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 /* Some nice accessors for the vmpressure. */
252 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
255 memcg = root_mem_cgroup;
256 return &memcg->vmpressure;
259 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
261 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
264 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
266 return (memcg == root_mem_cgroup);
269 #ifdef CONFIG_MEMCG_KMEM
271 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
272 * The main reason for not using cgroup id for this:
273 * this works better in sparse environments, where we have a lot of memcgs,
274 * but only a few kmem-limited. Or also, if we have, for instance, 200
275 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
276 * 200 entry array for that.
278 * The current size of the caches array is stored in memcg_nr_cache_ids. It
279 * will double each time we have to increase it.
281 static DEFINE_IDA(memcg_cache_ida);
282 int memcg_nr_cache_ids;
284 /* Protects memcg_nr_cache_ids */
285 static DECLARE_RWSEM(memcg_cache_ids_sem);
287 void memcg_get_cache_ids(void)
289 down_read(&memcg_cache_ids_sem);
292 void memcg_put_cache_ids(void)
294 up_read(&memcg_cache_ids_sem);
298 * MIN_SIZE is different than 1, because we would like to avoid going through
299 * the alloc/free process all the time. In a small machine, 4 kmem-limited
300 * cgroups is a reasonable guess. In the future, it could be a parameter or
301 * tunable, but that is strictly not necessary.
303 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
304 * this constant directly from cgroup, but it is understandable that this is
305 * better kept as an internal representation in cgroup.c. In any case, the
306 * cgrp_id space is not getting any smaller, and we don't have to necessarily
307 * increase ours as well if it increases.
309 #define MEMCG_CACHES_MIN_SIZE 4
310 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
313 * A lot of the calls to the cache allocation functions are expected to be
314 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
315 * conditional to this static branch, we'll have to allow modules that does
316 * kmem_cache_alloc and the such to see this symbol as well
318 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
319 EXPORT_SYMBOL(memcg_kmem_enabled_key);
321 struct workqueue_struct *memcg_kmem_cache_wq;
323 static int memcg_shrinker_map_size;
324 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
326 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
328 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
331 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
332 int size, int old_size)
334 struct memcg_shrinker_map *new, *old;
337 lockdep_assert_held(&memcg_shrinker_map_mutex);
340 old = rcu_dereference_protected(
341 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
342 /* Not yet online memcg */
346 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
350 /* Set all old bits, clear all new bits */
351 memset(new->map, (int)0xff, old_size);
352 memset((void *)new->map + old_size, 0, size - old_size);
354 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
355 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
361 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
363 struct mem_cgroup_per_node *pn;
364 struct memcg_shrinker_map *map;
367 if (mem_cgroup_is_root(memcg))
371 pn = mem_cgroup_nodeinfo(memcg, nid);
372 map = rcu_dereference_protected(pn->shrinker_map, true);
375 rcu_assign_pointer(pn->shrinker_map, NULL);
379 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
381 struct memcg_shrinker_map *map;
382 int nid, size, ret = 0;
384 if (mem_cgroup_is_root(memcg))
387 mutex_lock(&memcg_shrinker_map_mutex);
388 size = memcg_shrinker_map_size;
390 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
392 memcg_free_shrinker_maps(memcg);
396 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
398 mutex_unlock(&memcg_shrinker_map_mutex);
403 int memcg_expand_shrinker_maps(int new_id)
405 int size, old_size, ret = 0;
406 struct mem_cgroup *memcg;
408 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
409 old_size = memcg_shrinker_map_size;
410 if (size <= old_size)
413 mutex_lock(&memcg_shrinker_map_mutex);
414 if (!root_mem_cgroup)
417 for_each_mem_cgroup(memcg) {
418 if (mem_cgroup_is_root(memcg))
420 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
426 memcg_shrinker_map_size = size;
427 mutex_unlock(&memcg_shrinker_map_mutex);
430 #else /* CONFIG_MEMCG_KMEM */
431 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
435 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
436 #endif /* CONFIG_MEMCG_KMEM */
439 * mem_cgroup_css_from_page - css of the memcg associated with a page
440 * @page: page of interest
442 * If memcg is bound to the default hierarchy, css of the memcg associated
443 * with @page is returned. The returned css remains associated with @page
444 * until it is released.
446 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
449 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
451 struct mem_cgroup *memcg;
453 memcg = page->mem_cgroup;
455 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
456 memcg = root_mem_cgroup;
462 * page_cgroup_ino - return inode number of the memcg a page is charged to
465 * Look up the closest online ancestor of the memory cgroup @page is charged to
466 * and return its inode number or 0 if @page is not charged to any cgroup. It
467 * is safe to call this function without holding a reference to @page.
469 * Note, this function is inherently racy, because there is nothing to prevent
470 * the cgroup inode from getting torn down and potentially reallocated a moment
471 * after page_cgroup_ino() returns, so it only should be used by callers that
472 * do not care (such as procfs interfaces).
474 ino_t page_cgroup_ino(struct page *page)
476 struct mem_cgroup *memcg;
477 unsigned long ino = 0;
480 memcg = READ_ONCE(page->mem_cgroup);
481 while (memcg && !(memcg->css.flags & CSS_ONLINE))
482 memcg = parent_mem_cgroup(memcg);
484 ino = cgroup_ino(memcg->css.cgroup);
489 static struct mem_cgroup_per_node *
490 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
492 int nid = page_to_nid(page);
494 return memcg->nodeinfo[nid];
497 static struct mem_cgroup_tree_per_node *
498 soft_limit_tree_node(int nid)
500 return soft_limit_tree.rb_tree_per_node[nid];
503 static struct mem_cgroup_tree_per_node *
504 soft_limit_tree_from_page(struct page *page)
506 int nid = page_to_nid(page);
508 return soft_limit_tree.rb_tree_per_node[nid];
511 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
512 struct mem_cgroup_tree_per_node *mctz,
513 unsigned long new_usage_in_excess)
515 struct rb_node **p = &mctz->rb_root.rb_node;
516 struct rb_node *parent = NULL;
517 struct mem_cgroup_per_node *mz_node;
518 bool rightmost = true;
523 mz->usage_in_excess = new_usage_in_excess;
524 if (!mz->usage_in_excess)
528 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
530 if (mz->usage_in_excess < mz_node->usage_in_excess) {
536 * We can't avoid mem cgroups that are over their soft
537 * limit by the same amount
539 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
544 mctz->rb_rightmost = &mz->tree_node;
546 rb_link_node(&mz->tree_node, parent, p);
547 rb_insert_color(&mz->tree_node, &mctz->rb_root);
551 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
552 struct mem_cgroup_tree_per_node *mctz)
557 if (&mz->tree_node == mctz->rb_rightmost)
558 mctz->rb_rightmost = rb_prev(&mz->tree_node);
560 rb_erase(&mz->tree_node, &mctz->rb_root);
564 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
565 struct mem_cgroup_tree_per_node *mctz)
569 spin_lock_irqsave(&mctz->lock, flags);
570 __mem_cgroup_remove_exceeded(mz, mctz);
571 spin_unlock_irqrestore(&mctz->lock, flags);
574 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
576 unsigned long nr_pages = page_counter_read(&memcg->memory);
577 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
578 unsigned long excess = 0;
580 if (nr_pages > soft_limit)
581 excess = nr_pages - soft_limit;
586 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
588 unsigned long excess;
589 struct mem_cgroup_per_node *mz;
590 struct mem_cgroup_tree_per_node *mctz;
592 mctz = soft_limit_tree_from_page(page);
596 * Necessary to update all ancestors when hierarchy is used.
597 * because their event counter is not touched.
599 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
600 mz = mem_cgroup_page_nodeinfo(memcg, page);
601 excess = soft_limit_excess(memcg);
603 * We have to update the tree if mz is on RB-tree or
604 * mem is over its softlimit.
606 if (excess || mz->on_tree) {
609 spin_lock_irqsave(&mctz->lock, flags);
610 /* if on-tree, remove it */
612 __mem_cgroup_remove_exceeded(mz, mctz);
614 * Insert again. mz->usage_in_excess will be updated.
615 * If excess is 0, no tree ops.
617 __mem_cgroup_insert_exceeded(mz, mctz, excess);
618 spin_unlock_irqrestore(&mctz->lock, flags);
623 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
625 struct mem_cgroup_tree_per_node *mctz;
626 struct mem_cgroup_per_node *mz;
630 mz = mem_cgroup_nodeinfo(memcg, nid);
631 mctz = soft_limit_tree_node(nid);
633 mem_cgroup_remove_exceeded(mz, mctz);
637 static struct mem_cgroup_per_node *
638 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
640 struct mem_cgroup_per_node *mz;
644 if (!mctz->rb_rightmost)
645 goto done; /* Nothing to reclaim from */
647 mz = rb_entry(mctz->rb_rightmost,
648 struct mem_cgroup_per_node, tree_node);
650 * Remove the node now but someone else can add it back,
651 * we will to add it back at the end of reclaim to its correct
652 * position in the tree.
654 __mem_cgroup_remove_exceeded(mz, mctz);
655 if (!soft_limit_excess(mz->memcg) ||
656 !css_tryget_online(&mz->memcg->css))
662 static struct mem_cgroup_per_node *
663 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
665 struct mem_cgroup_per_node *mz;
667 spin_lock_irq(&mctz->lock);
668 mz = __mem_cgroup_largest_soft_limit_node(mctz);
669 spin_unlock_irq(&mctz->lock);
673 static unsigned long memcg_sum_events(struct mem_cgroup *memcg,
676 return atomic_long_read(&memcg->events[event]);
679 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
681 bool compound, int nr_pages)
684 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
685 * counted as CACHE even if it's on ANON LRU.
688 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
690 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
691 if (PageSwapBacked(page))
692 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
696 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
697 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
700 /* pagein of a big page is an event. So, ignore page size */
702 __count_memcg_events(memcg, PGPGIN, 1);
704 __count_memcg_events(memcg, PGPGOUT, 1);
705 nr_pages = -nr_pages; /* for event */
708 __this_cpu_add(memcg->stat_cpu->nr_page_events, nr_pages);
711 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
712 int nid, unsigned int lru_mask)
714 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
715 unsigned long nr = 0;
718 VM_BUG_ON((unsigned)nid >= nr_node_ids);
721 if (!(BIT(lru) & lru_mask))
723 nr += mem_cgroup_get_lru_size(lruvec, lru);
728 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
729 unsigned int lru_mask)
731 unsigned long nr = 0;
734 for_each_node_state(nid, N_MEMORY)
735 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
739 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
740 enum mem_cgroup_events_target target)
742 unsigned long val, next;
744 val = __this_cpu_read(memcg->stat_cpu->nr_page_events);
745 next = __this_cpu_read(memcg->stat_cpu->targets[target]);
746 /* from time_after() in jiffies.h */
747 if ((long)(next - val) < 0) {
749 case MEM_CGROUP_TARGET_THRESH:
750 next = val + THRESHOLDS_EVENTS_TARGET;
752 case MEM_CGROUP_TARGET_SOFTLIMIT:
753 next = val + SOFTLIMIT_EVENTS_TARGET;
755 case MEM_CGROUP_TARGET_NUMAINFO:
756 next = val + NUMAINFO_EVENTS_TARGET;
761 __this_cpu_write(memcg->stat_cpu->targets[target], next);
768 * Check events in order.
771 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
773 /* threshold event is triggered in finer grain than soft limit */
774 if (unlikely(mem_cgroup_event_ratelimit(memcg,
775 MEM_CGROUP_TARGET_THRESH))) {
777 bool do_numainfo __maybe_unused;
779 do_softlimit = mem_cgroup_event_ratelimit(memcg,
780 MEM_CGROUP_TARGET_SOFTLIMIT);
782 do_numainfo = mem_cgroup_event_ratelimit(memcg,
783 MEM_CGROUP_TARGET_NUMAINFO);
785 mem_cgroup_threshold(memcg);
786 if (unlikely(do_softlimit))
787 mem_cgroup_update_tree(memcg, page);
789 if (unlikely(do_numainfo))
790 atomic_inc(&memcg->numainfo_events);
795 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
798 * mm_update_next_owner() may clear mm->owner to NULL
799 * if it races with swapoff, page migration, etc.
800 * So this can be called with p == NULL.
805 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
807 EXPORT_SYMBOL(mem_cgroup_from_task);
810 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
811 * @mm: mm from which memcg should be extracted. It can be NULL.
813 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
814 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
817 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
819 struct mem_cgroup *memcg;
821 if (mem_cgroup_disabled())
827 * Page cache insertions can happen withou an
828 * actual mm context, e.g. during disk probing
829 * on boot, loopback IO, acct() writes etc.
832 memcg = root_mem_cgroup;
834 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
835 if (unlikely(!memcg))
836 memcg = root_mem_cgroup;
838 } while (!css_tryget_online(&memcg->css));
842 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
845 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
846 * @page: page from which memcg should be extracted.
848 * Obtain a reference on page->memcg and returns it if successful. Otherwise
849 * root_mem_cgroup is returned.
851 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
853 struct mem_cgroup *memcg = page->mem_cgroup;
855 if (mem_cgroup_disabled())
859 if (!memcg || !css_tryget_online(&memcg->css))
860 memcg = root_mem_cgroup;
864 EXPORT_SYMBOL(get_mem_cgroup_from_page);
867 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
869 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
871 if (unlikely(current->active_memcg)) {
872 struct mem_cgroup *memcg = root_mem_cgroup;
875 if (css_tryget_online(¤t->active_memcg->css))
876 memcg = current->active_memcg;
880 return get_mem_cgroup_from_mm(current->mm);
884 * mem_cgroup_iter - iterate over memory cgroup hierarchy
885 * @root: hierarchy root
886 * @prev: previously returned memcg, NULL on first invocation
887 * @reclaim: cookie for shared reclaim walks, NULL for full walks
889 * Returns references to children of the hierarchy below @root, or
890 * @root itself, or %NULL after a full round-trip.
892 * Caller must pass the return value in @prev on subsequent
893 * invocations for reference counting, or use mem_cgroup_iter_break()
894 * to cancel a hierarchy walk before the round-trip is complete.
896 * Reclaimers can specify a node and a priority level in @reclaim to
897 * divide up the memcgs in the hierarchy among all concurrent
898 * reclaimers operating on the same node and priority.
900 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
901 struct mem_cgroup *prev,
902 struct mem_cgroup_reclaim_cookie *reclaim)
904 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
905 struct cgroup_subsys_state *css = NULL;
906 struct mem_cgroup *memcg = NULL;
907 struct mem_cgroup *pos = NULL;
909 if (mem_cgroup_disabled())
913 root = root_mem_cgroup;
915 if (prev && !reclaim)
918 if (!root->use_hierarchy && root != root_mem_cgroup) {
927 struct mem_cgroup_per_node *mz;
929 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
930 iter = &mz->iter[reclaim->priority];
932 if (prev && reclaim->generation != iter->generation)
936 pos = READ_ONCE(iter->position);
937 if (!pos || css_tryget(&pos->css))
940 * css reference reached zero, so iter->position will
941 * be cleared by ->css_released. However, we should not
942 * rely on this happening soon, because ->css_released
943 * is called from a work queue, and by busy-waiting we
944 * might block it. So we clear iter->position right
947 (void)cmpxchg(&iter->position, pos, NULL);
955 css = css_next_descendant_pre(css, &root->css);
958 * Reclaimers share the hierarchy walk, and a
959 * new one might jump in right at the end of
960 * the hierarchy - make sure they see at least
961 * one group and restart from the beginning.
969 * Verify the css and acquire a reference. The root
970 * is provided by the caller, so we know it's alive
971 * and kicking, and don't take an extra reference.
973 memcg = mem_cgroup_from_css(css);
975 if (css == &root->css)
986 * The position could have already been updated by a competing
987 * thread, so check that the value hasn't changed since we read
988 * it to avoid reclaiming from the same cgroup twice.
990 (void)cmpxchg(&iter->position, pos, memcg);
998 reclaim->generation = iter->generation;
1004 if (prev && prev != root)
1005 css_put(&prev->css);
1011 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1012 * @root: hierarchy root
1013 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1015 void mem_cgroup_iter_break(struct mem_cgroup *root,
1016 struct mem_cgroup *prev)
1019 root = root_mem_cgroup;
1020 if (prev && prev != root)
1021 css_put(&prev->css);
1024 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1026 struct mem_cgroup *memcg = dead_memcg;
1027 struct mem_cgroup_reclaim_iter *iter;
1028 struct mem_cgroup_per_node *mz;
1032 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1033 for_each_node(nid) {
1034 mz = mem_cgroup_nodeinfo(memcg, nid);
1035 for (i = 0; i <= DEF_PRIORITY; i++) {
1036 iter = &mz->iter[i];
1037 cmpxchg(&iter->position,
1045 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1046 * @memcg: hierarchy root
1047 * @fn: function to call for each task
1048 * @arg: argument passed to @fn
1050 * This function iterates over tasks attached to @memcg or to any of its
1051 * descendants and calls @fn for each task. If @fn returns a non-zero
1052 * value, the function breaks the iteration loop and returns the value.
1053 * Otherwise, it will iterate over all tasks and return 0.
1055 * This function must not be called for the root memory cgroup.
