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 #ifdef CONFIG_MEMCG_KMEM
266 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
267 * The main reason for not using cgroup id for this:
268 * this works better in sparse environments, where we have a lot of memcgs,
269 * but only a few kmem-limited. Or also, if we have, for instance, 200
270 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
271 * 200 entry array for that.
273 * The current size of the caches array is stored in memcg_nr_cache_ids. It
274 * will double each time we have to increase it.
276 static DEFINE_IDA(memcg_cache_ida);
277 int memcg_nr_cache_ids;
279 /* Protects memcg_nr_cache_ids */
280 static DECLARE_RWSEM(memcg_cache_ids_sem);
282 void memcg_get_cache_ids(void)
284 down_read(&memcg_cache_ids_sem);
287 void memcg_put_cache_ids(void)
289 up_read(&memcg_cache_ids_sem);
293 * MIN_SIZE is different than 1, because we would like to avoid going through
294 * the alloc/free process all the time. In a small machine, 4 kmem-limited
295 * cgroups is a reasonable guess. In the future, it could be a parameter or
296 * tunable, but that is strictly not necessary.
298 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
299 * this constant directly from cgroup, but it is understandable that this is
300 * better kept as an internal representation in cgroup.c. In any case, the
301 * cgrp_id space is not getting any smaller, and we don't have to necessarily
302 * increase ours as well if it increases.
304 #define MEMCG_CACHES_MIN_SIZE 4
305 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
308 * A lot of the calls to the cache allocation functions are expected to be
309 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
310 * conditional to this static branch, we'll have to allow modules that does
311 * kmem_cache_alloc and the such to see this symbol as well
313 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
314 EXPORT_SYMBOL(memcg_kmem_enabled_key);
316 struct workqueue_struct *memcg_kmem_cache_wq;
318 static int memcg_shrinker_map_size;
319 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
321 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
323 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
326 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
327 int size, int old_size)
329 struct memcg_shrinker_map *new, *old;
332 lockdep_assert_held(&memcg_shrinker_map_mutex);
335 old = rcu_dereference_protected(
336 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
337 /* Not yet online memcg */
341 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
345 /* Set all old bits, clear all new bits */
346 memset(new->map, (int)0xff, old_size);
347 memset((void *)new->map + old_size, 0, size - old_size);
349 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
350 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
356 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
358 struct mem_cgroup_per_node *pn;
359 struct memcg_shrinker_map *map;
362 if (mem_cgroup_is_root(memcg))
366 pn = mem_cgroup_nodeinfo(memcg, nid);
367 map = rcu_dereference_protected(pn->shrinker_map, true);
370 rcu_assign_pointer(pn->shrinker_map, NULL);
374 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
376 struct memcg_shrinker_map *map;
377 int nid, size, ret = 0;
379 if (mem_cgroup_is_root(memcg))
382 mutex_lock(&memcg_shrinker_map_mutex);
383 size = memcg_shrinker_map_size;
385 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
387 memcg_free_shrinker_maps(memcg);
391 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
393 mutex_unlock(&memcg_shrinker_map_mutex);
398 int memcg_expand_shrinker_maps(int new_id)
400 int size, old_size, ret = 0;
401 struct mem_cgroup *memcg;
403 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
404 old_size = memcg_shrinker_map_size;
405 if (size <= old_size)
408 mutex_lock(&memcg_shrinker_map_mutex);
409 if (!root_mem_cgroup)
412 for_each_mem_cgroup(memcg) {
413 if (mem_cgroup_is_root(memcg))
415 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
421 memcg_shrinker_map_size = size;
422 mutex_unlock(&memcg_shrinker_map_mutex);
426 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
428 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
429 struct memcg_shrinker_map *map;
432 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
433 /* Pairs with smp mb in shrink_slab() */
434 smp_mb__before_atomic();
435 set_bit(shrinker_id, map->map);
440 #else /* CONFIG_MEMCG_KMEM */
441 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
445 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
446 #endif /* CONFIG_MEMCG_KMEM */
449 * mem_cgroup_css_from_page - css of the memcg associated with a page
450 * @page: page of interest
452 * If memcg is bound to the default hierarchy, css of the memcg associated
453 * with @page is returned. The returned css remains associated with @page
454 * until it is released.
456 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
459 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
461 struct mem_cgroup *memcg;
463 memcg = page->mem_cgroup;
465 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
466 memcg = root_mem_cgroup;
472 * page_cgroup_ino - return inode number of the memcg a page is charged to
475 * Look up the closest online ancestor of the memory cgroup @page is charged to
476 * and return its inode number or 0 if @page is not charged to any cgroup. It
477 * is safe to call this function without holding a reference to @page.
479 * Note, this function is inherently racy, because there is nothing to prevent
480 * the cgroup inode from getting torn down and potentially reallocated a moment
481 * after page_cgroup_ino() returns, so it only should be used by callers that
482 * do not care (such as procfs interfaces).
484 ino_t page_cgroup_ino(struct page *page)
486 struct mem_cgroup *memcg;
487 unsigned long ino = 0;
490 memcg = READ_ONCE(page->mem_cgroup);
491 while (memcg && !(memcg->css.flags & CSS_ONLINE))
492 memcg = parent_mem_cgroup(memcg);
494 ino = cgroup_ino(memcg->css.cgroup);
499 static struct mem_cgroup_per_node *
500 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
502 int nid = page_to_nid(page);
504 return memcg->nodeinfo[nid];
507 static struct mem_cgroup_tree_per_node *
508 soft_limit_tree_node(int nid)
510 return soft_limit_tree.rb_tree_per_node[nid];
513 static struct mem_cgroup_tree_per_node *
514 soft_limit_tree_from_page(struct page *page)
516 int nid = page_to_nid(page);
518 return soft_limit_tree.rb_tree_per_node[nid];
521 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
522 struct mem_cgroup_tree_per_node *mctz,
523 unsigned long new_usage_in_excess)
525 struct rb_node **p = &mctz->rb_root.rb_node;
526 struct rb_node *parent = NULL;
527 struct mem_cgroup_per_node *mz_node;
528 bool rightmost = true;
533 mz->usage_in_excess = new_usage_in_excess;
534 if (!mz->usage_in_excess)
538 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
540 if (mz->usage_in_excess < mz_node->usage_in_excess) {
546 * We can't avoid mem cgroups that are over their soft
547 * limit by the same amount
549 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
554 mctz->rb_rightmost = &mz->tree_node;
556 rb_link_node(&mz->tree_node, parent, p);
557 rb_insert_color(&mz->tree_node, &mctz->rb_root);
561 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
562 struct mem_cgroup_tree_per_node *mctz)
567 if (&mz->tree_node == mctz->rb_rightmost)
568 mctz->rb_rightmost = rb_prev(&mz->tree_node);
570 rb_erase(&mz->tree_node, &mctz->rb_root);
574 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
575 struct mem_cgroup_tree_per_node *mctz)
579 spin_lock_irqsave(&mctz->lock, flags);
580 __mem_cgroup_remove_exceeded(mz, mctz);
581 spin_unlock_irqrestore(&mctz->lock, flags);
584 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
586 unsigned long nr_pages = page_counter_read(&memcg->memory);
587 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
588 unsigned long excess = 0;
590 if (nr_pages > soft_limit)
591 excess = nr_pages - soft_limit;
596 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
598 unsigned long excess;
599 struct mem_cgroup_per_node *mz;
600 struct mem_cgroup_tree_per_node *mctz;
602 mctz = soft_limit_tree_from_page(page);
606 * Necessary to update all ancestors when hierarchy is used.
607 * because their event counter is not touched.
609 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
610 mz = mem_cgroup_page_nodeinfo(memcg, page);
611 excess = soft_limit_excess(memcg);
613 * We have to update the tree if mz is on RB-tree or
614 * mem is over its softlimit.
616 if (excess || mz->on_tree) {
619 spin_lock_irqsave(&mctz->lock, flags);
620 /* if on-tree, remove it */
622 __mem_cgroup_remove_exceeded(mz, mctz);
624 * Insert again. mz->usage_in_excess will be updated.
625 * If excess is 0, no tree ops.
627 __mem_cgroup_insert_exceeded(mz, mctz, excess);
628 spin_unlock_irqrestore(&mctz->lock, flags);
633 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
635 struct mem_cgroup_tree_per_node *mctz;
636 struct mem_cgroup_per_node *mz;
640 mz = mem_cgroup_nodeinfo(memcg, nid);
641 mctz = soft_limit_tree_node(nid);
643 mem_cgroup_remove_exceeded(mz, mctz);
647 static struct mem_cgroup_per_node *
648 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
650 struct mem_cgroup_per_node *mz;
654 if (!mctz->rb_rightmost)
655 goto done; /* Nothing to reclaim from */
657 mz = rb_entry(mctz->rb_rightmost,
658 struct mem_cgroup_per_node, tree_node);
660 * Remove the node now but someone else can add it back,
661 * we will to add it back at the end of reclaim to its correct
662 * position in the tree.
664 __mem_cgroup_remove_exceeded(mz, mctz);
665 if (!soft_limit_excess(mz->memcg) ||
666 !css_tryget_online(&mz->memcg->css))
672 static struct mem_cgroup_per_node *
673 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
675 struct mem_cgroup_per_node *mz;
677 spin_lock_irq(&mctz->lock);
678 mz = __mem_cgroup_largest_soft_limit_node(mctz);
679 spin_unlock_irq(&mctz->lock);
683 static unsigned long memcg_sum_events(struct mem_cgroup *memcg,
686 return atomic_long_read(&memcg->events[event]);
689 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
691 bool compound, int nr_pages)
694 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
695 * counted as CACHE even if it's on ANON LRU.
698 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
700 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
701 if (PageSwapBacked(page))
702 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
706 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
707 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
710 /* pagein of a big page is an event. So, ignore page size */
712 __count_memcg_events(memcg, PGPGIN, 1);
714 __count_memcg_events(memcg, PGPGOUT, 1);
715 nr_pages = -nr_pages; /* for event */
718 __this_cpu_add(memcg->stat_cpu->nr_page_events, nr_pages);
721 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
722 int nid, unsigned int lru_mask)
724 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
725 unsigned long nr = 0;
728 VM_BUG_ON((unsigned)nid >= nr_node_ids);
731 if (!(BIT(lru) & lru_mask))
733 nr += mem_cgroup_get_lru_size(lruvec, lru);
738 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
739 unsigned int lru_mask)
741 unsigned long nr = 0;
744 for_each_node_state(nid, N_MEMORY)
745 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
749 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
750 enum mem_cgroup_events_target target)
752 unsigned long val, next;
754 val = __this_cpu_read(memcg->stat_cpu->nr_page_events);
755 next = __this_cpu_read(memcg->stat_cpu->targets[target]);
756 /* from time_after() in jiffies.h */
757 if ((long)(next - val) < 0) {
759 case MEM_CGROUP_TARGET_THRESH:
760 next = val + THRESHOLDS_EVENTS_TARGET;
762 case MEM_CGROUP_TARGET_SOFTLIMIT:
763 next = val + SOFTLIMIT_EVENTS_TARGET;
765 case MEM_CGROUP_TARGET_NUMAINFO:
766 next = val + NUMAINFO_EVENTS_TARGET;
771 __this_cpu_write(memcg->stat_cpu->targets[target], next);
778 * Check events in order.
781 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
783 /* threshold event is triggered in finer grain than soft limit */
784 if (unlikely(mem_cgroup_event_ratelimit(memcg,
785 MEM_CGROUP_TARGET_THRESH))) {
787 bool do_numainfo __maybe_unused;
789 do_softlimit = mem_cgroup_event_ratelimit(memcg,
790 MEM_CGROUP_TARGET_SOFTLIMIT);
792 do_numainfo = mem_cgroup_event_ratelimit(memcg,
793 MEM_CGROUP_TARGET_NUMAINFO);
795 mem_cgroup_threshold(memcg);
796 if (unlikely(do_softlimit))
797 mem_cgroup_update_tree(memcg, page);
799 if (unlikely(do_numainfo))
800 atomic_inc(&memcg->numainfo_events);
805 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
808 * mm_update_next_owner() may clear mm->owner to NULL
809 * if it races with swapoff, page migration, etc.
810 * So this can be called with p == NULL.
815 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
817 EXPORT_SYMBOL(mem_cgroup_from_task);
820 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
821 * @mm: mm from which memcg should be extracted. It can be NULL.
823 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
824 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
827 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
829 struct mem_cgroup *memcg;
831 if (mem_cgroup_disabled())
837 * Page cache insertions can happen withou an
838 * actual mm context, e.g. during disk probing
839 * on boot, loopback IO, acct() writes etc.
842 memcg = root_mem_cgroup;
844 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
845 if (unlikely(!memcg))
846 memcg = root_mem_cgroup;
848 } while (!css_tryget_online(&memcg->css));
852 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
855 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
856 * @page: page from which memcg should be extracted.
858 * Obtain a reference on page->memcg and returns it if successful. Otherwise
859 * root_mem_cgroup is returned.
861 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
863 struct mem_cgroup *memcg = page->mem_cgroup;
865 if (mem_cgroup_disabled())
869 if (!memcg || !css_tryget_online(&memcg->css))
870 memcg = root_mem_cgroup;
874 EXPORT_SYMBOL(get_mem_cgroup_from_page);
877 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
879 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
881 if (unlikely(current->active_memcg)) {
882 struct mem_cgroup *memcg = root_mem_cgroup;
885 if (css_tryget_online(¤t->active_memcg->css))
886 memcg = current->active_memcg;
890 return get_mem_cgroup_from_mm(current->mm);
894 * mem_cgroup_iter - iterate over memory cgroup hierarchy
895 * @root: hierarchy root
896 * @prev: previously returned memcg, NULL on first invocation
897 * @reclaim: cookie for shared reclaim walks, NULL for full walks
899 * Returns references to children of the hierarchy below @root, or
900 * @root itself, or %NULL after a full round-trip.
902 * Caller must pass the return value in @prev on subsequent
903 * invocations for reference counting, or use mem_cgroup_iter_break()
904 * to cancel a hierarchy walk before the round-trip is complete.
906 * Reclaimers can specify a node and a priority level in @reclaim to
907 * divide up the memcgs in the hierarchy among all concurrent
908 * reclaimers operating on the same node and priority.
