() or the caller must have isolated/allocated it.
1125 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1127 struct mem_cgroup_per_node *mz;
1128 struct mem_cgroup *memcg;
1129 struct lruvec *lruvec;
1131 if (mem_cgroup_disabled()) {
1132 lruvec = &pgdat->lruvec;
1136 memcg = page->mem_cgroup;
1138 * Swapcache readahead pages are added to the LRU - and
1139 * possibly migrated - before they are charged.
1142 memcg = root_mem_cgroup;
1144 mz = mem_cgroup_page_nodeinfo(memcg, page);
1145 lruvec = &mz->lruvec;
1148 * Since a node can be onlined after the mem_cgroup was created,
1149 * we have to be prepared to initialize lruvec->zone here;
1150 * and if offlined then reonlined, we need to reinitialize it.
1152 if (unlikely(lruvec->pgdat != pgdat))
1153 lruvec->pgdat = pgdat;
1158 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1159 * @lruvec: mem_cgroup per zone lru vector
1160 * @lru: index of lru list the page is sitting on
1161 * @zid: zone id of the accounted pages
1162 * @nr_pages: positive when adding or negative when removing
1164 * This function must be called under lru_lock, just before a page is added
1165 * to or just after a page is removed from an lru list (that ordering being
1166 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1168 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1169 int zid, int nr_pages)
1171 struct mem_cgroup_per_node *mz;
1172 unsigned long *lru_size;
1175 if (mem_cgroup_disabled())
1178 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1179 lru_size = &mz->lru_zone_size[zid][lru];
1182 *lru_size += nr_pages;
1185 if (WARN_ONCE(size < 0,
1186 "%s(%p, %d, %d): lru_size %ld\n",
1187 __func__, lruvec, lru, nr_pages, size)) {
1193 *lru_size += nr_pages;
1196 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1198 struct mem_cgroup *task_memcg;
1199 struct task_struct *p;
1202 p = find_lock_task_mm(task);
1204 task_memcg = get_mem_cgroup_from_mm(p->mm);
1208 * All threads may have already detached their mm's, but the oom
1209 * killer still needs to detect if they have already been oom
1210 * killed to prevent needlessly killing additional tasks.
1213 task_memcg = mem_cgroup_from_task(task);
1214 css_get(&task_memcg->css);
1217 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1218 css_put(&task_memcg->css);
1223 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1224 * @memcg: the memory cgroup
1226 * Returns the maximum amount of memory @mem can be charged with, in
1229 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1231 unsigned long margin = 0;
1232 unsigned long count;
1233 unsigned long limit;
1235 count = page_counter_read(&memcg->memory);
1236 limit = READ_ONCE(memcg->memory.max);
1238 margin = limit - count;
1240 if (do_memsw_account()) {
1241 count = page_counter_read(&memcg->memsw);
1242 limit = READ_ONCE(memcg->memsw.max);
1244 margin = min(margin, limit - count);
1253 * A routine for checking "mem" is under move_account() or not.
1255 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1256 * moving cgroups. This is for waiting at high-memory pressure
1259 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1261 struct mem_cgroup *from;
1262 struct mem_cgroup *to;
1265 * Unlike task_move routines, we access mc.to, mc.from not under
1266 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1268 spin_lock(&mc.lock);
1274 ret = mem_cgroup_is_descendant(from, memcg) ||
1275 mem_cgroup_is_descendant(to, memcg);
1277 spin_unlock(&mc.lock);
1281 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1283 if (mc.moving_task && current != mc.moving_task) {
1284 if (mem_cgroup_under_move(memcg)) {
1286 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1287 /* moving charge context might have finished. */
1290 finish_wait(&mc.waitq, &wait);
1297 static const unsigned int memcg1_stats[] = {
1308 static const char *const memcg1_stat_names[] = {
1319 #define K(x) ((x) << (PAGE_SHIFT-10))
1321 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1322 * @memcg: The memory cgroup that went over limit
1323 * @p: Task that is going to be killed
1325 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1328 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1330 struct mem_cgroup *iter;
1336 pr_info("Task in ");
1337 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1338 pr_cont(" killed as a result of limit of ");
1340 pr_info("Memory limit reached of cgroup ");
1343 pr_cont_cgroup_path(memcg->css.cgroup);
1348 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1349 K((u64)page_counter_read(&memcg->memory)),
1350 K((u64)memcg->memory.max), memcg->memory.failcnt);
1351 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1352 K((u64)page_counter_read(&memcg->memsw)),
1353 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1354 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1355 K((u64)page_counter_read(&memcg->kmem)),
1356 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1358 for_each_mem_cgroup_tree(iter, memcg) {
1359 pr_info("Memory cgroup stats for ");
1360 pr_cont_cgroup_path(iter->css.cgroup);
1363 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1364 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1366 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1367 K(memcg_page_state(iter, memcg1_stats[i])));
1370 for (i = 0; i < NR_LRU_LISTS; i++)
1371 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1372 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1379 * Return the memory (and swap, if configured) limit for a memcg.
1381 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1385 max = memcg->memory.max;
1386 if (mem_cgroup_swappiness(memcg)) {
1387 unsigned long memsw_max;
1388 unsigned long swap_max;
1390 memsw_max = memcg->memsw.max;
1391 swap_max = memcg->swap.max;
1392 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1393 max = min(max + swap_max, memsw_max);
1398 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1401 struct oom_control oc = {
1405 .gfp_mask = gfp_mask,
1410 if (mutex_lock_killable(&oom_lock))
1413 * A few threads which were not waiting at mutex_lock_killable() can
1414 * fail to bail out. Therefore, check again after holding oom_lock.
1416 ret = should_force_charge() || out_of_memory(&oc);
1417 mutex_unlock(&oom_lock);
1421 #if MAX_NUMNODES > 1
1424 * test_mem_cgroup_node_reclaimable
1425 * @memcg: the target memcg
1426 * @nid: the node ID to be checked.
1427 * @noswap : specify true here if the user wants flle only information.
1429 * This function returns whether the specified memcg contains any
1430 * reclaimable pages on a node. Returns true if there are any reclaimable
1431 * pages in the node.
1433 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1434 int nid, bool noswap)
1436 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1438 if (noswap || !total_swap_pages)
1440 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1447 * Always updating the nodemask is not very good - even if we have an empty
1448 * list or the wrong list here, we can start from some node and traverse all
1449 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1452 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1456 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1457 * pagein/pageout changes since the last update.
1459 if (!atomic_read(&memcg->numainfo_events))
1461 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1464 /* make a nodemask where this memcg uses memory from */
1465 memcg->scan_nodes = node_states[N_MEMORY];
1467 for_each_node_mask(nid, node_states[N_MEMORY]) {
1469 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1470 node_clear(nid, memcg->scan_nodes);
1473 atomic_set(&memcg->numainfo_events, 0);
1474 atomic_set(&memcg->numainfo_updating, 0);
1478 * Selecting a node where we start reclaim from. Because what we need is just
1479 * reducing usage counter, start from anywhere is O,K. Considering
1480 * memory reclaim from current node, there are pros. and cons.
1482 * Freeing memory from current node means freeing memory from a node which
1483 * we'll use or we've used. So, it may make LRU bad. And if several threads
1484 * hit limits, it will see a contention on a node. But freeing from remote
1485 * node means more costs for memory reclaim because of memory latency.
1487 * Now, we use round-robin. Better algorithm is welcomed.
1489 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1493 mem_cgroup_may_update_nodemask(memcg);
1494 node = memcg->last_scanned_node;
1496 node = next_node_in(node, memcg->scan_nodes);
1498 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1499 * last time it really checked all the LRUs due to rate limiting.
1500 * Fallback to the current node in that case for simplicity.
1502 if (unlikely(node == MAX_NUMNODES))
1503 node = numa_node_id();
1505 memcg->last_scanned_node = node;
1509 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1515 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1518 unsigned long *total_scanned)
1520 struct mem_cgroup *victim = NULL;
1523 unsigned long excess;
1524 unsigned long nr_scanned;
1525 struct mem_cgroup_reclaim_cookie reclaim = {
1530 excess = soft_limit_excess(root_memcg);
1533 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1538 * If we have not been able to reclaim
1539 * anything, it might because there are
1540 * no reclaimable pages under this hierarchy
1545 * We want to do more targeted reclaim.
1546 * excess >> 2 is not to excessive so as to
1547 * reclaim too much, nor too less that we keep
1548 * coming back to reclaim from this cgroup
1550 if (total >= (excess >> 2) ||
1551 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1556 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1557 pgdat, &nr_scanned);
1558 *total_scanned += nr_scanned;
1559 if (!soft_limit_excess(root_memcg))
1562 mem_cgroup_iter_break(root_memcg, victim);
1566 #ifdef CONFIG_LOCKDEP
1567 static struct lockdep_map memcg_oom_lock_dep_map = {
1568 .name = "memcg_oom_lock",
1572 static DEFINE_SPINLOCK(memcg_oom_lock);
1575 * Check OOM-Killer is already running under our hierarchy.
1576 * If someone is running, return false.
1578 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1580 struct mem_cgroup *iter, *failed = NULL;
1582 spin_lock(&memcg_oom_lock);
1584 for_each_mem_cgroup_tree(iter, memcg) {
1585 if (iter->oom_lock) {
1587 * this subtree of our hierarchy is already locked
1588 * so we cannot give a lock.
1591 mem_cgroup_iter_break(memcg, iter);
1594 iter->oom_lock = true;
1599 * OK, we failed to lock the whole subtree so we have
1600 * to clean up what we set up to the failing subtree
1602 for_each_mem_cgroup_tree(iter, memcg) {
1603 if (iter == failed) {
1604 mem_cgroup_iter_break(memcg, iter);
1607 iter->oom_lock = false;
1610 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1612 spin_unlock(&memcg_oom_lock);
1617 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1619 struct mem_cgroup *iter;
1621 spin_lock(&memcg_oom_lock);
1622 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1623 for_each_mem_cgroup_tree(iter, memcg)
1624 iter->oom_lock = false;
1625 spin_unlock(&memcg_oom_lock);
1628 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1630 struct mem_cgroup *iter;
1632 spin_lock(&memcg_oom_lock);
1633 for_each_mem_cgroup_tree(iter, memcg)
1635 spin_unlock(&memcg_oom_lock);
1638 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1640 struct mem_cgroup *iter;
1643 * When a new child is created while the hierarchy is under oom,
1644 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1646 spin_lock(&memcg_oom_lock);
1647 for_each_mem_cgroup_tree(iter, memcg)
1648 if (iter->under_oom > 0)
1650 spin_unlock(&memcg_oom_lock);
1653 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1655 struct oom_wait_info {
1656 struct mem_cgroup *memcg;
1657 wait_queue_entry_t wait;
1660 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1661 unsigned mode, int sync, void *arg)
1663 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1664 struct mem_cgroup *oom_wait_memcg;
1665 struct oom_wait_info *oom_wait_info;
1667 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1668 oom_wait_memcg = oom_wait_info->memcg;
1670 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1671 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1673 return autoremove_wake_function(wait, mode, sync, arg);
1676 static void memcg_oom_recover(struct mem_cgroup *memcg)
1679 * For the following lockless ->under_oom test, the only required
1680 * guarantee is that it must see the state asserted by an OOM when
1681 * this function is called as a result of userland actions
1682 * triggered by the notification of the OOM. This is trivially
1683 * achieved by invoking mem_cgroup_mark_under_oom() before
1684 * triggering notification.
1686 if (memcg && memcg->under_oom)
1687 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1697 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1699 enum oom_status ret;
1702 if (order > PAGE_ALLOC_COSTLY_ORDER)
1706 * We are in the middle of the charge context here, so we
1707 * don't want to block when potentially sitting on a callstack
1708 * that holds all kinds of filesystem and mm locks.
1710 * cgroup1 allows disabling the OOM killer and waiting for outside
1711 * handling until the charge can succeed; remember the context and put
1712 * the task to sleep at the end of the page fault when all locks are
1715 * On the other hand, in-kernel OOM killer allows for an async victim
1716 * memory reclaim (oom_reaper) and that means that we are not solely
1717 * relying on the oom victim to make a forward progress and we can
1718 * invoke the oom killer here.
1720 * Please note that mem_cgroup_out_of_memory might fail to find a
1721 * victim and then we have to bail out from the charge path.
1723 if (memcg->oom_kill_disable) {
1724 if (!current->in_user_fault)
1726 css_get(&memcg->css);
1727 current->memcg_in_oom = memcg;
1728 current->memcg_oom_gfp_mask = mask;
1729 current->memcg_oom_order = order;
1734 mem_cgroup_mark_under_oom(memcg);
1736 locked = mem_cgroup_oom_trylock(memcg);
1739 mem_cgroup_oom_notify(memcg);
1741 mem_cgroup_unmark_under_oom(memcg);
1742 if (mem_cgroup_out_of_memory(memcg, mask, order))
1748 mem_cgroup_oom_unlock(memcg);
1754 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1755 * @handle: actually kill/wait or just clean up the OOM state
1757 * This has to be called at the end of a page fault if the memcg OOM
1758 * handler was enabled.
1760 * Memcg supports userspace OOM handling where failed allocations must
1761 * sleep on a waitqueue until the userspace task resolves the
1762 * situation. Sleeping directly in the charge context with all kinds
1763 * of locks held is not a good idea, instead we remember an OOM state
1764 * in the task and mem_cgroup_oom_synchronize() has to be called at
1765 * the end of the page fault to complete the OOM handling.
1767 * Returns %true if an ongoing memcg OOM situation was detected and
1768 * completed, %false otherwise.
1770 bool mem_cgroup_oom_synchronize(bool handle)
1772 struct mem_cgroup *memcg = current->memcg_in_oom;
1773 struct oom_wait_info owait;
1776 /* OOM is global, do not handle */
1783 owait.memcg = memcg;
1784 owait.wait.flags = 0;
1785 owait.wait.func = memcg_oom_wake_function;
1786 owait.wait.private = current;
1787 INIT_LIST_HEAD(&owait.wait.entry);
1789 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1790 mem_cgroup_mark_under_oom(memcg);
1792 locked = mem_cgroup_oom_trylock(memcg);
1795 mem_cgroup_oom_notify(memcg);
1797 if (locked && !memcg->oom_kill_disable) {
1798 mem_cgroup_unmark_under_oom(memcg);
1799 finish_wait(&memcg_oom_waitq, &owait.wait);
1800 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1801 current->memcg_oom_order);
1804 mem_cgroup_unmark_under_oom(memcg);
1805 finish_wait(&memcg_oom_waitq, &owait.wait);
1809 mem_cgroup_oom_unlock(memcg);
1811 * There is no guarantee that an OOM-lock contender
1812 * sees the wakeups triggered by the OOM kill
1813 * uncharges. Wake any sleepers explicitely.
1815 memcg_oom_recover(memcg);
1818 current->memcg_in_oom = NULL;
1819 css_put(&memcg->css);
1824 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1825 * @victim: task to be killed by the OOM killer
1826 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1828 * Returns a pointer to a memory cgroup, which has to be cleaned up
1829 * by killing all belonging OOM-killable tasks.
1831 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1833 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1834 struct mem_cgroup *oom_domain)
1836 struct mem_cgroup *oom_group = NULL;
1837 struct mem_cgroup *memcg;
1839 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1843 oom_domain = root_mem_cgroup;
1847 memcg = mem_cgroup_from_task(victim);
1848 if (memcg == root_mem_cgroup)
1852 * Traverse the memory cgroup hierarchy from the victim task's
1853 * cgroup up to the OOMing cgroup (or root) to find the
1854 * highest-level memory cgroup with oom.group set.