1057 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1058 int (*fn)(struct task_struct *, void *), void *arg)
1060 struct mem_cgroup *iter;
1063 BUG_ON(memcg == root_mem_cgroup);
1065 for_each_mem_cgroup_tree(iter, memcg) {
1066 struct css_task_iter it;
1067 struct task_struct *task;
1069 css_task_iter_start(&iter->css, 0, &it);
1070 while (!ret && (task = css_task_iter_next(&it)))
1071 ret = fn(task, arg);
1072 css_task_iter_end(&it);
1074 mem_cgroup_iter_break(memcg, iter);
1082 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1084 * @pgdat: pgdat of the page
1086 * This function is only safe when following the LRU page isolation
1087 * and putback protocol: the LRU lock must be held, and the page must
1088 * either be PageLRU() or the caller must have isolated/allocated it.
1090 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1092 struct mem_cgroup_per_node *mz;
1093 struct mem_cgroup *memcg;
1094 struct lruvec *lruvec;
1096 if (mem_cgroup_disabled()) {
1097 lruvec = &pgdat->lruvec;
1101 memcg = page->mem_cgroup;
1103 * Swapcache readahead pages are added to the LRU - and
1104 * possibly migrated - before they are charged.
1107 memcg = root_mem_cgroup;
1109 mz = mem_cgroup_page_nodeinfo(memcg, page);
1110 lruvec = &mz->lruvec;
1113 * Since a node can be onlined after the mem_cgroup was created,
1114 * we have to be prepared to initialize lruvec->zone here;
1115 * and if offlined then reonlined, we need to reinitialize it.
1117 if (unlikely(lruvec->pgdat != pgdat))
1118 lruvec->pgdat = pgdat;
1123 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1124 * @lruvec: mem_cgroup per zone lru vector
1125 * @lru: index of lru list the page is sitting on
1126 * @zid: zone id of the accounted pages
1127 * @nr_pages: positive when adding or negative when removing
1129 * This function must be called under lru_lock, just before a page is added
1130 * to or just after a page is removed from an lru list (that ordering being
1131 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1133 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1134 int zid, int nr_pages)
1136 struct mem_cgroup_per_node *mz;
1137 unsigned long *lru_size;
1140 if (mem_cgroup_disabled())
1143 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1144 lru_size = &mz->lru_zone_size[zid][lru];
1147 *lru_size += nr_pages;
1150 if (WARN_ONCE(size < 0,
1151 "%s(%p, %d, %d): lru_size %ld\n",
1152 __func__, lruvec, lru, nr_pages, size)) {
1158 *lru_size += nr_pages;
1161 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1163 struct mem_cgroup *task_memcg;
1164 struct task_struct *p;
1167 p = find_lock_task_mm(task);
1169 task_memcg = get_mem_cgroup_from_mm(p->mm);
1173 * All threads may have already detached their mm's, but the oom
1174 * killer still needs to detect if they have already been oom
1175 * killed to prevent needlessly killing additional tasks.
1178 task_memcg = mem_cgroup_from_task(task);
1179 css_get(&task_memcg->css);
1182 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1183 css_put(&task_memcg->css);
1188 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1189 * @memcg: the memory cgroup
1191 * Returns the maximum amount of memory @mem can be charged with, in
1194 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1196 unsigned long margin = 0;
1197 unsigned long count;
1198 unsigned long limit;
1200 count = page_counter_read(&memcg->memory);
1201 limit = READ_ONCE(memcg->memory.max);
1203 margin = limit - count;
1205 if (do_memsw_account()) {
1206 count = page_counter_read(&memcg->memsw);
1207 limit = READ_ONCE(memcg->memsw.max);
1209 margin = min(margin, limit - count);
1218 * A routine for checking "mem" is under move_account() or not.
1220 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1221 * moving cgroups. This is for waiting at high-memory pressure
1224 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1226 struct mem_cgroup *from;
1227 struct mem_cgroup *to;
1230 * Unlike task_move routines, we access mc.to, mc.from not under
1231 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1233 spin_lock(&mc.lock);
1239 ret = mem_cgroup_is_descendant(from, memcg) ||
1240 mem_cgroup_is_descendant(to, memcg);
1242 spin_unlock(&mc.lock);
1246 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1248 if (mc.moving_task && current != mc.moving_task) {
1249 if (mem_cgroup_under_move(memcg)) {
1251 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1252 /* moving charge context might have finished. */
1255 finish_wait(&mc.waitq, &wait);
1262 static const unsigned int memcg1_stats[] = {
1273 static const char *const memcg1_stat_names[] = {
1284 #define K(x) ((x) << (PAGE_SHIFT-10))
1286 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1287 * @memcg: The memory cgroup that went over limit
1288 * @p: Task that is going to be killed
1290 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1293 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1295 struct mem_cgroup *iter;
1301 pr_info("Task in ");
1302 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1303 pr_cont(" killed as a result of limit of ");
1305 pr_info("Memory limit reached of cgroup ");
1308 pr_cont_cgroup_path(memcg->css.cgroup);
1313 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1314 K((u64)page_counter_read(&memcg->memory)),
1315 K((u64)memcg->memory.max), memcg->memory.failcnt);
1316 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1317 K((u64)page_counter_read(&memcg->memsw)),
1318 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1319 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1320 K((u64)page_counter_read(&memcg->kmem)),
1321 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1323 for_each_mem_cgroup_tree(iter, memcg) {
1324 pr_info("Memory cgroup stats for ");
1325 pr_cont_cgroup_path(iter->css.cgroup);
1328 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1329 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1331 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1332 K(memcg_page_state(iter, memcg1_stats[i])));
1335 for (i = 0; i < NR_LRU_LISTS; i++)
1336 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1337 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1344 * Return the memory (and swap, if configured) limit for a memcg.
1346 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1350 max = memcg->memory.max;
1351 if (mem_cgroup_swappiness(memcg)) {
1352 unsigned long memsw_max;
1353 unsigned long swap_max;
1355 memsw_max = memcg->memsw.max;
1356 swap_max = memcg->swap.max;
1357 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1358 max = min(max + swap_max, memsw_max);
1363 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1366 struct oom_control oc = {
1370 .gfp_mask = gfp_mask,
1375 mutex_lock(&oom_lock);
1376 ret = out_of_memory(&oc);
1377 mutex_unlock(&oom_lock);
1381 #if MAX_NUMNODES > 1
1384 * test_mem_cgroup_node_reclaimable
1385 * @memcg: the target memcg
1386 * @nid: the node ID to be checked.
1387 * @noswap : specify true here if the user wants flle only information.
1389 * This function returns whether the specified memcg contains any
1390 * reclaimable pages on a node. Returns true if there are any reclaimable
1391 * pages in the node.
1393 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1394 int nid, bool noswap)
1396 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1398 if (noswap || !total_swap_pages)
1400 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1407 * Always updating the nodemask is not very good - even if we have an empty
1408 * list or the wrong list here, we can start from some node and traverse all
1409 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1412 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1416 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1417 * pagein/pageout changes since the last update.
1419 if (!atomic_read(&memcg->numainfo_events))
1421 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1424 /* make a nodemask where this memcg uses memory from */
1425 memcg->scan_nodes = node_states[N_MEMORY];
1427 for_each_node_mask(nid, node_states[N_MEMORY]) {
1429 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1430 node_clear(nid, memcg->scan_nodes);
1433 atomic_set(&memcg->numainfo_events, 0);
1434 atomic_set(&memcg->numainfo_updating, 0);
1438 * Selecting a node where we start reclaim from. Because what we need is just
1439 * reducing usage counter, start from anywhere is O,K. Considering
1440 * memory reclaim from current node, there are pros. and cons.
1442 * Freeing memory from current node means freeing memory from a node which
1443 * we'll use or we've used. So, it may make LRU bad. And if several threads
1444 * hit limits, it will see a contention on a node. But freeing from remote
1445 * node means more costs for memory reclaim because of memory latency.
1447 * Now, we use round-robin. Better algorithm is welcomed.
1449 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1453 mem_cgroup_may_update_nodemask(memcg);
1454 node = memcg->last_scanned_node;
1456 node = next_node_in(node, memcg->scan_nodes);
1458 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1459 * last time it really checked all the LRUs due to rate limiting.
1460 * Fallback to the current node in that case for simplicity.
1462 if (unlikely(node == MAX_NUMNODES))
1463 node = numa_node_id();
1465 memcg->last_scanned_node = node;
1469 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1475 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1478 unsigned long *total_scanned)
1480 struct mem_cgroup *victim = NULL;
1483 unsigned long excess;
1484 unsigned long nr_scanned;
1485 struct mem_cgroup_reclaim_cookie reclaim = {
1490 excess = soft_limit_excess(root_memcg);
1493 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1498 * If we have not been able to reclaim
1499 * anything, it might because there are
1500 * no reclaimable pages under this hierarchy
1505 * We want to do more targeted reclaim.
1506 * excess >> 2 is not to excessive so as to
1507 * reclaim too much, nor too less that we keep
1508 * coming back to reclaim from this cgroup
1510 if (total >= (excess >> 2) ||
1511 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1516 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1517 pgdat, &nr_scanned);
1518 *total_scanned += nr_scanned;
1519 if (!soft_limit_excess(root_memcg))
1522 mem_cgroup_iter_break(root_memcg, victim);
1526 #ifdef CONFIG_LOCKDEP
1527 static struct lockdep_map memcg_oom_lock_dep_map = {
1528 .name = "memcg_oom_lock",
1532 static DEFINE_SPINLOCK(memcg_oom_lock);
1535 * Check OOM-Killer is already running under our hierarchy.
1536 * If someone is running, return false.
1538 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1540 struct mem_cgroup *iter, *failed = NULL;
1542 spin_lock(&memcg_oom_lock);
1544 for_each_mem_cgroup_tree(iter, memcg) {
1545 if (iter->oom_lock) {
1547 * this subtree of our hierarchy is already locked
1548 * so we cannot give a lock.
1551 mem_cgroup_iter_break(memcg, iter);
1554 iter->oom_lock = true;
1559 * OK, we failed to lock the whole subtree so we have
1560 * to clean up what we set up to the failing subtree
1562 for_each_mem_cgroup_tree(iter, memcg) {
1563 if (iter == failed) {
1564 mem_cgroup_iter_break(memcg, iter);
1567 iter->oom_lock = false;
1570 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1572 spin_unlock(&memcg_oom_lock);
1577 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1579 struct mem_cgroup *iter;
1581 spin_lock(&memcg_oom_lock);
1582 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1583 for_each_mem_cgroup_tree(iter, memcg)
1584 iter->oom_lock = false;
1585 spin_unlock(&memcg_oom_lock);
1588 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1590 struct mem_cgroup *iter;
1592 spin_lock(&memcg_oom_lock);
1593 for_each_mem_cgroup_tree(iter, memcg)
1595 spin_unlock(&memcg_oom_lock);
1598 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1600 struct mem_cgroup *iter;
1603 * When a new child is created while the hierarchy is under oom,
1604 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1606 spin_lock(&memcg_oom_lock);
1607 for_each_mem_cgroup_tree(iter, memcg)
1608 if (iter->under_oom > 0)
1610 spin_unlock(&memcg_oom_lock);
1613 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1615 struct oom_wait_info {
1616 struct mem_cgroup *memcg;
1617 wait_queue_entry_t wait;
1620 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1621 unsigned mode, int sync, void *arg)
1623 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1624 struct mem_cgroup *oom_wait_memcg;
1625 struct oom_wait_info *oom_wait_info;
1627 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1628 oom_wait_memcg = oom_wait_info->memcg;
1630 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1631 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1633 return autoremove_wake_function(wait, mode, sync, arg);
1636 static void memcg_oom_recover(struct mem_cgroup *memcg)
1639 * For the following lockless ->under_oom test, the only required
1640 * guarantee is that it must see the state asserted by an OOM when
1641 * this function is called as a result of userland actions
1642 * triggered by the notification of the OOM. This is trivially
1643 * achieved by invoking mem_cgroup_mark_under_oom() before
1644 * triggering notification.
1646 if (memcg && memcg->under_oom)
1647 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1657 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1659 if (order > PAGE_ALLOC_COSTLY_ORDER)
1663 * We are in the middle of the charge context here, so we
1664 * don't want to block when potentially sitting on a callstack
1665 * that holds all kinds of filesystem and mm locks.
1667 * cgroup1 allows disabling the OOM killer and waiting for outside
1668 * handling until the charge can succeed; remember the context and put
1669 * the task to sleep at the end of the page fault when all locks are
1672 * On the other hand, in-kernel OOM killer allows for an async victim
1673 * memory reclaim (oom_reaper) and that means that we are not solely
1674 * relying on the oom victim to make a forward progress and we can
1675 * invoke the oom killer here.
1677 * Please note that mem_cgroup_out_of_memory might fail to find a
1678 * victim and then we have to bail out from the charge path.
1680 if (memcg->oom_kill_disable) {
1681 if (!current->in_user_fault)
1683 css_get(&memcg->css);
1684 current->memcg_in_oom = memcg;
1685 current->memcg_oom_gfp_mask = mask;
1686 current->memcg_oom_order = order;
1691 if (mem_cgroup_out_of_memory(memcg, mask, order))
1694 WARN(1,"Memory cgroup charge failed because of no reclaimable memory! "
1695 "This looks like a misconfiguration or a kernel bug.");
1700 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1701 * @handle: actually kill/wait or just clean up the OOM state
1703 * This has to be called at the end of a page fault if the memcg OOM
1704 * handler was enabled.
1706 * Memcg supports userspace OOM handling where failed allocations must
1707 * sleep on a waitqueue until the userspace task resolves the
1708 * situation. Sleeping directly in the charge context with all kinds
1709 * of locks held is not a good idea, instead we remember an OOM state
1710 * in the task and mem_cgroup_oom_synchronize() has to be called at
1711 * the end of the page fault to complete the OOM handling.
1713 * Returns %true if an ongoing memcg OOM situation was detected and
1714 * completed, %false otherwise.
1716 bool mem_cgroup_oom_synchronize(bool handle)
1718 struct mem_cgroup *memcg = current->memcg_in_oom;
1719 struct oom_wait_info owait;
1722 /* OOM is global, do not handle */
1729 owait.memcg = memcg;
1730 owait.wait.flags = 0;
1731 owait.wait.func = memcg_oom_wake_function;
1732 owait.wait.private = current;
1733 INIT_LIST_HEAD(&owait.wait.entry);
1735 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1736 mem_cgroup_mark_under_oom(memcg);
1738 locked = mem_cgroup_oom_trylock(memcg);
1741 mem_cgroup_oom_notify(memcg);
1743 if (locked && !memcg->oom_kill_disable) {
1744 mem_cgroup_unmark_under_oom(memcg);
1745 finish_wait(&memcg_oom_waitq, &owait.wait);
1746 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1747 current->memcg_oom_order);
1750 mem_cgroup_unmark_under_oom(memcg);
1751 finish_wait(&memcg_oom_waitq, &owait.wait);
1755 mem_cgroup_oom_unlock(memcg);
1757 * There is no guarantee that an OOM-lock contender
1758 * sees the wakeups triggered by the OOM kill
1759 * uncharges. Wake any sleepers explicitely.
1761 memcg_oom_recover(memcg);
1764 current->memcg_in_oom = NULL;
1765 css_put(&memcg->css);
1770 * lock_page_memcg - lock a page->mem_cgroup binding
1773 * This function protects unlocked LRU pages from being moved to
1776 * It ensures lifetime of the returned memcg. Caller is responsible
1777 * for the lifetime of the page; __unlock_page_memcg() is available
1778 * when @page might get freed inside the locked section.
1780 struct mem_cgroup *lock_page_memcg(struct page *page)
1782 struct mem_cgroup *memcg;
1783 unsigned long flags;
1786 * The RCU lock is held throughout the transaction. The fast
1787 * path can get away without acquiring the memcg->move_lock
1788 * because page moving starts with an RCU grace period.
1790 * The RCU lock also protects the memcg from being freed when
1791 * the page state that is going to change is the only thing
1792 * preventing the page itself from being freed. E.g. writeback
1793 * doesn't hold a page reference and relies on PG_writeback to
1794 * keep off truncation, migration and so forth.
1798 if (mem_cgroup_disabled())
1801 memcg = page->mem_cgroup;
1802 if (unlikely(!memcg))
1805 if (atomic_read(&memcg->moving_account) <= 0)
1808 spin_lock_irqsave(&memcg->move_lock, flags);
1809 if (memcg != page->mem_cgroup) {
1810 spin_unlock_irqrestore(&memcg->move_lock, flags);
1815 * When charge migration first begins, we can have locked and
1816 * unlocked page stat updates happening concurrently. Track
1817 * the task who has the lock for unlock_page_memcg().
1819 memcg->move_lock_task = current;
1820 memcg->move_lock_flags = flags;
1824 EXPORT_SYMBOL(lock_page_memcg);
1827 * __unlock_page_memcg - unlock and unpin a memcg
1830 * Unlock and unpin a memcg returned by lock_page_memcg().
1832 void __unlock_page_memcg(struct mem_cgroup *memcg)
1834 if (memcg && memcg->move_lock_task == current) {
1835 unsigned long flags = memcg->move_lock_flags;
1837 memcg->move_lock_task = NULL;
1838 memcg->move_lock_flags = 0;
1840 spin_unlock_irqrestore(&memcg->move_lock, flags);
1847 * unlock_page_memcg - unlock a page->mem_cgroup binding
1850 void unlock_page_memcg(struct page *page)
1852 __unlock_page_memcg(page->mem_cgroup);
1854 EXPORT_SYMBOL(unlock_page_memcg);
1856 struct memcg_stock_pcp {
1857 struct mem_cgroup *cached; /* this never be root cgroup */
1858 unsigned int nr_pages;
1859 struct work_struct work;
1860 unsigned long flags;
1861 #define FLUSHING_CACHED_CHARGE 0
1863 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1864 static DEFINE_MUTEX(percpu_charge_mutex);
1867 * consume_stock: Try to consume stocked charge on this cpu.