910 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
911 struct mem_cgroup *prev,
912 struct mem_cgroup_reclaim_cookie *reclaim)
914 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
915 struct cgroup_subsys_state *css = NULL;
916 struct mem_cgroup *memcg = NULL;
917 struct mem_cgroup *pos = NULL;
919 if (mem_cgroup_disabled())
923 root = root_mem_cgroup;
925 if (prev && !reclaim)
928 if (!root->use_hierarchy && root != root_mem_cgroup) {
937 struct mem_cgroup_per_node *mz;
939 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
940 iter = &mz->iter[reclaim->priority];
942 if (prev && reclaim->generation != iter->generation)
946 pos = READ_ONCE(iter->position);
947 if (!pos || css_tryget(&pos->css))
950 * css reference reached zero, so iter->position will
951 * be cleared by ->css_released. However, we should not
952 * rely on this happening soon, because ->css_released
953 * is called from a work queue, and by busy-waiting we
954 * might block it. So we clear iter->position right
957 (void)cmpxchg(&iter->position, pos, NULL);
965 css = css_next_descendant_pre(css, &root->css);
968 * Reclaimers share the hierarchy walk, and a
969 * new one might jump in right at the end of
970 * the hierarchy - make sure they see at least
971 * one group and restart from the beginning.
979 * Verify the css and acquire a reference. The root
980 * is provided by the caller, so we know it's alive
981 * and kicking, and don't take an extra reference.
983 memcg = mem_cgroup_from_css(css);
985 if (css == &root->css)
996 * The position could have already been updated by a competing
997 * thread, so check that the value hasn't changed since we read
998 * it to avoid reclaiming from the same cgroup twice.
1000 (void)cmpxchg(&iter->position, pos, memcg);
1008 reclaim->generation = iter->generation;
1014 if (prev && prev != root)
1015 css_put(&prev->css);
1021 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1022 * @root: hierarchy root
1023 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1025 void mem_cgroup_iter_break(struct mem_cgroup *root,
1026 struct mem_cgroup *prev)
1029 root = root_mem_cgroup;
1030 if (prev && prev != root)
1031 css_put(&prev->css);
1034 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1036 struct mem_cgroup *memcg = dead_memcg;
1037 struct mem_cgroup_reclaim_iter *iter;
1038 struct mem_cgroup_per_node *mz;
1042 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1043 for_each_node(nid) {
1044 mz = mem_cgroup_nodeinfo(memcg, nid);
1045 for (i = 0; i <= DEF_PRIORITY; i++) {
1046 iter = &mz->iter[i];
1047 cmpxchg(&iter->position,
1055 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1056 * @memcg: hierarchy root
1057 * @fn: function to call for each task
1058 * @arg: argument passed to @fn
1060 * This function iterates over tasks attached to @memcg or to any of its
1061 * descendants and calls @fn for each task. If @fn returns a non-zero
1062 * value, the function breaks the iteration loop and returns the value.
1063 * Otherwise, it will iterate over all tasks and return 0.
1065 * This function must not be called for the root memory cgroup.
1067 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1068 int (*fn)(struct task_struct *, void *), void *arg)
1070 struct mem_cgroup *iter;
1073 BUG_ON(memcg == root_mem_cgroup);
1075 for_each_mem_cgroup_tree(iter, memcg) {
1076 struct css_task_iter it;
1077 struct task_struct *task;
1079 css_task_iter_start(&iter->css, 0, &it);
1080 while (!ret && (task = css_task_iter_next(&it)))
1081 ret = fn(task, arg);
1082 css_task_iter_end(&it);
1084 mem_cgroup_iter_break(memcg, iter);
1092 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1094 * @pgdat: pgdat of the page
1096 * This function is only safe when following the LRU page isolation
1097 * and putback protocol: the LRU lock must be held, and the page must
1098 * either be PageLRU() or the caller must have isolated/allocated it.
1100 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1102 struct mem_cgroup_per_node *mz;
1103 struct mem_cgroup *memcg;
1104 struct lruvec *lruvec;
1106 if (mem_cgroup_disabled()) {
1107 lruvec = &pgdat->lruvec;
1111 memcg = page->mem_cgroup;
1113 * Swapcache readahead pages are added to the LRU - and
1114 * possibly migrated - before they are charged.
1117 memcg = root_mem_cgroup;
1119 mz = mem_cgroup_page_nodeinfo(memcg, page);
1120 lruvec = &mz->lruvec;
1123 * Since a node can be onlined after the mem_cgroup was created,
1124 * we have to be prepared to initialize lruvec->zone here;
1125 * and if offlined then reonlined, we need to reinitialize it.
1127 if (unlikely(lruvec->pgdat != pgdat))
1128 lruvec->pgdat = pgdat;
1133 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1134 * @lruvec: mem_cgroup per zone lru vector
1135 * @lru: index of lru list the page is sitting on
1136 * @zid: zone id of the accounted pages
1137 * @nr_pages: positive when adding or negative when removing
1139 * This function must be called under lru_lock, just before a page is added
1140 * to or just after a page is removed from an lru list (that ordering being
1141 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1143 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1144 int zid, int nr_pages)
1146 struct mem_cgroup_per_node *mz;
1147 unsigned long *lru_size;
1150 if (mem_cgroup_disabled())
1153 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1154 lru_size = &mz->lru_zone_size[zid][lru];
1157 *lru_size += nr_pages;
1160 if (WARN_ONCE(size < 0,
1161 "%s(%p, %d, %d): lru_size %ld\n",
1162 __func__, lruvec, lru, nr_pages, size)) {
1168 *lru_size += nr_pages;
1171 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1173 struct mem_cgroup *task_memcg;
1174 struct task_struct *p;
1177 p = find_lock_task_mm(task);
1179 task_memcg = get_mem_cgroup_from_mm(p->mm);
1183 * All threads may have already detached their mm's, but the oom
1184 * killer still needs to detect if they have already been oom
1185 * killed to prevent needlessly killing additional tasks.
1188 task_memcg = mem_cgroup_from_task(task);
1189 css_get(&task_memcg->css);
1192 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1193 css_put(&task_memcg->css);
1198 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1199 * @memcg: the memory cgroup
1201 * Returns the maximum amount of memory @mem can be charged with, in
1204 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1206 unsigned long margin = 0;
1207 unsigned long count;
1208 unsigned long limit;
1210 count = page_counter_read(&memcg->memory);
1211 limit = READ_ONCE(memcg->memory.max);
1213 margin = limit - count;
1215 if (do_memsw_account()) {
1216 count = page_counter_read(&memcg->memsw);
1217 limit = READ_ONCE(memcg->memsw.max);
1219 margin = min(margin, limit - count);
1228 * A routine for checking "mem" is under move_account() or not.
1230 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1231 * moving cgroups. This is for waiting at high-memory pressure
1234 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1236 struct mem_cgroup *from;
1237 struct mem_cgroup *to;
1240 * Unlike task_move routines, we access mc.to, mc.from not under
1241 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1243 spin_lock(&mc.lock);
1249 ret = mem_cgroup_is_descendant(from, memcg) ||
1250 mem_cgroup_is_descendant(to, memcg);
1252 spin_unlock(&mc.lock);
1256 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1258 if (mc.moving_task && current != mc.moving_task) {
1259 if (mem_cgroup_under_move(memcg)) {
1261 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1262 /* moving charge context might have finished. */
1265 finish_wait(&mc.waitq, &wait);
1272 static const unsigned int memcg1_stats[] = {
1283 static const char *const memcg1_stat_names[] = {
1294 #define K(x) ((x) << (PAGE_SHIFT-10))
1296 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1297 * @memcg: The memory cgroup that went over limit
1298 * @p: Task that is going to be killed
1300 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1303 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1305 struct mem_cgroup *iter;
1311 pr_info("Task in ");
1312 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1313 pr_cont(" killed as a result of limit of ");
1315 pr_info("Memory limit reached of cgroup ");
1318 pr_cont_cgroup_path(memcg->css.cgroup);
1323 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1324 K((u64)page_counter_read(&memcg->memory)),
1325 K((u64)memcg->memory.max), memcg->memory.failcnt);
1326 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1327 K((u64)page_counter_read(&memcg->memsw)),
1328 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1329 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1330 K((u64)page_counter_read(&memcg->kmem)),
1331 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1333 for_each_mem_cgroup_tree(iter, memcg) {
1334 pr_info("Memory cgroup stats for ");
1335 pr_cont_cgroup_path(iter->css.cgroup);
1338 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1339 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1341 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1342 K(memcg_page_state(iter, memcg1_stats[i])));
1345 for (i = 0; i < NR_LRU_LISTS; i++)
1346 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1347 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1354 * Return the memory (and swap, if configured) limit for a memcg.
1356 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1360 max = memcg->memory.max;
1361 if (mem_cgroup_swappiness(memcg)) {
1362 unsigned long memsw_max;
1363 unsigned long swap_max;
1365 memsw_max = memcg->memsw.max;
1366 swap_max = memcg->swap.max;
1367 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1368 max = min(max + swap_max, memsw_max);
1373 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1376 struct oom_control oc = {
1380 .gfp_mask = gfp_mask,
1385 mutex_lock(&oom_lock);
1386 ret = out_of_memory(&oc);
1387 mutex_unlock(&oom_lock);
1391 #if MAX_NUMNODES > 1
1394 * test_mem_cgroup_node_reclaimable
1395 * @memcg: the target memcg
1396 * @nid: the node ID to be checked.
1397 * @noswap : specify true here if the user wants flle only information.
1399 * This function returns whether the specified memcg contains any
1400 * reclaimable pages on a node. Returns true if there are any reclaimable
1401 * pages in the node.
1403 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1404 int nid, bool noswap)
1406 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1408 if (noswap || !total_swap_pages)
1410 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1417 * Always updating the nodemask is not very good - even if we have an empty
1418 * list or the wrong list here, we can start from some node and traverse all
1419 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1422 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1426 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1427 * pagein/pageout changes since the last update.
1429 if (!atomic_read(&memcg->numainfo_events))
1431 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1434 /* make a nodemask where this memcg uses memory from */
1435 memcg->scan_nodes = node_states[N_MEMORY];
1437 for_each_node_mask(nid, node_states[N_MEMORY]) {
1439 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1440 node_clear(nid, memcg->scan_nodes);
1443 atomic_set(&memcg->numainfo_events, 0);
1444 atomic_set(&memcg->numainfo_updating, 0);
1448 * Selecting a node where we start reclaim from. Because what we need is just
1449 * reducing usage counter, start from anywhere is O,K. Considering
1450 * memory reclaim from current node, there are pros. and cons.
1452 * Freeing memory from current node means freeing memory from a node which
1453 * we'll use or we've used. So, it may make LRU bad. And if several threads
1454 * hit limits, it will see a contention on a node. But freeing from remote
1455 * node means more costs for memory reclaim because of memory latency.
1457 * Now, we use round-robin. Better algorithm is welcomed.
1459 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1463 mem_cgroup_may_update_nodemask(memcg);
1464 node = memcg->last_scanned_node;
1466 node = next_node_in(node, memcg->scan_nodes);
1468 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1469 * last time it really checked all the LRUs due to rate limiting.
1470 * Fallback to the current node in that case for simplicity.
1472 if (unlikely(node == MAX_NUMNODES))
1473 node = numa_node_id();
1475 memcg->last_scanned_node = node;
1479 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1485 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1488 unsigned long *total_scanned)
1490 struct mem_cgroup *victim = NULL;
1493 unsigned long excess;
1494 unsigned long nr_scanned;
1495 struct mem_cgroup_reclaim_cookie reclaim = {
1500 excess = soft_limit_excess(root_memcg);
1503 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1508 * If we have not been able to reclaim
1509 * anything, it might because there are
1510 * no reclaimable pages under this hierarchy
1515 * We want to do more targeted reclaim.
1516 * excess >> 2 is not to excessive so as to
1517 * reclaim too much, nor too less that we keep
1518 * coming back to reclaim from this cgroup
1520 if (total >= (excess >> 2) ||
1521 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1526 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1527 pgdat, &nr_scanned);
1528 *total_scanned += nr_scanned;
1529 if (!soft_limit_excess(root_memcg))
1532 mem_cgroup_iter_break(root_memcg, victim);
1536 #ifdef CONFIG_LOCKDEP
1537 static struct lockdep_map memcg_oom_lock_dep_map = {
1538 .name = "memcg_oom_lock",
1542 static DEFINE_SPINLOCK(memcg_oom_lock);
1545 * Check OOM-Killer is already running under our hierarchy.
1546 * If someone is running, return false.
1548 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1550 struct mem_cgroup *iter, *failed = NULL;
1552 spin_lock(&memcg_oom_lock);
1554 for_each_mem_cgroup_tree(iter, memcg) {
1555 if (iter->oom_lock) {
1557 * this subtree of our hierarchy is already locked
1558 * so we cannot give a lock.
1561 mem_cgroup_iter_break(memcg, iter);
1564 iter->oom_lock = true;
1569 * OK, we failed to lock the whole subtree so we have
1570 * to clean up what we set up to the failing subtree
1572 for_each_mem_cgroup_tree(iter, memcg) {
1573 if (iter == failed) {
1574 mem_cgroup_iter_break(memcg, iter);
1577 iter->oom_lock = false;
1580 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1582 spin_unlock(&memcg_oom_lock);
1587 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1589 struct mem_cgroup *iter;
1591 spin_lock(&memcg_oom_lock);
1592 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1593 for_each_mem_cgroup_tree(iter, memcg)
1594 iter->oom_lock = false;
1595 spin_unlock(&memcg_oom_lock);
1598 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1600 struct mem_cgroup *iter;
1602 spin_lock(&memcg_oom_lock);
1603 for_each_mem_cgroup_tree(iter, memcg)
1605 spin_unlock(&memcg_oom_lock);
1608 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1610 struct mem_cgroup *iter;
1613 * When a new child is created while the hierarchy is under oom,
1614 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1616 spin_lock(&memcg_oom_lock);
1617 for_each_mem_cgroup_tree(iter, memcg)
1618 if (iter->under_oom > 0)
1620 spin_unlock(&memcg_oom_lock);
1623 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1625 struct oom_wait_info {
1626 struct mem_cgroup *memcg;
1627 wait_queue_entry_t wait;
1630 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1631 unsigned mode, int sync, void *arg)
1633 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1634 struct mem_cgroup *oom_wait_memcg;
1635 struct oom_wait_info *oom_wait_info;
1637 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1638 oom_wait_memcg = oom_wait_info->memcg;
1640 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1641 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1643 return autoremove_wake_function(wait, mode, sync, arg);
1646 static void memcg_oom_recover(struct mem_cgroup *memcg)
1649 * For the following lockless ->under_oom test, the only required
1650 * guarantee is that it must see the state asserted by an OOM when
1651 * this function is called as a result of userland actions
1652 * triggered by the notification of the OOM. This is trivially
1653 * achieved by invoking mem_cgroup_mark_under_oom() before
1654 * triggering notification.