1856 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1857 if (memcg->oom_group)
1860 if (memcg == oom_domain)
1865 css_get(&oom_group->css);
1872 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1874 pr_info("Tasks in ");
1875 pr_cont_cgroup_path(memcg->css.cgroup);
1876 pr_cont(" are going to be killed due to memory.oom.group set\n");
1880 * lock_page_memcg - lock a page->mem_cgroup binding
1883 * This function protects unlocked LRU pages from being moved to
1886 * It ensures lifetime of the returned memcg. Caller is responsible
1887 * for the lifetime of the page; __unlock_page_memcg() is available
1888 * when @page might get freed inside the locked section.
1890 struct mem_cgroup *lock_page_memcg(struct page *page)
1892 struct mem_cgroup *memcg;
1893 unsigned long flags;
1896 * The RCU lock is held throughout the transaction. The fast
1897 * path can get away without acquiring the memcg->move_lock
1898 * because page moving starts with an RCU grace period.
1900 * The RCU lock also protects the memcg from being freed when
1901 * the page state that is going to change is the only thing
1902 * preventing the page itself from being freed. E.g. writeback
1903 * doesn't hold a page reference and relies on PG_writeback to
1904 * keep off truncation, migration and so forth.
1908 if (mem_cgroup_disabled())
1911 memcg = page->mem_cgroup;
1912 if (unlikely(!memcg))
1915 if (atomic_read(&memcg->moving_account) <= 0)
1918 spin_lock_irqsave(&memcg->move_lock, flags);
1919 if (memcg != page->mem_cgroup) {
1920 spin_unlock_irqrestore(&memcg->move_lock, flags);
1925 * When charge migration first begins, we can have locked and
1926 * unlocked page stat updates happening concurrently. Track
1927 * the task who has the lock for unlock_page_memcg().
1929 memcg->move_lock_task = current;
1930 memcg->move_lock_flags = flags;
1934 EXPORT_SYMBOL(lock_page_memcg);
1937 * __unlock_page_memcg - unlock and unpin a memcg
1940 * Unlock and unpin a memcg returned by lock_page_memcg().
1942 void __unlock_page_memcg(struct mem_cgroup *memcg)
1944 if (memcg && memcg->move_lock_task == current) {
1945 unsigned long flags = memcg->move_lock_flags;
1947 memcg->move_lock_task = NULL;
1948 memcg->move_lock_flags = 0;
1950 spin_unlock_irqrestore(&memcg->move_lock, flags);
1957 * unlock_page_memcg - unlock a page->mem_cgroup binding
1960 void unlock_page_memcg(struct page *page)
1962 __unlock_page_memcg(page->mem_cgroup);
1964 EXPORT_SYMBOL(unlock_page_memcg);
1966 struct memcg_stock_pcp {
1967 struct mem_cgroup *cached; /* this never be root cgroup */
1968 unsigned int nr_pages;
1969 struct work_struct work;
1970 unsigned long flags;
1971 #define FLUSHING_CACHED_CHARGE 0
1973 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1974 static DEFINE_MUTEX(percpu_charge_mutex);
1977 * consume_stock: Try to consume stocked charge on this cpu.
1978 * @memcg: memcg to consume from.
1979 * @nr_pages: how many pages to charge.
1981 * The charges will only happen if @memcg matches the current cpu's memcg
1982 * stock, and at least @nr_pages are available in that stock. Failure to
1983 * service an allocation will refill the stock.
1985 * returns true if successful, false otherwise.
1987 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1989 struct memcg_stock_pcp *stock;
1990 unsigned long flags;
1993 if (nr_pages > MEMCG_CHARGE_BATCH)
1996 local_irq_save(flags);
1998 stock = this_cpu_ptr(&memcg_stock);
1999 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2000 stock->nr_pages -= nr_pages;
2004 local_irq_restore(flags);
2010 * Returns stocks cached in percpu and reset cached information.
2012 static void drain_stock(struct memcg_stock_pcp *stock)
2014 struct mem_cgroup *old = stock->cached;
2016 if (stock->nr_pages) {
2017 page_counter_uncharge(&old->memory, stock->nr_pages);
2018 if (do_memsw_account())
2019 page_counter_uncharge(&old->memsw, stock->nr_pages);
2020 css_put_many(&old->css, stock->nr_pages);
2021 stock->nr_pages = 0;
2023 stock->cached = NULL;
2026 static void drain_local_stock(struct work_struct *dummy)
2028 struct memcg_stock_pcp *stock;
2029 unsigned long flags;
2032 * The only protection from memory hotplug vs. drain_stock races is
2033 * that we always operate on local CPU stock here with IRQ disabled
2035 local_irq_save(flags);
2037 stock = this_cpu_ptr(&memcg_stock);
2039 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2041 local_irq_restore(flags);
2045 * Cache charges(val) to local per_cpu area.
2046 * This will be consumed by consume_stock() function, later.
2048 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2050 struct memcg_stock_pcp *stock;
2051 unsigned long flags;
2053 local_irq_save(flags);
2055 stock = this_cpu_ptr(&memcg_stock);
2056 if (stock->cached != memcg) { /* reset if necessary */
2058 stock->cached = memcg;
2060 stock->nr_pages += nr_pages;
2062 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2065 local_irq_restore(flags);
2069 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2070 * of the hierarchy under it.
2072 static void drain_all_stock(struct mem_cgroup *root_memcg)
2076 /* If someone's already draining, avoid adding running more workers. */
2077 if (!mutex_trylock(&percpu_charge_mutex))
2080 * Notify other cpus that system-wide "drain" is running
2081 * We do not care about races with the cpu hotplug because cpu down
2082 * as well as workers from this path always operate on the local
2083 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2086 for_each_online_cpu(cpu) {
2087 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2088 struct mem_cgroup *memcg;
2090 memcg = stock->cached;
2091 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2093 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2094 css_put(&memcg->css);
2097 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2099 drain_local_stock(&stock->work);
2101 schedule_work_on(cpu, &stock->work);
2103 css_put(&memcg->css);
2106 mutex_unlock(&percpu_charge_mutex);
2109 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2111 struct memcg_stock_pcp *stock;
2112 struct mem_cgroup *memcg;
2114 stock = &per_cpu(memcg_stock, cpu);
2117 for_each_mem_cgroup(memcg) {
2120 for (i = 0; i < MEMCG_NR_STAT; i++) {
2124 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
2126 atomic_long_add(x, &memcg->stat[i]);
2128 if (i >= NR_VM_NODE_STAT_ITEMS)
2131 for_each_node(nid) {
2132 struct mem_cgroup_per_node *pn;
2134 pn = mem_cgroup_nodeinfo(memcg, nid);
2135 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2137 atomic_long_add(x, &pn->lruvec_stat[i]);
2141 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2144 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
2146 atomic_long_add(x, &memcg->events[i]);
2153 static void reclaim_high(struct mem_cgroup *memcg,
2154 unsigned int nr_pages,
2158 if (page_counter_read(&memcg->memory) <= memcg->high)
2160 memcg_memory_event(memcg, MEMCG_HIGH);
2161 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2162 } while ((memcg = parent_mem_cgroup(memcg)));
2165 static void high_work_func(struct work_struct *work)
2167 struct mem_cgroup *memcg;
2169 memcg = container_of(work, struct mem_cgroup, high_work);
2170 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2174 * Scheduled by try_charge() to be executed from the userland return path
2175 * and reclaims memory over the high limit.
2177 void mem_cgroup_handle_over_high(void)
2179 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2180 struct mem_cgroup *memcg;
2182 if (likely(!nr_pages))
2185 memcg = get_mem_cgroup_from_mm(current->mm);
2186 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2187 css_put(&memcg->css);
2188 current->memcg_nr_pages_over_high = 0;
2191 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2192 unsigned int nr_pages)
2194 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2195 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2196 struct mem_cgroup *mem_over_limit;
2197 struct page_counter *counter;
2198 unsigned long nr_reclaimed;
2199 bool may_swap = true;
2200 bool drained = false;
2202 enum oom_status oom_status;
2204 if (mem_cgroup_is_root(memcg))
2207 if (consume_stock(memcg, nr_pages))
2210 if (!do_memsw_account() ||
2211 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2212 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2214 if (do_memsw_account())
2215 page_counter_uncharge(&memcg->memsw, batch);
2216 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2218 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2222 if (batch > nr_pages) {
2228 * Memcg doesn't have a dedicated reserve for atomic
2229 * allocations. But like the global atomic pool, we need to
2230 * put the burden of reclaim on regular allocation requests
2231 * and let these go through as privileged allocations.
2233 if (gfp_mask & __GFP_ATOMIC)
2237 * Unlike in global OOM situations, memcg is not in a physical
2238 * memory shortage. Allow dying and OOM-killed tasks to
2239 * bypass the last charges so that they can exit quickly and
2240 * free their memory.
2242 if (unlikely(should_force_charge()))
2246 * Prevent unbounded recursion when reclaim operations need to
2247 * allocate memory. This might exceed the limits temporarily,
2248 * but we prefer facilitating memory reclaim and getting back
2249 * under the limit over triggering OOM kills in these cases.
2251 if (unlikely(current->flags & PF_MEMALLOC))
2254 if (unlikely(task_in_memcg_oom(current)))
2257 if (!gfpflags_allow_blocking(gfp_mask))
2260 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2262 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2263 gfp_mask, may_swap);
2265 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2269 drain_all_stock(mem_over_limit);
2274 if (gfp_mask & __GFP_NORETRY)
2277 * Even though the limit is exceeded at this point, reclaim
2278 * may have been able to free some pages. Retry the charge
2279 * before killing the task.
2281 * Only for regular pages, though: huge pages are rather
2282 * unlikely to succeed so close to the limit, and we fall back
2283 * to regular pages anyway in case of failure.
2285 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2288 * At task move, charge accounts can be doubly counted. So, it's
2289 * better to wait until the end of task_move if something is going on.
2291 if (mem_cgroup_wait_acct_move(mem_over_limit))
2297 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2300 if (gfp_mask & __GFP_NOFAIL)
2303 if (fatal_signal_pending(current))
2306 memcg_memory_event(mem_over_limit, MEMCG_OOM);
2309 * keep retrying as long as the memcg oom killer is able to make
2310 * a forward progress or bypass the charge if the oom killer
2311 * couldn't make any progress.
2313 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2314 get_order(nr_pages * PAGE_SIZE));
2315 switch (oom_status) {
2317 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2326 if (!(gfp_mask & __GFP_NOFAIL))
2330 * The allocation either can't fail or will lead to more memory
2331 * being freed very soon. Allow memory usage go over the limit
2332 * temporarily by force charging it.
2334 page_counter_charge(&memcg->memory, nr_pages);
2335 if (do_memsw_account())
2336 page_counter_charge(&memcg->memsw, nr_pages);
2337 css_get_many(&memcg->css, nr_pages);
2342 css_get_many(&memcg->css, batch);
2343 if (batch > nr_pages)
2344 refill_stock(memcg, batch - nr_pages);
2347 * If the hierarchy is above the normal consumption range, schedule
2348 * reclaim on returning to userland. We can perform reclaim here
2349 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2350 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2351 * not recorded as it most likely matches current's and won't
2352 * change in the meantime. As high limit is checked again before
2353 * reclaim, the cost of mismatch is negligible.
2356 if (page_counter_read(&memcg->memory) > memcg->high) {
2357 /* Don't bother a random interrupted task */
2358 if (in_interrupt()) {
2359 schedule_work(&memcg->high_work);
2362 current->memcg_nr_pages_over_high += batch;
2363 set_notify_resume(current);
2366 } while ((memcg = parent_mem_cgroup(memcg)));
2371 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2373 if (mem_cgroup_is_root(memcg))
2376 page_counter_uncharge(&memcg->memory, nr_pages);
2377 if (do_memsw_account())
2378 page_counter_uncharge(&memcg->memsw, nr_pages);
2380 css_put_many(&memcg->css, nr_pages);
2383 static void lock_page_lru(struct page *page, int *isolated)
2385 struct zone *zone = page_zone(page);
2387 spin_lock_irq(zone_lru_lock(zone));
2388 if (PageLRU(page)) {
2389 struct lruvec *lruvec;
2391 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2393 del_page_from_lru_list(page, lruvec, page_lru(page));
2399 static void unlock_page_lru(struct page *page, int isolated)
2401 struct zone *zone = page_zone(page);
2404 struct lruvec *lruvec;
2406 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2407 VM_BUG_ON_PAGE(PageLRU(page), page);
2409 add_page_to_lru_list(page, lruvec, page_lru(page));
2411 spin_unlock_irq(zone_lru_lock(zone));
2414 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2419 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2422 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2423 * may already be on some other mem_cgroup's LRU. Take care of it.
2426 lock_page_lru(page, &isolated);
2429 * Nobody should be changing or seriously looking at
2430 * page->mem_cgroup at this point:
2432 * - the page is uncharged
2434 * - the page is off-LRU
2436 * - an anonymous fault has exclusive page access, except for
2437 * a locked page table
2439 * - a page cache insertion, a swapin fault, or a migration
2440 * have the page locked
2442 page->mem_cgroup = memcg;
2445 unlock_page_lru(page, isolated);
2448 #ifdef CONFIG_MEMCG_KMEM
2449 static int memcg_alloc_cache_id(void)
2454 id = ida_simple_get(&memcg_cache_ida,
2455 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2459 if (id < memcg_nr_cache_ids)
2463 * There's no space for the new id in memcg_caches arrays,
2464 * so we have to grow them.
2466 down_write(&memcg_cache_ids_sem);
2468 size = 2 * (id + 1);
2469 if (size < MEMCG_CACHES_MIN_SIZE)
2470 size = MEMCG_CACHES_MIN_SIZE;
2471 else if (size > MEMCG_CACHES_MAX_SIZE)
2472 size = MEMCG_CACHES_MAX_SIZE;
2474 err = memcg_update_all_caches(size);
2476 err = memcg_update_all_list_lrus(size);
2478 memcg_nr_cache_ids = size;
2480 up_write(&memcg_cache_ids_sem);
2483 ida_simple_remove(&memcg_cache_ida, id);
2489 static void memcg_free_cache_id(int id)
2491 ida_simple_remove(&memcg_cache_ida, id);
2494 struct memcg_kmem_cache_create_work {
2495 struct mem_cgroup *memcg;
2496 struct kmem_cache *cachep;
2497 struct work_struct work;
2500 static void memcg_kmem_cache_create_func(struct work_struct *w)
2502 struct memcg_kmem_cache_create_work *cw =
2503 container_of(w, struct memcg_kmem_cache_create_work, work);
2504 struct mem_cgroup *memcg = cw->memcg;
2505 struct kmem_cache *cachep = cw->cachep;
2507 memcg_create_kmem_cache(memcg, cachep);
2509 css_put(&memcg->css);
2514 * Enqueue the creation of a per-memcg kmem_cache.
2516 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2517 struct kmem_cache *cachep)
2519 struct memcg_kmem_cache_create_work *cw;
2521 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2525 css_get(&memcg->css);
2528 cw->cachep = cachep;
2529 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2531 queue_work(memcg_kmem_cache_wq, &cw->work);
2534 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2535 struct kmem_cache *cachep)
2538 * We need to stop accounting when we kmalloc, because if the
2539 * corresponding kmalloc cache is not yet created, the first allocation
2540 * in __memcg_schedule_kmem_cache_create will recurse.
2542 * However, it is better to enclose the whole function. Depending on
2543 * the debugging options enabled, INIT_WORK(), for instance, can
2544 * trigger an allocation. This too, will make us recurse. Because at
2545 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2546 * the safest choice is to do it like this, wrapping the whole function.
2548 current->memcg_kmem_skip_account = 1;
2549 __memcg_schedule_kmem_cache_create(memcg, cachep);
2550 current->memcg_kmem_skip_account = 0;
2553 static inline bool memcg_kmem_bypass(void)
2555 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2561 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2562 * @cachep: the original global kmem cache
2564 * Return the kmem_cache we're supposed to use for a slab allocation.