1868 * @memcg: memcg to consume from.
1869 * @nr_pages: how many pages to charge.
1871 * The charges will only happen if @memcg matches the current cpu's memcg
1872 * stock, and at least @nr_pages are available in that stock. Failure to
1873 * service an allocation will refill the stock.
1875 * returns true if successful, false otherwise.
1877 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1879 struct memcg_stock_pcp *stock;
1880 unsigned long flags;
1883 if (nr_pages > MEMCG_CHARGE_BATCH)
1886 local_irq_save(flags);
1888 stock = this_cpu_ptr(&memcg_stock);
1889 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1890 stock->nr_pages -= nr_pages;
1894 local_irq_restore(flags);
1900 * Returns stocks cached in percpu and reset cached information.
1902 static void drain_stock(struct memcg_stock_pcp *stock)
1904 struct mem_cgroup *old = stock->cached;
1906 if (stock->nr_pages) {
1907 page_counter_uncharge(&old->memory, stock->nr_pages);
1908 if (do_memsw_account())
1909 page_counter_uncharge(&old->memsw, stock->nr_pages);
1910 css_put_many(&old->css, stock->nr_pages);
1911 stock->nr_pages = 0;
1913 stock->cached = NULL;
1916 static void drain_local_stock(struct work_struct *dummy)
1918 struct memcg_stock_pcp *stock;
1919 unsigned long flags;
1922 * The only protection from memory hotplug vs. drain_stock races is
1923 * that we always operate on local CPU stock here with IRQ disabled
1925 local_irq_save(flags);
1927 stock = this_cpu_ptr(&memcg_stock);
1929 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1931 local_irq_restore(flags);
1935 * Cache charges(val) to local per_cpu area.
1936 * This will be consumed by consume_stock() function, later.
1938 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1940 struct memcg_stock_pcp *stock;
1941 unsigned long flags;
1943 local_irq_save(flags);
1945 stock = this_cpu_ptr(&memcg_stock);
1946 if (stock->cached != memcg) { /* reset if necessary */
1948 stock->cached = memcg;
1950 stock->nr_pages += nr_pages;
1952 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
1955 local_irq_restore(flags);
1959 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1960 * of the hierarchy under it.
1962 static void drain_all_stock(struct mem_cgroup *root_memcg)
1966 /* If someone's already draining, avoid adding running more workers. */
1967 if (!mutex_trylock(&percpu_charge_mutex))
1970 * Notify other cpus that system-wide "drain" is running
1971 * We do not care about races with the cpu hotplug because cpu down
1972 * as well as workers from this path always operate on the local
1973 * per-cpu data. CPU up doesn't touch memcg_stock at all.
1976 for_each_online_cpu(cpu) {
1977 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1978 struct mem_cgroup *memcg;
1980 memcg = stock->cached;
1981 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
1983 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
1984 css_put(&memcg->css);
1987 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1989 drain_local_stock(&stock->work);
1991 schedule_work_on(cpu, &stock->work);
1993 css_put(&memcg->css);
1996 mutex_unlock(&percpu_charge_mutex);
1999 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2001 struct memcg_stock_pcp *stock;
2002 struct mem_cgroup *memcg;
2004 stock = &per_cpu(memcg_stock, cpu);
2007 for_each_mem_cgroup(memcg) {
2010 for (i = 0; i < MEMCG_NR_STAT; i++) {
2014 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
2016 atomic_long_add(x, &memcg->stat[i]);
2018 if (i >= NR_VM_NODE_STAT_ITEMS)
2021 for_each_node(nid) {
2022 struct mem_cgroup_per_node *pn;
2024 pn = mem_cgroup_nodeinfo(memcg, nid);
2025 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2027 atomic_long_add(x, &pn->lruvec_stat[i]);
2031 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2034 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
2036 atomic_long_add(x, &memcg->events[i]);
2043 static void reclaim_high(struct mem_cgroup *memcg,
2044 unsigned int nr_pages,
2048 if (page_counter_read(&memcg->memory) <= memcg->high)
2050 memcg_memory_event(memcg, MEMCG_HIGH);
2051 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2052 } while ((memcg = parent_mem_cgroup(memcg)));
2055 static void high_work_func(struct work_struct *work)
2057 struct mem_cgroup *memcg;
2059 memcg = container_of(work, struct mem_cgroup, high_work);
2060 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2064 * Scheduled by try_charge() to be executed from the userland return path
2065 * and reclaims memory over the high limit.
2067 void mem_cgroup_handle_over_high(void)
2069 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2070 struct mem_cgroup *memcg;
2072 if (likely(!nr_pages))
2075 memcg = get_mem_cgroup_from_mm(current->mm);
2076 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2077 css_put(&memcg->css);
2078 current->memcg_nr_pages_over_high = 0;
2081 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2082 unsigned int nr_pages)
2084 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2085 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2086 struct mem_cgroup *mem_over_limit;
2087 struct page_counter *counter;
2088 unsigned long nr_reclaimed;
2089 bool may_swap = true;
2090 bool drained = false;
2092 enum oom_status oom_status;
2094 if (mem_cgroup_is_root(memcg))
2097 if (consume_stock(memcg, nr_pages))
2100 if (!do_memsw_account() ||
2101 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2102 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2104 if (do_memsw_account())
2105 page_counter_uncharge(&memcg->memsw, batch);
2106 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2108 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2112 if (batch > nr_pages) {
2118 * Unlike in global OOM situations, memcg is not in a physical
2119 * memory shortage. Allow dying and OOM-killed tasks to
2120 * bypass the last charges so that they can exit quickly and
2121 * free their memory.
2123 if (unlikely(tsk_is_oom_victim(current) ||
2124 fatal_signal_pending(current) ||
2125 current->flags & PF_EXITING))
2129 * Prevent unbounded recursion when reclaim operations need to
2130 * allocate memory. This might exceed the limits temporarily,
2131 * but we prefer facilitating memory reclaim and getting back
2132 * under the limit over triggering OOM kills in these cases.
2134 if (unlikely(current->flags & PF_MEMALLOC))
2137 if (unlikely(task_in_memcg_oom(current)))
2140 if (!gfpflags_allow_blocking(gfp_mask))
2143 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2145 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2146 gfp_mask, may_swap);
2148 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2152 drain_all_stock(mem_over_limit);
2157 if (gfp_mask & __GFP_NORETRY)
2160 * Even though the limit is exceeded at this point, reclaim
2161 * may have been able to free some pages. Retry the charge
2162 * before killing the task.
2164 * Only for regular pages, though: huge pages are rather
2165 * unlikely to succeed so close to the limit, and we fall back
2166 * to regular pages anyway in case of failure.
2168 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2171 * At task move, charge accounts can be doubly counted. So, it's
2172 * better to wait until the end of task_move if something is going on.
2174 if (mem_cgroup_wait_acct_move(mem_over_limit))
2180 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2183 if (gfp_mask & __GFP_NOFAIL)
2186 if (fatal_signal_pending(current))
2189 memcg_memory_event(mem_over_limit, MEMCG_OOM);
2192 * keep retrying as long as the memcg oom killer is able to make
2193 * a forward progress or bypass the charge if the oom killer
2194 * couldn't make any progress.
2196 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2197 get_order(nr_pages * PAGE_SIZE));
2198 switch (oom_status) {
2200 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2209 if (!(gfp_mask & __GFP_NOFAIL))
2213 * The allocation either can't fail or will lead to more memory
2214 * being freed very soon. Allow memory usage go over the limit
2215 * temporarily by force charging it.
2217 page_counter_charge(&memcg->memory, nr_pages);
2218 if (do_memsw_account())
2219 page_counter_charge(&memcg->memsw, nr_pages);
2220 css_get_many(&memcg->css, nr_pages);
2225 css_get_many(&memcg->css, batch);
2226 if (batch > nr_pages)
2227 refill_stock(memcg, batch - nr_pages);
2230 * If the hierarchy is above the normal consumption range, schedule
2231 * reclaim on returning to userland. We can perform reclaim here
2232 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2233 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2234 * not recorded as it most likely matches current's and won't
2235 * change in the meantime. As high limit is checked again before
2236 * reclaim, the cost of mismatch is negligible.
2239 if (page_counter_read(&memcg->memory) > memcg->high) {
2240 /* Don't bother a random interrupted task */
2241 if (in_interrupt()) {
2242 schedule_work(&memcg->high_work);
2245 current->memcg_nr_pages_over_high += batch;
2246 set_notify_resume(current);
2249 } while ((memcg = parent_mem_cgroup(memcg)));
2254 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2256 if (mem_cgroup_is_root(memcg))
2259 page_counter_uncharge(&memcg->memory, nr_pages);
2260 if (do_memsw_account())
2261 page_counter_uncharge(&memcg->memsw, nr_pages);
2263 css_put_many(&memcg->css, nr_pages);
2266 static void lock_page_lru(struct page *page, int *isolated)
2268 struct zone *zone = page_zone(page);
2270 spin_lock_irq(zone_lru_lock(zone));
2271 if (PageLRU(page)) {
2272 struct lruvec *lruvec;
2274 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2276 del_page_from_lru_list(page, lruvec, page_lru(page));
2282 static void unlock_page_lru(struct page *page, int isolated)
2284 struct zone *zone = page_zone(page);
2287 struct lruvec *lruvec;
2289 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2290 VM_BUG_ON_PAGE(PageLRU(page), page);
2292 add_page_to_lru_list(page, lruvec, page_lru(page));
2294 spin_unlock_irq(zone_lru_lock(zone));
2297 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2302 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2305 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2306 * may already be on some other mem_cgroup's LRU. Take care of it.
2309 lock_page_lru(page, &isolated);
2312 * Nobody should be changing or seriously looking at
2313 * page->mem_cgroup at this point:
2315 * - the page is uncharged
2317 * - the page is off-LRU
2319 * - an anonymous fault has exclusive page access, except for
2320 * a locked page table
2322 * - a page cache insertion, a swapin fault, or a migration
2323 * have the page locked
2325 page->mem_cgroup = memcg;
2328 unlock_page_lru(page, isolated);
2331 #ifdef CONFIG_MEMCG_KMEM
2332 static int memcg_alloc_cache_id(void)
2337 id = ida_simple_get(&memcg_cache_ida,
2338 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2342 if (id < memcg_nr_cache_ids)
2346 * There's no space for the new id in memcg_caches arrays,
2347 * so we have to grow them.
2349 down_write(&memcg_cache_ids_sem);
2351 size = 2 * (id + 1);
2352 if (size < MEMCG_CACHES_MIN_SIZE)
2353 size = MEMCG_CACHES_MIN_SIZE;
2354 else if (size > MEMCG_CACHES_MAX_SIZE)
2355 size = MEMCG_CACHES_MAX_SIZE;
2357 err = memcg_update_all_caches(size);
2359 err = memcg_update_all_list_lrus(size);
2361 memcg_nr_cache_ids = size;
2363 up_write(&memcg_cache_ids_sem);
2366 ida_simple_remove(&memcg_cache_ida, id);
2372 static void memcg_free_cache_id(int id)
2374 ida_simple_remove(&memcg_cache_ida, id);
2377 struct memcg_kmem_cache_create_work {
2378 struct mem_cgroup *memcg;
2379 struct kmem_cache *cachep;
2380 struct work_struct work;
2383 static void memcg_kmem_cache_create_func(struct work_struct *w)
2385 struct memcg_kmem_cache_create_work *cw =
2386 container_of(w, struct memcg_kmem_cache_create_work, work);
2387 struct mem_cgroup *memcg = cw->memcg;
2388 struct kmem_cache *cachep = cw->cachep;
2390 memcg_create_kmem_cache(memcg, cachep);
2392 css_put(&memcg->css);
2397 * Enqueue the creation of a per-memcg kmem_cache.
2399 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2400 struct kmem_cache *cachep)
2402 struct memcg_kmem_cache_create_work *cw;
2404 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2408 css_get(&memcg->css);
2411 cw->cachep = cachep;
2412 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2414 queue_work(memcg_kmem_cache_wq, &cw->work);
2417 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2418 struct kmem_cache *cachep)
2421 * We need to stop accounting when we kmalloc, because if the
2422 * corresponding kmalloc cache is not yet created, the first allocation
2423 * in __memcg_schedule_kmem_cache_create will recurse.
2425 * However, it is better to enclose the whole function. Depending on
2426 * the debugging options enabled, INIT_WORK(), for instance, can
2427 * trigger an allocation. This too, will make us recurse. Because at
2428 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2429 * the safest choice is to do it like this, wrapping the whole function.
2431 current->memcg_kmem_skip_account = 1;
2432 __memcg_schedule_kmem_cache_create(memcg, cachep);
2433 current->memcg_kmem_skip_account = 0;
2436 static inline bool memcg_kmem_bypass(void)
2438 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2444 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2445 * @cachep: the original global kmem cache
2447 * Return the kmem_cache we're supposed to use for a slab allocation.
2448 * We try to use the current memcg's version of the cache.
2450 * If the cache does not exist yet, if we are the first user of it, we
2451 * create it asynchronously in a workqueue and let the current allocation
2452 * go through with the original cache.
2454 * This function takes a reference to the cache it returns to assure it
2455 * won't get destroyed while we are working with it. Once the caller is
2456 * done with it, memcg_kmem_put_cache() must be called to release the
2459 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2461 struct mem_cgroup *memcg;
2462 struct kmem_cache *memcg_cachep;
2465 VM_BUG_ON(!is_root_cache(cachep));
2467 if (memcg_kmem_bypass())
2470 if (current->memcg_kmem_skip_account)
2473 memcg = get_mem_cgroup_from_current();
2474 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2478 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2479 if (likely(memcg_cachep))
2480 return memcg_cachep;
2483 * If we are in a safe context (can wait, and not in interrupt
2484 * context), we could be be predictable and return right away.
2485 * This would guarantee that the allocation being performed
2486 * already belongs in the new cache.
2488 * However, there are some clashes that can arrive from locking.
2489 * For instance, because we acquire the slab_mutex while doing
2490 * memcg_create_kmem_cache, this means no further allocation
2491 * could happen with the slab_mutex held. So it's better to
2494 memcg_schedule_kmem_cache_create(memcg, cachep);
2496 css_put(&memcg->css);
2501 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2502 * @cachep: the cache returned by memcg_kmem_get_cache
2504 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2506 if (!is_root_cache(cachep))
2507 css_put(&cachep->memcg_params.memcg->css);
2511 * memcg_kmem_charge_memcg: charge a kmem page
2512 * @page: page to charge
2513 * @gfp: reclaim mode
2514 * @order: allocation order
2515 * @memcg: memory cgroup to charge
2517 * Returns 0 on success, an error code on failure.
2519 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2520 struct mem_cgroup *memcg)
2522 unsigned int nr_pages = 1 << order;
2523 struct page_counter *counter;
2526 ret = try_charge(memcg, gfp, nr_pages);
2530 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2531 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2532 cancel_charge(memcg, nr_pages);
2536 page->mem_cgroup = memcg;
2542 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2543 * @page: page to charge
2544 * @gfp: reclaim mode
2545 * @order: allocation order
2547 * Returns 0 on success, an error code on failure.
2549 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2551 struct mem_cgroup *memcg;
2554 if (memcg_kmem_bypass())
2557 memcg = get_mem_cgroup_from_current();
2558 if (!mem_cgroup_is_root(memcg)) {
2559 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2561 __SetPageKmemcg(page);
2563 css_put(&memcg->css);
2567 * memcg_kmem_uncharge: uncharge a kmem page
2568 * @page: page to uncharge
2569 * @order: allocation order
2571 void memcg_kmem_uncharge(struct page *page, int order)
2573 struct mem_cgroup *memcg = page->mem_cgroup;
2574 unsigned int nr_pages = 1 << order;
2579 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2581 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2582 page_counter_uncharge(&memcg->kmem, nr_pages);
2584 page_counter_uncharge(&memcg->memory, nr_pages);
2585 if (do_memsw_account())
2586 page_counter_uncharge(&memcg->memsw, nr_pages);
2588 page->mem_cgroup = NULL;
2590 /* slab pages do not have PageKmemcg flag set */
2591 if (PageKmemcg(page))
2592 __ClearPageKmemcg(page);
2594 css_put_many(&memcg->css, nr_pages);
2596 #endif /* CONFIG_MEMCG_KMEM */
2598 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2601 * Because tail pages are not marked as "used", set it. We're under
2602 * zone_lru_lock and migration entries setup in all page mappings.
2604 void mem_cgroup_split_huge_fixup(struct page *head)
2608 if (mem_cgroup_disabled())
2611 for (i = 1; i < HPAGE_PMD_NR; i++)
2612 head[i].mem_cgroup = head->mem_cgroup;
2614 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2616 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2618 #ifdef CONFIG_MEMCG_SWAP
2620 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2621 * @entry: swap entry to be moved
2622 * @from: mem_cgroup which the entry is moved from
2623 * @to: mem_cgroup which the entry is moved to
2625 * It succeeds only when the swap_cgroup's record for this entry is the same
2626 * as the mem_cgroup's id of @from.
2628 * Returns 0 on success, -EINVAL on failure.
2630 * The caller must have charged to @to, IOW, called page_counter_charge() about
2631 * both res and memsw, and called css_get().