1656 if (memcg && memcg->under_oom)
1657 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1667 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1669 if (order > PAGE_ALLOC_COSTLY_ORDER)
1673 * We are in the middle of the charge context here, so we
1674 * don't want to block when potentially sitting on a callstack
1675 * that holds all kinds of filesystem and mm locks.
1677 * cgroup1 allows disabling the OOM killer and waiting for outside
1678 * handling until the charge can succeed; remember the context and put
1679 * the task to sleep at the end of the page fault when all locks are
1682 * On the other hand, in-kernel OOM killer allows for an async victim
1683 * memory reclaim (oom_reaper) and that means that we are not solely
1684 * relying on the oom victim to make a forward progress and we can
1685 * invoke the oom killer here.
1687 * Please note that mem_cgroup_out_of_memory might fail to find a
1688 * victim and then we have to bail out from the charge path.
1690 if (memcg->oom_kill_disable) {
1691 if (!current->in_user_fault)
1693 css_get(&memcg->css);
1694 current->memcg_in_oom = memcg;
1695 current->memcg_oom_gfp_mask = mask;
1696 current->memcg_oom_order = order;
1701 if (mem_cgroup_out_of_memory(memcg, mask, order))
1708 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1709 * @handle: actually kill/wait or just clean up the OOM state
1711 * This has to be called at the end of a page fault if the memcg OOM
1712 * handler was enabled.
1714 * Memcg supports userspace OOM handling where failed allocations must
1715 * sleep on a waitqueue until the userspace task resolves the
1716 * situation. Sleeping directly in the charge context with all kinds
1717 * of locks held is not a good idea, instead we remember an OOM state
1718 * in the task and mem_cgroup_oom_synchronize() has to be called at
1719 * the end of the page fault to complete the OOM handling.
1721 * Returns %true if an ongoing memcg OOM situation was detected and
1722 * completed, %false otherwise.
1724 bool mem_cgroup_oom_synchronize(bool handle)
1726 struct mem_cgroup *memcg = current->memcg_in_oom;
1727 struct oom_wait_info owait;
1730 /* OOM is global, do not handle */
1737 owait.memcg = memcg;
1738 owait.wait.flags = 0;
1739 owait.wait.func = memcg_oom_wake_function;
1740 owait.wait.private = current;
1741 INIT_LIST_HEAD(&owait.wait.entry);
1743 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1744 mem_cgroup_mark_under_oom(memcg);
1746 locked = mem_cgroup_oom_trylock(memcg);
1749 mem_cgroup_oom_notify(memcg);
1751 if (locked && !memcg->oom_kill_disable) {
1752 mem_cgroup_unmark_under_oom(memcg);
1753 finish_wait(&memcg_oom_waitq, &owait.wait);
1754 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1755 current->memcg_oom_order);
1758 mem_cgroup_unmark_under_oom(memcg);
1759 finish_wait(&memcg_oom_waitq, &owait.wait);
1763 mem_cgroup_oom_unlock(memcg);
1765 * There is no guarantee that an OOM-lock contender
1766 * sees the wakeups triggered by the OOM kill
1767 * uncharges. Wake any sleepers explicitely.
1769 memcg_oom_recover(memcg);
1772 current->memcg_in_oom = NULL;
1773 css_put(&memcg->css);
1778 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1779 * @victim: task to be killed by the OOM killer
1780 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1782 * Returns a pointer to a memory cgroup, which has to be cleaned up
1783 * by killing all belonging OOM-killable tasks.
1785 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1787 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1788 struct mem_cgroup *oom_domain)
1790 struct mem_cgroup *oom_group = NULL;
1791 struct mem_cgroup *memcg;
1793 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1797 oom_domain = root_mem_cgroup;
1801 memcg = mem_cgroup_from_task(victim);
1802 if (memcg == root_mem_cgroup)
1806 * Traverse the memory cgroup hierarchy from the victim task's
1807 * cgroup up to the OOMing cgroup (or root) to find the
1808 * highest-level memory cgroup with oom.group set.
1810 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1811 if (memcg->oom_group)
1814 if (memcg == oom_domain)
1819 css_get(&oom_group->css);
1826 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1828 pr_info("Tasks in ");
1829 pr_cont_cgroup_path(memcg->css.cgroup);
1830 pr_cont(" are going to be killed due to memory.oom.group set\n");
1834 * lock_page_memcg - lock a page->mem_cgroup binding
1837 * This function protects unlocked LRU pages from being moved to
1840 * It ensures lifetime of the returned memcg. Caller is responsible
1841 * for the lifetime of the page; __unlock_page_memcg() is available
1842 * when @page might get freed inside the locked section.
1844 struct mem_cgroup *lock_page_memcg(struct page *page)
1846 struct mem_cgroup *memcg;
1847 unsigned long flags;
1850 * The RCU lock is held throughout the transaction. The fast
1851 * path can get away without acquiring the memcg->move_lock
1852 * because page moving starts with an RCU grace period.
1854 * The RCU lock also protects the memcg from being freed when
1855 * the page state that is going to change is the only thing
1856 * preventing the page itself from being freed. E.g. writeback
1857 * doesn't hold a page reference and relies on PG_writeback to
1858 * keep off truncation, migration and so forth.
1862 if (mem_cgroup_disabled())
1865 memcg = page->mem_cgroup;
1866 if (unlikely(!memcg))
1869 if (atomic_read(&memcg->moving_account) <= 0)
1872 spin_lock_irqsave(&memcg->move_lock, flags);
1873 if (memcg != page->mem_cgroup) {
1874 spin_unlock_irqrestore(&memcg->move_lock, flags);
1879 * When charge migration first begins, we can have locked and
1880 * unlocked page stat updates happening concurrently. Track
1881 * the task who has the lock for unlock_page_memcg().
1883 memcg->move_lock_task = current;
1884 memcg->move_lock_flags = flags;
1888 EXPORT_SYMBOL(lock_page_memcg);
1891 * __unlock_page_memcg - unlock and unpin a memcg
1894 * Unlock and unpin a memcg returned by lock_page_memcg().
1896 void __unlock_page_memcg(struct mem_cgroup *memcg)
1898 if (memcg && memcg->move_lock_task == current) {
1899 unsigned long flags = memcg->move_lock_flags;
1901 memcg->move_lock_task = NULL;
1902 memcg->move_lock_flags = 0;
1904 spin_unlock_irqrestore(&memcg->move_lock, flags);
1911 * unlock_page_memcg - unlock a page->mem_cgroup binding
1914 void unlock_page_memcg(struct page *page)
1916 __unlock_page_memcg(page->mem_cgroup);
1918 EXPORT_SYMBOL(unlock_page_memcg);
1920 struct memcg_stock_pcp {
1921 struct mem_cgroup *cached; /* this never be root cgroup */
1922 unsigned int nr_pages;
1923 struct work_struct work;
1924 unsigned long flags;
1925 #define FLUSHING_CACHED_CHARGE 0
1927 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1928 static DEFINE_MUTEX(percpu_charge_mutex);
1931 * consume_stock: Try to consume stocked charge on this cpu.
1932 * @memcg: memcg to consume from.
1933 * @nr_pages: how many pages to charge.
1935 * The charges will only happen if @memcg matches the current cpu's memcg
1936 * stock, and at least @nr_pages are available in that stock. Failure to
1937 * service an allocation will refill the stock.
1939 * returns true if successful, false otherwise.
1941 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1943 struct memcg_stock_pcp *stock;
1944 unsigned long flags;
1947 if (nr_pages > MEMCG_CHARGE_BATCH)
1950 local_irq_save(flags);
1952 stock = this_cpu_ptr(&memcg_stock);
1953 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1954 stock->nr_pages -= nr_pages;
1958 local_irq_restore(flags);
1964 * Returns stocks cached in percpu and reset cached information.
1966 static void drain_stock(struct memcg_stock_pcp *stock)
1968 struct mem_cgroup *old = stock->cached;
1970 if (stock->nr_pages) {
1971 page_counter_uncharge(&old->memory, stock->nr_pages);
1972 if (do_memsw_account())
1973 page_counter_uncharge(&old->memsw, stock->nr_pages);
1974 css_put_many(&old->css, stock->nr_pages);
1975 stock->nr_pages = 0;
1977 stock->cached = NULL;
1980 static void drain_local_stock(struct work_struct *dummy)
1982 struct memcg_stock_pcp *stock;
1983 unsigned long flags;
1986 * The only protection from memory hotplug vs. drain_stock races is
1987 * that we always operate on local CPU stock here with IRQ disabled
1989 local_irq_save(flags);
1991 stock = this_cpu_ptr(&memcg_stock);
1993 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1995 local_irq_restore(flags);
1999 * Cache charges(val) to local per_cpu area.
2000 * This will be consumed by consume_stock() function, later.
2002 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2004 struct memcg_stock_pcp *stock;
2005 unsigned long flags;
2007 local_irq_save(flags);
2009 stock = this_cpu_ptr(&memcg_stock);
2010 if (stock->cached != memcg) { /* reset if necessary */
2012 stock->cached = memcg;
2014 stock->nr_pages += nr_pages;
2016 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2019 local_irq_restore(flags);
2023 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2024 * of the hierarchy under it.
2026 static void drain_all_stock(struct mem_cgroup *root_memcg)
2030 /* If someone's already draining, avoid adding running more workers. */
2031 if (!mutex_trylock(&percpu_charge_mutex))
2034 * Notify other cpus that system-wide "drain" is running
2035 * We do not care about races with the cpu hotplug because cpu down
2036 * as well as workers from this path always operate on the local
2037 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2040 for_each_online_cpu(cpu) {
2041 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2042 struct mem_cgroup *memcg;
2044 memcg = stock->cached;
2045 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2047 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2048 css_put(&memcg->css);
2051 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2053 drain_local_stock(&stock->work);
2055 schedule_work_on(cpu, &stock->work);
2057 css_put(&memcg->css);
2060 mutex_unlock(&percpu_charge_mutex);
2063 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2065 struct memcg_stock_pcp *stock;
2066 struct mem_cgroup *memcg;
2068 stock = &per_cpu(memcg_stock, cpu);
2071 for_each_mem_cgroup(memcg) {
2074 for (i = 0; i < MEMCG_NR_STAT; i++) {
2078 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
2080 atomic_long_add(x, &memcg->stat[i]);
2082 if (i >= NR_VM_NODE_STAT_ITEMS)
2085 for_each_node(nid) {
2086 struct mem_cgroup_per_node *pn;
2088 pn = mem_cgroup_nodeinfo(memcg, nid);
2089 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2091 atomic_long_add(x, &pn->lruvec_stat[i]);
2095 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2098 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
2100 atomic_long_add(x, &memcg->events[i]);
2107 static void reclaim_high(struct mem_cgroup *memcg,
2108 unsigned int nr_pages,
2112 if (page_counter_read(&memcg->memory) <= memcg->high)
2114 memcg_memory_event(memcg, MEMCG_HIGH);
2115 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2116 } while ((memcg = parent_mem_cgroup(memcg)));
2119 static void high_work_func(struct work_struct *work)
2121 struct mem_cgroup *memcg;
2123 memcg = container_of(work, struct mem_cgroup, high_work);
2124 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2128 * Scheduled by try_charge() to be executed from the userland return path
2129 * and reclaims memory over the high limit.
2131 void mem_cgroup_handle_over_high(void)
2133 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2134 struct mem_cgroup *memcg;
2136 if (likely(!nr_pages))
2139 memcg = get_mem_cgroup_from_mm(current->mm);
2140 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2141 css_put(&memcg->css);
2142 current->memcg_nr_pages_over_high = 0;
2145 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2146 unsigned int nr_pages)
2148 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2149 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2150 struct mem_cgroup *mem_over_limit;
2151 struct page_counter *counter;
2152 unsigned long nr_reclaimed;
2153 bool may_swap = true;
2154 bool drained = false;
2156 enum oom_status oom_status;
2158 if (mem_cgroup_is_root(memcg))
2161 if (consume_stock(memcg, nr_pages))
2164 if (!do_memsw_account() ||
2165 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2166 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2168 if (do_memsw_account())
2169 page_counter_uncharge(&memcg->memsw, batch);
2170 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2172 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2176 if (batch > nr_pages) {
2182 * Unlike in global OOM situations, memcg is not in a physical
2183 * memory shortage. Allow dying and OOM-killed tasks to
2184 * bypass the last charges so that they can exit quickly and
2185 * free their memory.
2187 if (unlikely(tsk_is_oom_victim(current) ||
2188 fatal_signal_pending(current) ||
2189 current->flags & PF_EXITING))
2193 * Prevent unbounded recursion when reclaim operations need to
2194 * allocate memory. This might exceed the limits temporarily,
2195 * but we prefer facilitating memory reclaim and getting back
2196 * under the limit over triggering OOM kills in these cases.
2198 if (unlikely(current->flags & PF_MEMALLOC))
2201 if (unlikely(task_in_memcg_oom(current)))
2204 if (!gfpflags_allow_blocking(gfp_mask))
2207 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2209 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2210 gfp_mask, may_swap);
2212 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2216 drain_all_stock(mem_over_limit);
2221 if (gfp_mask & __GFP_NORETRY)
2224 * Even though the limit is exceeded at this point, reclaim
2225 * may have been able to free some pages. Retry the charge
2226 * before killing the task.
2228 * Only for regular pages, though: huge pages are rather
2229 * unlikely to succeed so close to the limit, and we fall back
2230 * to regular pages anyway in case of failure.
2232 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2235 * At task move, charge accounts can be doubly counted. So, it's
2236 * better to wait until the end of task_move if something is going on.
2238 if (mem_cgroup_wait_acct_move(mem_over_limit))
2244 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2247 if (gfp_mask & __GFP_NOFAIL)
2250 if (fatal_signal_pending(current))
2253 memcg_memory_event(mem_over_limit, MEMCG_OOM);
2256 * keep retrying as long as the memcg oom killer is able to make
2257 * a forward progress or bypass the charge if the oom killer
2258 * couldn't make any progress.
2260 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2261 get_order(nr_pages * PAGE_SIZE));
2262 switch (oom_status) {
2264 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2273 if (!(gfp_mask & __GFP_NOFAIL))
2277 * The allocation either can't fail or will lead to more memory
2278 * being freed very soon. Allow memory usage go over the limit
2279 * temporarily by force charging it.