2565 * We try to use the current memcg's version of the cache.
2567 * If the cache does not exist yet, if we are the first user of it, we
2568 * create it asynchronously in a workqueue and let the current allocation
2569 * go through with the original cache.
2571 * This function takes a reference to the cache it returns to assure it
2572 * won't get destroyed while we are working with it. Once the caller is
2573 * done with it, memcg_kmem_put_cache() must be called to release the
2576 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2578 struct mem_cgroup *memcg;
2579 struct kmem_cache *memcg_cachep;
2582 VM_BUG_ON(!is_root_cache(cachep));
2584 if (memcg_kmem_bypass())
2587 if (current->memcg_kmem_skip_account)
2590 memcg = get_mem_cgroup_from_current();
2591 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2595 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2596 if (likely(memcg_cachep))
2597 return memcg_cachep;
2600 * If we are in a safe context (can wait, and not in interrupt
2601 * context), we could be be predictable and return right away.
2602 * This would guarantee that the allocation being performed
2603 * already belongs in the new cache.
2605 * However, there are some clashes that can arrive from locking.
2606 * For instance, because we acquire the slab_mutex while doing
2607 * memcg_create_kmem_cache, this means no further allocation
2608 * could happen with the slab_mutex held. So it's better to
2611 memcg_schedule_kmem_cache_create(memcg, cachep);
2613 css_put(&memcg->css);
2618 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2619 * @cachep: the cache returned by memcg_kmem_get_cache
2621 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2623 if (!is_root_cache(cachep))
2624 css_put(&cachep->memcg_params.memcg->css);
2628 * memcg_kmem_charge_memcg: charge a kmem page
2629 * @page: page to charge
2630 * @gfp: reclaim mode
2631 * @order: allocation order
2632 * @memcg: memory cgroup to charge
2634 * Returns 0 on success, an error code on failure.
2636 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2637 struct mem_cgroup *memcg)
2639 unsigned int nr_pages = 1 << order;
2640 struct page_counter *counter;
2643 ret = try_charge(memcg, gfp, nr_pages);
2647 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2648 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2651 * Enforce __GFP_NOFAIL allocation because callers are not
2652 * prepared to see failures and likely do not have any failure
2655 if (gfp & __GFP_NOFAIL) {
2656 page_counter_charge(&memcg->kmem, nr_pages);
2659 cancel_charge(memcg, nr_pages);
2663 page->mem_cgroup = memcg;
2669 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2670 * @page: page to charge
2671 * @gfp: reclaim mode
2672 * @order: allocation order
2674 * Returns 0 on success, an error code on failure.
2676 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2678 struct mem_cgroup *memcg;
2681 if (mem_cgroup_disabled() || memcg_kmem_bypass())
2684 memcg = get_mem_cgroup_from_current();
2685 if (!mem_cgroup_is_root(memcg)) {
2686 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2688 __SetPageKmemcg(page);
2690 css_put(&memcg->css);
2694 * memcg_kmem_uncharge: uncharge a kmem page
2695 * @page: page to uncharge
2696 * @order: allocation order
2698 void memcg_kmem_uncharge(struct page *page, int order)
2700 struct mem_cgroup *memcg = page->mem_cgroup;
2701 unsigned int nr_pages = 1 << order;
2706 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2708 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2709 page_counter_uncharge(&memcg->kmem, nr_pages);
2711 page_counter_uncharge(&memcg->memory, nr_pages);
2712 if (do_memsw_account())
2713 page_counter_uncharge(&memcg->memsw, nr_pages);
2715 page->mem_cgroup = NULL;
2717 /* slab pages do not have PageKmemcg flag set */
2718 if (PageKmemcg(page))
2719 __ClearPageKmemcg(page);
2721 css_put_many(&memcg->css, nr_pages);
2723 #endif /* CONFIG_MEMCG_KMEM */
2725 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2728 * Because tail pages are not marked as "used", set it. We're under
2729 * zone_lru_lock and migration entries setup in all page mappings.
2731 void mem_cgroup_split_huge_fixup(struct page *head)
2735 if (mem_cgroup_disabled())
2738 for (i = 1; i < HPAGE_PMD_NR; i++)
2739 head[i].mem_cgroup = head->mem_cgroup;
2741 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2743 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2745 #ifdef CONFIG_MEMCG_SWAP
2747 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2748 * @entry: swap entry to be moved
2749 * @from: mem_cgroup which the entry is moved from
2750 * @to: mem_cgroup which the entry is moved to
2752 * It succeeds only when the swap_cgroup's record for this entry is the same
2753 * as the mem_cgroup's id of @from.
2755 * Returns 0 on success, -EINVAL on failure.
2757 * The caller must have charged to @to, IOW, called page_counter_charge() about
2758 * both res and memsw, and called css_get().
2760 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2761 struct mem_cgroup *from, struct mem_cgroup *to)
2763 unsigned short old_id, new_id;
2765 old_id = mem_cgroup_id(from);
2766 new_id = mem_cgroup_id(to);
2768 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2769 mod_memcg_state(from, MEMCG_SWAP, -1);
2770 mod_memcg_state(to, MEMCG_SWAP, 1);
2776 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2777 struct mem_cgroup *from, struct mem_cgroup *to)
2783 static DEFINE_MUTEX(memcg_max_mutex);
2785 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2786 unsigned long max, bool memsw)
2788 bool enlarge = false;
2789 bool drained = false;
2791 bool limits_invariant;
2792 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2795 if (signal_pending(current)) {
2800 mutex_lock(&memcg_max_mutex);
2802 * Make sure that the new limit (memsw or memory limit) doesn't
2803 * break our basic invariant rule memory.max <= memsw.max.
2805 limits_invariant = memsw ? max >= memcg->memory.max :
2806 max <= memcg->memsw.max;
2807 if (!limits_invariant) {
2808 mutex_unlock(&memcg_max_mutex);
2812 if (max > counter->max)
2814 ret = page_counter_set_max(counter, max);
2815 mutex_unlock(&memcg_max_mutex);
2821 drain_all_stock(memcg);
2826 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2827 GFP_KERNEL, !memsw)) {
2833 if (!ret && enlarge)
2834 memcg_oom_recover(memcg);
2839 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2841 unsigned long *total_scanned)
2843 unsigned long nr_reclaimed = 0;
2844 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2845 unsigned long reclaimed;
2847 struct mem_cgroup_tree_per_node *mctz;
2848 unsigned long excess;
2849 unsigned long nr_scanned;
2854 mctz = soft_limit_tree_node(pgdat->node_id);
2857 * Do not even bother to check the largest node if the root
2858 * is empty. Do it lockless to prevent lock bouncing. Races
2859 * are acceptable as soft limit is best effort anyway.
2861 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2865 * This loop can run a while, specially if mem_cgroup's continuously
2866 * keep exceeding their soft limit and putting the system under
2873 mz = mem_cgroup_largest_soft_limit_node(mctz);
2878 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2879 gfp_mask, &nr_scanned);
2880 nr_reclaimed += reclaimed;
2881 *total_scanned += nr_scanned;
2882 spin_lock_irq(&mctz->lock);
2883 __mem_cgroup_remove_exceeded(mz, mctz);
2886 * If we failed to reclaim anything from this memory cgroup
2887 * it is time to move on to the next cgroup
2891 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2893 excess = soft_limit_excess(mz->memcg);
2895 * One school of thought says that we should not add
2896 * back the node to the tree if reclaim returns 0.
2897 * But our reclaim could return 0, simply because due
2898 * to priority we are exposing a smaller subset of
2899 * memory to reclaim from. Consider this as a longer
2902 /* If excess == 0, no tree ops */
2903 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2904 spin_unlock_irq(&mctz->lock);
2905 css_put(&mz->memcg->css);
2908 * Could not reclaim anything and there are no more
2909 * mem cgroups to try or we seem to be looping without
2910 * reclaiming anything.
2912 if (!nr_reclaimed &&
2914 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2916 } while (!nr_reclaimed);
2918 css_put(&next_mz->memcg->css);
2919 return nr_reclaimed;
2923 * Test whether @memcg has children, dead or alive. Note that this
2924 * function doesn't care whether @memcg has use_hierarchy enabled and
2925 * returns %true if there are child csses according to the cgroup
2926 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2928 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2933 ret = css_next_child(NULL, &memcg->css);
2939 * Reclaims as many pages from the given memcg as possible.
2941 * Caller is responsible for holding css reference for memcg.
2943 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2945 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2947 /* we call try-to-free pages for make this cgroup empty */
2948 lru_add_drain_all();
2950 drain_all_stock(memcg);
2952 /* try to free all pages in this cgroup */
2953 while (nr_retries && page_counter_read(&memcg->memory)) {
2956 if (signal_pending(current))
2959 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2963 /* maybe some writeback is necessary */
2964 congestion_wait(BLK_RW_ASYNC, HZ/10);
2972 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2973 char *buf, size_t nbytes,
2976 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2978 if (mem_cgroup_is_root(memcg))
2980 return mem_cgroup_force_empty(memcg) ?: nbytes;
2983 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2986 return mem_cgroup_from_css(css)->use_hierarchy;
2989 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2990 struct cftype *cft, u64 val)
2993 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2994 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2996 if (memcg->use_hierarchy == val)
3000 * If parent's use_hierarchy is set, we can't make any modifications
3001 * in the child subtrees. If it is unset, then the change can
3002 * occur, provided the current cgroup has no children.
3004 * For the root cgroup, parent_mem is NULL, we allow value to be
3005 * set if there are no children.
3007 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3008 (val == 1 || val == 0)) {
3009 if (!memcg_has_children(memcg))
3010 memcg->use_hierarchy = val;
3019 struct accumulated_stats {
3020 unsigned long stat[MEMCG_NR_STAT];
3021 unsigned long events[NR_VM_EVENT_ITEMS];
3022 unsigned long lru_pages[NR_LRU_LISTS];
3023 const unsigned int *stats_array;
3024 const unsigned int *events_array;
3029 static void accumulate_memcg_tree(struct mem_cgroup *memcg,
3030 struct accumulated_stats *acc)
3032 struct mem_cgroup *mi;
3035 for_each_mem_cgroup_tree(mi, memcg) {
3036 for (i = 0; i < acc->stats_size; i++)
3037 acc->stat[i] += memcg_page_state(mi,
3038 acc->stats_array ? acc->stats_array[i] : i);
3040 for (i = 0; i < acc->events_size; i++)
3041 acc->events[i] += memcg_sum_events(mi,
3042 acc->events_array ? acc->events_array[i] : i);
3044 for (i = 0; i < NR_LRU_LISTS; i++)
3045 acc->lru_pages[i] +=
3046 mem_cgroup_nr_lru_pages(mi, BIT(i));
3050 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3052 unsigned long val = 0;
3054 if (mem_cgroup_is_root(memcg)) {
3055 struct mem_cgroup *iter;
3057 for_each_mem_cgroup_tree(iter, memcg) {
3058 val += memcg_page_state(iter, MEMCG_CACHE);
3059 val += memcg_page_state(iter, MEMCG_RSS);
3061 val += memcg_page_state(iter, MEMCG_SWAP);
3065 val = page_counter_read(&memcg->memory);
3067 val = page_counter_read(&memcg->memsw);
3080 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3083 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3084 struct page_counter *counter;
3086 switch (MEMFILE_TYPE(cft->private)) {
3088 counter = &memcg->memory;
3091 counter = &memcg->memsw;
3094 counter = &memcg->kmem;
3097 counter = &memcg->tcpmem;
3103 switch (MEMFILE_ATTR(cft->private)) {
3105 if (counter == &memcg->memory)
3106 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3107 if (counter == &memcg->memsw)
3108 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3109 return (u64)page_counter_read(counter) * PAGE_SIZE;
3111 return (u64)counter->max * PAGE_SIZE;
3113 return (u64)counter->watermark * PAGE_SIZE;
3115 return counter->failcnt;
3116 case RES_SOFT_LIMIT:
3117 return (u64)memcg->soft_limit * PAGE_SIZE;
3123 #ifdef CONFIG_MEMCG_KMEM
3124 static int memcg_online_kmem(struct mem_cgroup *memcg)
3128 if (cgroup_memory_nokmem)
3131 BUG_ON(memcg->kmemcg_id >= 0);
3132 BUG_ON(memcg->kmem_state);
3134 memcg_id = memcg_alloc_cache_id();
3138 static_branch_inc(&memcg_kmem_enabled_key);
3140 * A memory cgroup is considered kmem-online as soon as it gets
3141 * kmemcg_id. Setting the id after enabling static branching will
3142 * guarantee no one starts accounting before all call sites are
3145 memcg->kmemcg_id = memcg_id;
3146 memcg->kmem_state = KMEM_ONLINE;
3147 INIT_LIST_HEAD(&memcg->kmem_caches);
3152 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3154 struct cgroup_subsys_state *css;
3155 struct mem_cgroup *parent, *child;
3158 if (memcg->kmem_state != KMEM_ONLINE)
3161 * Clear the online state before clearing memcg_caches array
3162 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3163 * guarantees that no cache will be created for this cgroup
3164 * after we are done (see memcg_create_kmem_cache()).
3166 memcg->kmem_state = KMEM_ALLOCATED;
3168 memcg_deactivate_kmem_caches(memcg);
3170 kmemcg_id = memcg->kmemcg_id;
3171 BUG_ON(kmemcg_id < 0);
3173 parent = parent_mem_cgroup(memcg);
3175 parent = root_mem_cgroup;
3178 * Change kmemcg_id of this cgroup and all its descendants to the
3179 * parent's id, and then move all entries from this cgroup's list_lrus
3180 * to ones of the parent. After we have finished, all list_lrus
3181 * corresponding to this cgroup are guaranteed to remain empty. The
3182 * ordering is imposed by list_lru_node->lock taken by
3183 * memcg_drain_all_list_lrus().
3185 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3186 css_for_each_descendant_pre(css, &memcg->css) {
3187 child = mem_cgroup_from_css(css);
3188 BUG_ON(child->kmemcg_id != kmemcg_id);
3189 child->kmemcg_id = parent->kmemcg_id;
3190 if (!memcg->use_hierarchy)
3195 memcg_drain_all_list_lrus(kmemcg_id, parent);
3197 memcg_free_cache_id(kmemcg_id);
3200 static void memcg_free_kmem(struct mem_cgroup *memcg)
3202 /* css_alloc() failed, offlining didn't happen */
3203 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3204 memcg_offline_kmem(memcg);
3206 if (memcg->kmem_state == KMEM_ALLOCATED) {
3207 memcg_destroy_kmem_caches(memcg);
3208 static_branch_dec(&memcg_kmem_enabled_key);
3209 WARN_ON(page_counter_read(&memcg->kmem));
3213 static int memcg_online_kmem(struct mem_cgroup *memcg)
3217 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3220 static void memcg_free_kmem(struct mem_cgroup *memcg)
3223 #endif /* CONFIG_MEMCG_KMEM */
3225 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3230 mutex_lock(&memcg_max_mutex);
3231 ret = page_counter_set_max(&memcg->kmem, max);
3232 mutex_unlock(&memcg_max_mutex);
3236 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3240 mutex_lock(&memcg_max_mutex);
3242 ret = page_counter_set_max(&memcg->tcpmem, max);
3246 if (!memcg->tcpmem_active) {
3248 * The active flag needs to be written after the static_key
3249 * update. This is what guarantees that the socket activation
3250 * function is the last one to run. See mem_cgroup_sk_alloc()
3251 * for details, and note that we don't mark any socket as
3252 * belonging to this memcg until that flag is up.
3254 * We need to do this, because static_keys will span multiple
3255 * sites, but we can't control their order. If we mark a socket
3256 * as accounted, but the accounting functions are not patched in
3257 * yet, we'll lose accounting.