2633 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2634 struct mem_cgroup *from, struct mem_cgroup *to)
2636 unsigned short old_id, new_id;
2638 old_id = mem_cgroup_id(from);
2639 new_id = mem_cgroup_id(to);
2641 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2642 mod_memcg_state(from, MEMCG_SWAP, -1);
2643 mod_memcg_state(to, MEMCG_SWAP, 1);
2649 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2650 struct mem_cgroup *from, struct mem_cgroup *to)
2656 static DEFINE_MUTEX(memcg_max_mutex);
2658 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2659 unsigned long max, bool memsw)
2661 bool enlarge = false;
2662 bool drained = false;
2664 bool limits_invariant;
2665 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2668 if (signal_pending(current)) {
2673 mutex_lock(&memcg_max_mutex);
2675 * Make sure that the new limit (memsw or memory limit) doesn't
2676 * break our basic invariant rule memory.max <= memsw.max.
2678 limits_invariant = memsw ? max >= memcg->memory.max :
2679 max <= memcg->memsw.max;
2680 if (!limits_invariant) {
2681 mutex_unlock(&memcg_max_mutex);
2685 if (max > counter->max)
2687 ret = page_counter_set_max(counter, max);
2688 mutex_unlock(&memcg_max_mutex);
2694 drain_all_stock(memcg);
2699 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2700 GFP_KERNEL, !memsw)) {
2706 if (!ret && enlarge)
2707 memcg_oom_recover(memcg);
2712 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2714 unsigned long *total_scanned)
2716 unsigned long nr_reclaimed = 0;
2717 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2718 unsigned long reclaimed;
2720 struct mem_cgroup_tree_per_node *mctz;
2721 unsigned long excess;
2722 unsigned long nr_scanned;
2727 mctz = soft_limit_tree_node(pgdat->node_id);
2730 * Do not even bother to check the largest node if the root
2731 * is empty. Do it lockless to prevent lock bouncing. Races
2732 * are acceptable as soft limit is best effort anyway.
2734 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2738 * This loop can run a while, specially if mem_cgroup's continuously
2739 * keep exceeding their soft limit and putting the system under
2746 mz = mem_cgroup_largest_soft_limit_node(mctz);
2751 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2752 gfp_mask, &nr_scanned);
2753 nr_reclaimed += reclaimed;
2754 *total_scanned += nr_scanned;
2755 spin_lock_irq(&mctz->lock);
2756 __mem_cgroup_remove_exceeded(mz, mctz);
2759 * If we failed to reclaim anything from this memory cgroup
2760 * it is time to move on to the next cgroup
2764 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2766 excess = soft_limit_excess(mz->memcg);
2768 * One school of thought says that we should not add
2769 * back the node to the tree if reclaim returns 0.
2770 * But our reclaim could return 0, simply because due
2771 * to priority we are exposing a smaller subset of
2772 * memory to reclaim from. Consider this as a longer
2775 /* If excess == 0, no tree ops */
2776 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2777 spin_unlock_irq(&mctz->lock);
2778 css_put(&mz->memcg->css);
2781 * Could not reclaim anything and there are no more
2782 * mem cgroups to try or we seem to be looping without
2783 * reclaiming anything.
2785 if (!nr_reclaimed &&
2787 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2789 } while (!nr_reclaimed);
2791 css_put(&next_mz->memcg->css);
2792 return nr_reclaimed;
2796 * Test whether @memcg has children, dead or alive. Note that this
2797 * function doesn't care whether @memcg has use_hierarchy enabled and
2798 * returns %true if there are child csses according to the cgroup
2799 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2801 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2806 ret = css_next_child(NULL, &memcg->css);
2812 * Reclaims as many pages from the given memcg as possible.
2814 * Caller is responsible for holding css reference for memcg.
2816 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2818 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2820 /* we call try-to-free pages for make this cgroup empty */
2821 lru_add_drain_all();
2823 drain_all_stock(memcg);
2825 /* try to free all pages in this cgroup */
2826 while (nr_retries && page_counter_read(&memcg->memory)) {
2829 if (signal_pending(current))
2832 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2836 /* maybe some writeback is necessary */
2837 congestion_wait(BLK_RW_ASYNC, HZ/10);
2845 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2846 char *buf, size_t nbytes,
2849 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2851 if (mem_cgroup_is_root(memcg))
2853 return mem_cgroup_force_empty(memcg) ?: nbytes;
2856 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2859 return mem_cgroup_from_css(css)->use_hierarchy;
2862 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2863 struct cftype *cft, u64 val)
2866 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2867 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2869 if (memcg->use_hierarchy == val)
2873 * If parent's use_hierarchy is set, we can't make any modifications
2874 * in the child subtrees. If it is unset, then the change can
2875 * occur, provided the current cgroup has no children.
2877 * For the root cgroup, parent_mem is NULL, we allow value to be
2878 * set if there are no children.
2880 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2881 (val == 1 || val == 0)) {
2882 if (!memcg_has_children(memcg))
2883 memcg->use_hierarchy = val;
2892 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2894 struct mem_cgroup *iter;
2897 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2899 for_each_mem_cgroup_tree(iter, memcg) {
2900 for (i = 0; i < MEMCG_NR_STAT; i++)
2901 stat[i] += memcg_page_state(iter, i);
2905 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2907 struct mem_cgroup *iter;
2910 memset(events, 0, sizeof(*events) * NR_VM_EVENT_ITEMS);
2912 for_each_mem_cgroup_tree(iter, memcg) {
2913 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
2914 events[i] += memcg_sum_events(iter, i);
2918 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2920 unsigned long val = 0;
2922 if (mem_cgroup_is_root(memcg)) {
2923 struct mem_cgroup *iter;
2925 for_each_mem_cgroup_tree(iter, memcg) {
2926 val += memcg_page_state(iter, MEMCG_CACHE);
2927 val += memcg_page_state(iter, MEMCG_RSS);
2929 val += memcg_page_state(iter, MEMCG_SWAP);
2933 val = page_counter_read(&memcg->memory);
2935 val = page_counter_read(&memcg->memsw);
2948 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2951 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2952 struct page_counter *counter;
2954 switch (MEMFILE_TYPE(cft->private)) {
2956 counter = &memcg->memory;
2959 counter = &memcg->memsw;
2962 counter = &memcg->kmem;
2965 counter = &memcg->tcpmem;
2971 switch (MEMFILE_ATTR(cft->private)) {
2973 if (counter == &memcg->memory)
2974 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2975 if (counter == &memcg->memsw)
2976 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2977 return (u64)page_counter_read(counter) * PAGE_SIZE;
2979 return (u64)counter->max * PAGE_SIZE;
2981 return (u64)counter->watermark * PAGE_SIZE;
2983 return counter->failcnt;
2984 case RES_SOFT_LIMIT:
2985 return (u64)memcg->soft_limit * PAGE_SIZE;
2991 #ifdef CONFIG_MEMCG_KMEM
2992 static int memcg_online_kmem(struct mem_cgroup *memcg)
2996 if (cgroup_memory_nokmem)
2999 BUG_ON(memcg->kmemcg_id >= 0);
3000 BUG_ON(memcg->kmem_state);
3002 memcg_id = memcg_alloc_cache_id();
3006 static_branch_inc(&memcg_kmem_enabled_key);
3008 * A memory cgroup is considered kmem-online as soon as it gets
3009 * kmemcg_id. Setting the id after enabling static branching will
3010 * guarantee no one starts accounting before all call sites are
3013 memcg->kmemcg_id = memcg_id;
3014 memcg->kmem_state = KMEM_ONLINE;
3015 INIT_LIST_HEAD(&memcg->kmem_caches);
3020 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3022 struct cgroup_subsys_state *css;
3023 struct mem_cgroup *parent, *child;
3026 if (memcg->kmem_state != KMEM_ONLINE)
3029 * Clear the online state before clearing memcg_caches array
3030 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3031 * guarantees that no cache will be created for this cgroup
3032 * after we are done (see memcg_create_kmem_cache()).
3034 memcg->kmem_state = KMEM_ALLOCATED;
3036 memcg_deactivate_kmem_caches(memcg);
3038 kmemcg_id = memcg->kmemcg_id;
3039 BUG_ON(kmemcg_id < 0);
3041 parent = parent_mem_cgroup(memcg);
3043 parent = root_mem_cgroup;
3046 * Change kmemcg_id of this cgroup and all its descendants to the
3047 * parent's id, and then move all entries from this cgroup's list_lrus
3048 * to ones of the parent. After we have finished, all list_lrus
3049 * corresponding to this cgroup are guaranteed to remain empty. The
3050 * ordering is imposed by list_lru_node->lock taken by
3051 * memcg_drain_all_list_lrus().
3053 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3054 css_for_each_descendant_pre(css, &memcg->css) {
3055 child = mem_cgroup_from_css(css);
3056 BUG_ON(child->kmemcg_id != kmemcg_id);
3057 child->kmemcg_id = parent->kmemcg_id;
3058 if (!memcg->use_hierarchy)
3063 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3065 memcg_free_cache_id(kmemcg_id);
3068 static void memcg_free_kmem(struct mem_cgroup *memcg)
3070 /* css_alloc() failed, offlining didn't happen */
3071 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3072 memcg_offline_kmem(memcg);
3074 if (memcg->kmem_state == KMEM_ALLOCATED) {
3075 memcg_destroy_kmem_caches(memcg);
3076 static_branch_dec(&memcg_kmem_enabled_key);
3077 WARN_ON(page_counter_read(&memcg->kmem));
3081 static int memcg_online_kmem(struct mem_cgroup *memcg)
3085 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3088 static void memcg_free_kmem(struct mem_cgroup *memcg)
3091 #endif /* CONFIG_MEMCG_KMEM */
3093 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3098 mutex_lock(&memcg_max_mutex);
3099 ret = page_counter_set_max(&memcg->kmem, max);
3100 mutex_unlock(&memcg_max_mutex);
3104 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3108 mutex_lock(&memcg_max_mutex);
3110 ret = page_counter_set_max(&memcg->tcpmem, max);
3114 if (!memcg->tcpmem_active) {
3116 * The active flag needs to be written after the static_key
3117 * update. This is what guarantees that the socket activation
3118 * function is the last one to run. See mem_cgroup_sk_alloc()
3119 * for details, and note that we don't mark any socket as
3120 * belonging to this memcg until that flag is up.
3122 * We need to do this, because static_keys will span multiple
3123 * sites, but we can't control their order. If we mark a socket
3124 * as accounted, but the accounting functions are not patched in
3125 * yet, we'll lose accounting.
3127 * We never race with the readers in mem_cgroup_sk_alloc(),
3128 * because when this value change, the code to process it is not
3131 static_branch_inc(&memcg_sockets_enabled_key);
3132 memcg->tcpmem_active = true;
3135 mutex_unlock(&memcg_max_mutex);
3140 * The user of this function is...
3143 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3144 char *buf, size_t nbytes, loff_t off)
3146 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3147 unsigned long nr_pages;
3150 buf = strstrip(buf);
3151 ret = page_counter_memparse(buf, "-1", &nr_pages);
3155 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3157 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3161 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3163 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3166 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3169 ret = memcg_update_kmem_max(memcg, nr_pages);
3172 ret = memcg_update_tcp_max(memcg, nr_pages);
3176 case RES_SOFT_LIMIT:
3177 memcg->soft_limit = nr_pages;
3181 return ret ?: nbytes;
3184 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3185 size_t nbytes, loff_t off)
3187 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3188 struct page_counter *counter;
3190 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3192 counter = &memcg->memory;
3195 counter = &memcg->memsw;
3198 counter = &memcg->kmem;
3201 counter = &memcg->tcpmem;
3207 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3209 page_counter_reset_watermark(counter);
3212 counter->failcnt = 0;
3221 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3224 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3228 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3229 struct cftype *cft, u64 val)
3231 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3233 if (val & ~MOVE_MASK)
3237 * No kind of locking is needed in here, because ->can_attach() will
3238 * check this value once in the beginning of the process, and then carry
3239 * on with stale data. This means that changes to this value will only
3240 * affect task migrations starting after the change.
3242 memcg->move_charge_at_immigrate = val;
3246 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3247 struct cftype *cft, u64 val)
3254 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3258 unsigned int lru_mask;
3261 static const struct numa_stat stats[] = {
3262 { "total", LRU_ALL },
3263 { "file", LRU_ALL_FILE },
3264 { "anon", LRU_ALL_ANON },
3265 { "unevictable", BIT(LRU_UNEVICTABLE) },
3267 const struct numa_stat *stat;
3270 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3272 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3273 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3274 seq_printf(m, "%s=%lu", stat->name, nr);
3275 for_each_node_state(nid, N_MEMORY) {
3276 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3278 seq_printf(m, " N%d=%lu", nid, nr);
3283 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3284 struct mem_cgroup *iter;
3287 for_each_mem_cgroup_tree(iter, memcg)
3288 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3289 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3290 for_each_node_state(nid, N_MEMORY) {
3292 for_each_mem_cgroup_tree(iter, memcg)
3293 nr += mem_cgroup_node_nr_lru_pages(
3294 iter, nid, stat->lru_mask);
3295 seq_printf(m, " N%d=%lu", nid, nr);
3302 #endif /* CONFIG_NUMA */
3304 /* Universal VM events cgroup1 shows, original sort order */
3305 static const unsigned int memcg1_events[] = {
3312 static const char *const memcg1_event_names[] = {
3319 static int memcg_stat_show(struct seq_file *m, void *v)
3321 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3322 unsigned long memory, memsw;
3323 struct mem_cgroup *mi;
3326 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3327 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3329 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3330 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3332 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3333 memcg_page_state(memcg, memcg1_stats[i]) *
3337 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3338 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3339 memcg_sum_events(memcg, memcg1_events[i]));
3341 for (i = 0; i < NR_LRU_LISTS; i++)
3342 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3343 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3345 /* Hierarchical information */
3346 memory = memsw = PAGE_COUNTER_MAX;
3347 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3348 memory = min(memory, mi->memory.max);
3349 memsw = min(memsw, mi->memsw.max);
3351 seq_printf(m, "hierarchical_memory_limit %llu\n",
3352 (u64)memory * PAGE_SIZE);
3353 if (do_memsw_account())
3354 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3355 (u64)memsw * PAGE_SIZE);
3357 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3358 unsigned long long val = 0;
3360 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3362 for_each_mem_cgroup_tree(mi, memcg)
3363 val += memcg_page_state(mi, memcg1_stats[i]) *
3365 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i], val);
3368 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++) {
3369 unsigned long long val = 0;
3371 for_each_mem_cgroup_tree(mi, memcg)
3372 val += memcg_sum_events(mi, memcg1_events[i]);
3373 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i], val);
3376 for (i = 0; i < NR_LRU_LISTS; i++) {
3377 unsigned long long val = 0;
3379 for_each_mem_cgroup_tree(mi, memcg)
3380 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3381 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3384 #ifdef CONFIG_DEBUG_VM
3387 struct mem_cgroup_per_node *mz;
3388 struct zone_reclaim_stat *rstat;
3389 unsigned long recent_rotated[2] = {0, 0};
3390 unsigned long recent_scanned[2] = {0, 0};
3392 for_each_online_pgdat(pgdat) {
3393 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3394 rstat = &mz->lruvec.reclaim_stat;
3396 recent_rotated[0] += rstat->recent_rotated[0];
3397 recent_rotated[1] += rstat->recent_rotated[1];
3398 recent_scanned[0] += rstat->recent_scanned[0];
3399 recent_scanned[1] += rstat->recent_scanned[1];
3401 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3402 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3403 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3404 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3411 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3414 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3416 return mem_cgroup_swappiness(memcg);
3419 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3420 struct cftype *cft, u64 val)
3422 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3428 memcg->swappiness = val;
3430 vm_swappiness = val;
3435 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3437 struct mem_cgroup_threshold_ary *t;
3438 unsigned long usage;
3443 t = rcu_dereference(memcg->thresholds.primary);
3445 t = rcu_dereference(memcg->memsw_thresholds.primary);
3450 usage = mem_cgroup_usage(memcg, swap);
3453 * current_threshold points to threshold just below or equal to usage.
3454 * If it's not true, a threshold was crossed after last
3455 * call of __mem_cgroup_threshold().
3457 i = t->current_threshold;
3460 * Iterate backward over array of thresholds starting from
3461 * current_threshold and check if a threshold is crossed.
3462 * If none of thresholds below usage is crossed, we read
3463 * only one element of the array here.
3465 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3466 eventfd_signal(t->entries[i].eventfd, 1);
3468 /* i = current_threshold + 1 */
3472 * Iterate forward over array of thresholds starting from
3473 * current_threshold+1 and check if a threshold is crossed.
3474 * If none of thresholds above usage is crossed, we read
3475 * only one element of the array here.