2281 page_counter_charge(&memcg->memory, nr_pages);
2282 if (do_memsw_account())
2283 page_counter_charge(&memcg->memsw, nr_pages);
2284 css_get_many(&memcg->css, nr_pages);
2289 css_get_many(&memcg->css, batch);
2290 if (batch > nr_pages)
2291 refill_stock(memcg, batch - nr_pages);
2294 * If the hierarchy is above the normal consumption range, schedule
2295 * reclaim on returning to userland. We can perform reclaim here
2296 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2297 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2298 * not recorded as it most likely matches current's and won't
2299 * change in the meantime. As high limit is checked again before
2300 * reclaim, the cost of mismatch is negligible.
2303 if (page_counter_read(&memcg->memory) > memcg->high) {
2304 /* Don't bother a random interrupted task */
2305 if (in_interrupt()) {
2306 schedule_work(&memcg->high_work);
2309 current->memcg_nr_pages_over_high += batch;
2310 set_notify_resume(current);
2313 } while ((memcg = parent_mem_cgroup(memcg)));
2318 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2320 if (mem_cgroup_is_root(memcg))
2323 page_counter_uncharge(&memcg->memory, nr_pages);
2324 if (do_memsw_account())
2325 page_counter_uncharge(&memcg->memsw, nr_pages);
2327 css_put_many(&memcg->css, nr_pages);
2330 static void lock_page_lru(struct page *page, int *isolated)
2332 struct zone *zone = page_zone(page);
2334 spin_lock_irq(zone_lru_lock(zone));
2335 if (PageLRU(page)) {
2336 struct lruvec *lruvec;
2338 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2340 del_page_from_lru_list(page, lruvec, page_lru(page));
2346 static void unlock_page_lru(struct page *page, int isolated)
2348 struct zone *zone = page_zone(page);
2351 struct lruvec *lruvec;
2353 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2354 VM_BUG_ON_PAGE(PageLRU(page), page);
2356 add_page_to_lru_list(page, lruvec, page_lru(page));
2358 spin_unlock_irq(zone_lru_lock(zone));
2361 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2366 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2369 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2370 * may already be on some other mem_cgroup's LRU. Take care of it.
2373 lock_page_lru(page, &isolated);
2376 * Nobody should be changing or seriously looking at
2377 * page->mem_cgroup at this point:
2379 * - the page is uncharged
2381 * - the page is off-LRU
2383 * - an anonymous fault has exclusive page access, except for
2384 * a locked page table
2386 * - a page cache insertion, a swapin fault, or a migration
2387 * have the page locked
2389 page->mem_cgroup = memcg;
2392 unlock_page_lru(page, isolated);
2395 #ifdef CONFIG_MEMCG_KMEM
2396 static int memcg_alloc_cache_id(void)
2401 id = ida_simple_get(&memcg_cache_ida,
2402 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2406 if (id < memcg_nr_cache_ids)
2410 * There's no space for the new id in memcg_caches arrays,
2411 * so we have to grow them.
2413 down_write(&memcg_cache_ids_sem);
2415 size = 2 * (id + 1);
2416 if (size < MEMCG_CACHES_MIN_SIZE)
2417 size = MEMCG_CACHES_MIN_SIZE;
2418 else if (size > MEMCG_CACHES_MAX_SIZE)
2419 size = MEMCG_CACHES_MAX_SIZE;
2421 err = memcg_update_all_caches(size);
2423 err = memcg_update_all_list_lrus(size);
2425 memcg_nr_cache_ids = size;
2427 up_write(&memcg_cache_ids_sem);
2430 ida_simple_remove(&memcg_cache_ida, id);
2436 static void memcg_free_cache_id(int id)
2438 ida_simple_remove(&memcg_cache_ida, id);
2441 struct memcg_kmem_cache_create_work {
2442 struct mem_cgroup *memcg;
2443 struct kmem_cache *cachep;
2444 struct work_struct work;
2447 static void memcg_kmem_cache_create_func(struct work_struct *w)
2449 struct memcg_kmem_cache_create_work *cw =
2450 container_of(w, struct memcg_kmem_cache_create_work, work);
2451 struct mem_cgroup *memcg = cw->memcg;
2452 struct kmem_cache *cachep = cw->cachep;
2454 memcg_create_kmem_cache(memcg, cachep);
2456 css_put(&memcg->css);
2461 * Enqueue the creation of a per-memcg kmem_cache.
2463 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2464 struct kmem_cache *cachep)
2466 struct memcg_kmem_cache_create_work *cw;
2468 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2472 css_get(&memcg->css);
2475 cw->cachep = cachep;
2476 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2478 queue_work(memcg_kmem_cache_wq, &cw->work);
2481 static inline bool memcg_kmem_bypass(void)
2483 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2489 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2490 * @cachep: the original global kmem cache
2492 * Return the kmem_cache we're supposed to use for a slab allocation.
2493 * We try to use the current memcg's version of the cache.
2495 * If the cache does not exist yet, if we are the first user of it, we
2496 * create it asynchronously in a workqueue and let the current allocation
2497 * go through with the original cache.
2499 * This function takes a reference to the cache it returns to assure it
2500 * won't get destroyed while we are working with it. Once the caller is
2501 * done with it, memcg_kmem_put_cache() must be called to release the
2504 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2506 struct mem_cgroup *memcg;
2507 struct kmem_cache *memcg_cachep;
2510 VM_BUG_ON(!is_root_cache(cachep));
2512 if (memcg_kmem_bypass())
2515 memcg = get_mem_cgroup_from_current();
2516 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2520 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2521 if (likely(memcg_cachep))
2522 return memcg_cachep;
2525 * If we are in a safe context (can wait, and not in interrupt
2526 * context), we could be be predictable and return right away.
2527 * This would guarantee that the allocation being performed
2528 * already belongs in the new cache.
2530 * However, there are some clashes that can arrive from locking.
2531 * For instance, because we acquire the slab_mutex while doing
2532 * memcg_create_kmem_cache, this means no further allocation
2533 * could happen with the slab_mutex held. So it's better to
2536 memcg_schedule_kmem_cache_create(memcg, cachep);
2538 css_put(&memcg->css);
2543 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2544 * @cachep: the cache returned by memcg_kmem_get_cache
2546 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2548 if (!is_root_cache(cachep))
2549 css_put(&cachep->memcg_params.memcg->css);
2553 * memcg_kmem_charge_memcg: charge a kmem page
2554 * @page: page to charge
2555 * @gfp: reclaim mode
2556 * @order: allocation order
2557 * @memcg: memory cgroup to charge
2559 * Returns 0 on success, an error code on failure.
2561 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2562 struct mem_cgroup *memcg)
2564 unsigned int nr_pages = 1 << order;
2565 struct page_counter *counter;
2568 ret = try_charge(memcg, gfp, nr_pages);
2572 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2573 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2574 cancel_charge(memcg, nr_pages);
2578 page->mem_cgroup = memcg;
2584 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2585 * @page: page to charge
2586 * @gfp: reclaim mode
2587 * @order: allocation order
2589 * Returns 0 on success, an error code on failure.
2591 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2593 struct mem_cgroup *memcg;
2596 if (memcg_kmem_bypass())
2599 memcg = get_mem_cgroup_from_current();
2600 if (!mem_cgroup_is_root(memcg)) {
2601 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2603 __SetPageKmemcg(page);
2605 css_put(&memcg->css);
2609 * memcg_kmem_uncharge: uncharge a kmem page
2610 * @page: page to uncharge
2611 * @order: allocation order
2613 void memcg_kmem_uncharge(struct page *page, int order)
2615 struct mem_cgroup *memcg = page->mem_cgroup;
2616 unsigned int nr_pages = 1 << order;
2621 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2623 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2624 page_counter_uncharge(&memcg->kmem, nr_pages);
2626 page_counter_uncharge(&memcg->memory, nr_pages);
2627 if (do_memsw_account())
2628 page_counter_uncharge(&memcg->memsw, nr_pages);
2630 page->mem_cgroup = NULL;
2632 /* slab pages do not have PageKmemcg flag set */
2633 if (PageKmemcg(page))
2634 __ClearPageKmemcg(page);
2636 css_put_many(&memcg->css, nr_pages);
2638 #endif /* CONFIG_MEMCG_KMEM */
2640 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2643 * Because tail pages are not marked as "used", set it. We're under
2644 * zone_lru_lock and migration entries setup in all page mappings.
2646 void mem_cgroup_split_huge_fixup(struct page *head)
2650 if (mem_cgroup_disabled())
2653 for (i = 1; i < HPAGE_PMD_NR; i++)
2654 head[i].mem_cgroup = head->mem_cgroup;
2656 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2658 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2660 #ifdef CONFIG_MEMCG_SWAP
2662 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2663 * @entry: swap entry to be moved
2664 * @from: mem_cgroup which the entry is moved from
2665 * @to: mem_cgroup which the entry is moved to
2667 * It succeeds only when the swap_cgroup's record for this entry is the same
2668 * as the mem_cgroup's id of @from.
2670 * Returns 0 on success, -EINVAL on failure.
2672 * The caller must have charged to @to, IOW, called page_counter_charge() about
2673 * both res and memsw, and called css_get().
2675 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2676 struct mem_cgroup *from, struct mem_cgroup *to)
2678 unsigned short old_id, new_id;
2680 old_id = mem_cgroup_id(from);
2681 new_id = mem_cgroup_id(to);
2683 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2684 mod_memcg_state(from, MEMCG_SWAP, -1);
2685 mod_memcg_state(to, MEMCG_SWAP, 1);
2691 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2692 struct mem_cgroup *from, struct mem_cgroup *to)
2698 static DEFINE_MUTEX(memcg_max_mutex);
2700 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2701 unsigned long max, bool memsw)
2703 bool enlarge = false;
2704 bool drained = false;
2706 bool limits_invariant;
2707 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2710 if (signal_pending(current)) {
2715 mutex_lock(&memcg_max_mutex);
2717 * Make sure that the new limit (memsw or memory limit) doesn't
2718 * break our basic invariant rule memory.max <= memsw.max.
2720 limits_invariant = memsw ? max >= memcg->memory.max :
2721 max <= memcg->memsw.max;
2722 if (!limits_invariant) {
2723 mutex_unlock(&memcg_max_mutex);
2727 if (max > counter->max)
2729 ret = page_counter_set_max(counter, max);
2730 mutex_unlock(&memcg_max_mutex);
2736 drain_all_stock(memcg);
2741 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2742 GFP_KERNEL, !memsw)) {
2748 if (!ret && enlarge)
2749 memcg_oom_recover(memcg);
2754 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2756 unsigned long *total_scanned)
2758 unsigned long nr_reclaimed = 0;
2759 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2760 unsigned long reclaimed;
2762 struct mem_cgroup_tree_per_node *mctz;
2763 unsigned long excess;
2764 unsigned long nr_scanned;
2769 mctz = soft_limit_tree_node(pgdat->node_id);
2772 * Do not even bother to check the largest node if the root
2773 * is empty. Do it lockless to prevent lock bouncing. Races
2774 * are acceptable as soft limit is best effort anyway.
2776 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2780 * This loop can run a while, specially if mem_cgroup's continuously
2781 * keep exceeding their soft limit and putting the system under
2788 mz = mem_cgroup_largest_soft_limit_node(mctz);
2793 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2794 gfp_mask, &nr_scanned);
2795 nr_reclaimed += reclaimed;
2796 *total_scanned += nr_scanned;
2797 spin_lock_irq(&mctz->lock);
2798 __mem_cgroup_remove_exceeded(mz, mctz);
2801 * If we failed to reclaim anything from this memory cgroup
2802 * it is time to move on to the next cgroup
2806 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2808 excess = soft_limit_excess(mz->memcg);
2810 * One school of thought says that we should not add
2811 * back the node to the tree if reclaim returns 0.
2812 * But our reclaim could return 0, simply because due
2813 * to priority we are exposing a smaller subset of
2814 * memory to reclaim from. Consider this as a longer
2817 /* If excess == 0, no tree ops */
2818 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2819 spin_unlock_irq(&mctz->lock);
2820 css_put(&mz->memcg->css);
2823 * Could not reclaim anything and there are no more
2824 * mem cgroups to try or we seem to be looping without
2825 * reclaiming anything.
2827 if (!nr_reclaimed &&
2829 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2831 } while (!nr_reclaimed);
2833 css_put(&next_mz->memcg->css);
2834 return nr_reclaimed;
2838 * Test whether @memcg has children, dead or alive. Note that this
2839 * function doesn't care whether @memcg has use_hierarchy enabled and
2840 * returns %true if there are child csses according to the cgroup
2841 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2843 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2848 ret = css_next_child(NULL, &memcg->css);
2854 * Reclaims as many pages from the given memcg as possible.
2856 * Caller is responsible for holding css reference for memcg.
2858 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2860 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2862 /* we call try-to-free pages for make this cgroup empty */
2863 lru_add_drain_all();
2865 drain_all_stock(memcg);
2867 /* try to free all pages in this cgroup */
2868 while (nr_retries && page_counter_read(&memcg->memory)) {
2871 if (signal_pending(current))
2874 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2878 /* maybe some writeback is necessary */
2879 congestion_wait(BLK_RW_ASYNC, HZ/10);
2887 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2888 char *buf, size_t nbytes,
2891 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2893 if (mem_cgroup_is_root(memcg))
2895 return mem_cgroup_force_empty(memcg) ?: nbytes;
2898 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2901 return mem_cgroup_from_css(css)->use_hierarchy;
2904 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2905 struct cftype *cft, u64 val)
2908 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2909 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2911 if (memcg->use_hierarchy == val)
2915 * If parent's use_hierarchy is set, we can't make any modifications
2916 * in the child subtrees. If it is unset, then the change can
2917 * occur, provided the current cgroup has no children.
2919 * For the root cgroup, parent_mem is NULL, we allow value to be
2920 * set if there are no children.