3259 * We never race with the readers in mem_cgroup_sk_alloc(),
3260 * because when this value change, the code to process it is not
3263 static_branch_inc(&memcg_sockets_enabled_key);
3264 memcg->tcpmem_active = true;
3267 mutex_unlock(&memcg_max_mutex);
3272 * The user of this function is...
3275 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3276 char *buf, size_t nbytes, loff_t off)
3278 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3279 unsigned long nr_pages;
3282 buf = strstrip(buf);
3283 ret = page_counter_memparse(buf, "-1", &nr_pages);
3287 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3289 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3293 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3295 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3298 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3301 ret = memcg_update_kmem_max(memcg, nr_pages);
3304 ret = memcg_update_tcp_max(memcg, nr_pages);
3308 case RES_SOFT_LIMIT:
3309 memcg->soft_limit = nr_pages;
3313 return ret ?: nbytes;
3316 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3317 size_t nbytes, loff_t off)
3319 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3320 struct page_counter *counter;
3322 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3324 counter = &memcg->memory;
3327 counter = &memcg->memsw;
3330 counter = &memcg->kmem;
3333 counter = &memcg->tcpmem;
3339 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3341 page_counter_reset_watermark(counter);
3344 counter->failcnt = 0;
3353 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3356 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3360 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3361 struct cftype *cft, u64 val)
3363 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3365 if (val & ~MOVE_MASK)
3369 * No kind of locking is needed in here, because ->can_attach() will
3370 * check this value once in the beginning of the process, and then carry
3371 * on with stale data. This means that changes to this value will only
3372 * affect task migrations starting after the change.
3374 memcg->move_charge_at_immigrate = val;
3378 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3379 struct cftype *cft, u64 val)
3386 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3390 unsigned int lru_mask;
3393 static const struct numa_stat stats[] = {
3394 { "total", LRU_ALL },
3395 { "file", LRU_ALL_FILE },
3396 { "anon", LRU_ALL_ANON },
3397 { "unevictable", BIT(LRU_UNEVICTABLE) },
3399 const struct numa_stat *stat;
3402 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3404 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3405 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3406 seq_printf(m, "%s=%lu", stat->name, nr);
3407 for_each_node_state(nid, N_MEMORY) {
3408 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3410 seq_printf(m, " N%d=%lu", nid, nr);
3415 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3416 struct mem_cgroup *iter;
3419 for_each_mem_cgroup_tree(iter, memcg)
3420 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3421 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3422 for_each_node_state(nid, N_MEMORY) {
3424 for_each_mem_cgroup_tree(iter, memcg)
3425 nr += mem_cgroup_node_nr_lru_pages(
3426 iter, nid, stat->lru_mask);
3427 seq_printf(m, " N%d=%lu", nid, nr);
3434 #endif /* CONFIG_NUMA */
3436 /* Universal VM events cgroup1 shows, original sort order */
3437 static const unsigned int memcg1_events[] = {
3444 static const char *const memcg1_event_names[] = {
3451 static int memcg_stat_show(struct seq_file *m, void *v)
3453 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3454 unsigned long memory, memsw;
3455 struct mem_cgroup *mi;
3457 struct accumulated_stats acc;
3459 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3460 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3462 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3463 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3465 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3466 memcg_page_state(memcg, memcg1_stats[i]) *
3470 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3471 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3472 memcg_sum_events(memcg, memcg1_events[i]));
3474 for (i = 0; i < NR_LRU_LISTS; i++)
3475 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3476 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3478 /* Hierarchical information */
3479 memory = memsw = PAGE_COUNTER_MAX;
3480 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3481 memory = min(memory, mi->memory.max);
3482 memsw = min(memsw, mi->memsw.max);
3484 seq_printf(m, "hierarchical_memory_limit %llu\n",
3485 (u64)memory * PAGE_SIZE);
3486 if (do_memsw_account())
3487 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3488 (u64)memsw * PAGE_SIZE);
3490 memset(&acc, 0, sizeof(acc));
3491 acc.stats_size = ARRAY_SIZE(memcg1_stats);
3492 acc.stats_array = memcg1_stats;
3493 acc.events_size = ARRAY_SIZE(memcg1_events);
3494 acc.events_array = memcg1_events;
3495 accumulate_memcg_tree(memcg, &acc);
3497 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3498 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3500 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3501 (u64)acc.stat[i] * PAGE_SIZE);
3504 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3505 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3506 (u64)acc.events[i]);
3508 for (i = 0; i < NR_LRU_LISTS; i++)
3509 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3510 (u64)acc.lru_pages[i] * PAGE_SIZE);
3512 #ifdef CONFIG_DEBUG_VM
3515 struct mem_cgroup_per_node *mz;
3516 struct zone_reclaim_stat *rstat;
3517 unsigned long recent_rotated[2] = {0, 0};
3518 unsigned long recent_scanned[2] = {0, 0};
3520 for_each_online_pgdat(pgdat) {
3521 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3522 rstat = &mz->lruvec.reclaim_stat;
3524 recent_rotated[0] += rstat->recent_rotated[0];
3525 recent_rotated[1] += rstat->recent_rotated[1];
3526 recent_scanned[0] += rstat->recent_scanned[0];
3527 recent_scanned[1] += rstat->recent_scanned[1];
3529 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3530 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3531 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3532 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3539 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3542 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3544 return mem_cgroup_swappiness(memcg);
3547 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3548 struct cftype *cft, u64 val)
3550 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3556 memcg->swappiness = val;
3558 vm_swappiness = val;
3563 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3565 struct mem_cgroup_threshold_ary *t;
3566 unsigned long usage;
3571 t = rcu_dereference(memcg->thresholds.primary);
3573 t = rcu_dereference(memcg->memsw_thresholds.primary);
3578 usage = mem_cgroup_usage(memcg, swap);
3581 * current_threshold points to threshold just below or equal to usage.
3582 * If it's not true, a threshold was crossed after last
3583 * call of __mem_cgroup_threshold().
3585 i = t->current_threshold;
3588 * Iterate backward over array of thresholds starting from
3589 * current_threshold and check if a threshold is crossed.
3590 * If none of thresholds below usage is crossed, we read
3591 * only one element of the array here.
3593 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3594 eventfd_signal(t->entries[i].eventfd, 1);
3596 /* i = current_threshold + 1 */
3600 * Iterate forward over array of thresholds starting from
3601 * current_threshold+1 and check if a threshold is crossed.
3602 * If none of thresholds above usage is crossed, we read
3603 * only one element of the array here.
3605 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3606 eventfd_signal(t->entries[i].eventfd, 1);
3608 /* Update current_threshold */
3609 t->current_threshold = i - 1;
3614 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3617 __mem_cgroup_threshold(memcg, false);
3618 if (do_memsw_account())
3619 __mem_cgroup_threshold(memcg, true);
3621 memcg = parent_mem_cgroup(memcg);
3625 static int compare_thresholds(const void *a, const void *b)
3627 const struct mem_cgroup_threshold *_a = a;
3628 const struct mem_cgroup_threshold *_b = b;
3630 if (_a->threshold > _b->threshold)
3633 if (_a->threshold < _b->threshold)
3639 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3641 struct mem_cgroup_eventfd_list *ev;
3643 spin_lock(&memcg_oom_lock);
3645 list_for_each_entry(ev, &memcg->oom_notify, list)
3646 eventfd_signal(ev->eventfd, 1);
3648 spin_unlock(&memcg_oom_lock);
3652 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3654 struct mem_cgroup *iter;
3656 for_each_mem_cgroup_tree(iter, memcg)
3657 mem_cgroup_oom_notify_cb(iter);
3660 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3661 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3663 struct mem_cgroup_thresholds *thresholds;
3664 struct mem_cgroup_threshold_ary *new;
3665 unsigned long threshold;
3666 unsigned long usage;
3669 ret = page_counter_memparse(args, "-1", &threshold);
3673 mutex_lock(&memcg->thresholds_lock);
3676 thresholds = &memcg->thresholds;
3677 usage = mem_cgroup_usage(memcg, false);
3678 } else if (type == _MEMSWAP) {
3679 thresholds = &memcg->memsw_thresholds;
3680 usage = mem_cgroup_usage(memcg, true);
3684 /* Check if a threshold crossed before adding a new one */
3685 if (thresholds->primary)
3686 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3688 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3690 /* Allocate memory for new array of thresholds */
3691 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3699 /* Copy thresholds (if any) to new array */
3700 if (thresholds->primary) {
3701 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3702 sizeof(struct mem_cgroup_threshold));
3705 /* Add new threshold */
3706 new->entries[size - 1].eventfd = eventfd;
3707 new->entries[size - 1].threshold = threshold;
3709 /* Sort thresholds. Registering of new threshold isn't time-critical */
3710 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3711 compare_thresholds, NULL);
3713 /* Find current threshold */
3714 new->current_threshold = -1;
3715 for (i = 0; i < size; i++) {
3716 if (new->entries[i].threshold <= usage) {
3718 * new->current_threshold will not be used until
3719 * rcu_assign_pointer(), so it's safe to increment
3722 ++new->current_threshold;
3727 /* Free old spare buffer and save old primary buffer as spare */
3728 kfree(thresholds->spare);
3729 thresholds->spare = thresholds->primary;
3731 rcu_assign_pointer(thresholds->primary, new);
3733 /* To be sure that nobody uses thresholds */
3737 mutex_unlock(&memcg->thresholds_lock);
3742 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3743 struct eventfd_ctx *eventfd, const char *args)
3745 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3748 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3749 struct eventfd_ctx *eventfd, const char *args)
3751 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3754 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3755 struct eventfd_ctx *eventfd, enum res_type type)
3757 struct mem_cgroup_thresholds *thresholds;
3758 struct mem_cgroup_threshold_ary *new;
3759 unsigned long usage;
3762 mutex_lock(&memcg->thresholds_lock);
3765 thresholds = &memcg->thresholds;
3766 usage = mem_cgroup_usage(memcg, false);
3767 } else if (type == _MEMSWAP) {
3768 thresholds = &memcg->memsw_thresholds;
3769 usage = mem_cgroup_usage(memcg, true);
3773 if (!thresholds->primary)
3776 /* Check if a threshold crossed before removing */
3777 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3779 /* Calculate new number of threshold */
3781 for (i = 0; i < thresholds->primary->size; i++) {
3782 if (thresholds->primary->entries[i].eventfd != eventfd)
3786 new = thresholds->spare;
3788 /* Set thresholds array to NULL if we don't have thresholds */
3797 /* Copy thresholds and find current threshold */
3798 new->current_threshold = -1;
3799 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3800 if (thresholds->primary->entries[i].eventfd == eventfd)
3803 new->entries[j] = thresholds->primary->entries[i];
3804 if (new->entries[j].threshold <= usage) {
3806 * new->current_threshold will not be used
3807 * until rcu_assign_pointer(), so it's safe to increment
3810 ++new->current_threshold;
3816 /* Swap primary and spare array */
3817 thresholds->spare = thresholds->primary;
3819 rcu_assign_pointer(thresholds->primary, new);
3821 /* To be sure that nobody uses thresholds */
3824 /* If all events are unregistered, free the spare array */
3826 kfree(thresholds->spare);
3827 thresholds->spare = NULL;
3830 mutex_unlock(&memcg->thresholds_lock);
3833 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3834 struct eventfd_ctx *eventfd)
3836 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3839 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3840 struct eventfd_ctx *eventfd)
3842 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3845 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3846 struct eventfd_ctx *eventfd, const char *args)
3848 struct mem_cgroup_eventfd_list *event;
3850 event = kmalloc(sizeof(*event), GFP_KERNEL);
3854 spin_lock(&memcg_oom_lock);
3856 event->eventfd = eventfd;
3857 list_add(&event->list, &memcg->oom_notify);
3859 /* already in OOM ? */
3860 if (memcg->under_oom)
3861 eventfd_signal(eventfd, 1);
3862 spin_unlock(&memcg_oom_lock);
3867 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3868 struct eventfd_ctx *eventfd)
3870 struct mem_cgroup_eventfd_list *ev, *tmp;
3872 spin_lock(&memcg_oom_lock);
3874 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3875 if (ev->eventfd == eventfd) {
3876 list_del(&ev->list);
3881 spin_unlock(&memcg_oom_lock);
3884 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3886 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3888 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3889 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3890 seq_printf(sf, "oom_kill %lu\n",
3891 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3895 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3896 struct cftype *cft, u64 val)
3898 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3900 /* cannot set to root cgroup and only 0 and 1 are allowed */
3901 if (!css->parent || !((val == 0) || (val == 1)))
3904 memcg->oom_kill_disable = val;
3906 memcg_oom_recover(memcg);
3911 #ifdef CONFIG_CGROUP_WRITEBACK
3913 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3915 return wb_domain_init(&memcg->cgwb_domain, gfp);
3918 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3920 wb_domain_exit(&memcg->cgwb_domain);
3923 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3925 wb_domain_size_changed(&memcg->cgwb_domain);
3928 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3930 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3932 if (!memcg->css.parent)
3935 return &memcg->cgwb_domain;
3939 * idx can be of type enum memcg_stat_item or node_stat_item.
3940 * Keep in sync with memcg_exact_page().
3942 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
3944 long x = atomic_long_read(&memcg->stat[idx]);
3947 for_each_online_cpu(cpu)
3948 x += per_cpu_ptr(memcg->stat_cpu, cpu)->count[idx];
3955 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3956 * @wb: bdi_writeback in question
3957 * @pfilepages: out parameter for number of file pages
3958 * @pheadroom: out parameter for number of allocatable pages according to memcg
3959 * @pdirty: out parameter for number of dirty pages
3960 * @pwriteback: out parameter for number of pages under writeback
3962 * Determine the numbers of file, headroom, dirty, and writeback pages in
3963 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3964 * is a bit more involved.
3966 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3967 * headroom is calculated as the lowest headroom of itself and the
3968 * ancestors. Note that this doesn't consider the actual amount of
3969 * available memory in the system. The caller should further cap
3970 * *@pheadroom accordingly.
3972 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3973 unsigned long *pheadroom, unsigned long *pdirty,
3974 unsigned long *pwriteback)
3976 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3977 struct mem_cgroup *parent;
3979 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
3981 /* this should eventually include NR_UNSTABLE_NFS */
3982 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
3983 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3984 (1 << LRU_ACTIVE_FILE));
3985 *pheadroom = PAGE_COUNTER_MAX;
3987 while ((parent = parent_mem_cgroup(memcg))) {
3988 unsigned long ceiling = min(memcg->memory.max, memcg->high);
3989 unsigned long used = page_counter_read(&memcg->memory);
3991 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3996 #else /* CONFIG_CGROUP_WRITEBACK */
3998 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4003 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4007 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4011 #endif /* CONFIG_CGROUP_WRITEBACK */
4014 * DO NOT USE IN NEW FILES.
4016 * "cgroup.event_control" implementation.
4018 * This is way over-engineered. It tries to support fully configurable
4019 * events for each user. Such level of flexibility is completely
4020 * unnecessary especially in the light of the planned unified hierarchy.
4022 * Please deprecate this and replace with something simpler if at all
4027 * Unregister event and free resources.
4029 * Gets called from workqueue.
4031 static void memcg_event_remove(struct work_struct *work)
4033 struct mem_cgroup_event *event =
4034 container_of(work, struct mem_cgroup_event, remove);
4035 struct mem_cgroup *memcg = event->memcg;
4037 remove_wait_queue(event->wqh, &event->wait);
4039 event->unregister_event(memcg, event->eventfd);
4041 /* Notify userspace the event is going away. */
4042 eventfd_signal(event->eventfd, 1);
4044 eventfd_ctx_put(event->eventfd);
4046 css_put(&memcg->css);
4050 * Gets called on EPOLLHUP on eventfd when user closes it.