3477 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3478 eventfd_signal(t->entries[i].eventfd, 1);
3480 /* Update current_threshold */
3481 t->current_threshold = i - 1;
3486 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3489 __mem_cgroup_threshold(memcg, false);
3490 if (do_memsw_account())
3491 __mem_cgroup_threshold(memcg, true);
3493 memcg = parent_mem_cgroup(memcg);
3497 static int compare_thresholds(const void *a, const void *b)
3499 const struct mem_cgroup_threshold *_a = a;
3500 const struct mem_cgroup_threshold *_b = b;
3502 if (_a->threshold > _b->threshold)
3505 if (_a->threshold < _b->threshold)
3511 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3513 struct mem_cgroup_eventfd_list *ev;
3515 spin_lock(&memcg_oom_lock);
3517 list_for_each_entry(ev, &memcg->oom_notify, list)
3518 eventfd_signal(ev->eventfd, 1);
3520 spin_unlock(&memcg_oom_lock);
3524 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3526 struct mem_cgroup *iter;
3528 for_each_mem_cgroup_tree(iter, memcg)
3529 mem_cgroup_oom_notify_cb(iter);
3532 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3533 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3535 struct mem_cgroup_thresholds *thresholds;
3536 struct mem_cgroup_threshold_ary *new;
3537 unsigned long threshold;
3538 unsigned long usage;
3541 ret = page_counter_memparse(args, "-1", &threshold);
3545 mutex_lock(&memcg->thresholds_lock);
3548 thresholds = &memcg->thresholds;
3549 usage = mem_cgroup_usage(memcg, false);
3550 } else if (type == _MEMSWAP) {
3551 thresholds = &memcg->memsw_thresholds;
3552 usage = mem_cgroup_usage(memcg, true);
3556 /* Check if a threshold crossed before adding a new one */
3557 if (thresholds->primary)
3558 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3560 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3562 /* Allocate memory for new array of thresholds */
3563 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3571 /* Copy thresholds (if any) to new array */
3572 if (thresholds->primary) {
3573 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3574 sizeof(struct mem_cgroup_threshold));
3577 /* Add new threshold */
3578 new->entries[size - 1].eventfd = eventfd;
3579 new->entries[size - 1].threshold = threshold;
3581 /* Sort thresholds. Registering of new threshold isn't time-critical */
3582 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3583 compare_thresholds, NULL);
3585 /* Find current threshold */
3586 new->current_threshold = -1;
3587 for (i = 0; i < size; i++) {
3588 if (new->entries[i].threshold <= usage) {
3590 * new->current_threshold will not be used until
3591 * rcu_assign_pointer(), so it's safe to increment
3594 ++new->current_threshold;
3599 /* Free old spare buffer and save old primary buffer as spare */
3600 kfree(thresholds->spare);
3601 thresholds->spare = thresholds->primary;
3603 rcu_assign_pointer(thresholds->primary, new);
3605 /* To be sure that nobody uses thresholds */
3609 mutex_unlock(&memcg->thresholds_lock);
3614 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3615 struct eventfd_ctx *eventfd, const char *args)
3617 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3620 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3621 struct eventfd_ctx *eventfd, const char *args)
3623 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3626 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3627 struct eventfd_ctx *eventfd, enum res_type type)
3629 struct mem_cgroup_thresholds *thresholds;
3630 struct mem_cgroup_threshold_ary *new;
3631 unsigned long usage;
3634 mutex_lock(&memcg->thresholds_lock);
3637 thresholds = &memcg->thresholds;
3638 usage = mem_cgroup_usage(memcg, false);
3639 } else if (type == _MEMSWAP) {
3640 thresholds = &memcg->memsw_thresholds;
3641 usage = mem_cgroup_usage(memcg, true);
3645 if (!thresholds->primary)
3648 /* Check if a threshold crossed before removing */
3649 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3651 /* Calculate new number of threshold */
3653 for (i = 0; i < thresholds->primary->size; i++) {
3654 if (thresholds->primary->entries[i].eventfd != eventfd)
3658 new = thresholds->spare;
3660 /* Set thresholds array to NULL if we don't have thresholds */
3669 /* Copy thresholds and find current threshold */
3670 new->current_threshold = -1;
3671 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3672 if (thresholds->primary->entries[i].eventfd == eventfd)
3675 new->entries[j] = thresholds->primary->entries[i];
3676 if (new->entries[j].threshold <= usage) {
3678 * new->current_threshold will not be used
3679 * until rcu_assign_pointer(), so it's safe to increment
3682 ++new->current_threshold;
3688 /* Swap primary and spare array */
3689 thresholds->spare = thresholds->primary;
3691 rcu_assign_pointer(thresholds->primary, new);
3693 /* To be sure that nobody uses thresholds */
3696 /* If all events are unregistered, free the spare array */
3698 kfree(thresholds->spare);
3699 thresholds->spare = NULL;
3702 mutex_unlock(&memcg->thresholds_lock);
3705 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3706 struct eventfd_ctx *eventfd)
3708 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3711 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3712 struct eventfd_ctx *eventfd)
3714 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3717 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3718 struct eventfd_ctx *eventfd, const char *args)
3720 struct mem_cgroup_eventfd_list *event;
3722 event = kmalloc(sizeof(*event), GFP_KERNEL);
3726 spin_lock(&memcg_oom_lock);
3728 event->eventfd = eventfd;
3729 list_add(&event->list, &memcg->oom_notify);
3731 /* already in OOM ? */
3732 if (memcg->under_oom)
3733 eventfd_signal(eventfd, 1);
3734 spin_unlock(&memcg_oom_lock);
3739 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3740 struct eventfd_ctx *eventfd)
3742 struct mem_cgroup_eventfd_list *ev, *tmp;
3744 spin_lock(&memcg_oom_lock);
3746 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3747 if (ev->eventfd == eventfd) {
3748 list_del(&ev->list);
3753 spin_unlock(&memcg_oom_lock);
3756 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3758 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3760 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3761 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3762 seq_printf(sf, "oom_kill %lu\n",
3763 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3767 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3768 struct cftype *cft, u64 val)
3770 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3772 /* cannot set to root cgroup and only 0 and 1 are allowed */
3773 if (!css->parent || !((val == 0) || (val == 1)))
3776 memcg->oom_kill_disable = val;
3778 memcg_oom_recover(memcg);
3783 #ifdef CONFIG_CGROUP_WRITEBACK
3785 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3787 return wb_domain_init(&memcg->cgwb_domain, gfp);
3790 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3792 wb_domain_exit(&memcg->cgwb_domain);
3795 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3797 wb_domain_size_changed(&memcg->cgwb_domain);
3800 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3802 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3804 if (!memcg->css.parent)
3807 return &memcg->cgwb_domain;
3811 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3812 * @wb: bdi_writeback in question
3813 * @pfilepages: out parameter for number of file pages
3814 * @pheadroom: out parameter for number of allocatable pages according to memcg
3815 * @pdirty: out parameter for number of dirty pages
3816 * @pwriteback: out parameter for number of pages under writeback
3818 * Determine the numbers of file, headroom, dirty, and writeback pages in
3819 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3820 * is a bit more involved.
3822 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3823 * headroom is calculated as the lowest headroom of itself and the
3824 * ancestors. Note that this doesn't consider the actual amount of
3825 * available memory in the system. The caller should further cap
3826 * *@pheadroom accordingly.
3828 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3829 unsigned long *pheadroom, unsigned long *pdirty,
3830 unsigned long *pwriteback)
3832 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3833 struct mem_cgroup *parent;
3835 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
3837 /* this should eventually include NR_UNSTABLE_NFS */
3838 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
3839 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3840 (1 << LRU_ACTIVE_FILE));
3841 *pheadroom = PAGE_COUNTER_MAX;
3843 while ((parent = parent_mem_cgroup(memcg))) {
3844 unsigned long ceiling = min(memcg->memory.max, memcg->high);
3845 unsigned long used = page_counter_read(&memcg->memory);
3847 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3852 #else /* CONFIG_CGROUP_WRITEBACK */
3854 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3859 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3863 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3867 #endif /* CONFIG_CGROUP_WRITEBACK */
3870 * DO NOT USE IN NEW FILES.
3872 * "cgroup.event_control" implementation.
3874 * This is way over-engineered. It tries to support fully configurable
3875 * events for each user. Such level of flexibility is completely
3876 * unnecessary especially in the light of the planned unified hierarchy.
3878 * Please deprecate this and replace with something simpler if at all
3883 * Unregister event and free resources.
3885 * Gets called from workqueue.
3887 static void memcg_event_remove(struct work_struct *work)
3889 struct mem_cgroup_event *event =
3890 container_of(work, struct mem_cgroup_event, remove);
3891 struct mem_cgroup *memcg = event->memcg;
3893 remove_wait_queue(event->wqh, &event->wait);
3895 event->unregister_event(memcg, event->eventfd);
3897 /* Notify userspace the event is going away. */
3898 eventfd_signal(event->eventfd, 1);
3900 eventfd_ctx_put(event->eventfd);
3902 css_put(&memcg->css);
3906 * Gets called on EPOLLHUP on eventfd when user closes it.
3908 * Called with wqh->lock held and interrupts disabled.
3910 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
3911 int sync, void *key)
3913 struct mem_cgroup_event *event =
3914 container_of(wait, struct mem_cgroup_event, wait);
3915 struct mem_cgroup *memcg = event->memcg;
3916 __poll_t flags = key_to_poll(key);
3918 if (flags & EPOLLHUP) {
3920 * If the event has been detached at cgroup removal, we
3921 * can simply return knowing the other side will cleanup
3924 * We can't race against event freeing since the other
3925 * side will require wqh->lock via remove_wait_queue(),
3928 spin_lock(&memcg->event_list_lock);
3929 if (!list_empty(&event->list)) {
3930 list_del_init(&event->list);
3932 * We are in atomic context, but cgroup_event_remove()
3933 * may sleep, so we have to call it in workqueue.
3935 schedule_work(&event->remove);
3937 spin_unlock(&memcg->event_list_lock);
3943 static void memcg_event_ptable_queue_proc(struct file *file,
3944 wait_queue_head_t *wqh, poll_table *pt)
3946 struct mem_cgroup_event *event =
3947 container_of(pt, struct mem_cgroup_event, pt);
3950 add_wait_queue(wqh, &event->wait);
3954 * DO NOT USE IN NEW FILES.
3956 * Parse input and register new cgroup event handler.
3958 * Input must be in format '<event_fd> <control_fd> <args>'.
3959 * Interpretation of args is defined by control file implementation.
3961 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3962 char *buf, size_t nbytes, loff_t off)
3964 struct cgroup_subsys_state *css = of_css(of);
3965 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3966 struct mem_cgroup_event *event;
3967 struct cgroup_subsys_state *cfile_css;
3968 unsigned int efd, cfd;
3975 buf = strstrip(buf);
3977 efd = simple_strtoul(buf, &endp, 10);
3982 cfd = simple_strtoul(buf, &endp, 10);
3983 if ((*endp != ' ') && (*endp != '\0'))
3987 event = kzalloc(sizeof(*event), GFP_KERNEL);
3991 event->memcg = memcg;
3992 INIT_LIST_HEAD(&event->list);
3993 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3994 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3995 INIT_WORK(&event->remove, memcg_event_remove);
4003 event->eventfd = eventfd_ctx_fileget(efile.file);
4004 if (IS_ERR(event->eventfd)) {
4005 ret = PTR_ERR(event->eventfd);
4012 goto out_put_eventfd;
4015 /* the process need read permission on control file */
4016 /* AV: shouldn't we check that it's been opened for read instead? */
4017 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4022 * Determine the event callbacks and set them in @event. This used
4023 * to be done via struct cftype but cgroup core no longer knows
4024 * about these events. The following is crude but the whole thing
4025 * is for compatibility anyway.
4027 * DO NOT ADD NEW FILES.
4029 name = cfile.file->f_path.dentry->d_name.name;
4031 if (!strcmp(name, "memory.usage_in_bytes")) {
4032 event->register_event = mem_cgroup_usage_register_event;
4033 event->unregister_event = mem_cgroup_usage_unregister_event;
4034 } else if (!strcmp(name, "memory.oom_control")) {
4035 event->register_event = mem_cgroup_oom_register_event;
4036 event->unregister_event = mem_cgroup_oom_unregister_event;
4037 } else if (!strcmp(name, "memory.pressure_level")) {
4038 event->register_event = vmpressure_register_event;
4039 event->unregister_event = vmpressure_unregister_event;
4040 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4041 event->register_event = memsw_cgroup_usage_register_event;
4042 event->unregister_event = memsw_cgroup_usage_unregister_event;
4049 * Verify @cfile should belong to @css. Also, remaining events are
4050 * automatically removed on cgroup destruction but the removal is
4051 * asynchronous, so take an extra ref on @css.
4053 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4054 &memory_cgrp_subsys);
4056 if (IS_ERR(cfile_css))
4058 if (cfile_css != css) {
4063 ret = event->register_event(memcg, event->eventfd, buf);
4067 vfs_poll(efile.file, &event->pt);
4069 spin_lock(&memcg->event_list_lock);
4070 list_add(&event->list, &memcg->event_list);
4071 spin_unlock(&memcg->event_list_lock);
4083 eventfd_ctx_put(event->eventfd);
4092 static struct cftype mem_cgroup_legacy_files[] = {
4094 .name = "usage_in_bytes",
4095 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4096 .read_u64 = mem_cgroup_read_u64,
4099 .name = "max_usage_in_bytes",
4100 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4101 .write = mem_cgroup_reset,
4102 .read_u64 = mem_cgroup_read_u64,
4105 .name = "limit_in_bytes",
4106 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4107 .write = mem_cgroup_write,
4108 .read_u64 = mem_cgroup_read_u64,
4111 .name = "soft_limit_in_bytes",
4112 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4113 .write = mem_cgroup_write,
4114 .read_u64 = mem_cgroup_read_u64,
4118 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4119 .write = mem_cgroup_reset,
4120 .read_u64 = mem_cgroup_read_u64,
4124 .seq_show = memcg_stat_show,
4127 .name = "force_empty",
4128 .write = mem_cgroup_force_empty_write,
4131 .name = "use_hierarchy",
4132 .write_u64 = mem_cgroup_hierarchy_write,
4133 .read_u64 = mem_cgroup_hierarchy_read,
4136 .name = "cgroup.event_control", /* XXX: for compat */
4137 .write = memcg_write_event_control,
4138 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4141 .name = "swappiness",
4142 .read_u64 = mem_cgroup_swappiness_read,
4143 .write_u64 = mem_cgroup_swappiness_write,
4146 .name = "move_charge_at_immigrate",
4147 .read_u64 = mem_cgroup_move_charge_read,
4148 .write_u64 = mem_cgroup_move_charge_write,
4151 .name = "oom_control",
4152 .seq_show = mem_cgroup_oom_control_read,
4153 .write_u64 = mem_cgroup_oom_control_write,
4154 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4157 .name = "pressure_level",
4161 .name = "numa_stat",
4162 .seq_show = memcg_numa_stat_show,
4166 .name = "kmem.limit_in_bytes",
4167 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4168 .write = mem_cgroup_write,
4169 .read_u64 = mem_cgroup_read_u64,
4172 .name = "kmem.usage_in_bytes",
4173 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4174 .read_u64 = mem_cgroup_read_u64,
4177 .name = "kmem.failcnt",
4178 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4179 .write = mem_cgroup_reset,
4180 .read_u64 = mem_cgroup_read_u64,
4183 .name = "kmem.max_usage_in_bytes",
4184 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4185 .write = mem_cgroup_reset,
4186 .read_u64 = mem_cgroup_read_u64,
4188 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4190 .name = "kmem.slabinfo",
4191 .seq_start = memcg_slab_start,
4192 .seq_next = memcg_slab_next,
4193 .seq_stop = memcg_slab_stop,
4194 .seq_show = memcg_slab_show,
4198 .name = "kmem.tcp.limit_in_bytes",
4199 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4200 .write = mem_cgroup_write,
4201 .read_u64 = mem_cgroup_read_u64,
4204 .name = "kmem.tcp.usage_in_bytes",
4205 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4206 .read_u64 = mem_cgroup_read_u64,
4209 .name = "kmem.tcp.failcnt",
4210 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4211 .write = mem_cgroup_reset,
4212 .read_u64 = mem_cgroup_read_u64,
4215 .name = "kmem.tcp.max_usage_in_bytes",
4216 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4217 .write = mem_cgroup_reset,
4218 .read_u64 = mem_cgroup_read_u64,
4220 { }, /* terminate */
4224 * Private memory cgroup IDR
4226 * Swap-out records and page cache shadow entries need to store memcg
4227 * references in constrained space, so we maintain an ID space that is
4228 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4229 * memory-controlled cgroups to 64k.
4231 * However, there usually are many references to the oflline CSS after
4232 * the cgroup has been destroyed, such as page cache or reclaimable
4233 * slab objects, that don't need to hang on to the ID. We want to keep
4234 * those dead CSS from occupying IDs, or we might quickly exhaust the
4235 * relatively small ID space and prevent the creation of new cgroups
4236 * even when there are much fewer than 64k cgroups - possibly none.
4238 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4239 * be freed and recycled when it's no longer needed, which is usually
4240 * when the CSS is offlined.
4242 * The only exception to that are records of swapped out tmpfs/shmem
4243 * pages that need to be attributed to live ancestors on swapin. But
4244 * those references are manageable from userspace.
4247 static DEFINE_IDR(mem_cgroup_idr);
4249 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4251 if (memcg->id.id > 0) {
4252 idr_remove(&mem_cgroup_idr, memcg->id.id);
4257 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4259 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4260 atomic_add(n, &memcg->id.ref);
4263 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4265 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4266 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4267 mem_cgroup_id_remove(memcg);
4269 /* Memcg ID pins CSS */
4270 css_put(&memcg->css);
4274 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4276 mem_cgroup_id_get_many(memcg, 1);
4279 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4281 mem_cgroup_id_put_many(memcg, 1);
4285 * mem_cgroup_from_id - look up a memcg from a memcg id
4286 * @id: the memcg id to look up
4288 * Caller must hold rcu_read_lock().
4290 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4292 WARN_ON_ONCE(!rcu_read_lock_held());
4293 return idr_find(&mem_cgroup_idr, id);
4296 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4298 struct mem_cgroup_per_node *pn;
4301 * This routine is called against possible nodes.
4302 * But it's BUG to call kmalloc() against offline node.