2922 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2923 (val == 1 || val == 0)) {
2924 if (!memcg_has_children(memcg))
2925 memcg->use_hierarchy = val;
2934 struct accumulated_stats {
2935 unsigned long stat[MEMCG_NR_STAT];
2936 unsigned long events[NR_VM_EVENT_ITEMS];
2937 unsigned long lru_pages[NR_LRU_LISTS];
2938 const unsigned int *stats_array;
2939 const unsigned int *events_array;
2944 static void accumulate_memcg_tree(struct mem_cgroup *memcg,
2945 struct accumulated_stats *acc)
2947 struct mem_cgroup *mi;
2950 for_each_mem_cgroup_tree(mi, memcg) {
2951 for (i = 0; i < acc->stats_size; i++)
2952 acc->stat[i] += memcg_page_state(mi,
2953 acc->stats_array ? acc->stats_array[i] : i);
2955 for (i = 0; i < acc->events_size; i++)
2956 acc->events[i] += memcg_sum_events(mi,
2957 acc->events_array ? acc->events_array[i] : i);
2959 for (i = 0; i < NR_LRU_LISTS; i++)
2960 acc->lru_pages[i] +=
2961 mem_cgroup_nr_lru_pages(mi, BIT(i));
2965 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2967 unsigned long val = 0;
2969 if (mem_cgroup_is_root(memcg)) {
2970 struct mem_cgroup *iter;
2972 for_each_mem_cgroup_tree(iter, memcg) {
2973 val += memcg_page_state(iter, MEMCG_CACHE);
2974 val += memcg_page_state(iter, MEMCG_RSS);
2976 val += memcg_page_state(iter, MEMCG_SWAP);
2980 val = page_counter_read(&memcg->memory);
2982 val = page_counter_read(&memcg->memsw);
2995 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2998 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2999 struct page_counter *counter;
3001 switch (MEMFILE_TYPE(cft->private)) {
3003 counter = &memcg->memory;
3006 counter = &memcg->memsw;
3009 counter = &memcg->kmem;
3012 counter = &memcg->tcpmem;
3018 switch (MEMFILE_ATTR(cft->private)) {
3020 if (counter == &memcg->memory)
3021 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3022 if (counter == &memcg->memsw)
3023 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3024 return (u64)page_counter_read(counter) * PAGE_SIZE;
3026 return (u64)counter->max * PAGE_SIZE;
3028 return (u64)counter->watermark * PAGE_SIZE;
3030 return counter->failcnt;
3031 case RES_SOFT_LIMIT:
3032 return (u64)memcg->soft_limit * PAGE_SIZE;
3038 #ifdef CONFIG_MEMCG_KMEM
3039 static int memcg_online_kmem(struct mem_cgroup *memcg)
3043 if (cgroup_memory_nokmem)
3046 BUG_ON(memcg->kmemcg_id >= 0);
3047 BUG_ON(memcg->kmem_state);
3049 memcg_id = memcg_alloc_cache_id();
3053 static_branch_inc(&memcg_kmem_enabled_key);
3055 * A memory cgroup is considered kmem-online as soon as it gets
3056 * kmemcg_id. Setting the id after enabling static branching will
3057 * guarantee no one starts accounting before all call sites are
3060 memcg->kmemcg_id = memcg_id;
3061 memcg->kmem_state = KMEM_ONLINE;
3062 INIT_LIST_HEAD(&memcg->kmem_caches);
3067 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3069 struct cgroup_subsys_state *css;
3070 struct mem_cgroup *parent, *child;
3073 if (memcg->kmem_state != KMEM_ONLINE)
3076 * Clear the online state before clearing memcg_caches array
3077 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3078 * guarantees that no cache will be created for this cgroup
3079 * after we are done (see memcg_create_kmem_cache()).
3081 memcg->kmem_state = KMEM_ALLOCATED;
3083 memcg_deactivate_kmem_caches(memcg);
3085 kmemcg_id = memcg->kmemcg_id;
3086 BUG_ON(kmemcg_id < 0);
3088 parent = parent_mem_cgroup(memcg);
3090 parent = root_mem_cgroup;
3093 * Change kmemcg_id of this cgroup and all its descendants to the
3094 * parent's id, and then move all entries from this cgroup's list_lrus
3095 * to ones of the parent. After we have finished, all list_lrus
3096 * corresponding to this cgroup are guaranteed to remain empty. The
3097 * ordering is imposed by list_lru_node->lock taken by
3098 * memcg_drain_all_list_lrus().
3100 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3101 css_for_each_descendant_pre(css, &memcg->css) {
3102 child = mem_cgroup_from_css(css);
3103 BUG_ON(child->kmemcg_id != kmemcg_id);
3104 child->kmemcg_id = parent->kmemcg_id;
3105 if (!memcg->use_hierarchy)
3110 memcg_drain_all_list_lrus(kmemcg_id, parent);
3112 memcg_free_cache_id(kmemcg_id);
3115 static void memcg_free_kmem(struct mem_cgroup *memcg)
3117 /* css_alloc() failed, offlining didn't happen */
3118 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3119 memcg_offline_kmem(memcg);
3121 if (memcg->kmem_state == KMEM_ALLOCATED) {
3122 memcg_destroy_kmem_caches(memcg);
3123 static_branch_dec(&memcg_kmem_enabled_key);
3124 WARN_ON(page_counter_read(&memcg->kmem));
3128 static int memcg_online_kmem(struct mem_cgroup *memcg)
3132 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3135 static void memcg_free_kmem(struct mem_cgroup *memcg)
3138 #endif /* CONFIG_MEMCG_KMEM */
3140 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3145 mutex_lock(&memcg_max_mutex);
3146 ret = page_counter_set_max(&memcg->kmem, max);
3147 mutex_unlock(&memcg_max_mutex);
3151 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3155 mutex_lock(&memcg_max_mutex);
3157 ret = page_counter_set_max(&memcg->tcpmem, max);
3161 if (!memcg->tcpmem_active) {
3163 * The active flag needs to be written after the static_key
3164 * update. This is what guarantees that the socket activation
3165 * function is the last one to run. See mem_cgroup_sk_alloc()
3166 * for details, and note that we don't mark any socket as
3167 * belonging to this memcg until that flag is up.
3169 * We need to do this, because static_keys will span multiple
3170 * sites, but we can't control their order. If we mark a socket
3171 * as accounted, but the accounting functions are not patched in
3172 * yet, we'll lose accounting.
3174 * We never race with the readers in mem_cgroup_sk_alloc(),
3175 * because when this value change, the code to process it is not
3178 static_branch_inc(&memcg_sockets_enabled_key);
3179 memcg->tcpmem_active = true;
3182 mutex_unlock(&memcg_max_mutex);
3187 * The user of this function is...
3190 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3191 char *buf, size_t nbytes, loff_t off)
3193 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3194 unsigned long nr_pages;
3197 buf = strstrip(buf);
3198 ret = page_counter_memparse(buf, "-1", &nr_pages);
3202 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3204 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3208 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3210 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3213 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3216 ret = memcg_update_kmem_max(memcg, nr_pages);
3219 ret = memcg_update_tcp_max(memcg, nr_pages);
3223 case RES_SOFT_LIMIT:
3224 memcg->soft_limit = nr_pages;
3228 return ret ?: nbytes;
3231 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3232 size_t nbytes, loff_t off)
3234 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3235 struct page_counter *counter;
3237 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3239 counter = &memcg->memory;
3242 counter = &memcg->memsw;
3245 counter = &memcg->kmem;
3248 counter = &memcg->tcpmem;
3254 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3256 page_counter_reset_watermark(counter);
3259 counter->failcnt = 0;
3268 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3271 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3275 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3276 struct cftype *cft, u64 val)
3278 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3280 if (val & ~MOVE_MASK)
3284 * No kind of locking is needed in here, because ->can_attach() will
3285 * check this value once in the beginning of the process, and then carry
3286 * on with stale data. This means that changes to this value will only
3287 * affect task migrations starting after the change.
3289 memcg->move_charge_at_immigrate = val;
3293 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3294 struct cftype *cft, u64 val)
3301 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3305 unsigned int lru_mask;
3308 static const struct numa_stat stats[] = {
3309 { "total", LRU_ALL },
3310 { "file", LRU_ALL_FILE },
3311 { "anon", LRU_ALL_ANON },
3312 { "unevictable", BIT(LRU_UNEVICTABLE) },
3314 const struct numa_stat *stat;
3317 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3319 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3320 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3321 seq_printf(m, "%s=%lu", stat->name, nr);
3322 for_each_node_state(nid, N_MEMORY) {
3323 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3325 seq_printf(m, " N%d=%lu", nid, nr);
3330 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3331 struct mem_cgroup *iter;
3334 for_each_mem_cgroup_tree(iter, memcg)
3335 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3336 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3337 for_each_node_state(nid, N_MEMORY) {
3339 for_each_mem_cgroup_tree(iter, memcg)
3340 nr += mem_cgroup_node_nr_lru_pages(
3341 iter, nid, stat->lru_mask);
3342 seq_printf(m, " N%d=%lu", nid, nr);
3349 #endif /* CONFIG_NUMA */
3351 /* Universal VM events cgroup1 shows, original sort order */
3352 static const unsigned int memcg1_events[] = {
3359 static const char *const memcg1_event_names[] = {
3366 static int memcg_stat_show(struct seq_file *m, void *v)
3368 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3369 unsigned long memory, memsw;
3370 struct mem_cgroup *mi;
3372 struct accumulated_stats acc;
3374 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3375 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3377 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3378 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3380 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3381 memcg_page_state(memcg, memcg1_stats[i]) *
3385 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3386 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3387 memcg_sum_events(memcg, memcg1_events[i]));
3389 for (i = 0; i < NR_LRU_LISTS; i++)
3390 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3391 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3393 /* Hierarchical information */
3394 memory = memsw = PAGE_COUNTER_MAX;
3395 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3396 memory = min(memory, mi->memory.max);
3397 memsw = min(memsw, mi->memsw.max);
3399 seq_printf(m, "hierarchical_memory_limit %llu\n",
3400 (u64)memory * PAGE_SIZE);
3401 if (do_memsw_account())
3402 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3403 (u64)memsw * PAGE_SIZE);
3405 memset(&acc, 0, sizeof(acc));
3406 acc.stats_size = ARRAY_SIZE(memcg1_stats);
3407 acc.stats_array = memcg1_stats;
3408 acc.events_size = ARRAY_SIZE(memcg1_events);
3409 acc.events_array = memcg1_events;
3410 accumulate_memcg_tree(memcg, &acc);
3412 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3413 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3415 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3416 (u64)acc.stat[i] * PAGE_SIZE);
3419 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3420 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3421 (u64)acc.events[i]);
3423 for (i = 0; i < NR_LRU_LISTS; i++)
3424 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3425 (u64)acc.lru_pages[i] * PAGE_SIZE);
3427 #ifdef CONFIG_DEBUG_VM
3430 struct mem_cgroup_per_node *mz;
3431 struct zone_reclaim_stat *rstat;
3432 unsigned long recent_rotated[2] = {0, 0};
3433 unsigned long recent_scanned[2] = {0, 0};
3435 for_each_online_pgdat(pgdat) {
3436 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3437 rstat = &mz->lruvec.reclaim_stat;
3439 recent_rotated[0] += rstat->recent_rotated[0];
3440 recent_rotated[1] += rstat->recent_rotated[1];
3441 recent_scanned[0] += rstat->recent_scanned[0];
3442 recent_scanned[1] += rstat->recent_scanned[1];
3444 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3445 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3446 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3447 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3454 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3457 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3459 return mem_cgroup_swappiness(memcg);
3462 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3463 struct cftype *cft, u64 val)
3465 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3471 memcg->swappiness = val;
3473 vm_swappiness = val;
3478 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3480 struct mem_cgroup_threshold_ary *t;
3481 unsigned long usage;
3486 t = rcu_dereference(memcg->thresholds.primary);
3488 t = rcu_dereference(memcg->memsw_thresholds.primary);
3493 usage = mem_cgroup_usage(memcg, swap);
3496 * current_threshold points to threshold just below or equal to usage.
3497 * If it's not true, a threshold was crossed after last
3498 * call of __mem_cgroup_threshold().
3500 i = t->current_threshold;
3503 * Iterate backward over array of thresholds starting from
3504 * current_threshold and check if a threshold is crossed.
3505 * If none of thresholds below usage is crossed, we read
3506 * only one element of the array here.
3508 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3509 eventfd_signal(t->entries[i].eventfd, 1);
3511 /* i = current_threshold + 1 */
3515 * Iterate forward over array of thresholds starting from
3516 * current_threshold+1 and check if a threshold is crossed.
3517 * If none of thresholds above usage is crossed, we read
3518 * only one element of the array here.