4052 * Called with wqh->lock held and interrupts disabled.
4054 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4055 int sync, void *key)
4057 struct mem_cgroup_event *event =
4058 container_of(wait, struct mem_cgroup_event, wait);
4059 struct mem_cgroup *memcg = event->memcg;
4060 __poll_t flags = key_to_poll(key);
4062 if (flags & EPOLLHUP) {
4064 * If the event has been detached at cgroup removal, we
4065 * can simply return knowing the other side will cleanup
4068 * We can't race against event freeing since the other
4069 * side will require wqh->lock via remove_wait_queue(),
4072 spin_lock(&memcg->event_list_lock);
4073 if (!list_empty(&event->list)) {
4074 list_del_init(&event->list);
4076 * We are in atomic context, but cgroup_event_remove()
4077 * may sleep, so we have to call it in workqueue.
4079 schedule_work(&event->remove);
4081 spin_unlock(&memcg->event_list_lock);
4087 static void memcg_event_ptable_queue_proc(struct file *file,
4088 wait_queue_head_t *wqh, poll_table *pt)
4090 struct mem_cgroup_event *event =
4091 container_of(pt, struct mem_cgroup_event, pt);
4094 add_wait_queue(wqh, &event->wait);
4098 * DO NOT USE IN NEW FILES.
4100 * Parse input and register new cgroup event handler.
4102 * Input must be in format '<event_fd> <control_fd> <args>'.
4103 * Interpretation of args is defined by control file implementation.
4105 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4106 char *buf, size_t nbytes, loff_t off)
4108 struct cgroup_subsys_state *css = of_css(of);
4109 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4110 struct mem_cgroup_event *event;
4111 struct cgroup_subsys_state *cfile_css;
4112 unsigned int efd, cfd;
4119 buf = strstrip(buf);
4121 efd = simple_strtoul(buf, &endp, 10);
4126 cfd = simple_strtoul(buf, &endp, 10);
4127 if ((*endp != ' ') && (*endp != '\0'))
4131 event = kzalloc(sizeof(*event), GFP_KERNEL);
4135 event->memcg = memcg;
4136 INIT_LIST_HEAD(&event->list);
4137 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4138 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4139 INIT_WORK(&event->remove, memcg_event_remove);
4147 event->eventfd = eventfd_ctx_fileget(efile.file);
4148 if (IS_ERR(event->eventfd)) {
4149 ret = PTR_ERR(event->eventfd);
4156 goto out_put_eventfd;
4159 /* the process need read permission on control file */
4160 /* AV: shouldn't we check that it's been opened for read instead? */
4161 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4166 * Determine the event callbacks and set them in @event. This used
4167 * to be done via struct cftype but cgroup core no longer knows
4168 * about these events. The following is crude but the whole thing
4169 * is for compatibility anyway.
4171 * DO NOT ADD NEW FILES.
4173 name = cfile.file->f_path.dentry->d_name.name;
4175 if (!strcmp(name, "memory.usage_in_bytes")) {
4176 event->register_event = mem_cgroup_usage_register_event;
4177 event->unregister_event = mem_cgroup_usage_unregister_event;
4178 } else if (!strcmp(name, "memory.oom_control")) {
4179 event->register_event = mem_cgroup_oom_register_event;
4180 event->unregister_event = mem_cgroup_oom_unregister_event;
4181 } else if (!strcmp(name, "memory.pressure_level")) {
4182 event->register_event = vmpressure_register_event;
4183 event->unregister_event = vmpressure_unregister_event;
4184 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4185 event->register_event = memsw_cgroup_usage_register_event;
4186 event->unregister_event = memsw_cgroup_usage_unregister_event;
4193 * Verify @cfile should belong to @css. Also, remaining events are
4194 * automatically removed on cgroup destruction but the removal is
4195 * asynchronous, so take an extra ref on @css.
4197 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4198 &memory_cgrp_subsys);
4200 if (IS_ERR(cfile_css))
4202 if (cfile_css != css) {
4207 ret = event->register_event(memcg, event->eventfd, buf);
4211 vfs_poll(efile.file, &event->pt);
4213 spin_lock(&memcg->event_list_lock);
4214 list_add(&event->list, &memcg->event_list);
4215 spin_unlock(&memcg->event_list_lock);
4227 eventfd_ctx_put(event->eventfd);
4236 static struct cftype mem_cgroup_legacy_files[] = {
4238 .name = "usage_in_bytes",
4239 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4240 .read_u64 = mem_cgroup_read_u64,
4243 .name = "max_usage_in_bytes",
4244 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4245 .write = mem_cgroup_reset,
4246 .read_u64 = mem_cgroup_read_u64,
4249 .name = "limit_in_bytes",
4250 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4251 .write = mem_cgroup_write,
4252 .read_u64 = mem_cgroup_read_u64,
4255 .name = "soft_limit_in_bytes",
4256 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4257 .write = mem_cgroup_write,
4258 .read_u64 = mem_cgroup_read_u64,
4262 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4263 .write = mem_cgroup_reset,
4264 .read_u64 = mem_cgroup_read_u64,
4268 .seq_show = memcg_stat_show,
4271 .name = "force_empty",
4272 .write = mem_cgroup_force_empty_write,
4275 .name = "use_hierarchy",
4276 .write_u64 = mem_cgroup_hierarchy_write,
4277 .read_u64 = mem_cgroup_hierarchy_read,
4280 .name = "cgroup.event_control", /* XXX: for compat */
4281 .write = memcg_write_event_control,
4282 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4285 .name = "swappiness",
4286 .read_u64 = mem_cgroup_swappiness_read,
4287 .write_u64 = mem_cgroup_swappiness_write,
4290 .name = "move_charge_at_immigrate",
4291 .read_u64 = mem_cgroup_move_charge_read,
4292 .write_u64 = mem_cgroup_move_charge_write,
4295 .name = "oom_control",
4296 .seq_show = mem_cgroup_oom_control_read,
4297 .write_u64 = mem_cgroup_oom_control_write,
4298 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4301 .name = "pressure_level",
4305 .name = "numa_stat",
4306 .seq_show = memcg_numa_stat_show,
4310 .name = "kmem.limit_in_bytes",
4311 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4312 .write = mem_cgroup_write,
4313 .read_u64 = mem_cgroup_read_u64,
4316 .name = "kmem.usage_in_bytes",
4317 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4318 .read_u64 = mem_cgroup_read_u64,
4321 .name = "kmem.failcnt",
4322 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4323 .write = mem_cgroup_reset,
4324 .read_u64 = mem_cgroup_read_u64,
4327 .name = "kmem.max_usage_in_bytes",
4328 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4329 .write = mem_cgroup_reset,
4330 .read_u64 = mem_cgroup_read_u64,
4332 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4334 .name = "kmem.slabinfo",
4335 .seq_start = memcg_slab_start,
4336 .seq_next = memcg_slab_next,
4337 .seq_stop = memcg_slab_stop,
4338 .seq_show = memcg_slab_show,
4342 .name = "kmem.tcp.limit_in_bytes",
4343 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4344 .write = mem_cgroup_write,
4345 .read_u64 = mem_cgroup_read_u64,
4348 .name = "kmem.tcp.usage_in_bytes",
4349 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4350 .read_u64 = mem_cgroup_read_u64,
4353 .name = "kmem.tcp.failcnt",
4354 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4355 .write = mem_cgroup_reset,
4356 .read_u64 = mem_cgroup_read_u64,
4359 .name = "kmem.tcp.max_usage_in_bytes",
4360 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4361 .write = mem_cgroup_reset,
4362 .read_u64 = mem_cgroup_read_u64,
4364 { }, /* terminate */
4368 * Private memory cgroup IDR
4370 * Swap-out records and page cache shadow entries need to store memcg
4371 * references in constrained space, so we maintain an ID space that is
4372 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4373 * memory-controlled cgroups to 64k.
4375 * However, there usually are many references to the oflline CSS after
4376 * the cgroup has been destroyed, such as page cache or reclaimable
4377 * slab objects, that don't need to hang on to the ID. We want to keep
4378 * those dead CSS from occupying IDs, or we might quickly exhaust the
4379 * relatively small ID space and prevent the creation of new cgroups
4380 * even when there are much fewer than 64k cgroups - possibly none.
4382 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4383 * be freed and recycled when it's no longer needed, which is usually
4384 * when the CSS is offlined.
4386 * The only exception to that are records of swapped out tmpfs/shmem
4387 * pages that need to be attributed to live ancestors on swapin. But
4388 * those references are manageable from userspace.
4391 static DEFINE_IDR(mem_cgroup_idr);
4393 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4395 if (memcg->id.id > 0) {
4396 idr_remove(&mem_cgroup_idr, memcg->id.id);
4401 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4403 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4404 atomic_add(n, &memcg->id.ref);
4407 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4409 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4410 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4411 mem_cgroup_id_remove(memcg);
4413 /* Memcg ID pins CSS */
4414 css_put(&memcg->css);
4418 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4420 mem_cgroup_id_get_many(memcg, 1);
4423 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4425 mem_cgroup_id_put_many(memcg, 1);
4429 * mem_cgroup_from_id - look up a memcg from a memcg id
4430 * @id: the memcg id to look up
4432 * Caller must hold rcu_read_lock().
4434 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4436 WARN_ON_ONCE(!rcu_read_lock_held());
4437 return idr_find(&mem_cgroup_idr, id);
4440 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4442 struct mem_cgroup_per_node *pn;
4445 * This routine is called against possible nodes.
4446 * But it's BUG to call kmalloc() against offline node.
4448 * TODO: this routine can waste much memory for nodes which will
4449 * never be onlined. It's better to use memory hotplug callback
4452 if (!node_state(node, N_NORMAL_MEMORY))
4454 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4458 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4459 if (!pn->lruvec_stat_cpu) {
4464 lruvec_init(&pn->lruvec);
4465 pn->usage_in_excess = 0;
4466 pn->on_tree = false;
4469 memcg->nodeinfo[node] = pn;
4473 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4475 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4480 free_percpu(pn->lruvec_stat_cpu);
4484 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4489 free_mem_cgroup_per_node_info(memcg, node);
4490 free_percpu(memcg->stat_cpu);
4494 static void mem_cgroup_free(struct mem_cgroup *memcg)
4496 memcg_wb_domain_exit(memcg);
4497 __mem_cgroup_free(memcg);
4500 static struct mem_cgroup *mem_cgroup_alloc(void)
4502 struct mem_cgroup *memcg;
4506 size = sizeof(struct mem_cgroup);
4507 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4509 memcg = kzalloc(size, GFP_KERNEL);
4513 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4514 1, MEM_CGROUP_ID_MAX,
4516 if (memcg->id.id < 0)
4519 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4520 if (!memcg->stat_cpu)
4524 if (alloc_mem_cgroup_per_node_info(memcg, node))
4527 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4530 INIT_WORK(&memcg->high_work, high_work_func);
4531 memcg->last_scanned_node = MAX_NUMNODES;
4532 INIT_LIST_HEAD(&memcg->oom_notify);
4533 mutex_init(&memcg->thresholds_lock);
4534 spin_lock_init(&memcg->move_lock);
4535 vmpressure_init(&memcg->vmpressure);
4536 INIT_LIST_HEAD(&memcg->event_list);
4537 spin_lock_init(&memcg->event_list_lock);
4538 memcg->socket_pressure = jiffies;
4539 #ifdef CONFIG_MEMCG_KMEM
4540 memcg->kmemcg_id = -1;
4542 #ifdef CONFIG_CGROUP_WRITEBACK
4543 INIT_LIST_HEAD(&memcg->cgwb_list);
4545 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4548 mem_cgroup_id_remove(memcg);
4549 __mem_cgroup_free(memcg);
4553 static struct cgroup_subsys_state * __ref
4554 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4556 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4557 struct mem_cgroup *memcg;
4558 long error = -ENOMEM;
4560 memcg = mem_cgroup_alloc();
4562 return ERR_PTR(error);
4564 memcg->high = PAGE_COUNTER_MAX;
4565 memcg->soft_limit = PAGE_COUNTER_MAX;
4567 memcg->swappiness = mem_cgroup_swappiness(parent);
4568 memcg->oom_kill_disable = parent->oom_kill_disable;
4570 if (parent && parent->use_hierarchy) {
4571 memcg->use_hierarchy = true;
4572 page_counter_init(&memcg->memory, &parent->memory);
4573 page_counter_init(&memcg->swap, &parent->swap);
4574 page_counter_init(&memcg->memsw, &parent->memsw);
4575 page_counter_init(&memcg->kmem, &parent->kmem);
4576 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4578 page_counter_init(&memcg->memory, NULL);
4579 page_counter_init(&memcg->swap, NULL);
4580 page_counter_init(&memcg->memsw, NULL);
4581 page_counter_init(&memcg->kmem, NULL);
4582 page_counter_init(&memcg->tcpmem, NULL);
4584 * Deeper hierachy with use_hierarchy == false doesn't make
4585 * much sense so let cgroup subsystem know about this
4586 * unfortunate state in our controller.
4588 if (parent != root_mem_cgroup)
4589 memory_cgrp_subsys.broken_hierarchy = true;
4592 /* The following stuff does not apply to the root */
4594 root_mem_cgroup = memcg;
4598 error = memcg_online_kmem(memcg);
4602 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4603 static_branch_inc(&memcg_sockets_enabled_key);
4607 mem_cgroup_id_remove(memcg);
4608 mem_cgroup_free(memcg);
4609 return ERR_PTR(-ENOMEM);
4612 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4614 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4617 * A memcg must be visible for memcg_expand_shrinker_maps()
4618 * by the time the maps are allocated. So, we allocate maps
4619 * here, when for_each_mem_cgroup() can't skip it.
4621 if (memcg_alloc_shrinker_maps(memcg)) {
4622 mem_cgroup_id_remove(memcg);
4626 /* Online state pins memcg ID, memcg ID pins CSS */
4627 atomic_set(&memcg->id.ref, 1);
4632 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4634 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4635 struct mem_cgroup_event *event, *tmp;
4638 * Unregister events and notify userspace.
4639 * Notify userspace about cgroup removing only after rmdir of cgroup
4640 * directory to avoid race between userspace and kernelspace.
4642 spin_lock(&memcg->event_list_lock);
4643 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4644 list_del_init(&event->list);
4645 schedule_work(&event->remove);
4647 spin_unlock(&memcg->event_list_lock);
4649 page_counter_set_min(&memcg->memory, 0);
4650 page_counter_set_low(&memcg->memory, 0);
4652 memcg_offline_kmem(memcg);
4653 wb_memcg_offline(memcg);
4655 mem_cgroup_id_put(memcg);
4658 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4660 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4662 invalidate_reclaim_iterators(memcg);
4665 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4667 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4669 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4670 static_branch_dec(&memcg_sockets_enabled_key);
4672 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4673 static_branch_dec(&memcg_sockets_enabled_key);
4675 vmpressure_cleanup(&memcg->vmpressure);
4676 cancel_work_sync(&memcg->high_work);
4677 mem_cgroup_remove_from_trees(memcg);
4678 memcg_free_shrinker_maps(memcg);
4679 memcg_free_kmem(memcg);
4680 mem_cgroup_free(memcg);
4684 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4685 * @css: the target css
4687 * Reset the states of the mem_cgroup associated with @css. This is
4688 * invoked when the userland requests disabling on the default hierarchy
4689 * but the memcg is pinned through dependency. The memcg should stop
4690 * applying policies and should revert to the vanilla state as it may be
4691 * made visible again.
4693 * The current implementation only resets the essential configurations.
4694 * This needs to be expanded to cover all the visible parts.