4304 * TODO: this routine can waste much memory for nodes which will
4305 * never be onlined. It's better to use memory hotplug callback
4308 if (!node_state(node, N_NORMAL_MEMORY))
4310 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4314 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4315 if (!pn->lruvec_stat_cpu) {
4320 lruvec_init(&pn->lruvec);
4321 pn->usage_in_excess = 0;
4322 pn->on_tree = false;
4325 memcg->nodeinfo[node] = pn;
4329 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4331 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4336 free_percpu(pn->lruvec_stat_cpu);
4340 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4345 free_mem_cgroup_per_node_info(memcg, node);
4346 free_percpu(memcg->stat_cpu);
4350 static void mem_cgroup_free(struct mem_cgroup *memcg)
4352 memcg_wb_domain_exit(memcg);
4353 __mem_cgroup_free(memcg);
4356 static struct mem_cgroup *mem_cgroup_alloc(void)
4358 struct mem_cgroup *memcg;
4362 size = sizeof(struct mem_cgroup);
4363 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4365 memcg = kzalloc(size, GFP_KERNEL);
4369 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4370 1, MEM_CGROUP_ID_MAX,
4372 if (memcg->id.id < 0)
4375 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4376 if (!memcg->stat_cpu)
4380 if (alloc_mem_cgroup_per_node_info(memcg, node))
4383 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4386 INIT_WORK(&memcg->high_work, high_work_func);
4387 memcg->last_scanned_node = MAX_NUMNODES;
4388 INIT_LIST_HEAD(&memcg->oom_notify);
4389 mutex_init(&memcg->thresholds_lock);
4390 spin_lock_init(&memcg->move_lock);
4391 vmpressure_init(&memcg->vmpressure);
4392 INIT_LIST_HEAD(&memcg->event_list);
4393 spin_lock_init(&memcg->event_list_lock);
4394 memcg->socket_pressure = jiffies;
4395 #ifdef CONFIG_MEMCG_KMEM
4396 memcg->kmemcg_id = -1;
4398 #ifdef CONFIG_CGROUP_WRITEBACK
4399 INIT_LIST_HEAD(&memcg->cgwb_list);
4401 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4404 mem_cgroup_id_remove(memcg);
4405 __mem_cgroup_free(memcg);
4409 static struct cgroup_subsys_state * __ref
4410 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4412 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4413 struct mem_cgroup *memcg;
4414 long error = -ENOMEM;
4416 memcg = mem_cgroup_alloc();
4418 return ERR_PTR(error);
4420 memcg->high = PAGE_COUNTER_MAX;
4421 memcg->soft_limit = PAGE_COUNTER_MAX;
4423 memcg->swappiness = mem_cgroup_swappiness(parent);
4424 memcg->oom_kill_disable = parent->oom_kill_disable;
4426 if (parent && parent->use_hierarchy) {
4427 memcg->use_hierarchy = true;
4428 page_counter_init(&memcg->memory, &parent->memory);
4429 page_counter_init(&memcg->swap, &parent->swap);
4430 page_counter_init(&memcg->memsw, &parent->memsw);
4431 page_counter_init(&memcg->kmem, &parent->kmem);
4432 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4434 page_counter_init(&memcg->memory, NULL);
4435 page_counter_init(&memcg->swap, NULL);
4436 page_counter_init(&memcg->memsw, NULL);
4437 page_counter_init(&memcg->kmem, NULL);
4438 page_counter_init(&memcg->tcpmem, NULL);
4440 * Deeper hierachy with use_hierarchy == false doesn't make
4441 * much sense so let cgroup subsystem know about this
4442 * unfortunate state in our controller.
4444 if (parent != root_mem_cgroup)
4445 memory_cgrp_subsys.broken_hierarchy = true;
4448 /* The following stuff does not apply to the root */
4450 root_mem_cgroup = memcg;
4454 error = memcg_online_kmem(memcg);
4458 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4459 static_branch_inc(&memcg_sockets_enabled_key);
4463 mem_cgroup_id_remove(memcg);
4464 mem_cgroup_free(memcg);
4465 return ERR_PTR(-ENOMEM);
4468 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4470 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4473 * A memcg must be visible for memcg_expand_shrinker_maps()
4474 * by the time the maps are allocated. So, we allocate maps
4475 * here, when for_each_mem_cgroup() can't skip it.
4477 if (memcg_alloc_shrinker_maps(memcg)) {
4478 mem_cgroup_id_remove(memcg);
4482 /* Online state pins memcg ID, memcg ID pins CSS */
4483 atomic_set(&memcg->id.ref, 1);
4488 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4490 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4491 struct mem_cgroup_event *event, *tmp;
4494 * Unregister events and notify userspace.
4495 * Notify userspace about cgroup removing only after rmdir of cgroup
4496 * directory to avoid race between userspace and kernelspace.
4498 spin_lock(&memcg->event_list_lock);
4499 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4500 list_del_init(&event->list);
4501 schedule_work(&event->remove);
4503 spin_unlock(&memcg->event_list_lock);
4505 page_counter_set_min(&memcg->memory, 0);
4506 page_counter_set_low(&memcg->memory, 0);
4508 memcg_offline_kmem(memcg);
4509 wb_memcg_offline(memcg);
4511 mem_cgroup_id_put(memcg);
4514 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4516 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4518 invalidate_reclaim_iterators(memcg);
4521 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4523 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4525 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4526 static_branch_dec(&memcg_sockets_enabled_key);
4528 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4529 static_branch_dec(&memcg_sockets_enabled_key);
4531 vmpressure_cleanup(&memcg->vmpressure);
4532 cancel_work_sync(&memcg->high_work);
4533 mem_cgroup_remove_from_trees(memcg);
4534 memcg_free_shrinker_maps(memcg);
4535 memcg_free_kmem(memcg);
4536 mem_cgroup_free(memcg);
4540 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4541 * @css: the target css
4543 * Reset the states of the mem_cgroup associated with @css. This is
4544 * invoked when the userland requests disabling on the default hierarchy
4545 * but the memcg is pinned through dependency. The memcg should stop
4546 * applying policies and should revert to the vanilla state as it may be
4547 * made visible again.
4549 * The current implementation only resets the essential configurations.
4550 * This needs to be expanded to cover all the visible parts.
4552 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4554 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4556 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4557 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4558 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4559 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4560 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4561 page_counter_set_min(&memcg->memory, 0);
4562 page_counter_set_low(&memcg->memory, 0);
4563 memcg->high = PAGE_COUNTER_MAX;
4564 memcg->soft_limit = PAGE_COUNTER_MAX;
4565 memcg_wb_domain_size_changed(memcg);
4569 /* Handlers for move charge at task migration. */
4570 static int mem_cgroup_do_precharge(unsigned long count)
4574 /* Try a single bulk charge without reclaim first, kswapd may wake */
4575 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4577 mc.precharge += count;
4581 /* Try charges one by one with reclaim, but do not retry */
4583 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4597 enum mc_target_type {
4604 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4605 unsigned long addr, pte_t ptent)
4607 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4609 if (!page || !page_mapped(page))
4611 if (PageAnon(page)) {
4612 if (!(mc.flags & MOVE_ANON))
4615 if (!(mc.flags & MOVE_FILE))
4618 if (!get_page_unless_zero(page))
4624 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4625 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4626 pte_t ptent, swp_entry_t *entry)
4628 struct page *page = NULL;
4629 swp_entry_t ent = pte_to_swp_entry(ptent);
4631 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4635 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4636 * a device and because they are not accessible by CPU they are store
4637 * as special swap entry in the CPU page table.
4639 if (is_device_private_entry(ent)) {
4640 page = device_private_entry_to_page(ent);
4642 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4643 * a refcount of 1 when free (unlike normal page)
4645 if (!page_ref_add_unless(page, 1, 1))
4651 * Because lookup_swap_cache() updates some statistics counter,
4652 * we call find_get_page() with swapper_space directly.
4654 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4655 if (do_memsw_account())
4656 entry->val = ent.val;
4661 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4662 pte_t ptent, swp_entry_t *entry)
4668 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4669 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4671 struct page *page = NULL;
4672 struct address_space *mapping;
4675 if (!vma->vm_file) /* anonymous vma */
4677 if (!(mc.flags & MOVE_FILE))
4680 mapping = vma->vm_file->f_mapping;
4681 pgoff = linear_page_index(vma, addr);
4683 /* page is moved even if it's not RSS of this task(page-faulted). */
4685 /* shmem/tmpfs may report page out on swap: account for that too. */
4686 if (shmem_mapping(mapping)) {
4687 page = find_get_entry(mapping, pgoff);
4688 if (radix_tree_exceptional_entry(page)) {
4689 swp_entry_t swp = radix_to_swp_entry(page);
4690 if (do_memsw_account())
4692 page = find_get_page(swap_address_space(swp),
4696 page = find_get_page(mapping, pgoff);
4698 page = find_get_page(mapping, pgoff);
4704 * mem_cgroup_move_account - move account of the page
4706 * @compound: charge the page as compound or small page
4707 * @from: mem_cgroup which the page is moved from.
4708 * @to: mem_cgroup which the page is moved to. @from != @to.
4710 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4712 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4715 static int mem_cgroup_move_account(struct page *page,
4717 struct mem_cgroup *from,
4718 struct mem_cgroup *to)
4720 unsigned long flags;
4721 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4725 VM_BUG_ON(from == to);
4726 VM_BUG_ON_PAGE(PageLRU(page), page);
4727 VM_BUG_ON(compound && !PageTransHuge(page));
4730 * Prevent mem_cgroup_migrate() from looking at
4731 * page->mem_cgroup of its source page while we change it.
4734 if (!trylock_page(page))
4738 if (page->mem_cgroup != from)
4741 anon = PageAnon(page);
4743 spin_lock_irqsave(&from->move_lock, flags);
4745 if (!anon && page_mapped(page)) {
4746 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4747 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4751 * move_lock grabbed above and caller set from->moving_account, so
4752 * mod_memcg_page_state will serialize updates to PageDirty.
4753 * So mapping should be stable for dirty pages.
4755 if (!anon && PageDirty(page)) {
4756 struct address_space *mapping = page_mapping(page);
4758 if (mapping_cap_account_dirty(mapping)) {
4759 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4760 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4764 if (PageWriteback(page)) {
4765 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4766 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4770 * It is safe to change page->mem_cgroup here because the page
4771 * is referenced, charged, and isolated - we can't race with
4772 * uncharging, charging, migration, or LRU putback.
4775 /* caller should have done css_get */
4776 page->mem_cgroup = to;
4777 spin_unlock_irqrestore(&from->move_lock, flags);
4781 local_irq_disable();
4782 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4783 memcg_check_events(to, page);
4784 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4785 memcg_check_events(from, page);
4794 * get_mctgt_type - get target type of moving charge
4795 * @vma: the vma the pte to be checked belongs
4796 * @addr: the address corresponding to the pte to be checked
4797 * @ptent: the pte to be checked
4798 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4801 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4802 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4803 * move charge. if @target is not NULL, the page is stored in target->page
4804 * with extra refcnt got(Callers should handle it).
4805 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4806 * target for charge migration. if @target is not NULL, the entry is stored
4808 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4809 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4810 * For now we such page is charge like a regular page would be as for all
4811 * intent and purposes it is just special memory taking the place of a
4814 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4816 * Called with pte lock held.
4819 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4820 unsigned long addr, pte_t ptent, union mc_target *target)
4822 struct page *page = NULL;
4823 enum mc_target_type ret = MC_TARGET_NONE;
4824 swp_entry_t ent = { .val = 0 };
4826 if (pte_present(ptent))
4827 page = mc_handle_present_pte(vma, addr, ptent);
4828 else if (is_swap_pte(ptent))
4829 page = mc_handle_swap_pte(vma, ptent, &ent);
4830 else if (pte_none(ptent))
4831 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4833 if (!page && !ent.val)
4837 * Do only loose check w/o serialization.
4838 * mem_cgroup_move_account() checks the page is valid or
4839 * not under LRU exclusion.
4841 if (page->mem_cgroup == mc.from) {
4842 ret = MC_TARGET_PAGE;
4843 if (is_device_private_page(page) ||
4844 is_device_public_page(page))
4845 ret = MC_TARGET_DEVICE;
4847 target->page = page;
4849 if (!ret || !target)
4853 * There is a swap entry and a page doesn't exist or isn't charged.
4854 * But we cannot move a tail-page in a THP.
4856 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
4857 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4858 ret = MC_TARGET_SWAP;
4865 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4867 * We don't consider PMD mapped swapping or file mapped pages because THP does
4868 * not support them for now.
4869 * Caller should make sure that pmd_trans_huge(pmd) is true.
4871 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4872 unsigned long addr, pmd_t pmd, union mc_target *target)
4874 struct page *page = NULL;
4875 enum mc_target_type ret = MC_TARGET_NONE;
4877 if (unlikely(is_swap_pmd(pmd))) {
4878 VM_BUG_ON(thp_migration_supported() &&
4879 !is_pmd_migration_entry(pmd));
4882 page = pmd_page(pmd);
4883 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4884 if (!(mc.flags & MOVE_ANON))
4886 if (page->mem_cgroup == mc.from) {
4887 ret = MC_TARGET_PAGE;
4890 target->page = page;
4896 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4897 unsigned long addr, pmd_t pmd, union mc_target *target)
4899 return MC_TARGET_NONE;
4903 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4904 unsigned long addr, unsigned long end,
4905 struct mm_walk *walk)
4907 struct vm_area_struct *vma = walk->vma;
4911 ptl = pmd_trans_huge_lock(pmd, vma);
4914 * Note their can not be MC_TARGET_DEVICE for now as we do not
4915 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
4916 * MEMORY_DEVICE_PRIVATE but this might change.
4918 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4919 mc.precharge += HPAGE_PMD_NR;
4924 if (pmd_trans_unstable(pmd))
4926 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4927 for (; addr != end; pte++, addr += PAGE_SIZE)
4928 if (get_mctgt_type(vma, addr, *pte, NULL))
4929 mc.precharge++; /* increment precharge temporarily */
4930 pte_unmap_unlock(pte - 1, ptl);
4936 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4938 unsigned long precharge;
4940 struct mm_walk mem_cgroup_count_precharge_walk = {
4941 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4944 down_read(&mm->mmap_sem);
4945 walk_page_range(0, mm->highest_vm_end,
4946 &mem_cgroup_count_precharge_walk);
4947 up_read(&mm->mmap_sem);
4949 precharge = mc.precharge;
4955 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4957 unsigned long precharge = mem_cgroup_count_precharge(mm);
4959 VM_BUG_ON(mc.moving_task);
4960 mc.moving_task = current;
4961 return mem_cgroup_do_precharge(precharge);
4964 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4965 static void __mem_cgroup_clear_mc(void)
4967 struct mem_cgroup *from = mc.from;
4968 struct mem_cgroup *to = mc.to;
4970 /* we must uncharge all the leftover precharges from mc.to */
4972 cancel_charge(mc.to, mc.precharge);
4976 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4977 * we must uncharge here.
4979 if (mc.moved_charge) {
4980 cancel_charge(mc.from, mc.moved_charge);
4981 mc.moved_charge = 0;
4983 /* we must fixup refcnts and charges */
4984 if (mc.moved_swap) {
4985 /* uncharge swap account from the old cgroup */
4986 if (!mem_cgroup_is_root(mc.from))
4987 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4989 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4992 * we charged both to->memory and to->memsw, so we
4993 * should uncharge to->memory.
4995 if (!mem_cgroup_is_root(mc.to))
4996 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4998 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4999 css_put_many(&mc.to->css, mc.moved_swap);
5003 memcg_oom_recover(from);
5004 memcg_oom_recover(to);
5005 wake_up_all(&mc.waitq);
5008 static void mem_cgroup_clear_mc(void)
5010 struct mm_struct *mm = mc.mm;
5013 * we must clear moving_task before waking up waiters at the end of
5016 mc.moving_task = NULL;
5017 __mem_cgroup_clear_mc();
5018 spin_lock(&mc.lock);
5022 spin_unlock(&mc.lock);
5027 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5029 struct cgroup_subsys_state *css;
5030 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5031 struct mem_cgroup *from;
5032 struct task_struct *leader, *p;
5033 struct mm_struct *mm;
5034 unsigned long move_flags;
5037 /* charge immigration isn't supported on the default hierarchy */
5038 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5042 * Multi-process migrations only happen on the default hierarchy
5043 * where charge immigration is not used. Perform charge
5044 * immigration if @tset contains a leader and whine if there are
5048 cgroup_taskset_for_each_leader(leader, css, tset) {
5051 memcg = mem_cgroup_from_css(css);
5057 * We are now commited to this value whatever it is. Changes in this
5058 * tunable will only affect upcoming migrations, not the current one.
5059 * So we need to save it, and keep it going.