3520 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3521 eventfd_signal(t->entries[i].eventfd, 1);
3523 /* Update current_threshold */
3524 t->current_threshold = i - 1;
3529 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3532 __mem_cgroup_threshold(memcg, false);
3533 if (do_memsw_account())
3534 __mem_cgroup_threshold(memcg, true);
3536 memcg = parent_mem_cgroup(memcg);
3540 static int compare_thresholds(const void *a, const void *b)
3542 const struct mem_cgroup_threshold *_a = a;
3543 const struct mem_cgroup_threshold *_b = b;
3545 if (_a->threshold > _b->threshold)
3548 if (_a->threshold < _b->threshold)
3554 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3556 struct mem_cgroup_eventfd_list *ev;
3558 spin_lock(&memcg_oom_lock);
3560 list_for_each_entry(ev, &memcg->oom_notify, list)
3561 eventfd_signal(ev->eventfd, 1);
3563 spin_unlock(&memcg_oom_lock);
3567 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3569 struct mem_cgroup *iter;
3571 for_each_mem_cgroup_tree(iter, memcg)
3572 mem_cgroup_oom_notify_cb(iter);
3575 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3576 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3578 struct mem_cgroup_thresholds *thresholds;
3579 struct mem_cgroup_threshold_ary *new;
3580 unsigned long threshold;
3581 unsigned long usage;
3584 ret = page_counter_memparse(args, "-1", &threshold);
3588 mutex_lock(&memcg->thresholds_lock);
3591 thresholds = &memcg->thresholds;
3592 usage = mem_cgroup_usage(memcg, false);
3593 } else if (type == _MEMSWAP) {
3594 thresholds = &memcg->memsw_thresholds;
3595 usage = mem_cgroup_usage(memcg, true);
3599 /* Check if a threshold crossed before adding a new one */
3600 if (thresholds->primary)
3601 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3603 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3605 /* Allocate memory for new array of thresholds */
3606 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3614 /* Copy thresholds (if any) to new array */
3615 if (thresholds->primary) {
3616 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3617 sizeof(struct mem_cgroup_threshold));
3620 /* Add new threshold */
3621 new->entries[size - 1].eventfd = eventfd;
3622 new->entries[size - 1].threshold = threshold;
3624 /* Sort thresholds. Registering of new threshold isn't time-critical */
3625 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3626 compare_thresholds, NULL);
3628 /* Find current threshold */
3629 new->current_threshold = -1;
3630 for (i = 0; i < size; i++) {
3631 if (new->entries[i].threshold <= usage) {
3633 * new->current_threshold will not be used until
3634 * rcu_assign_pointer(), so it's safe to increment
3637 ++new->current_threshold;
3642 /* Free old spare buffer and save old primary buffer as spare */
3643 kfree(thresholds->spare);
3644 thresholds->spare = thresholds->primary;
3646 rcu_assign_pointer(thresholds->primary, new);
3648 /* To be sure that nobody uses thresholds */
3652 mutex_unlock(&memcg->thresholds_lock);
3657 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3658 struct eventfd_ctx *eventfd, const char *args)
3660 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3663 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3664 struct eventfd_ctx *eventfd, const char *args)
3666 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3669 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3670 struct eventfd_ctx *eventfd, enum res_type type)
3672 struct mem_cgroup_thresholds *thresholds;
3673 struct mem_cgroup_threshold_ary *new;
3674 unsigned long usage;
3677 mutex_lock(&memcg->thresholds_lock);
3680 thresholds = &memcg->thresholds;
3681 usage = mem_cgroup_usage(memcg, false);
3682 } else if (type == _MEMSWAP) {
3683 thresholds = &memcg->memsw_thresholds;
3684 usage = mem_cgroup_usage(memcg, true);
3688 if (!thresholds->primary)
3691 /* Check if a threshold crossed before removing */
3692 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3694 /* Calculate new number of threshold */
3696 for (i = 0; i < thresholds->primary->size; i++) {
3697 if (thresholds->primary->entries[i].eventfd != eventfd)
3701 new = thresholds->spare;
3703 /* Set thresholds array to NULL if we don't have thresholds */
3712 /* Copy thresholds and find current threshold */
3713 new->current_threshold = -1;
3714 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3715 if (thresholds->primary->entries[i].eventfd == eventfd)
3718 new->entries[j] = thresholds->primary->entries[i];
3719 if (new->entries[j].threshold <= usage) {
3721 * new->current_threshold will not be used
3722 * until rcu_assign_pointer(), so it's safe to increment
3725 ++new->current_threshold;
3731 /* Swap primary and spare array */
3732 thresholds->spare = thresholds->primary;
3734 rcu_assign_pointer(thresholds->primary, new);
3736 /* To be sure that nobody uses thresholds */
3739 /* If all events are unregistered, free the spare array */
3741 kfree(thresholds->spare);
3742 thresholds->spare = NULL;
3745 mutex_unlock(&memcg->thresholds_lock);
3748 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3749 struct eventfd_ctx *eventfd)
3751 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3754 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3755 struct eventfd_ctx *eventfd)
3757 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3760 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3761 struct eventfd_ctx *eventfd, const char *args)
3763 struct mem_cgroup_eventfd_list *event;
3765 event = kmalloc(sizeof(*event), GFP_KERNEL);
3769 spin_lock(&memcg_oom_lock);
3771 event->eventfd = eventfd;
3772 list_add(&event->list, &memcg->oom_notify);
3774 /* already in OOM ? */
3775 if (memcg->under_oom)
3776 eventfd_signal(eventfd, 1);
3777 spin_unlock(&memcg_oom_lock);
3782 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3783 struct eventfd_ctx *eventfd)
3785 struct mem_cgroup_eventfd_list *ev, *tmp;
3787 spin_lock(&memcg_oom_lock);
3789 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3790 if (ev->eventfd == eventfd) {
3791 list_del(&ev->list);
3796 spin_unlock(&memcg_oom_lock);
3799 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3801 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3803 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3804 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3805 seq_printf(sf, "oom_kill %lu\n",
3806 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3810 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3811 struct cftype *cft, u64 val)
3813 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3815 /* cannot set to root cgroup and only 0 and 1 are allowed */
3816 if (!css->parent || !((val == 0) || (val == 1)))
3819 memcg->oom_kill_disable = val;
3821 memcg_oom_recover(memcg);
3826 #ifdef CONFIG_CGROUP_WRITEBACK
3828 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3830 return wb_domain_init(&memcg->cgwb_domain, gfp);
3833 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3835 wb_domain_exit(&memcg->cgwb_domain);
3838 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3840 wb_domain_size_changed(&memcg->cgwb_domain);
3843 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3845 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3847 if (!memcg->css.parent)
3850 return &memcg->cgwb_domain;
3854 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3855 * @wb: bdi_writeback in question
3856 * @pfilepages: out parameter for number of file pages
3857 * @pheadroom: out parameter for number of allocatable pages according to memcg
3858 * @pdirty: out parameter for number of dirty pages
3859 * @pwriteback: out parameter for number of pages under writeback
3861 * Determine the numbers of file, headroom, dirty, and writeback pages in
3862 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3863 * is a bit more involved.
3865 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3866 * headroom is calculated as the lowest headroom of itself and the
3867 * ancestors. Note that this doesn't consider the actual amount of
3868 * available memory in the system. The caller should further cap
3869 * *@pheadroom accordingly.
3871 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3872 unsigned long *pheadroom, unsigned long *pdirty,
3873 unsigned long *pwriteback)
3875 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3876 struct mem_cgroup *parent;
3878 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
3880 /* this should eventually include NR_UNSTABLE_NFS */
3881 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
3882 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3883 (1 << LRU_ACTIVE_FILE));
3884 *pheadroom = PAGE_COUNTER_MAX;
3886 while ((parent = parent_mem_cgroup(memcg))) {
3887 unsigned long ceiling = min(memcg->memory.max, memcg->high);
3888 unsigned long used = page_counter_read(&memcg->memory);
3890 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3895 #else /* CONFIG_CGROUP_WRITEBACK */
3897 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3902 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3906 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3910 #endif /* CONFIG_CGROUP_WRITEBACK */
3913 * DO NOT USE IN NEW FILES.
3915 * "cgroup.event_control" implementation.
3917 * This is way over-engineered. It tries to support fully configurable
3918 * events for each user. Such level of flexibility is completely
3919 * unnecessary especially in the light of the planned unified hierarchy.
3921 * Please deprecate this and replace with something simpler if at all
3926 * Unregister event and free resources.
3928 * Gets called from workqueue.
3930 static void memcg_event_remove(struct work_struct *work)
3932 struct mem_cgroup_event *event =
3933 container_of(work, struct mem_cgroup_event, remove);
3934 struct mem_cgroup *memcg = event->memcg;
3936 remove_wait_queue(event->wqh, &event->wait);
3938 event->unregister_event(memcg, event->eventfd);
3940 /* Notify userspace the event is going away. */
3941 eventfd_signal(event->eventfd, 1);
3943 eventfd_ctx_put(event->eventfd);
3945 css_put(&memcg->css);
3949 * Gets called on EPOLLHUP on eventfd when user closes it.
3951 * Called with wqh->lock held and interrupts disabled.
3953 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
3954 int sync, void *key)
3956 struct mem_cgroup_event *event =
3957 container_of(wait, struct mem_cgroup_event, wait);
3958 struct mem_cgroup *memcg = event->memcg;
3959 __poll_t flags = key_to_poll(key);
3961 if (flags & EPOLLHUP) {
3963 * If the event has been detached at cgroup removal, we
3964 * can simply return knowing the other side will cleanup
3967 * We can't race against event freeing since the other
3968 * side will require wqh->lock via remove_wait_queue(),
3971 spin_lock(&memcg->event_list_lock);
3972 if (!list_empty(&event->list)) {
3973 list_del_init(&event->list);
3975 * We are in atomic context, but cgroup_event_remove()
3976 * may sleep, so we have to call it in workqueue.
3978 schedule_work(&event->remove);
3980 spin_unlock(&memcg->event_list_lock);
3986 static void memcg_event_ptable_queue_proc(struct file *file,
3987 wait_queue_head_t *wqh, poll_table *pt)
3989 struct mem_cgroup_event *event =
3990 container_of(pt, struct mem_cgroup_event, pt);
3993 add_wait_queue(wqh, &event->wait);
3997 * DO NOT USE IN NEW FILES.
3999 * Parse input and register new cgroup event handler.
4001 * Input must be in format '<event_fd> <control_fd> <args>'.
4002 * Interpretation of args is defined by control file implementation.
4004 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4005 char *buf, size_t nbytes, loff_t off)
4007 struct cgroup_subsys_state *css = of_css(of);
4008 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4009 struct mem_cgroup_event *event;
4010 struct cgroup_subsys_state *cfile_css;
4011 unsigned int efd, cfd;
4018 buf = strstrip(buf);
4020 efd = simple_strtoul(buf, &endp, 10);
4025 cfd = simple_strtoul(buf, &endp, 10);
4026 if ((*endp != ' ') && (*endp != '\0'))
4030 event = kzalloc(sizeof(*event), GFP_KERNEL);
4034 event->memcg = memcg;
4035 INIT_LIST_HEAD(&event->list);
4036 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4037 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4038 INIT_WORK(&event->remove, memcg_event_remove);
4046 event->eventfd = eventfd_ctx_fileget(efile.file);
4047 if (IS_ERR(event->eventfd)) {
4048 ret = PTR_ERR(event->eventfd);
4055 goto out_put_eventfd;
4058 /* the process need read permission on control file */
4059 /* AV: shouldn't we check that it's been opened for read instead? */
4060 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4065 * Determine the event callbacks and set them in @event. This used
4066 * to be done via struct cftype but cgroup core no longer knows
4067 * about these events. The following is crude but the whole thing
4068 * is for compatibility anyway.
4070 * DO NOT ADD NEW FILES.
4072 name = cfile.file->f_path.dentry->d_name.name;
4074 if (!strcmp(name, "memory.usage_in_bytes")) {
4075 event->register_event = mem_cgroup_usage_register_event;
4076 event->unregister_event = mem_cgroup_usage_unregister_event;
4077 } else if (!strcmp(name, "memory.oom_control")) {
4078 event->register_event = mem_cgroup_oom_register_event;
4079 event->unregister_event = mem_cgroup_oom_unregister_event;
4080 } else if (!strcmp(name, "memory.pressure_level")) {
4081 event->register_event = vmpressure_register_event;
4082 event->unregister_event = vmpressure_unregister_event;
4083 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4084 event->register_event = memsw_cgroup_usage_register_event;
4085 event->unregister_event = memsw_cgroup_usage_unregister_event;
4092 * Verify @cfile should belong to @css. Also, remaining events are
4093 * automatically removed on cgroup destruction but the removal is
4094 * asynchronous, so take an extra ref on @css.
4096 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4097 &memory_cgrp_subsys);
4099 if (IS_ERR(cfile_css))
4101 if (cfile_css != css) {
4106 ret = event->register_event(memcg, event->eventfd, buf);
4110 vfs_poll(efile.file, &event->pt);
4112 spin_lock(&memcg->event_list_lock);
4113 list_add(&event->list, &memcg->event_list);
4114 spin_unlock(&memcg->event_list_lock);
4126 eventfd_ctx_put(event->eventfd);
4135 static struct cftype mem_cgroup_legacy_files[] = {
4137 .name = "usage_in_bytes",
4138 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4139 .read_u64 = mem_cgroup_read_u64,
4142 .name = "max_usage_in_bytes",
4143 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4144 .write = mem_cgroup_reset,
4145 .read_u64 = mem_cgroup_read_u64,
4148 .name = "limit_in_bytes",
4149 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4150 .write = mem_cgroup_write,
4151 .read_u64 = mem_cgroup_read_u64,
4154 .name = "soft_limit_in_bytes",
4155 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4156 .write = mem_cgroup_write,
4157 .read_u64 = mem_cgroup_read_u64,
4161 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4162 .write = mem_cgroup_reset,
4163 .read_u64 = mem_cgroup_read_u64,
4167 .seq_show = memcg_stat_show,
4170 .name = "force_empty",
4171 .write = mem_cgroup_force_empty_write,
4174 .name = "use_hierarchy",
4175 .write_u64 = mem_cgroup_hierarchy_write,
4176 .read_u64 = mem_cgroup_hierarchy_read,
4179 .name = "cgroup.event_control", /* XXX: for compat */
4180 .write = memcg_write_event_control,
4181 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4184 .name = "swappiness",
4185 .read_u64 = mem_cgroup_swappiness_read,
4186 .write_u64 = mem_cgroup_swappiness_write,
4189 .name = "move_charge_at_immigrate",
4190 .read_u64 = mem_cgroup_move_charge_read,
4191 .write_u64 = mem_cgroup_move_charge_write,
4194 .name = "oom_control",
4195 .seq_show = mem_cgroup_oom_control_read,
4196 .write_u64 = mem_cgroup_oom_control_write,
4197 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4200 .name = "pressure_level",
4204 .name = "numa_stat",
4205 .seq_show = memcg_numa_stat_show,
4209 .name = "kmem.limit_in_bytes",
4210 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4211 .write = mem_cgroup_write,
4212 .read_u64 = mem_cgroup_read_u64,
4215 .name = "kmem.usage_in_bytes",
4216 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4217 .read_u64 = mem_cgroup_read_u64,
4220 .name = "kmem.failcnt",
4221 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4222 .write = mem_cgroup_reset,
4223 .read_u64 = mem_cgroup_read_u64,
4226 .name = "kmem.max_usage_in_bytes",
4227 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4228 .write = mem_cgroup_reset,
4229 .read_u64 = mem_cgroup_read_u64,
4231 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4233 .name = "kmem.slabinfo",
4234 .seq_start = memcg_slab_start,
4235 .seq_next = memcg_slab_next,
4236 .seq_stop = memcg_slab_stop,
4237 .seq_show = memcg_slab_show,
4241 .name = "kmem.tcp.limit_in_bytes",
4242 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4243 .write = mem_cgroup_write,
4244 .read_u64 = mem_cgroup_read_u64,
4247 .name = "kmem.tcp.usage_in_bytes",
4248 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4249 .read_u64 = mem_cgroup_read_u64,
4252 .name = "kmem.tcp.failcnt",
4253 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4254 .write = mem_cgroup_reset,
4255 .read_u64 = mem_cgroup_read_u64,
4258 .name = "kmem.tcp.max_usage_in_bytes",
4259 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4260 .write = mem_cgroup_reset,
4261 .read_u64 = mem_cgroup_read_u64,
4263 { }, /* terminate */
4267 * Private memory cgroup IDR
4269 * Swap-out records and page cache shadow entries need to store memcg
4270 * references in constrained space, so we maintain an ID space that is
4271 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4272 * memory-controlled cgroups to 64k.