4696 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4698 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4700 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4701 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4702 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4703 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4704 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4705 page_counter_set_min(&memcg->memory, 0);
4706 page_counter_set_low(&memcg->memory, 0);
4707 memcg->high = PAGE_COUNTER_MAX;
4708 memcg->soft_limit = PAGE_COUNTER_MAX;
4709 memcg_wb_domain_size_changed(memcg);
4713 /* Handlers for move charge at task migration. */
4714 static int mem_cgroup_do_precharge(unsigned long count)
4718 /* Try a single bulk charge without reclaim first, kswapd may wake */
4719 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4721 mc.precharge += count;
4725 /* Try charges one by one with reclaim, but do not retry */
4727 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4741 enum mc_target_type {
4748 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4749 unsigned long addr, pte_t ptent)
4751 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4753 if (!page || !page_mapped(page))
4755 if (PageAnon(page)) {
4756 if (!(mc.flags & MOVE_ANON))
4759 if (!(mc.flags & MOVE_FILE))
4762 if (!get_page_unless_zero(page))
4768 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4769 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4770 pte_t ptent, swp_entry_t *entry)
4772 struct page *page = NULL;
4773 swp_entry_t ent = pte_to_swp_entry(ptent);
4775 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4779 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4780 * a device and because they are not accessible by CPU they are store
4781 * as special swap entry in the CPU page table.
4783 if (is_device_private_entry(ent)) {
4784 page = device_private_entry_to_page(ent);
4786 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4787 * a refcount of 1 when free (unlike normal page)
4789 if (!page_ref_add_unless(page, 1, 1))
4795 * Because lookup_swap_cache() updates some statistics counter,
4796 * we call find_get_page() with swapper_space directly.
4798 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4799 if (do_memsw_account())
4800 entry->val = ent.val;
4805 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4806 pte_t ptent, swp_entry_t *entry)
4812 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4813 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4815 struct page *page = NULL;
4816 struct address_space *mapping;
4819 if (!vma->vm_file) /* anonymous vma */
4821 if (!(mc.flags & MOVE_FILE))
4824 mapping = vma->vm_file->f_mapping;
4825 pgoff = linear_page_index(vma, addr);
4827 /* page is moved even if it's not RSS of this task(page-faulted). */
4829 /* shmem/tmpfs may report page out on swap: account for that too. */
4830 if (shmem_mapping(mapping)) {
4831 page = find_get_entry(mapping, pgoff);
4832 if (radix_tree_exceptional_entry(page)) {
4833 swp_entry_t swp = radix_to_swp_entry(page);
4834 if (do_memsw_account())
4836 page = find_get_page(swap_address_space(swp),
4840 page = find_get_page(mapping, pgoff);
4842 page = find_get_page(mapping, pgoff);
4848 * mem_cgroup_move_account - move account of the page
4850 * @compound: charge the page as compound or small page
4851 * @from: mem_cgroup which the page is moved from.
4852 * @to: mem_cgroup which the page is moved to. @from != @to.
4854 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4856 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4859 static int mem_cgroup_move_account(struct page *page,
4861 struct mem_cgroup *from,
4862 struct mem_cgroup *to)
4864 unsigned long flags;
4865 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4869 VM_BUG_ON(from == to);
4870 VM_BUG_ON_PAGE(PageLRU(page), page);
4871 VM_BUG_ON(compound && !PageTransHuge(page));
4874 * Prevent mem_cgroup_migrate() from looking at
4875 * page->mem_cgroup of its source page while we change it.
4878 if (!trylock_page(page))
4882 if (page->mem_cgroup != from)
4885 anon = PageAnon(page);
4887 spin_lock_irqsave(&from->move_lock, flags);
4889 if (!anon && page_mapped(page)) {
4890 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4891 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4895 * move_lock grabbed above and caller set from->moving_account, so
4896 * mod_memcg_page_state will serialize updates to PageDirty.
4897 * So mapping should be stable for dirty pages.
4899 if (!anon && PageDirty(page)) {
4900 struct address_space *mapping = page_mapping(page);
4902 if (mapping_cap_account_dirty(mapping)) {
4903 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4904 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4908 if (PageWriteback(page)) {
4909 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4910 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4914 * It is safe to change page->mem_cgroup here because the page
4915 * is referenced, charged, and isolated - we can't race with
4916 * uncharging, charging, migration, or LRU putback.
4919 /* caller should have done css_get */
4920 page->mem_cgroup = to;
4921 spin_unlock_irqrestore(&from->move_lock, flags);
4925 local_irq_disable();
4926 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4927 memcg_check_events(to, page);
4928 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4929 memcg_check_events(from, page);
4938 * get_mctgt_type - get target type of moving charge
4939 * @vma: the vma the pte to be checked belongs
4940 * @addr: the address corresponding to the pte to be checked
4941 * @ptent: the pte to be checked
4942 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4945 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4946 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4947 * move charge. if @target is not NULL, the page is stored in target->page
4948 * with extra refcnt got(Callers should handle it).
4949 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4950 * target for charge migration. if @target is not NULL, the entry is stored
4952 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4953 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4954 * For now we such page is charge like a regular page would be as for all
4955 * intent and purposes it is just special memory taking the place of a
4958 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4960 * Called with pte lock held.
4963 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4964 unsigned long addr, pte_t ptent, union mc_target *target)
4966 struct page *page = NULL;
4967 enum mc_target_type ret = MC_TARGET_NONE;
4968 swp_entry_t ent = { .val = 0 };
4970 if (pte_present(ptent))
4971 page = mc_handle_present_pte(vma, addr, ptent);
4972 else if (is_swap_pte(ptent))
4973 page = mc_handle_swap_pte(vma, ptent, &ent);
4974 else if (pte_none(ptent))
4975 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4977 if (!page && !ent.val)
4981 * Do only loose check w/o serialization.
4982 * mem_cgroup_move_account() checks the page is valid or
4983 * not under LRU exclusion.
4985 if (page->mem_cgroup == mc.from) {
4986 ret = MC_TARGET_PAGE;
4987 if (is_device_private_page(page) ||
4988 is_device_public_page(page))
4989 ret = MC_TARGET_DEVICE;
4991 target->page = page;
4993 if (!ret || !target)
4997 * There is a swap entry and a page doesn't exist or isn't charged.
4998 * But we cannot move a tail-page in a THP.
5000 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5001 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5002 ret = MC_TARGET_SWAP;
5009 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5011 * We don't consider PMD mapped swapping or file mapped pages because THP does
5012 * not support them for now.
5013 * Caller should make sure that pmd_trans_huge(pmd) is true.
5015 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5016 unsigned long addr, pmd_t pmd, union mc_target *target)
5018 struct page *page = NULL;
5019 enum mc_target_type ret = MC_TARGET_NONE;
5021 if (unlikely(is_swap_pmd(pmd))) {
5022 VM_BUG_ON(thp_migration_supported() &&
5023 !is_pmd_migration_entry(pmd));
5026 page = pmd_page(pmd);
5027 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5028 if (!(mc.flags & MOVE_ANON))
5030 if (page->mem_cgroup == mc.from) {
5031 ret = MC_TARGET_PAGE;
5034 target->page = page;
5040 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5041 unsigned long addr, pmd_t pmd, union mc_target *target)
5043 return MC_TARGET_NONE;
5047 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5048 unsigned long addr, unsigned long end,
5049 struct mm_walk *walk)
5051 struct vm_area_struct *vma = walk->vma;
5055 ptl = pmd_trans_huge_lock(pmd, vma);
5058 * Note their can not be MC_TARGET_DEVICE for now as we do not
5059 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5060 * MEMORY_DEVICE_PRIVATE but this might change.
5062 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5063 mc.precharge += HPAGE_PMD_NR;
5068 if (pmd_trans_unstable(pmd))
5070 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5071 for (; addr != end; pte++, addr += PAGE_SIZE)
5072 if (get_mctgt_type(vma, addr, *pte, NULL))
5073 mc.precharge++; /* increment precharge temporarily */
5074 pte_unmap_unlock(pte - 1, ptl);
5080 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5082 unsigned long precharge;
5084 struct mm_walk mem_cgroup_count_precharge_walk = {
5085 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5088 down_read(&mm->mmap_sem);
5089 walk_page_range(0, mm->highest_vm_end,
5090 &mem_cgroup_count_precharge_walk);
5091 up_read(&mm->mmap_sem);
5093 precharge = mc.precharge;
5099 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5101 unsigned long precharge = mem_cgroup_count_precharge(mm);
5103 VM_BUG_ON(mc.moving_task);
5104 mc.moving_task = current;
5105 return mem_cgroup_do_precharge(precharge);
5108 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5109 static void __mem_cgroup_clear_mc(void)
5111 struct mem_cgroup *from = mc.from;
5112 struct mem_cgroup *to = mc.to;
5114 /* we must uncharge all the leftover precharges from mc.to */
5116 cancel_charge(mc.to, mc.precharge);
5120 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5121 * we must uncharge here.
5123 if (mc.moved_charge) {
5124 cancel_charge(mc.from, mc.moved_charge);
5125 mc.moved_charge = 0;
5127 /* we must fixup refcnts and charges */
5128 if (mc.moved_swap) {
5129 /* uncharge swap account from the old cgroup */
5130 if (!mem_cgroup_is_root(mc.from))
5131 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5133 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5136 * we charged both to->memory and to->memsw, so we
5137 * should uncharge to->memory.
5139 if (!mem_cgroup_is_root(mc.to))
5140 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5142 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5143 css_put_many(&mc.to->css, mc.moved_swap);
5147 memcg_oom_recover(from);
5148 memcg_oom_recover(to);
5149 wake_up_all(&mc.waitq);
5152 static void mem_cgroup_clear_mc(void)
5154 struct mm_struct *mm = mc.mm;
5157 * we must clear moving_task before waking up waiters at the end of
5160 mc.moving_task = NULL;
5161 __mem_cgroup_clear_mc();
5162 spin_lock(&mc.lock);
5166 spin_unlock(&mc.lock);
5171 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5173 struct cgroup_subsys_state *css;
5174 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5175 struct mem_cgroup *from;
5176 struct task_struct *leader, *p;
5177 struct mm_struct *mm;
5178 unsigned long move_flags;
5181 /* charge immigration isn't supported on the default hierarchy */
5182 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5186 * Multi-process migrations only happen on the default hierarchy
5187 * where charge immigration is not used. Perform charge
5188 * immigration if @tset contains a leader and whine if there are
5192 cgroup_taskset_for_each_leader(leader, css, tset) {
5195 memcg = mem_cgroup_from_css(css);
5201 * We are now commited to this value whatever it is. Changes in this
5202 * tunable will only affect upcoming migrations, not the current one.
5203 * So we need to save it, and keep it going.
5205 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5209 from = mem_cgroup_from_task(p);
5211 VM_BUG_ON(from == memcg);
5213 mm = get_task_mm(p);
5216 /* We move charges only when we move a owner of the mm */
5217 if (mm->owner == p) {
5220 VM_BUG_ON(mc.precharge);
5221 VM_BUG_ON(mc.moved_charge);
5222 VM_BUG_ON(mc.moved_swap);
5224 spin_lock(&mc.lock);
5228 mc.flags = move_flags;
5229 spin_unlock(&mc.lock);
5230 /* We set mc.moving_task later */
5232 ret = mem_cgroup_precharge_mc(mm);
5234 mem_cgroup_clear_mc();
5241 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5244 mem_cgroup_clear_mc();
5247 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5248 unsigned long addr, unsigned long end,
5249 struct mm_walk *walk)
5252 struct vm_area_struct *vma = walk->vma;
5255 enum mc_target_type target_type;
5256 union mc_target target;
5259 ptl = pmd_trans_huge_lock(pmd, vma);
5261 if (mc.precharge < HPAGE_PMD_NR) {
5265 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5266 if (target_type == MC_TARGET_PAGE) {
5268 if (!isolate_lru_page(page)) {
5269 if (!mem_cgroup_move_account(page, true,
5271 mc.precharge -= HPAGE_PMD_NR;
5272 mc.moved_charge += HPAGE_PMD_NR;
5274 putback_lru_page(page);
5277 } else if (target_type == MC_TARGET_DEVICE) {
5279 if (!mem_cgroup_move_account(page, true,
5281 mc.precharge -= HPAGE_PMD_NR;
5282 mc.moved_charge += HPAGE_PMD_NR;
5290 if (pmd_trans_unstable(pmd))
5293 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5294 for (; addr != end; addr += PAGE_SIZE) {
5295 pte_t ptent = *(pte++);
5296 bool device = false;
5302 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5303 case MC_TARGET_DEVICE:
5306 case MC_TARGET_PAGE:
5309 * We can have a part of the split pmd here. Moving it
5310 * can be done but it would be too convoluted so simply
5311 * ignore such a partial THP and keep it in original
5312 * memcg. There should be somebody mapping the head.
5314 if (PageTransCompound(page))
5316 if (!device && isolate_lru_page(page))
5318 if (!mem_cgroup_move_account(page, false,
5321 /* we uncharge from mc.from later. */
5325 putback_lru_page(page);
5326 put: /* get_mctgt_type() gets the page */
5329 case MC_TARGET_SWAP:
5331 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5333 /* we fixup refcnts and charges later. */
5341 pte_unmap_unlock(pte - 1, ptl);
5346 * We have consumed all precharges we got in can_attach().
5347 * We try charge one by one, but don't do any additional
5348 * charges to mc.to if we have failed in charge once in attach()
5351 ret = mem_cgroup_do_precharge(1);
5359 static void mem_cgroup_move_charge(void)
5361 struct mm_walk mem_cgroup_move_charge_walk = {
5362 .pmd_entry = mem_cgroup_move_charge_pte_range,
5366 lru_add_drain_all();
5368 * Signal lock_page_memcg() to take the memcg's move_lock
5369 * while we're moving its pages to another memcg. Then wait
5370 * for already started RCU-only updates to finish.
5372 atomic_inc(&mc.from->moving_account);
5375 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5377 * Someone who are holding the mmap_sem might be waiting in
5378 * waitq. So we cancel all extra charges, wake up all waiters,
5379 * and retry. Because we cancel precharges, we might not be able
5380 * to move enough charges, but moving charge is a best-effort
5381 * feature anyway, so it wouldn't be a big problem.
5383 __mem_cgroup_clear_mc();
5388 * When we have consumed all precharges and failed in doing
5389 * additional charge, the page walk just aborts.
5391 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5393 up_read(&mc.mm->mmap_sem);
5394 atomic_dec(&mc.from->moving_account);
5397 static void mem_cgroup_move_task(void)
5400 mem_cgroup_move_charge();
5401 mem_cgroup_clear_mc();
5404 #else /* !CONFIG_MMU */
5405 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5409 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5412 static void mem_cgroup_move_task(void)
5418 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5419 * to verify whether we're attached to the default hierarchy on each mount
5422 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5425 * use_hierarchy is forced on the default hierarchy. cgroup core
5426 * guarantees that @root doesn't have any children, so turning it
5427 * on for the root memcg is enough.