5061 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5065 from = mem_cgroup_from_task(p);
5067 VM_BUG_ON(from == memcg);
5069 mm = get_task_mm(p);
5072 /* We move charges only when we move a owner of the mm */
5073 if (mm->owner == p) {
5076 VM_BUG_ON(mc.precharge);
5077 VM_BUG_ON(mc.moved_charge);
5078 VM_BUG_ON(mc.moved_swap);
5080 spin_lock(&mc.lock);
5084 mc.flags = move_flags;
5085 spin_unlock(&mc.lock);
5086 /* We set mc.moving_task later */
5088 ret = mem_cgroup_precharge_mc(mm);
5090 mem_cgroup_clear_mc();
5097 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5100 mem_cgroup_clear_mc();
5103 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5104 unsigned long addr, unsigned long end,
5105 struct mm_walk *walk)
5108 struct vm_area_struct *vma = walk->vma;
5111 enum mc_target_type target_type;
5112 union mc_target target;
5115 ptl = pmd_trans_huge_lock(pmd, vma);
5117 if (mc.precharge < HPAGE_PMD_NR) {
5121 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5122 if (target_type == MC_TARGET_PAGE) {
5124 if (!isolate_lru_page(page)) {
5125 if (!mem_cgroup_move_account(page, true,
5127 mc.precharge -= HPAGE_PMD_NR;
5128 mc.moved_charge += HPAGE_PMD_NR;
5130 putback_lru_page(page);
5133 } else if (target_type == MC_TARGET_DEVICE) {
5135 if (!mem_cgroup_move_account(page, true,
5137 mc.precharge -= HPAGE_PMD_NR;
5138 mc.moved_charge += HPAGE_PMD_NR;
5146 if (pmd_trans_unstable(pmd))
5149 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5150 for (; addr != end; addr += PAGE_SIZE) {
5151 pte_t ptent = *(pte++);
5152 bool device = false;
5158 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5159 case MC_TARGET_DEVICE:
5162 case MC_TARGET_PAGE:
5165 * We can have a part of the split pmd here. Moving it
5166 * can be done but it would be too convoluted so simply
5167 * ignore such a partial THP and keep it in original
5168 * memcg. There should be somebody mapping the head.
5170 if (PageTransCompound(page))
5172 if (!device && isolate_lru_page(page))
5174 if (!mem_cgroup_move_account(page, false,
5177 /* we uncharge from mc.from later. */
5181 putback_lru_page(page);
5182 put: /* get_mctgt_type() gets the page */
5185 case MC_TARGET_SWAP:
5187 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5189 /* we fixup refcnts and charges later. */
5197 pte_unmap_unlock(pte - 1, ptl);
5202 * We have consumed all precharges we got in can_attach().
5203 * We try charge one by one, but don't do any additional
5204 * charges to mc.to if we have failed in charge once in attach()
5207 ret = mem_cgroup_do_precharge(1);
5215 static void mem_cgroup_move_charge(void)
5217 struct mm_walk mem_cgroup_move_charge_walk = {
5218 .pmd_entry = mem_cgroup_move_charge_pte_range,
5222 lru_add_drain_all();
5224 * Signal lock_page_memcg() to take the memcg's move_lock
5225 * while we're moving its pages to another memcg. Then wait
5226 * for already started RCU-only updates to finish.
5228 atomic_inc(&mc.from->moving_account);
5231 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5233 * Someone who are holding the mmap_sem might be waiting in
5234 * waitq. So we cancel all extra charges, wake up all waiters,
5235 * and retry. Because we cancel precharges, we might not be able
5236 * to move enough charges, but moving charge is a best-effort
5237 * feature anyway, so it wouldn't be a big problem.
5239 __mem_cgroup_clear_mc();
5244 * When we have consumed all precharges and failed in doing
5245 * additional charge, the page walk just aborts.
5247 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5249 up_read(&mc.mm->mmap_sem);
5250 atomic_dec(&mc.from->moving_account);
5253 static void mem_cgroup_move_task(void)
5256 mem_cgroup_move_charge();
5257 mem_cgroup_clear_mc();
5260 #else /* !CONFIG_MMU */
5261 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5265 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5268 static void mem_cgroup_move_task(void)
5274 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5275 * to verify whether we're attached to the default hierarchy on each mount
5278 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5281 * use_hierarchy is forced on the default hierarchy. cgroup core
5282 * guarantees that @root doesn't have any children, so turning it
5283 * on for the root memcg is enough.
5285 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5286 root_mem_cgroup->use_hierarchy = true;
5288 root_mem_cgroup->use_hierarchy = false;
5291 static u64 memory_current_read(struct cgroup_subsys_state *css,
5294 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5296 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5299 static int memory_min_show(struct seq_file *m, void *v)
5301 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5302 unsigned long min = READ_ONCE(memcg->memory.min);
5304 if (min == PAGE_COUNTER_MAX)
5305 seq_puts(m, "max\n");
5307 seq_printf(m, "%llu\n", (u64)min * PAGE_SIZE);
5312 static ssize_t memory_min_write(struct kernfs_open_file *of,
5313 char *buf, size_t nbytes, loff_t off)
5315 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5319 buf = strstrip(buf);
5320 err = page_counter_memparse(buf, "max", &min);
5324 page_counter_set_min(&memcg->memory, min);
5329 static int memory_low_show(struct seq_file *m, void *v)
5331 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5332 unsigned long low = READ_ONCE(memcg->memory.low);
5334 if (low == PAGE_COUNTER_MAX)
5335 seq_puts(m, "max\n");
5337 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5342 static ssize_t memory_low_write(struct kernfs_open_file *of,
5343 char *buf, size_t nbytes, loff_t off)
5345 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5349 buf = strstrip(buf);
5350 err = page_counter_memparse(buf, "max", &low);
5354 page_counter_set_low(&memcg->memory, low);
5359 static int memory_high_show(struct seq_file *m, void *v)
5361 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5362 unsigned long high = READ_ONCE(memcg->high);
5364 if (high == PAGE_COUNTER_MAX)
5365 seq_puts(m, "max\n");
5367 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5372 static ssize_t memory_high_write(struct kernfs_open_file *of,
5373 char *buf, size_t nbytes, loff_t off)
5375 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5376 unsigned long nr_pages;
5380 buf = strstrip(buf);
5381 err = page_counter_memparse(buf, "max", &high);
5387 nr_pages = page_counter_read(&memcg->memory);
5388 if (nr_pages > high)
5389 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5392 memcg_wb_domain_size_changed(memcg);
5396 static int memory_max_show(struct seq_file *m, void *v)
5398 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5399 unsigned long max = READ_ONCE(memcg->memory.max);
5401 if (max == PAGE_COUNTER_MAX)
5402 seq_puts(m, "max\n");
5404 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5409 static ssize_t memory_max_write(struct kernfs_open_file *of,
5410 char *buf, size_t nbytes, loff_t off)
5412 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5413 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5414 bool drained = false;
5418 buf = strstrip(buf);
5419 err = page_counter_memparse(buf, "max", &max);
5423 xchg(&memcg->memory.max, max);
5426 unsigned long nr_pages = page_counter_read(&memcg->memory);
5428 if (nr_pages <= max)
5431 if (signal_pending(current)) {
5437 drain_all_stock(memcg);
5443 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5449 memcg_memory_event(memcg, MEMCG_OOM);
5450 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5454 memcg_wb_domain_size_changed(memcg);
5458 static int memory_events_show(struct seq_file *m, void *v)
5460 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5462 seq_printf(m, "low %lu\n",
5463 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5464 seq_printf(m, "high %lu\n",
5465 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5466 seq_printf(m, "max %lu\n",
5467 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5468 seq_printf(m, "oom %lu\n",
5469 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5470 seq_printf(m, "oom_kill %lu\n",
5471 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5476 static int memory_stat_show(struct seq_file *m, void *v)
5478 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5479 unsigned long stat[MEMCG_NR_STAT];
5480 unsigned long events[NR_VM_EVENT_ITEMS];
5484 * Provide statistics on the state of the memory subsystem as
5485 * well as cumulative event counters that show past behavior.
5487 * This list is ordered following a combination of these gradients:
5488 * 1) generic big picture -> specifics and details
5489 * 2) reflecting userspace activity -> reflecting kernel heuristics
5491 * Current memory state:
5494 tree_stat(memcg, stat);
5495 tree_events(memcg, events);
5497 seq_printf(m, "anon %llu\n",
5498 (u64)stat[MEMCG_RSS] * PAGE_SIZE);
5499 seq_printf(m, "file %llu\n",
5500 (u64)stat[MEMCG_CACHE] * PAGE_SIZE);
5501 seq_printf(m, "kernel_stack %llu\n",
5502 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5503 seq_printf(m, "slab %llu\n",
5504 (u64)(stat[NR_SLAB_RECLAIMABLE] +
5505 stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5506 seq_printf(m, "sock %llu\n",
5507 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5509 seq_printf(m, "shmem %llu\n",
5510 (u64)stat[NR_SHMEM] * PAGE_SIZE);
5511 seq_printf(m, "file_mapped %llu\n",
5512 (u64)stat[NR_FILE_MAPPED] * PAGE_SIZE);
5513 seq_printf(m, "file_dirty %llu\n",
5514 (u64)stat[NR_FILE_DIRTY] * PAGE_SIZE);
5515 seq_printf(m, "file_writeback %llu\n",
5516 (u64)stat[NR_WRITEBACK] * PAGE_SIZE);
5518 for (i = 0; i < NR_LRU_LISTS; i++) {
5519 struct mem_cgroup *mi;
5520 unsigned long val = 0;
5522 for_each_mem_cgroup_tree(mi, memcg)
5523 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5524 seq_printf(m, "%s %llu\n",
5525 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5528 seq_printf(m, "slab_reclaimable %llu\n",
5529 (u64)stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5530 seq_printf(m, "slab_unreclaimable %llu\n",
5531 (u64)stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5533 /* Accumulated memory events */
5535 seq_printf(m, "pgfault %lu\n", events[PGFAULT]);
5536 seq_printf(m, "pgmajfault %lu\n", events[PGMAJFAULT]);
5538 seq_printf(m, "pgrefill %lu\n", events[PGREFILL]);
5539 seq_printf(m, "pgscan %lu\n", events[PGSCAN_KSWAPD] +
5540 events[PGSCAN_DIRECT]);
5541 seq_printf(m, "pgsteal %lu\n", events[PGSTEAL_KSWAPD] +
5542 events[PGSTEAL_DIRECT]);
5543 seq_printf(m, "pgactivate %lu\n", events[PGACTIVATE]);
5544 seq_printf(m, "pgdeactivate %lu\n", events[PGDEACTIVATE]);
5545 seq_printf(m, "pglazyfree %lu\n", events[PGLAZYFREE]);
5546 seq_printf(m, "pglazyfreed %lu\n", events[PGLAZYFREED]);
5548 seq_printf(m, "workingset_refault %lu\n",
5549 stat[WORKINGSET_REFAULT]);
5550 seq_printf(m, "workingset_activate %lu\n",
5551 stat[WORKINGSET_ACTIVATE]);
5552 seq_printf(m, "workingset_nodereclaim %lu\n",
5553 stat[WORKINGSET_NODERECLAIM]);
5558 static struct cftype memory_files[] = {
5561 .flags = CFTYPE_NOT_ON_ROOT,
5562 .read_u64 = memory_current_read,
5566 .flags = CFTYPE_NOT_ON_ROOT,
5567 .seq_show = memory_min_show,
5568 .write = memory_min_write,
5572 .flags = CFTYPE_NOT_ON_ROOT,
5573 .seq_show = memory_low_show,
5574 .write = memory_low_write,
5578 .flags = CFTYPE_NOT_ON_ROOT,
5579 .seq_show = memory_high_show,
5580 .write = memory_high_write,
5584 .flags = CFTYPE_NOT_ON_ROOT,
5585 .seq_show = memory_max_show,
5586 .write = memory_max_write,
5590 .flags = CFTYPE_NOT_ON_ROOT,
5591 .file_offset = offsetof(struct mem_cgroup, events_file),
5592 .seq_show = memory_events_show,
5596 .flags = CFTYPE_NOT_ON_ROOT,
5597 .seq_show = memory_stat_show,
5602 struct cgroup_subsys memory_cgrp_subsys = {
5603 .css_alloc = mem_cgroup_css_alloc,
5604 .css_online = mem_cgroup_css_online,
5605 .css_offline = mem_cgroup_css_offline,
5606 .css_released = mem_cgroup_css_released,
5607 .css_free = mem_cgroup_css_free,
5608 .css_reset = mem_cgroup_css_reset,
5609 .can_attach = mem_cgroup_can_attach,
5610 .cancel_attach = mem_cgroup_cancel_attach,
5611 .post_attach = mem_cgroup_move_task,
5612 .bind = mem_cgroup_bind,
5613 .dfl_cftypes = memory_files,
5614 .legacy_cftypes = mem_cgroup_legacy_files,
5619 * mem_cgroup_protected - check if memory consumption is in the normal range
5620 * @root: the top ancestor of the sub-tree being checked
5621 * @memcg: the memory cgroup to check
5623 * WARNING: This function is not stateless! It can only be used as part
5624 * of a top-down tree iteration, not for isolated queries.
5626 * Returns one of the following:
5627 * MEMCG_PROT_NONE: cgroup memory is not protected
5628 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5629 * an unprotected supply of reclaimable memory from other cgroups.
5630 * MEMCG_PROT_MIN: cgroup memory is protected
5632 * @root is exclusive; it is never protected when looked at directly
5634 * To provide a proper hierarchical behavior, effective memory.min/low values
5635 * are used. Below is the description of how effective memory.low is calculated.
5636 * Effective memory.min values is calculated in the same way.
5638 * Effective memory.low is always equal or less than the original memory.low.
5639 * If there is no memory.low overcommittment (which is always true for
5640 * top-level memory cgroups), these two values are equal.
5641 * Otherwise, it's a part of parent's effective memory.low,
5642 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5643 * memory.low usages, where memory.low usage is the size of actually
5647 * elow = min( memory.low, parent->elow * ------------------ ),
5648 * siblings_low_usage
5650 * | memory.current, if memory.current < memory.low
5655 * Such definition of the effective memory.low provides the expected
5656 * hierarchical behavior: parent's memory.low value is limiting
5657 * children, unprotected memory is reclaimed first and cgroups,
5658 * which are not using their guarantee do not affect actual memory
5661 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5663 * A A/memory.low = 2G, A/memory.current = 6G
5665 * BC DE B/memory.low = 3G B/memory.current = 2G
5666 * C/memory.low = 1G C/memory.current = 2G
5667 * D/memory.low = 0 D/memory.current = 2G
5668 * E/memory.low = 10G E/memory.current = 0
5670 * and the memory pressure is applied, the following memory distribution
5671 * is expected (approximately):
5673 * A/memory.current = 2G
5675 * B/memory.current = 1.3G
5676 * C/memory.current = 0.6G
5677 * D/memory.current = 0
5678 * E/memory.current = 0
5680 * These calculations require constant tracking of the actual low usages
5681 * (see propagate_protected_usage()), as well as recursive calculation of
5682 * effective memory.low values. But as we do call mem_cgroup_protected()
5683 * path for each memory cgroup top-down from the reclaim,
5684 * it's possible to optimize this part, and save calculated elow
5685 * for next usage. This part is intentionally racy, but it's ok,
5686 * as memory.low is a best-effort mechanism.
5688 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5689 struct mem_cgroup *memcg)
5691 struct mem_cgroup *parent;
5692 unsigned long emin, parent_emin;
5693 unsigned long elow, parent_elow;
5694 unsigned long usage;
5696 if (mem_cgroup_disabled())
5697 return MEMCG_PROT_NONE;
5700 root = root_mem_cgroup;
5702 return MEMCG_PROT_NONE;
5704 usage = page_counter_read(&memcg->memory);
5706 return MEMCG_PROT_NONE;
5708 emin = memcg->memory.min;
5709 elow = memcg->memory.low;
5711 parent = parent_mem_cgroup(memcg);
5712 /* No parent means a non-hierarchical mode on v1 memcg */
5714 return MEMCG_PROT_NONE;
5719 parent_emin = READ_ONCE(parent->memory.emin);
5720 emin = min(emin, parent_emin);
5721 if (emin && parent_emin) {
5722 unsigned long min_usage, siblings_min_usage;
5724 min_usage = min(usage, memcg->memory.min);
5725 siblings_min_usage = atomic_long_read(
5726 &parent->memory.children_min_usage);
5728 if (min_usage && siblings_min_usage)
5729 emin = min(emin, parent_emin * min_usage /
5730 siblings_min_usage);
5733 parent_elow = READ_ONCE(parent->memory.elow);
5734 elow = min(elow, parent_elow);
5735 if (elow && parent_elow) {
5736 unsigned long low_usage, siblings_low_usage;
5738 low_usage = min(usage, memcg->memory.low);
5739 siblings_low_usage = atomic_long_read(
5740 &parent->memory.children_low_usage);
5742 if (low_usage && siblings_low_usage)
5743 elow = min(elow, parent_elow * low_usage /
5744 siblings_low_usage);
5748 memcg->memory.emin = emin;
5749 memcg->memory.elow = elow;
5752 return MEMCG_PROT_MIN;
5753 else if (usage <= elow)
5754 return MEMCG_PROT_LOW;
5756 return MEMCG_PROT_NONE;
5760 * mem_cgroup_try_charge - try charging a page
5761 * @page: page to charge
5762 * @mm: mm context of the victim
5763 * @gfp_mask: reclaim mode
5764 * @memcgp: charged memcg return
5765 * @compound: charge the page as compound or small page
5767 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5768 * pages according to @gfp_mask if necessary.
5770 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5771 * Otherwise, an error code is returned.
5773 * After page->mapping has been set up, the caller must finalize the
5774 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5775 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5777 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5778 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5781 struct mem_cgroup *memcg = NULL;
5782 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5785 if (mem_cgroup_disabled())
5788 if (PageSwapCache(page)) {
5790 * Every swap fault against a single page tries to charge the
5791 * page, bail as early as possible. shmem_unuse() encounters
5792 * already charged pages, too. The USED bit is protected by
5793 * the page lock, which serializes swap cache removal, which
5794 * in turn serializes uncharging.