4274 * However, there usually are many references to the oflline CSS after
4275 * the cgroup has been destroyed, such as page cache or reclaimable
4276 * slab objects, that don't need to hang on to the ID. We want to keep
4277 * those dead CSS from occupying IDs, or we might quickly exhaust the
4278 * relatively small ID space and prevent the creation of new cgroups
4279 * even when there are much fewer than 64k cgroups - possibly none.
4281 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4282 * be freed and recycled when it's no longer needed, which is usually
4283 * when the CSS is offlined.
4285 * The only exception to that are records of swapped out tmpfs/shmem
4286 * pages that need to be attributed to live ancestors on swapin. But
4287 * those references are manageable from userspace.
4290 static DEFINE_IDR(mem_cgroup_idr);
4292 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4294 if (memcg->id.id > 0) {
4295 idr_remove(&mem_cgroup_idr, memcg->id.id);
4300 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4302 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4303 atomic_add(n, &memcg->id.ref);
4306 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4308 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4309 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4310 mem_cgroup_id_remove(memcg);
4312 /* Memcg ID pins CSS */
4313 css_put(&memcg->css);
4317 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4319 mem_cgroup_id_get_many(memcg, 1);
4322 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4324 mem_cgroup_id_put_many(memcg, 1);
4328 * mem_cgroup_from_id - look up a memcg from a memcg id
4329 * @id: the memcg id to look up
4331 * Caller must hold rcu_read_lock().
4333 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4335 WARN_ON_ONCE(!rcu_read_lock_held());
4336 return idr_find(&mem_cgroup_idr, id);
4339 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4341 struct mem_cgroup_per_node *pn;
4344 * This routine is called against possible nodes.
4345 * But it's BUG to call kmalloc() against offline node.
4347 * TODO: this routine can waste much memory for nodes which will
4348 * never be onlined. It's better to use memory hotplug callback
4351 if (!node_state(node, N_NORMAL_MEMORY))
4353 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4357 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4358 if (!pn->lruvec_stat_cpu) {
4363 lruvec_init(&pn->lruvec);
4364 pn->usage_in_excess = 0;
4365 pn->on_tree = false;
4368 memcg->nodeinfo[node] = pn;
4372 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4374 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4379 free_percpu(pn->lruvec_stat_cpu);
4383 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4388 free_mem_cgroup_per_node_info(memcg, node);
4389 free_percpu(memcg->stat_cpu);
4393 static void mem_cgroup_free(struct mem_cgroup *memcg)
4395 memcg_wb_domain_exit(memcg);
4396 __mem_cgroup_free(memcg);
4399 static struct mem_cgroup *mem_cgroup_alloc(void)
4401 struct mem_cgroup *memcg;
4405 size = sizeof(struct mem_cgroup);
4406 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4408 memcg = kzalloc(size, GFP_KERNEL);
4412 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4413 1, MEM_CGROUP_ID_MAX,
4415 if (memcg->id.id < 0)
4418 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4419 if (!memcg->stat_cpu)
4423 if (alloc_mem_cgroup_per_node_info(memcg, node))
4426 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4429 INIT_WORK(&memcg->high_work, high_work_func);
4430 memcg->last_scanned_node = MAX_NUMNODES;
4431 INIT_LIST_HEAD(&memcg->oom_notify);
4432 mutex_init(&memcg->thresholds_lock);
4433 spin_lock_init(&memcg->move_lock);
4434 vmpressure_init(&memcg->vmpressure);
4435 INIT_LIST_HEAD(&memcg->event_list);
4436 spin_lock_init(&memcg->event_list_lock);
4437 memcg->socket_pressure = jiffies;
4438 #ifdef CONFIG_MEMCG_KMEM
4439 memcg->kmemcg_id = -1;
4441 #ifdef CONFIG_CGROUP_WRITEBACK
4442 INIT_LIST_HEAD(&memcg->cgwb_list);
4444 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4447 mem_cgroup_id_remove(memcg);
4448 __mem_cgroup_free(memcg);
4452 static struct cgroup_subsys_state * __ref
4453 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4455 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4456 struct mem_cgroup *memcg;
4457 long error = -ENOMEM;
4459 memcg = mem_cgroup_alloc();
4461 return ERR_PTR(error);
4463 memcg->high = PAGE_COUNTER_MAX;
4464 memcg->soft_limit = PAGE_COUNTER_MAX;
4466 memcg->swappiness = mem_cgroup_swappiness(parent);
4467 memcg->oom_kill_disable = parent->oom_kill_disable;
4469 if (parent && parent->use_hierarchy) {
4470 memcg->use_hierarchy = true;
4471 page_counter_init(&memcg->memory, &parent->memory);
4472 page_counter_init(&memcg->swap, &parent->swap);
4473 page_counter_init(&memcg->memsw, &parent->memsw);
4474 page_counter_init(&memcg->kmem, &parent->kmem);
4475 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4477 page_counter_init(&memcg->memory, NULL);
4478 page_counter_init(&memcg->swap, NULL);
4479 page_counter_init(&memcg->memsw, NULL);
4480 page_counter_init(&memcg->kmem, NULL);
4481 page_counter_init(&memcg->tcpmem, NULL);
4483 * Deeper hierachy with use_hierarchy == false doesn't make
4484 * much sense so let cgroup subsystem know about this
4485 * unfortunate state in our controller.
4487 if (parent != root_mem_cgroup)
4488 memory_cgrp_subsys.broken_hierarchy = true;
4491 /* The following stuff does not apply to the root */
4493 root_mem_cgroup = memcg;
4497 error = memcg_online_kmem(memcg);
4501 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4502 static_branch_inc(&memcg_sockets_enabled_key);
4506 mem_cgroup_id_remove(memcg);
4507 mem_cgroup_free(memcg);
4508 return ERR_PTR(-ENOMEM);
4511 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4513 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4516 * A memcg must be visible for memcg_expand_shrinker_maps()
4517 * by the time the maps are allocated. So, we allocate maps
4518 * here, when for_each_mem_cgroup() can't skip it.
4520 if (memcg_alloc_shrinker_maps(memcg)) {
4521 mem_cgroup_id_remove(memcg);
4525 /* Online state pins memcg ID, memcg ID pins CSS */
4526 atomic_set(&memcg->id.ref, 1);
4531 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4533 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4534 struct mem_cgroup_event *event, *tmp;
4537 * Unregister events and notify userspace.
4538 * Notify userspace about cgroup removing only after rmdir of cgroup
4539 * directory to avoid race between userspace and kernelspace.
4541 spin_lock(&memcg->event_list_lock);
4542 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4543 list_del_init(&event->list);
4544 schedule_work(&event->remove);
4546 spin_unlock(&memcg->event_list_lock);
4548 page_counter_set_min(&memcg->memory, 0);
4549 page_counter_set_low(&memcg->memory, 0);
4551 memcg_offline_kmem(memcg);
4552 wb_memcg_offline(memcg);
4554 drain_all_stock(memcg);
4556 mem_cgroup_id_put(memcg);
4559 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4561 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4563 invalidate_reclaim_iterators(memcg);
4566 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4568 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4570 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4571 static_branch_dec(&memcg_sockets_enabled_key);
4573 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4574 static_branch_dec(&memcg_sockets_enabled_key);
4576 vmpressure_cleanup(&memcg->vmpressure);
4577 cancel_work_sync(&memcg->high_work);
4578 mem_cgroup_remove_from_trees(memcg);
4579 memcg_free_shrinker_maps(memcg);
4580 memcg_free_kmem(memcg);
4581 mem_cgroup_free(memcg);
4585 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4586 * @css: the target css
4588 * Reset the states of the mem_cgroup associated with @css. This is
4589 * invoked when the userland requests disabling on the default hierarchy
4590 * but the memcg is pinned through dependency. The memcg should stop
4591 * applying policies and should revert to the vanilla state as it may be
4592 * made visible again.
4594 * The current implementation only resets the essential configurations.
4595 * This needs to be expanded to cover all the visible parts.
4597 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4599 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4601 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4602 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4603 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4604 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4605 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4606 page_counter_set_min(&memcg->memory, 0);
4607 page_counter_set_low(&memcg->memory, 0);
4608 memcg->high = PAGE_COUNTER_MAX;
4609 memcg->soft_limit = PAGE_COUNTER_MAX;
4610 memcg_wb_domain_size_changed(memcg);
4614 /* Handlers for move charge at task migration. */
4615 static int mem_cgroup_do_precharge(unsigned long count)
4619 /* Try a single bulk charge without reclaim first, kswapd may wake */
4620 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4622 mc.precharge += count;
4626 /* Try charges one by one with reclaim, but do not retry */
4628 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4642 enum mc_target_type {
4649 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4650 unsigned long addr, pte_t ptent)
4652 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4654 if (!page || !page_mapped(page))
4656 if (PageAnon(page)) {
4657 if (!(mc.flags & MOVE_ANON))
4660 if (!(mc.flags & MOVE_FILE))
4663 if (!get_page_unless_zero(page))
4669 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4670 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4671 pte_t ptent, swp_entry_t *entry)
4673 struct page *page = NULL;
4674 swp_entry_t ent = pte_to_swp_entry(ptent);
4676 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4680 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4681 * a device and because they are not accessible by CPU they are store
4682 * as special swap entry in the CPU page table.
4684 if (is_device_private_entry(ent)) {
4685 page = device_private_entry_to_page(ent);
4687 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4688 * a refcount of 1 when free (unlike normal page)
4690 if (!page_ref_add_unless(page, 1, 1))
4696 * Because lookup_swap_cache() updates some statistics counter,
4697 * we call find_get_page() with swapper_space directly.
4699 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4700 if (do_memsw_account())
4701 entry->val = ent.val;
4706 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4707 pte_t ptent, swp_entry_t *entry)
4713 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4714 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4716 struct page *page = NULL;
4717 struct address_space *mapping;
4720 if (!vma->vm_file) /* anonymous vma */
4722 if (!(mc.flags & MOVE_FILE))
4725 mapping = vma->vm_file->f_mapping;
4726 pgoff = linear_page_index(vma, addr);
4728 /* page is moved even if it's not RSS of this task(page-faulted). */
4730 /* shmem/tmpfs may report page out on swap: account for that too. */
4731 if (shmem_mapping(mapping)) {
4732 page = find_get_entry(mapping, pgoff);
4733 if (radix_tree_exceptional_entry(page)) {
4734 swp_entry_t swp = radix_to_swp_entry(page);
4735 if (do_memsw_account())
4737 page = find_get_page(swap_address_space(swp),
4741 page = find_get_page(mapping, pgoff);
4743 page = find_get_page(mapping, pgoff);
4749 * mem_cgroup_move_account - move account of the page
4751 * @compound: charge the page as compound or small page
4752 * @from: mem_cgroup which the page is moved from.
4753 * @to: mem_cgroup which the page is moved to. @from != @to.
4755 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4757 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4760 static int mem_cgroup_move_account(struct page *page,
4762 struct mem_cgroup *from,
4763 struct mem_cgroup *to)
4765 unsigned long flags;
4766 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4770 VM_BUG_ON(from == to);
4771 VM_BUG_ON_PAGE(PageLRU(page), page);
4772 VM_BUG_ON(compound && !PageTransHuge(page));
4775 * Prevent mem_cgroup_migrate() from looking at
4776 * page->mem_cgroup of its source page while we change it.
4779 if (!trylock_page(page))
4783 if (page->mem_cgroup != from)
4786 anon = PageAnon(page);
4788 spin_lock_irqsave(&from->move_lock, flags);
4790 if (!anon && page_mapped(page)) {
4791 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4792 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4796 * move_lock grabbed above and caller set from->moving_account, so
4797 * mod_memcg_page_state will serialize updates to PageDirty.
4798 * So mapping should be stable for dirty pages.
4800 if (!anon && PageDirty(page)) {
4801 struct address_space *mapping = page_mapping(page);
4803 if (mapping_cap_account_dirty(mapping)) {
4804 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4805 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4809 if (PageWriteback(page)) {
4810 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4811 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4815 * It is safe to change page->mem_cgroup here because the page
4816 * is referenced, charged, and isolated - we can't race with
4817 * uncharging, charging, migration, or LRU putback.
4820 /* caller should have done css_get */
4821 page->mem_cgroup = to;
4822 spin_unlock_irqrestore(&from->move_lock, flags);
4826 local_irq_disable();
4827 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4828 memcg_check_events(to, page);
4829 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4830 memcg_check_events(from, page);
4839 * get_mctgt_type - get target type of moving charge
4840 * @vma: the vma the pte to be checked belongs
4841 * @addr: the address corresponding to the pte to be checked
4842 * @ptent: the pte to be checked
4843 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4846 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4847 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4848 * move charge. if @target is not NULL, the page is stored in target->page
4849 * with extra refcnt got(Callers should handle it).
4850 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4851 * target for charge migration. if @target is not NULL, the entry is stored
4853 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4854 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4855 * For now we such page is charge like a regular page would be as for all
4856 * intent and purposes it is just special memory taking the place of a
4859 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4861 * Called with pte lock held.
4864 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4865 unsigned long addr, pte_t ptent, union mc_target *target)
4867 struct page *page = NULL;
4868 enum mc_target_type ret = MC_TARGET_NONE;
4869 swp_entry_t ent = { .val = 0 };
4871 if (pte_present(ptent))
4872 page = mc_handle_present_pte(vma, addr, ptent);
4873 else if (is_swap_pte(ptent))
4874 page = mc_handle_swap_pte(vma, ptent, &ent);
4875 else if (pte_none(ptent))
4876 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4878 if (!page && !ent.val)
4882 * Do only loose check w/o serialization.
4883 * mem_cgroup_move_account() checks the page is valid or
4884 * not under LRU exclusion.
4886 if (page->mem_cgroup == mc.from) {
4887 ret = MC_TARGET_PAGE;
4888 if (is_device_private_page(page) ||
4889 is_device_public_page(page))
4890 ret = MC_TARGET_DEVICE;
4892 target->page = page;
4894 if (!ret || !target)
4898 * There is a swap entry and a page doesn't exist or isn't charged.
4899 * But we cannot move a tail-page in a THP.
4901 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
4902 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4903 ret = MC_TARGET_SWAP;
4910 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4912 * We don't consider PMD mapped swapping or file mapped pages because THP does
4913 * not support them for now.