5429 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5430 root_mem_cgroup->use_hierarchy = true;
5432 root_mem_cgroup->use_hierarchy = false;
5435 static u64 memory_current_read(struct cgroup_subsys_state *css,
5438 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5440 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5443 static int memory_min_show(struct seq_file *m, void *v)
5445 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5446 unsigned long min = READ_ONCE(memcg->memory.min);
5448 if (min == PAGE_COUNTER_MAX)
5449 seq_puts(m, "max\n");
5451 seq_printf(m, "%llu\n", (u64)min * PAGE_SIZE);
5456 static ssize_t memory_min_write(struct kernfs_open_file *of,
5457 char *buf, size_t nbytes, loff_t off)
5459 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5463 buf = strstrip(buf);
5464 err = page_counter_memparse(buf, "max", &min);
5468 page_counter_set_min(&memcg->memory, min);
5473 static int memory_low_show(struct seq_file *m, void *v)
5475 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5476 unsigned long low = READ_ONCE(memcg->memory.low);
5478 if (low == PAGE_COUNTER_MAX)
5479 seq_puts(m, "max\n");
5481 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5486 static ssize_t memory_low_write(struct kernfs_open_file *of,
5487 char *buf, size_t nbytes, loff_t off)
5489 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5493 buf = strstrip(buf);
5494 err = page_counter_memparse(buf, "max", &low);
5498 page_counter_set_low(&memcg->memory, low);
5503 static int memory_high_show(struct seq_file *m, void *v)
5505 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5506 unsigned long high = READ_ONCE(memcg->high);
5508 if (high == PAGE_COUNTER_MAX)
5509 seq_puts(m, "max\n");
5511 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5516 static ssize_t memory_high_write(struct kernfs_open_file *of,
5517 char *buf, size_t nbytes, loff_t off)
5519 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5520 unsigned long nr_pages;
5524 buf = strstrip(buf);
5525 err = page_counter_memparse(buf, "max", &high);
5531 nr_pages = page_counter_read(&memcg->memory);
5532 if (nr_pages > high)
5533 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5536 memcg_wb_domain_size_changed(memcg);
5540 static int memory_max_show(struct seq_file *m, void *v)
5542 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5543 unsigned long max = READ_ONCE(memcg->memory.max);
5545 if (max == PAGE_COUNTER_MAX)
5546 seq_puts(m, "max\n");
5548 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5553 static ssize_t memory_max_write(struct kernfs_open_file *of,
5554 char *buf, size_t nbytes, loff_t off)
5556 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5557 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5558 bool drained = false;
5562 buf = strstrip(buf);
5563 err = page_counter_memparse(buf, "max", &max);
5567 xchg(&memcg->memory.max, max);
5570 unsigned long nr_pages = page_counter_read(&memcg->memory);
5572 if (nr_pages <= max)
5575 if (signal_pending(current)) {
5581 drain_all_stock(memcg);
5587 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5593 memcg_memory_event(memcg, MEMCG_OOM);
5594 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5598 memcg_wb_domain_size_changed(memcg);
5602 static int memory_events_show(struct seq_file *m, void *v)
5604 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5606 seq_printf(m, "low %lu\n",
5607 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5608 seq_printf(m, "high %lu\n",
5609 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5610 seq_printf(m, "max %lu\n",
5611 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5612 seq_printf(m, "oom %lu\n",
5613 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5614 seq_printf(m, "oom_kill %lu\n",
5615 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5620 static int memory_stat_show(struct seq_file *m, void *v)
5622 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5623 struct accumulated_stats acc;
5627 * Provide statistics on the state of the memory subsystem as
5628 * well as cumulative event counters that show past behavior.
5630 * This list is ordered following a combination of these gradients:
5631 * 1) generic big picture -> specifics and details
5632 * 2) reflecting userspace activity -> reflecting kernel heuristics
5634 * Current memory state:
5637 memset(&acc, 0, sizeof(acc));
5638 acc.stats_size = MEMCG_NR_STAT;
5639 acc.events_size = NR_VM_EVENT_ITEMS;
5640 accumulate_memcg_tree(memcg, &acc);
5642 seq_printf(m, "anon %llu\n",
5643 (u64)acc.stat[MEMCG_RSS] * PAGE_SIZE);
5644 seq_printf(m, "file %llu\n",
5645 (u64)acc.stat[MEMCG_CACHE] * PAGE_SIZE);
5646 seq_printf(m, "kernel_stack %llu\n",
5647 (u64)acc.stat[MEMCG_KERNEL_STACK_KB] * 1024);
5648 seq_printf(m, "slab %llu\n",
5649 (u64)(acc.stat[NR_SLAB_RECLAIMABLE] +
5650 acc.stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5651 seq_printf(m, "sock %llu\n",
5652 (u64)acc.stat[MEMCG_SOCK] * PAGE_SIZE);
5654 seq_printf(m, "shmem %llu\n",
5655 (u64)acc.stat[NR_SHMEM] * PAGE_SIZE);
5656 seq_printf(m, "file_mapped %llu\n",
5657 (u64)acc.stat[NR_FILE_MAPPED] * PAGE_SIZE);
5658 seq_printf(m, "file_dirty %llu\n",
5659 (u64)acc.stat[NR_FILE_DIRTY] * PAGE_SIZE);
5660 seq_printf(m, "file_writeback %llu\n",
5661 (u64)acc.stat[NR_WRITEBACK] * PAGE_SIZE);
5663 for (i = 0; i < NR_LRU_LISTS; i++)
5664 seq_printf(m, "%s %llu\n", mem_cgroup_lru_names[i],
5665 (u64)acc.lru_pages[i] * PAGE_SIZE);
5667 seq_printf(m, "slab_reclaimable %llu\n",
5668 (u64)acc.stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5669 seq_printf(m, "slab_unreclaimable %llu\n",
5670 (u64)acc.stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5672 /* Accumulated memory events */
5674 seq_printf(m, "pgfault %lu\n", acc.events[PGFAULT]);
5675 seq_printf(m, "pgmajfault %lu\n", acc.events[PGMAJFAULT]);
5677 seq_printf(m, "pgrefill %lu\n", acc.events[PGREFILL]);
5678 seq_printf(m, "pgscan %lu\n", acc.events[PGSCAN_KSWAPD] +
5679 acc.events[PGSCAN_DIRECT]);
5680 seq_printf(m, "pgsteal %lu\n", acc.events[PGSTEAL_KSWAPD] +
5681 acc.events[PGSTEAL_DIRECT]);
5682 seq_printf(m, "pgactivate %lu\n", acc.events[PGACTIVATE]);
5683 seq_printf(m, "pgdeactivate %lu\n", acc.events[PGDEACTIVATE]);
5684 seq_printf(m, "pglazyfree %lu\n", acc.events[PGLAZYFREE]);
5685 seq_printf(m, "pglazyfreed %lu\n", acc.events[PGLAZYFREED]);
5687 seq_printf(m, "workingset_refault %lu\n",
5688 acc.stat[WORKINGSET_REFAULT]);
5689 seq_printf(m, "workingset_activate %lu\n",
5690 acc.stat[WORKINGSET_ACTIVATE]);
5691 seq_printf(m, "workingset_nodereclaim %lu\n",
5692 acc.stat[WORKINGSET_NODERECLAIM]);
5697 static int memory_oom_group_show(struct seq_file *m, void *v)
5699 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5701 seq_printf(m, "%d\n", memcg->oom_group);
5706 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5707 char *buf, size_t nbytes, loff_t off)
5709 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5712 buf = strstrip(buf);
5716 ret = kstrtoint(buf, 0, &oom_group);
5720 if (oom_group != 0 && oom_group != 1)
5723 memcg->oom_group = oom_group;
5728 static struct cftype memory_files[] = {
5731 .flags = CFTYPE_NOT_ON_ROOT,
5732 .read_u64 = memory_current_read,
5736 .flags = CFTYPE_NOT_ON_ROOT,
5737 .seq_show = memory_min_show,
5738 .write = memory_min_write,
5742 .flags = CFTYPE_NOT_ON_ROOT,
5743 .seq_show = memory_low_show,
5744 .write = memory_low_write,
5748 .flags = CFTYPE_NOT_ON_ROOT,
5749 .seq_show = memory_high_show,
5750 .write = memory_high_write,
5754 .flags = CFTYPE_NOT_ON_ROOT,
5755 .seq_show = memory_max_show,
5756 .write = memory_max_write,
5760 .flags = CFTYPE_NOT_ON_ROOT,
5761 .file_offset = offsetof(struct mem_cgroup, events_file),
5762 .seq_show = memory_events_show,
5766 .flags = CFTYPE_NOT_ON_ROOT,
5767 .seq_show = memory_stat_show,
5770 .name = "oom.group",
5771 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5772 .seq_show = memory_oom_group_show,
5773 .write = memory_oom_group_write,
5778 struct cgroup_subsys memory_cgrp_subsys = {
5779 .css_alloc = mem_cgroup_css_alloc,
5780 .css_online = mem_cgroup_css_online,
5781 .css_offline = mem_cgroup_css_offline,
5782 .css_released = mem_cgroup_css_released,
5783 .css_free = mem_cgroup_css_free,
5784 .css_reset = mem_cgroup_css_reset,
5785 .can_attach = mem_cgroup_can_attach,
5786 .cancel_attach = mem_cgroup_cancel_attach,
5787 .post_attach = mem_cgroup_move_task,
5788 .bind = mem_cgroup_bind,
5789 .dfl_cftypes = memory_files,
5790 .legacy_cftypes = mem_cgroup_legacy_files,
5795 * mem_cgroup_protected - check if memory consumption is in the normal range
5796 * @root: the top ancestor of the sub-tree being checked
5797 * @memcg: the memory cgroup to check
5799 * WARNING: This function is not stateless! It can only be used as part
5800 * of a top-down tree iteration, not for isolated queries.
5802 * Returns one of the following:
5803 * MEMCG_PROT_NONE: cgroup memory is not protected
5804 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5805 * an unprotected supply of reclaimable memory from other cgroups.
5806 * MEMCG_PROT_MIN: cgroup memory is protected
5808 * @root is exclusive; it is never protected when looked at directly
5810 * To provide a proper hierarchical behavior, effective memory.min/low values
5811 * are used. Below is the description of how effective memory.low is calculated.
5812 * Effective memory.min values is calculated in the same way.
5814 * Effective memory.low is always equal or less than the original memory.low.
5815 * If there is no memory.low overcommittment (which is always true for
5816 * top-level memory cgroups), these two values are equal.
5817 * Otherwise, it's a part of parent's effective memory.low,
5818 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5819 * memory.low usages, where memory.low usage is the size of actually
5823 * elow = min( memory.low, parent->elow * ------------------ ),
5824 * siblings_low_usage
5826 * | memory.current, if memory.current < memory.low
5831 * Such definition of the effective memory.low provides the expected
5832 * hierarchical behavior: parent's memory.low value is limiting
5833 * children, unprotected memory is reclaimed first and cgroups,
5834 * which are not using their guarantee do not affect actual memory
5837 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5839 * A A/memory.low = 2G, A/memory.current = 6G
5841 * BC DE B/memory.low = 3G B/memory.current = 2G
5842 * C/memory.low = 1G C/memory.current = 2G
5843 * D/memory.low = 0 D/memory.current = 2G
5844 * E/memory.low = 10G E/memory.current = 0
5846 * and the memory pressure is applied, the following memory distribution
5847 * is expected (approximately):
5849 * A/memory.current = 2G
5851 * B/memory.current = 1.3G
5852 * C/memory.current = 0.6G
5853 * D/memory.current = 0
5854 * E/memory.current = 0
5856 * These calculations require constant tracking of the actual low usages
5857 * (see propagate_protected_usage()), as well as recursive calculation of
5858 * effective memory.low values. But as we do call mem_cgroup_protected()
5859 * path for each memory cgroup top-down from the reclaim,
5860 * it's possible to optimize this part, and save calculated elow
5861 * for next usage. This part is intentionally racy, but it's ok,
5862 * as memory.low is a best-effort mechanism.
5864 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5865 struct mem_cgroup *memcg)
5867 struct mem_cgroup *parent;
5868 unsigned long emin, parent_emin;
5869 unsigned long elow, parent_elow;
5870 unsigned long usage;
5872 if (mem_cgroup_disabled())
5873 return MEMCG_PROT_NONE;
5876 root = root_mem_cgroup;
5878 return MEMCG_PROT_NONE;
5880 usage = page_counter_read(&memcg->memory);
5882 return MEMCG_PROT_NONE;
5884 emin = memcg->memory.min;
5885 elow = memcg->memory.low;
5887 parent = parent_mem_cgroup(memcg);
5888 /* No parent means a non-hierarchical mode on v1 memcg */
5890 return MEMCG_PROT_NONE;
5895 parent_emin = READ_ONCE(parent->memory.emin);
5896 emin = min(emin, parent_emin);
5897 if (emin && parent_emin) {
5898 unsigned long min_usage, siblings_min_usage;
5900 min_usage = min(usage, memcg->memory.min);
5901 siblings_min_usage = atomic_long_read(
5902 &parent->memory.children_min_usage);
5904 if (min_usage && siblings_min_usage)
5905 emin = min(emin, parent_emin * min_usage /
5906 siblings_min_usage);
5909 parent_elow = READ_ONCE(parent->memory.elow);
5910 elow = min(elow, parent_elow);
5911 if (elow && parent_elow) {
5912 unsigned long low_usage, siblings_low_usage;
5914 low_usage = min(usage, memcg->memory.low);
5915 siblings_low_usage = atomic_long_read(
5916 &parent->memory.children_low_usage);
5918 if (low_usage && siblings_low_usage)
5919 elow = min(elow, parent_elow * low_usage /
5920 siblings_low_usage);
5924 memcg->memory.emin = emin;
5925 memcg->memory.elow = elow;
5928 return MEMCG_PROT_MIN;
5929 else if (usage <= elow)
5930 return MEMCG_PROT_LOW;
5932 return MEMCG_PROT_NONE;
5936 * mem_cgroup_try_charge - try charging a page
5937 * @page: page to charge
5938 * @mm: mm context of the victim
5939 * @gfp_mask: reclaim mode
5940 * @memcgp: charged memcg return
5941 * @compound: charge the page as compound or small page
5943 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5944 * pages according to @gfp_mask if necessary.
5946 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5947 * Otherwise, an error code is returned.
5949 * After page->mapping has been set up, the caller must finalize the
5950 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5951 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5953 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5954 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5957 struct mem_cgroup *memcg = NULL;
5958 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5961 if (mem_cgroup_disabled())
5964 if (PageSwapCache(page)) {
5966 * Every swap fault against a single page tries to charge the
5967 * page, bail as early as possible. shmem_unuse() encounters
5968 * already charged pages, too. The USED bit is protected by
5969 * the page lock, which serializes swap cache removal, which
5970 * in turn serializes uncharging.
5972 VM_BUG_ON_PAGE(!PageLocked(page), page);
5973 if (compound_head(page)->mem_cgroup)
5976 if (do_swap_account) {
5977 swp_entry_t ent = { .val = page_private(page), };
5978 unsigned short id = lookup_swap_cgroup_id(ent);
5981 memcg = mem_cgroup_from_id(id);
5982 if (memcg && !css_tryget_online(&memcg->css))
5989 memcg = get_mem_cgroup_from_mm(mm);
5991 ret = try_charge(memcg, gfp_mask, nr_pages);
5993 css_put(&memcg->css);
5999 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6000 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6003 struct mem_cgroup *memcg;
6006 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6008 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6013 * mem_cgroup_commit_charge - commit a page charge
6014 * @page: page to charge
6015 * @memcg: memcg to charge the page to
6016 * @lrucare: page might be on LRU already
6017 * @compound: charge the page as compound or small page
6019 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6020 * after page->mapping has been set up. This must happen atomically
6021 * as part of the page instantiation, i.e. under the page table lock
6022 * for anonymous pages, under the page lock for page and swap cache.
6024 * In addition, the page must not be on the LRU during the commit, to
6025 * prevent racing with task migration. If it might be, use @lrucare.
6027 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6029 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6030 bool lrucare, bool compound)
6032 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6034 VM_BUG_ON_PAGE(!page->mapping, page);
6035 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6037 if (mem_cgroup_disabled())
6040 * Swap faults will attempt to charge the same page multiple
6041 * times. But reuse_swap_page() might have removed the page
6042 * from swapcache already, so we can't check PageSwapCache().
6047 commit_charge(page, memcg, lrucare);
6049 local_irq_disable();
6050 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6051 memcg_check_events(memcg, page);
6054 if (do_memsw_account() && PageSwapCache(page)) {
6055 swp_entry_t entry = { .val = page_private(page) };
6057 * The swap entry might not get freed for a long time,
6058 * let's not wait for it. The page already received a
6059 * memory+swap charge, drop the swap entry duplicate.