5796 VM_BUG_ON_PAGE(!PageLocked(page), page);
5797 if (compound_head(page)->mem_cgroup)
5800 if (do_swap_account) {
5801 swp_entry_t ent = { .val = page_private(page), };
5802 unsigned short id = lookup_swap_cgroup_id(ent);
5805 memcg = mem_cgroup_from_id(id);
5806 if (memcg && !css_tryget_online(&memcg->css))
5813 memcg = get_mem_cgroup_from_mm(mm);
5815 ret = try_charge(memcg, gfp_mask, nr_pages);
5817 css_put(&memcg->css);
5823 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
5824 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5827 struct mem_cgroup *memcg;
5830 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
5832 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
5837 * mem_cgroup_commit_charge - commit a page charge
5838 * @page: page to charge
5839 * @memcg: memcg to charge the page to
5840 * @lrucare: page might be on LRU already
5841 * @compound: charge the page as compound or small page
5843 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5844 * after page->mapping has been set up. This must happen atomically
5845 * as part of the page instantiation, i.e. under the page table lock
5846 * for anonymous pages, under the page lock for page and swap cache.
5848 * In addition, the page must not be on the LRU during the commit, to
5849 * prevent racing with task migration. If it might be, use @lrucare.
5851 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5853 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5854 bool lrucare, bool compound)
5856 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5858 VM_BUG_ON_PAGE(!page->mapping, page);
5859 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5861 if (mem_cgroup_disabled())
5864 * Swap faults will attempt to charge the same page multiple
5865 * times. But reuse_swap_page() might have removed the page
5866 * from swapcache already, so we can't check PageSwapCache().
5871 commit_charge(page, memcg, lrucare);
5873 local_irq_disable();
5874 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5875 memcg_check_events(memcg, page);
5878 if (do_memsw_account() && PageSwapCache(page)) {
5879 swp_entry_t entry = { .val = page_private(page) };
5881 * The swap entry might not get freed for a long time,
5882 * let's not wait for it. The page already received a
5883 * memory+swap charge, drop the swap entry duplicate.
5885 mem_cgroup_uncharge_swap(entry, nr_pages);
5890 * mem_cgroup_cancel_charge - cancel a page charge
5891 * @page: page to charge
5892 * @memcg: memcg to charge the page to
5893 * @compound: charge the page as compound or small page
5895 * Cancel a charge transaction started by mem_cgroup_try_charge().
5897 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5900 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5902 if (mem_cgroup_disabled())
5905 * Swap faults will attempt to charge the same page multiple
5906 * times. But reuse_swap_page() might have removed the page
5907 * from swapcache already, so we can't check PageSwapCache().
5912 cancel_charge(memcg, nr_pages);
5915 struct uncharge_gather {
5916 struct mem_cgroup *memcg;
5917 unsigned long pgpgout;
5918 unsigned long nr_anon;
5919 unsigned long nr_file;
5920 unsigned long nr_kmem;
5921 unsigned long nr_huge;
5922 unsigned long nr_shmem;
5923 struct page *dummy_page;
5926 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
5928 memset(ug, 0, sizeof(*ug));
5931 static void uncharge_batch(const struct uncharge_gather *ug)
5933 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
5934 unsigned long flags;
5936 if (!mem_cgroup_is_root(ug->memcg)) {
5937 page_counter_uncharge(&ug->memcg->memory, nr_pages);
5938 if (do_memsw_account())
5939 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
5940 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
5941 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
5942 memcg_oom_recover(ug->memcg);
5945 local_irq_save(flags);
5946 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
5947 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
5948 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
5949 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
5950 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
5951 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
5952 memcg_check_events(ug->memcg, ug->dummy_page);
5953 local_irq_restore(flags);
5955 if (!mem_cgroup_is_root(ug->memcg))
5956 css_put_many(&ug->memcg->css, nr_pages);
5959 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
5961 VM_BUG_ON_PAGE(PageLRU(page), page);
5962 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
5963 !PageHWPoison(page) , page);
5965 if (!page->mem_cgroup)
5969 * Nobody should be changing or seriously looking at
5970 * page->mem_cgroup at this point, we have fully
5971 * exclusive access to the page.
5974 if (ug->memcg != page->mem_cgroup) {
5977 uncharge_gather_clear(ug);
5979 ug->memcg = page->mem_cgroup;
5982 if (!PageKmemcg(page)) {
5983 unsigned int nr_pages = 1;
5985 if (PageTransHuge(page)) {
5986 nr_pages <<= compound_order(page);
5987 ug->nr_huge += nr_pages;
5990 ug->nr_anon += nr_pages;
5992 ug->nr_file += nr_pages;
5993 if (PageSwapBacked(page))
5994 ug->nr_shmem += nr_pages;
5998 ug->nr_kmem += 1 << compound_order(page);
5999 __ClearPageKmemcg(page);
6002 ug->dummy_page = page;
6003 page->mem_cgroup = NULL;
6006 static void uncharge_list(struct list_head *page_list)
6008 struct uncharge_gather ug;
6009 struct list_head *next;
6011 uncharge_gather_clear(&ug);
6014 * Note that the list can be a single page->lru; hence the
6015 * do-while loop instead of a simple list_for_each_entry().
6017 next = page_list->next;
6021 page = list_entry(next, struct page, lru);
6022 next = page->lru.next;
6024 uncharge_page(page, &ug);
6025 } while (next != page_list);
6028 uncharge_batch(&ug);
6032 * mem_cgroup_uncharge - uncharge a page
6033 * @page: page to uncharge
6035 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6036 * mem_cgroup_commit_charge().
6038 void mem_cgroup_uncharge(struct page *page)
6040 struct uncharge_gather ug;
6042 if (mem_cgroup_disabled())
6045 /* Don't touch page->lru of any random page, pre-check: */
6046 if (!page->mem_cgroup)
6049 uncharge_gather_clear(&ug);
6050 uncharge_page(page, &ug);
6051 uncharge_batch(&ug);
6055 * mem_cgroup_uncharge_list - uncharge a list of page
6056 * @page_list: list of pages to uncharge
6058 * Uncharge a list of pages previously charged with
6059 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6061 void mem_cgroup_uncharge_list(struct list_head *page_list)
6063 if (mem_cgroup_disabled())
6066 if (!list_empty(page_list))
6067 uncharge_list(page_list);
6071 * mem_cgroup_migrate - charge a page's replacement
6072 * @oldpage: currently circulating page
6073 * @newpage: replacement page
6075 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6076 * be uncharged upon free.
6078 * Both pages must be locked, @newpage->mapping must be set up.
6080 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6082 struct mem_cgroup *memcg;
6083 unsigned int nr_pages;
6085 unsigned long flags;
6087 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6088 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6089 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6090 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6093 if (mem_cgroup_disabled())
6096 /* Page cache replacement: new page already charged? */
6097 if (newpage->mem_cgroup)
6100 /* Swapcache readahead pages can get replaced before being charged */
6101 memcg = oldpage->mem_cgroup;
6105 /* Force-charge the new page. The old one will be freed soon */
6106 compound = PageTransHuge(newpage);
6107 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6109 page_counter_charge(&memcg->memory, nr_pages);
6110 if (do_memsw_account())
6111 page_counter_charge(&memcg->memsw, nr_pages);
6112 css_get_many(&memcg->css, nr_pages);
6114 commit_charge(newpage, memcg, false);
6116 local_irq_save(flags);
6117 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6118 memcg_check_events(memcg, newpage);
6119 local_irq_restore(flags);
6122 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6123 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6125 void mem_cgroup_sk_alloc(struct sock *sk)
6127 struct mem_cgroup *memcg;
6129 if (!mem_cgroup_sockets_enabled)
6133 * Socket cloning can throw us here with sk_memcg already
6134 * filled. It won't however, necessarily happen from
6135 * process context. So the test for root memcg given
6136 * the current task's memcg won't help us in this case.
6138 * Respecting the original socket's memcg is a better
6139 * decision in this case.
6142 css_get(&sk->sk_memcg->css);
6147 memcg = mem_cgroup_from_task(current);
6148 if (memcg == root_mem_cgroup)
6150 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6152 if (css_tryget_online(&memcg->css))
6153 sk->sk_memcg = memcg;
6158 void mem_cgroup_sk_free(struct sock *sk)
6161 css_put(&sk->sk_memcg->css);
6165 * mem_cgroup_charge_skmem - charge socket memory
6166 * @memcg: memcg to charge
6167 * @nr_pages: number of pages to charge
6169 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6170 * @memcg's configured limit, %false if the charge had to be forced.
6172 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6174 gfp_t gfp_mask = GFP_KERNEL;
6176 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6177 struct page_counter *fail;
6179 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6180 memcg->tcpmem_pressure = 0;
6183 page_counter_charge(&memcg->tcpmem, nr_pages);
6184 memcg->tcpmem_pressure = 1;
6188 /* Don't block in the packet receive path */
6190 gfp_mask = GFP_NOWAIT;
6192 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6194 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6197 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6202 * mem_cgroup_uncharge_skmem - uncharge socket memory
6203 * @memcg: memcg to uncharge
6204 * @nr_pages: number of pages to uncharge
6206 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6208 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6209 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6213 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6215 refill_stock(memcg, nr_pages);
6218 static int __init cgroup_memory(char *s)
6222 while ((token = strsep(&s, ",")) != NULL) {
6225 if (!strcmp(token, "nosocket"))
6226 cgroup_memory_nosocket = true;
6227 if (!strcmp(token, "nokmem"))
6228 cgroup_memory_nokmem = true;
6232 __setup("cgroup.memory=", cgroup_memory);
6235 * subsys_initcall() for memory controller.
6237 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6238 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6239 * basically everything that doesn't depend on a specific mem_cgroup structure
6240 * should be initialized from here.
6242 static int __init mem_cgroup_init(void)
6246 #ifdef CONFIG_MEMCG_KMEM
6248 * Kmem cache creation is mostly done with the slab_mutex held,
6249 * so use a workqueue with limited concurrency to avoid stalling
6250 * all worker threads in case lots of cgroups are created and
6251 * destroyed simultaneously.
6253 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6254 BUG_ON(!memcg_kmem_cache_wq);
6257 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6258 memcg_hotplug_cpu_dead);
6260 for_each_possible_cpu(cpu)
6261 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6264 for_each_node(node) {
6265 struct mem_cgroup_tree_per_node *rtpn;
6267 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6268 node_online(node) ? node : NUMA_NO_NODE);
6270 rtpn->rb_root = RB_ROOT;
6271 rtpn->rb_rightmost = NULL;
6272 spin_lock_init(&rtpn->lock);
6273 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6278 subsys_initcall(mem_cgroup_init);
6280 #ifdef CONFIG_MEMCG_SWAP
6281 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6283 while (!atomic_inc_not_zero(&memcg->id.ref)) {
6285 * The root cgroup cannot be destroyed, so it's refcount must
6288 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6292 memcg = parent_mem_cgroup(memcg);
6294 memcg = root_mem_cgroup;
6300 * mem_cgroup_swapout - transfer a memsw charge to swap
6301 * @page: page whose memsw charge to transfer
6302 * @entry: swap entry to move the charge to
6304 * Transfer the memsw charge of @page to @entry.
6306 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6308 struct mem_cgroup *memcg, *swap_memcg;
6309 unsigned int nr_entries;
6310 unsigned short oldid;
6312 VM_BUG_ON_PAGE(PageLRU(page), page);
6313 VM_BUG_ON_PAGE(page_count(page), page);
6315 if (!do_memsw_account())
6318 memcg = page->mem_cgroup;
6320 /* Readahead page, never charged */
6325 * In case the memcg owning these pages has been offlined and doesn't
6326 * have an ID allocated to it anymore, charge the closest online
6327 * ancestor for the swap instead and transfer the memory+swap charge.
6329 swap_memcg = mem_cgroup_id_get_online(memcg);
6330 nr_entries = hpage_nr_pages(page);
6331 /* Get references for the tail pages, too */
6333 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6334 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6336 VM_BUG_ON_PAGE(oldid, page);
6337 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6339 page->mem_cgroup = NULL;
6341 if (!mem_cgroup_is_root(memcg))
6342 page_counter_uncharge(&memcg->memory, nr_entries);
6344 if (memcg != swap_memcg) {
6345 if (!mem_cgroup_is_root(swap_memcg))
6346 page_counter_charge(&swap_memcg->memsw, nr_entries);
6347 page_counter_uncharge(&memcg->memsw, nr_entries);
6351 * Interrupts should be disabled here because the caller holds the
6352 * i_pages lock which is taken with interrupts-off. It is
6353 * important here to have the interrupts disabled because it is the
6354 * only synchronisation we have for updating the per-CPU variables.
6356 VM_BUG_ON(!irqs_disabled());
6357 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6359 memcg_check_events(memcg, page);
6361 if (!mem_cgroup_is_root(memcg))
6362 css_put_many(&memcg->css, nr_entries);
6366 * mem_cgroup_try_charge_swap - try charging swap space for a page
6367 * @page: page being added to swap
6368 * @entry: swap entry to charge
6370 * Try to charge @page's memcg for the swap space at @entry.
6372 * Returns 0 on success, -ENOMEM on failure.
6374 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6376 unsigned int nr_pages = hpage_nr_pages(page);
6377 struct page_counter *counter;
6378 struct mem_cgroup *memcg;
6379 unsigned short oldid;
6381 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6384 memcg = page->mem_cgroup;
6386 /* Readahead page, never charged */
6391 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6395 memcg = mem_cgroup_id_get_online(memcg);
6397 if (!mem_cgroup_is_root(memcg) &&
6398 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6399 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6400 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6401 mem_cgroup_id_put(memcg);
6405 /* Get references for the tail pages, too */
6407 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6408 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6409 VM_BUG_ON_PAGE(oldid, page);
6410 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6416 * mem_cgroup_uncharge_swap - uncharge swap space
6417 * @entry: swap entry to uncharge
6418 * @nr_pages: the amount of swap space to uncharge
6420 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6422 struct mem_cgroup *memcg;
6425 if (!do_swap_account)
6428 id = swap_cgroup_record(entry, 0, nr_pages);
6430 memcg = mem_cgroup_from_id(id);
6432 if (!mem_cgroup_is_root(memcg)) {
6433 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6434 page_counter_uncharge(&memcg->swap, nr_pages);
6436 page_counter_uncharge(&memcg->memsw, nr_pages);
6438 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6439 mem_cgroup_id_put_many(memcg, nr_pages);
6444 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6446 long nr_swap_pages = get_nr_swap_pages();
6448 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6449 return nr_swap_pages;
6450 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6451 nr_swap_pages = min_t(long, nr_swap_pages,
6452 READ_ONCE(memcg->swap.max) -
6453 page_counter_read(&memcg->swap));
6454 return nr_swap_pages;
6457 bool mem_cgroup_swap_full(struct page *page)
6459 struct mem_cgroup *memcg;
6461 VM_BUG_ON_PAGE(!PageLocked(page), page);
6465 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6468 memcg = page->mem_cgroup;
6472 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6473 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6479 /* for remember boot option*/
6480 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6481 static int really_do_swap_account __initdata = 1;
6483 static int really_do_swap_account __initdata;
6486 static int __init enable_swap_account(char *s)
6488 if (!strcmp(s, "1"))
6489 really_do_swap_account = 1;
6490 else if (!strcmp(s, "0"))
6491 really_do_swap_account = 0;
6494 __setup("swapaccount=", enable_swap_account);
6496 static u64 swap_current_read(struct cgroup_subsys_state *css,
6499 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6501 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6504 static int swap_max_show(struct seq_file *m, void *v)
6506 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6507 unsigned long max = READ_ONCE(memcg->swap.max);
6509 if (max == PAGE_COUNTER_MAX)
6510 seq_puts(m, "max\n");
6512 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6517 static ssize_t swap_max_write(struct kernfs_open_file *of,
6518 char *buf, size_t nbytes, loff_t off)
6520 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6524 buf = strstrip(buf);
6525 err = page_counter_memparse(buf, "max", &max);
6529 xchg(&memcg->swap.max, max);
6534 static int swap_events_show(struct seq_file *m, void *v)
6536 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6538 seq_printf(m, "max %lu\n",
6539 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6540 seq_printf(m, "fail %lu\n",
6541 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6546 static struct cftype swap_files[] = {
6548 .name = "swap.current",
6549 .flags = CFTYPE_NOT_ON_ROOT,
6550 .read_u64 = swap_current_read,
6554 .flags = CFTYPE_NOT_ON_ROOT,
6555 .seq_show = swap_max_show,
6556 .write = swap_max_write,
6559 .name = "swap.events",
6560 .flags = CFTYPE_NOT_ON_ROOT,
6561 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6562 .seq_show = swap_events_show,
6567 static struct cftype memsw_cgroup_files[] = {
6569 .name = "memsw.usage_in_bytes",
6570 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6571 .read_u64 = mem_cgroup_read_u64,
6574 .name = "memsw.max_usage_in_bytes",
6575 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6576 .write = mem_cgroup_reset,
6577 .read_u64 = mem_cgroup_read_u64,
6580 .name = "memsw.limit_in_bytes",
6581 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6582 .write = mem_cgroup_write,
6583 .read_u64 = mem_cgroup_read_u64,
6586 .name = "memsw.failcnt",
6587 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6588 .write = mem_cgroup_reset,
6589 .read_u64 = mem_cgroup_read_u64,
6591 { }, /* terminate */
6594 static int __init mem_cgroup_swap_init(void)
6596 if (!mem_cgroup_disabled() && really_do_swap_account) {
6597 do_swap_account = 1;
6598 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6600 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6601 memsw_cgroup_files));
6605 subsys_initcall(mem_cgroup_swap_init);
6607 #endif /* CONFIG_MEMCG_SWAP */