6061 mem_cgroup_uncharge_swap(entry, nr_pages);
6066 * mem_cgroup_cancel_charge - cancel a page charge
6067 * @page: page to charge
6068 * @memcg: memcg to charge the page to
6069 * @compound: charge the page as compound or small page
6071 * Cancel a charge transaction started by mem_cgroup_try_charge().
6073 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6076 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6078 if (mem_cgroup_disabled())
6081 * Swap faults will attempt to charge the same page multiple
6082 * times. But reuse_swap_page() might have removed the page
6083 * from swapcache already, so we can't check PageSwapCache().
6088 cancel_charge(memcg, nr_pages);
6091 struct uncharge_gather {
6092 struct mem_cgroup *memcg;
6093 unsigned long pgpgout;
6094 unsigned long nr_anon;
6095 unsigned long nr_file;
6096 unsigned long nr_kmem;
6097 unsigned long nr_huge;
6098 unsigned long nr_shmem;
6099 struct page *dummy_page;
6102 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6104 memset(ug, 0, sizeof(*ug));
6107 static void uncharge_batch(const struct uncharge_gather *ug)
6109 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6110 unsigned long flags;
6112 if (!mem_cgroup_is_root(ug->memcg)) {
6113 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6114 if (do_memsw_account())
6115 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6116 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6117 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6118 memcg_oom_recover(ug->memcg);
6121 local_irq_save(flags);
6122 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6123 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6124 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6125 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6126 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6127 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
6128 memcg_check_events(ug->memcg, ug->dummy_page);
6129 local_irq_restore(flags);
6131 if (!mem_cgroup_is_root(ug->memcg))
6132 css_put_many(&ug->memcg->css, nr_pages);
6135 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6137 VM_BUG_ON_PAGE(PageLRU(page), page);
6138 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6139 !PageHWPoison(page) , page);
6141 if (!page->mem_cgroup)
6145 * Nobody should be changing or seriously looking at
6146 * page->mem_cgroup at this point, we have fully
6147 * exclusive access to the page.
6150 if (ug->memcg != page->mem_cgroup) {
6153 uncharge_gather_clear(ug);
6155 ug->memcg = page->mem_cgroup;
6158 if (!PageKmemcg(page)) {
6159 unsigned int nr_pages = 1;
6161 if (PageTransHuge(page)) {
6162 nr_pages <<= compound_order(page);
6163 ug->nr_huge += nr_pages;
6166 ug->nr_anon += nr_pages;
6168 ug->nr_file += nr_pages;
6169 if (PageSwapBacked(page))
6170 ug->nr_shmem += nr_pages;
6174 ug->nr_kmem += 1 << compound_order(page);
6175 __ClearPageKmemcg(page);
6178 ug->dummy_page = page;
6179 page->mem_cgroup = NULL;
6182 static void uncharge_list(struct list_head *page_list)
6184 struct uncharge_gather ug;
6185 struct list_head *next;
6187 uncharge_gather_clear(&ug);
6190 * Note that the list can be a single page->lru; hence the
6191 * do-while loop instead of a simple list_for_each_entry().
6193 next = page_list->next;
6197 page = list_entry(next, struct page, lru);
6198 next = page->lru.next;
6200 uncharge_page(page, &ug);
6201 } while (next != page_list);
6204 uncharge_batch(&ug);
6208 * mem_cgroup_uncharge - uncharge a page
6209 * @page: page to uncharge
6211 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6212 * mem_cgroup_commit_charge().
6214 void mem_cgroup_uncharge(struct page *page)
6216 struct uncharge_gather ug;
6218 if (mem_cgroup_disabled())
6221 /* Don't touch page->lru of any random page, pre-check: */
6222 if (!page->mem_cgroup)
6225 uncharge_gather_clear(&ug);
6226 uncharge_page(page, &ug);
6227 uncharge_batch(&ug);
6231 * mem_cgroup_uncharge_list - uncharge a list of page
6232 * @page_list: list of pages to uncharge
6234 * Uncharge a list of pages previously charged with
6235 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6237 void mem_cgroup_uncharge_list(struct list_head *page_list)
6239 if (mem_cgroup_disabled())
6242 if (!list_empty(page_list))
6243 uncharge_list(page_list);
6247 * mem_cgroup_migrate - charge a page's replacement
6248 * @oldpage: currently circulating page
6249 * @newpage: replacement page
6251 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6252 * be uncharged upon free.
6254 * Both pages must be locked, @newpage->mapping must be set up.
6256 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6258 struct mem_cgroup *memcg;
6259 unsigned int nr_pages;
6261 unsigned long flags;
6263 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6264 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6265 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6266 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6269 if (mem_cgroup_disabled())
6272 /* Page cache replacement: new page already charged? */
6273 if (newpage->mem_cgroup)
6276 /* Swapcache readahead pages can get replaced before being charged */
6277 memcg = oldpage->mem_cgroup;
6281 /* Force-charge the new page. The old one will be freed soon */
6282 compound = PageTransHuge(newpage);
6283 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6285 page_counter_charge(&memcg->memory, nr_pages);
6286 if (do_memsw_account())
6287 page_counter_charge(&memcg->memsw, nr_pages);
6288 css_get_many(&memcg->css, nr_pages);
6290 commit_charge(newpage, memcg, false);
6292 local_irq_save(flags);
6293 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6294 memcg_check_events(memcg, newpage);
6295 local_irq_restore(flags);
6298 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6299 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6301 void mem_cgroup_sk_alloc(struct sock *sk)
6303 struct mem_cgroup *memcg;
6305 if (!mem_cgroup_sockets_enabled)
6309 * Socket cloning can throw us here with sk_memcg already
6310 * filled. It won't however, necessarily happen from
6311 * process context. So the test for root memcg given
6312 * the current task's memcg won't help us in this case.
6314 * Respecting the original socket's memcg is a better
6315 * decision in this case.
6318 css_get(&sk->sk_memcg->css);
6323 memcg = mem_cgroup_from_task(current);
6324 if (memcg == root_mem_cgroup)
6326 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6328 if (css_tryget_online(&memcg->css))
6329 sk->sk_memcg = memcg;
6334 void mem_cgroup_sk_free(struct sock *sk)
6337 css_put(&sk->sk_memcg->css);
6341 * mem_cgroup_charge_skmem - charge socket memory
6342 * @memcg: memcg to charge
6343 * @nr_pages: number of pages to charge
6345 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6346 * @memcg's configured limit, %false if the charge had to be forced.
6348 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6350 gfp_t gfp_mask = GFP_KERNEL;
6352 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6353 struct page_counter *fail;
6355 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6356 memcg->tcpmem_pressure = 0;
6359 page_counter_charge(&memcg->tcpmem, nr_pages);
6360 memcg->tcpmem_pressure = 1;
6364 /* Don't block in the packet receive path */
6366 gfp_mask = GFP_NOWAIT;
6368 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6370 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6373 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6378 * mem_cgroup_uncharge_skmem - uncharge socket memory
6379 * @memcg: memcg to uncharge
6380 * @nr_pages: number of pages to uncharge
6382 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6384 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6385 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6389 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6391 refill_stock(memcg, nr_pages);
6394 static int __init cgroup_memory(char *s)
6398 while ((token = strsep(&s, ",")) != NULL) {
6401 if (!strcmp(token, "nosocket"))
6402 cgroup_memory_nosocket = true;
6403 if (!strcmp(token, "nokmem"))
6404 cgroup_memory_nokmem = true;
6408 __setup("cgroup.memory=", cgroup_memory);
6411 * subsys_initcall() for memory controller.
6413 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6414 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6415 * basically everything that doesn't depend on a specific mem_cgroup structure
6416 * should be initialized from here.
6418 static int __init mem_cgroup_init(void)
6422 #ifdef CONFIG_MEMCG_KMEM
6424 * Kmem cache creation is mostly done with the slab_mutex held,
6425 * so use a workqueue with limited concurrency to avoid stalling
6426 * all worker threads in case lots of cgroups are created and
6427 * destroyed simultaneously.
6429 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6430 BUG_ON(!memcg_kmem_cache_wq);
6433 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6434 memcg_hotplug_cpu_dead);
6436 for_each_possible_cpu(cpu)
6437 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6440 for_each_node(node) {
6441 struct mem_cgroup_tree_per_node *rtpn;
6443 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6444 node_online(node) ? node : NUMA_NO_NODE);
6446 rtpn->rb_root = RB_ROOT;
6447 rtpn->rb_rightmost = NULL;
6448 spin_lock_init(&rtpn->lock);
6449 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6454 subsys_initcall(mem_cgroup_init);
6456 #ifdef CONFIG_MEMCG_SWAP
6457 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6459 while (!atomic_inc_not_zero(&memcg->id.ref)) {
6461 * The root cgroup cannot be destroyed, so it's refcount must
6464 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6468 memcg = parent_mem_cgroup(memcg);
6470 memcg = root_mem_cgroup;
6476 * mem_cgroup_swapout - transfer a memsw charge to swap
6477 * @page: page whose memsw charge to transfer
6478 * @entry: swap entry to move the charge to
6480 * Transfer the memsw charge of @page to @entry.
6482 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6484 struct mem_cgroup *memcg, *swap_memcg;
6485 unsigned int nr_entries;
6486 unsigned short oldid;
6488 VM_BUG_ON_PAGE(PageLRU(page), page);
6489 VM_BUG_ON_PAGE(page_count(page), page);
6491 if (!do_memsw_account())
6494 memcg = page->mem_cgroup;
6496 /* Readahead page, never charged */
6501 * In case the memcg owning these pages has been offlined and doesn't
6502 * have an ID allocated to it anymore, charge the closest online
6503 * ancestor for the swap instead and transfer the memory+swap charge.
6505 swap_memcg = mem_cgroup_id_get_online(memcg);
6506 nr_entries = hpage_nr_pages(page);
6507 /* Get references for the tail pages, too */
6509 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6510 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6512 VM_BUG_ON_PAGE(oldid, page);
6513 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6515 page->mem_cgroup = NULL;
6517 if (!mem_cgroup_is_root(memcg))
6518 page_counter_uncharge(&memcg->memory, nr_entries);
6520 if (memcg != swap_memcg) {
6521 if (!mem_cgroup_is_root(swap_memcg))
6522 page_counter_charge(&swap_memcg->memsw, nr_entries);
6523 page_counter_uncharge(&memcg->memsw, nr_entries);
6527 * Interrupts should be disabled here because the caller holds the
6528 * i_pages lock which is taken with interrupts-off. It is
6529 * important here to have the interrupts disabled because it is the
6530 * only synchronisation we have for updating the per-CPU variables.
6532 VM_BUG_ON(!irqs_disabled());
6533 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6535 memcg_check_events(memcg, page);
6537 if (!mem_cgroup_is_root(memcg))
6538 css_put_many(&memcg->css, nr_entries);
6542 * mem_cgroup_try_charge_swap - try charging swap space for a page
6543 * @page: page being added to swap
6544 * @entry: swap entry to charge
6546 * Try to charge @page's memcg for the swap space at @entry.
6548 * Returns 0 on success, -ENOMEM on failure.
6550 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6552 unsigned int nr_pages = hpage_nr_pages(page);
6553 struct page_counter *counter;
6554 struct mem_cgroup *memcg;
6555 unsigned short oldid;
6557 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6560 memcg = page->mem_cgroup;
6562 /* Readahead page, never charged */
6567 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6571 memcg = mem_cgroup_id_get_online(memcg);
6573 if (!mem_cgroup_is_root(memcg) &&
6574 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6575 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6576 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6577 mem_cgroup_id_put(memcg);
6581 /* Get references for the tail pages, too */
6583 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6584 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6585 VM_BUG_ON_PAGE(oldid, page);
6586 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6592 * mem_cgroup_uncharge_swap - uncharge swap space
6593 * @entry: swap entry to uncharge
6594 * @nr_pages: the amount of swap space to uncharge
6596 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6598 struct mem_cgroup *memcg;
6601 if (!do_swap_account)
6604 id = swap_cgroup_record(entry, 0, nr_pages);
6606 memcg = mem_cgroup_from_id(id);
6608 if (!mem_cgroup_is_root(memcg)) {
6609 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6610 page_counter_uncharge(&memcg->swap, nr_pages);
6612 page_counter_uncharge(&memcg->memsw, nr_pages);
6614 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6615 mem_cgroup_id_put_many(memcg, nr_pages);
6620 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6622 long nr_swap_pages = get_nr_swap_pages();
6624 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6625 return nr_swap_pages;
6626 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6627 nr_swap_pages = min_t(long, nr_swap_pages,
6628 READ_ONCE(memcg->swap.max) -
6629 page_counter_read(&memcg->swap));
6630 return nr_swap_pages;
6633 bool mem_cgroup_swap_full(struct page *page)
6635 struct mem_cgroup *memcg;
6637 VM_BUG_ON_PAGE(!PageLocked(page), page);
6641 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6644 memcg = page->mem_cgroup;
6648 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6649 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6655 /* for remember boot option*/
6656 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6657 static int really_do_swap_account __initdata = 1;
6659 static int really_do_swap_account __initdata;
6662 static int __init enable_swap_account(char *s)
6664 if (!strcmp(s, "1"))
6665 really_do_swap_account = 1;
6666 else if (!strcmp(s, "0"))
6667 really_do_swap_account = 0;
6670 __setup("swapaccount=", enable_swap_account);
6672 static u64 swap_current_read(struct cgroup_subsys_state *css,
6675 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6677 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6680 static int swap_max_show(struct seq_file *m, void *v)
6682 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6683 unsigned long max = READ_ONCE(memcg->swap.max);
6685 if (max == PAGE_COUNTER_MAX)
6686 seq_puts(m, "max\n");
6688 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6693 static ssize_t swap_max_write(struct kernfs_open_file *of,
6694 char *buf, size_t nbytes, loff_t off)
6696 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6700 buf = strstrip(buf);
6701 err = page_counter_memparse(buf, "max", &max);
6705 xchg(&memcg->swap.max, max);
6710 static int swap_events_show(struct seq_file *m, void *v)
6712 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6714 seq_printf(m, "max %lu\n",
6715 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6716 seq_printf(m, "fail %lu\n",
6717 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6722 static struct cftype swap_files[] = {
6724 .name = "swap.current",
6725 .flags = CFTYPE_NOT_ON_ROOT,
6726 .read_u64 = swap_current_read,
6730 .flags = CFTYPE_NOT_ON_ROOT,
6731 .seq_show = swap_max_show,
6732 .write = swap_max_write,
6735 .name = "swap.events",
6736 .flags = CFTYPE_NOT_ON_ROOT,
6737 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6738 .seq_show = swap_events_show,
6743 static struct cftype memsw_cgroup_files[] = {
6745 .name = "memsw.usage_in_bytes",
6746 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6747 .read_u64 = mem_cgroup_read_u64,
6750 .name = "memsw.max_usage_in_bytes",
6751 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6752 .write = mem_cgroup_reset,
6753 .read_u64 = mem_cgroup_read_u64,
6756 .name = "memsw.limit_in_bytes",
6757 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6758 .write = mem_cgroup_write,
6759 .read_u64 = mem_cgroup_read_u64,
6762 .name = "memsw.failcnt",
6763 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6764 .write = mem_cgroup_reset,
6765 .read_u64 = mem_cgroup_read_u64,
6767 { }, /* terminate */
6770 static int __init mem_cgroup_swap_init(void)
6772 if (!mem_cgroup_disabled() && really_do_swap_account) {
6773 do_swap_account = 1;
6774 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6776 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6777 memsw_cgroup_files));
6781 subsys_initcall(mem_cgroup_swap_init);
6783 #endif /* CONFIG_MEMCG_SWAP */