4914 * Caller should make sure that pmd_trans_huge(pmd) is true.
4916 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4917 unsigned long addr, pmd_t pmd, union mc_target *target)
4919 struct page *page = NULL;
4920 enum mc_target_type ret = MC_TARGET_NONE;
4922 if (unlikely(is_swap_pmd(pmd))) {
4923 VM_BUG_ON(thp_migration_supported() &&
4924 !is_pmd_migration_entry(pmd));
4927 page = pmd_page(pmd);
4928 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4929 if (!(mc.flags & MOVE_ANON))
4931 if (page->mem_cgroup == mc.from) {
4932 ret = MC_TARGET_PAGE;
4935 target->page = page;
4941 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4942 unsigned long addr, pmd_t pmd, union mc_target *target)
4944 return MC_TARGET_NONE;
4948 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4949 unsigned long addr, unsigned long end,
4950 struct mm_walk *walk)
4952 struct vm_area_struct *vma = walk->vma;
4956 ptl = pmd_trans_huge_lock(pmd, vma);
4959 * Note their can not be MC_TARGET_DEVICE for now as we do not
4960 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
4961 * MEMORY_DEVICE_PRIVATE but this might change.
4963 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4964 mc.precharge += HPAGE_PMD_NR;
4969 if (pmd_trans_unstable(pmd))
4971 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4972 for (; addr != end; pte++, addr += PAGE_SIZE)
4973 if (get_mctgt_type(vma, addr, *pte, NULL))
4974 mc.precharge++; /* increment precharge temporarily */
4975 pte_unmap_unlock(pte - 1, ptl);
4981 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4983 unsigned long precharge;
4985 struct mm_walk mem_cgroup_count_precharge_walk = {
4986 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4989 down_read(&mm->mmap_sem);
4990 walk_page_range(0, mm->highest_vm_end,
4991 &mem_cgroup_count_precharge_walk);
4992 up_read(&mm->mmap_sem);
4994 precharge = mc.precharge;
5000 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5002 unsigned long precharge = mem_cgroup_count_precharge(mm);
5004 VM_BUG_ON(mc.moving_task);
5005 mc.moving_task = current;
5006 return mem_cgroup_do_precharge(precharge);
5009 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5010 static void __mem_cgroup_clear_mc(void)
5012 struct mem_cgroup *from = mc.from;
5013 struct mem_cgroup *to = mc.to;
5015 /* we must uncharge all the leftover precharges from mc.to */
5017 cancel_charge(mc.to, mc.precharge);
5021 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5022 * we must uncharge here.
5024 if (mc.moved_charge) {
5025 cancel_charge(mc.from, mc.moved_charge);
5026 mc.moved_charge = 0;
5028 /* we must fixup refcnts and charges */
5029 if (mc.moved_swap) {
5030 /* uncharge swap account from the old cgroup */
5031 if (!mem_cgroup_is_root(mc.from))
5032 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5034 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5037 * we charged both to->memory and to->memsw, so we
5038 * should uncharge to->memory.
5040 if (!mem_cgroup_is_root(mc.to))
5041 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5043 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5044 css_put_many(&mc.to->css, mc.moved_swap);
5048 memcg_oom_recover(from);
5049 memcg_oom_recover(to);
5050 wake_up_all(&mc.waitq);
5053 static void mem_cgroup_clear_mc(void)
5055 struct mm_struct *mm = mc.mm;
5058 * we must clear moving_task before waking up waiters at the end of
5061 mc.moving_task = NULL;
5062 __mem_cgroup_clear_mc();
5063 spin_lock(&mc.lock);
5067 spin_unlock(&mc.lock);
5072 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5074 struct cgroup_subsys_state *css;
5075 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5076 struct mem_cgroup *from;
5077 struct task_struct *leader, *p;
5078 struct mm_struct *mm;
5079 unsigned long move_flags;
5082 /* charge immigration isn't supported on the default hierarchy */
5083 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5087 * Multi-process migrations only happen on the default hierarchy
5088 * where charge immigration is not used. Perform charge
5089 * immigration if @tset contains a leader and whine if there are
5093 cgroup_taskset_for_each_leader(leader, css, tset) {
5096 memcg = mem_cgroup_from_css(css);
5102 * We are now commited to this value whatever it is. Changes in this
5103 * tunable will only affect upcoming migrations, not the current one.
5104 * So we need to save it, and keep it going.
5106 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5110 from = mem_cgroup_from_task(p);
5112 VM_BUG_ON(from == memcg);
5114 mm = get_task_mm(p);
5117 /* We move charges only when we move a owner of the mm */
5118 if (mm->owner == p) {
5121 VM_BUG_ON(mc.precharge);
5122 VM_BUG_ON(mc.moved_charge);
5123 VM_BUG_ON(mc.moved_swap);
5125 spin_lock(&mc.lock);
5129 mc.flags = move_flags;
5130 spin_unlock(&mc.lock);
5131 /* We set mc.moving_task later */
5133 ret = mem_cgroup_precharge_mc(mm);
5135 mem_cgroup_clear_mc();
5142 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5145 mem_cgroup_clear_mc();
5148 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5149 unsigned long addr, unsigned long end,
5150 struct mm_walk *walk)
5153 struct vm_area_struct *vma = walk->vma;
5156 enum mc_target_type target_type;
5157 union mc_target target;
5160 ptl = pmd_trans_huge_lock(pmd, vma);
5162 if (mc.precharge < HPAGE_PMD_NR) {
5166 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5167 if (target_type == MC_TARGET_PAGE) {
5169 if (!isolate_lru_page(page)) {
5170 if (!mem_cgroup_move_account(page, true,
5172 mc.precharge -= HPAGE_PMD_NR;
5173 mc.moved_charge += HPAGE_PMD_NR;
5175 putback_lru_page(page);
5178 } else if (target_type == MC_TARGET_DEVICE) {
5180 if (!mem_cgroup_move_account(page, true,
5182 mc.precharge -= HPAGE_PMD_NR;
5183 mc.moved_charge += HPAGE_PMD_NR;
5191 if (pmd_trans_unstable(pmd))
5194 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5195 for (; addr != end; addr += PAGE_SIZE) {
5196 pte_t ptent = *(pte++);
5197 bool device = false;
5203 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5204 case MC_TARGET_DEVICE:
5207 case MC_TARGET_PAGE:
5210 * We can have a part of the split pmd here. Moving it
5211 * can be done but it would be too convoluted so simply
5212 * ignore such a partial THP and keep it in original
5213 * memcg. There should be somebody mapping the head.
5215 if (PageTransCompound(page))
5217 if (!device && isolate_lru_page(page))
5219 if (!mem_cgroup_move_account(page, false,
5222 /* we uncharge from mc.from later. */
5226 putback_lru_page(page);
5227 put: /* get_mctgt_type() gets the page */
5230 case MC_TARGET_SWAP:
5232 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5234 /* we fixup refcnts and charges later. */
5242 pte_unmap_unlock(pte - 1, ptl);
5247 * We have consumed all precharges we got in can_attach().
5248 * We try charge one by one, but don't do any additional
5249 * charges to mc.to if we have failed in charge once in attach()
5252 ret = mem_cgroup_do_precharge(1);
5260 static void mem_cgroup_move_charge(void)
5262 struct mm_walk mem_cgroup_move_charge_walk = {
5263 .pmd_entry = mem_cgroup_move_charge_pte_range,
5267 lru_add_drain_all();
5269 * Signal lock_page_memcg() to take the memcg's move_lock
5270 * while we're moving its pages to another memcg. Then wait
5271 * for already started RCU-only updates to finish.
5273 atomic_inc(&mc.from->moving_account);
5276 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5278 * Someone who are holding the mmap_sem might be waiting in
5279 * waitq. So we cancel all extra charges, wake up all waiters,
5280 * and retry. Because we cancel precharges, we might not be able
5281 * to move enough charges, but moving charge is a best-effort
5282 * feature anyway, so it wouldn't be a big problem.
5284 __mem_cgroup_clear_mc();
5289 * When we have consumed all precharges and failed in doing
5290 * additional charge, the page walk just aborts.
5292 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5294 up_read(&mc.mm->mmap_sem);
5295 atomic_dec(&mc.from->moving_account);
5298 static void mem_cgroup_move_task(void)
5301 mem_cgroup_move_charge();
5302 mem_cgroup_clear_mc();
5305 #else /* !CONFIG_MMU */
5306 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5310 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5313 static void mem_cgroup_move_task(void)
5319 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5320 * to verify whether we're attached to the default hierarchy on each mount
5323 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5326 * use_hierarchy is forced on the default hierarchy. cgroup core
5327 * guarantees that @root doesn't have any children, so turning it
5328 * on for the root memcg is enough.
5330 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5331 root_mem_cgroup->use_hierarchy = true;
5333 root_mem_cgroup->use_hierarchy = false;
5336 static u64 memory_current_read(struct cgroup_subsys_state *css,
5339 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5341 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5344 static int memory_min_show(struct seq_file *m, void *v)
5346 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5347 unsigned long min = READ_ONCE(memcg->memory.min);
5349 if (min == PAGE_COUNTER_MAX)
5350 seq_puts(m, "max\n");
5352 seq_printf(m, "%llu\n", (u64)min * PAGE_SIZE);
5357 static ssize_t memory_min_write(struct kernfs_open_file *of,
5358 char *buf, size_t nbytes, loff_t off)
5360 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5364 buf = strstrip(buf);
5365 err = page_counter_memparse(buf, "max", &min);
5369 page_counter_set_min(&memcg->memory, min);
5374 static int memory_low_show(struct seq_file *m, void *v)
5376 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5377 unsigned long low = READ_ONCE(memcg->memory.low);
5379 if (low == PAGE_COUNTER_MAX)
5380 seq_puts(m, "max\n");
5382 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5387 static ssize_t memory_low_write(struct kernfs_open_file *of,
5388 char *buf, size_t nbytes, loff_t off)
5390 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5394 buf = strstrip(buf);
5395 err = page_counter_memparse(buf, "max", &low);
5399 page_counter_set_low(&memcg->memory, low);
5404 static int memory_high_show(struct seq_file *m, void *v)
5406 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5407 unsigned long high = READ_ONCE(memcg->high);
5409 if (high == PAGE_COUNTER_MAX)
5410 seq_puts(m, "max\n");
5412 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5417 static ssize_t memory_high_write(struct kernfs_open_file *of,
5418 char *buf, size_t nbytes, loff_t off)
5420 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5421 unsigned long nr_pages;
5425 buf = strstrip(buf);
5426 err = page_counter_memparse(buf, "max", &high);
5432 nr_pages = page_counter_read(&memcg->memory);
5433 if (nr_pages > high)
5434 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5437 memcg_wb_domain_size_changed(memcg);
5441 static int memory_max_show(struct seq_file *m, void *v)
5443 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5444 unsigned long max = READ_ONCE(memcg->memory.max);
5446 if (max == PAGE_COUNTER_MAX)
5447 seq_puts(m, "max\n");
5449 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5454 static ssize_t memory_max_write(struct kernfs_open_file *of,
5455 char *buf, size_t nbytes, loff_t off)
5457 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5458 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5459 bool drained = false;
5463 buf = strstrip(buf);
5464 err = page_counter_memparse(buf, "max", &max);
5468 xchg(&memcg->memory.max, max);
5471 unsigned long nr_pages = page_counter_read(&memcg->memory);
5473 if (nr_pages <= max)
5476 if (signal_pending(current)) {
5482 drain_all_stock(memcg);
5488 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5494 memcg_memory_event(memcg, MEMCG_OOM);
5495 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5499 memcg_wb_domain_size_changed(memcg);
5503 static int memory_events_show(struct seq_file *m, void *v)
5505 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5507 seq_printf(m, "low %lu\n",
5508 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5509 seq_printf(m, "high %lu\n",
5510 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5511 seq_printf(m, "max %lu\n",
5512 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5513 seq_printf(m, "oom %lu\n",
5514 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5515 seq_printf(m, "oom_kill %lu\n",
5516 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5521 static int memory_stat_show(struct seq_file *m, void *v)
5523 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5524 struct accumulated_stats acc;
5528 * Provide statistics on the state of the memory subsystem as
5529 * well as cumulative event counters that show past behavior.
5531 * This list is ordered following a combination of these gradients:
5532 * 1) generic big picture -> specifics and details
5533 * 2) reflecting userspace activity -> reflecting kernel heuristics
5535 * Current memory state:
5538 memset(&acc, 0, sizeof(acc));
5539 acc.stats_size = MEMCG_NR_STAT;
5540 acc.events_size = NR_VM_EVENT_ITEMS;
5541 accumulate_memcg_tree(memcg, &acc);
5543 seq_printf(m, "anon %llu\n",
5544 (u64)acc.stat[MEMCG_RSS] * PAGE_SIZE);
5545 seq_printf(m, "file %llu\n",
5546 (u64)acc.stat[MEMCG_CACHE] * PAGE_SIZE);
5547 seq_printf(m, "kernel_stack %llu\n",
5548 (u64)acc.stat[MEMCG_KERNEL_STACK_KB] * 1024);
5549 seq_printf(m, "slab %llu\n",
5550 (u64)(acc.stat[NR_SLAB_RECLAIMABLE] +
5551 acc.stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5552 seq_printf(m, "sock %llu\n",
5553 (u64)acc.stat[MEMCG_SOCK] * PAGE_SIZE);
5555 seq_printf(m, "shmem %llu\n",
5556 (u64)acc.stat[NR_SHMEM] * PAGE_SIZE);
5557 seq_printf(m, "file_mapped %llu\n",
5558 (u64)acc.stat[NR_FILE_MAPPED] * PAGE_SIZE);
5559 seq_printf(m, "file_dirty %llu\n",
5560 (u64)acc.stat[NR_FILE_DIRTY] * PAGE_SIZE);
5561 seq_printf(m, "file_writeback %llu\n",
5562 (u64)acc.stat[NR_WRITEBACK] * PAGE_SIZE);
5564 for (i = 0; i < NR_LRU_LISTS; i++)
5565 seq_printf(m, "%s %llu\n", mem_cgroup_lru_names[i],
5566 (u64)acc.lru_pages[i] * PAGE_SIZE);
5568 seq_printf(m, "slab_reclaimable %llu\n",
5569 (u64)acc.stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5570 seq_printf(m, "slab_unreclaimable %llu\n",
5571 (u64)acc.stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5573 /* Accumulated memory events */
5575 seq_printf(m, "pgfault %lu\n", acc.events[PGFAULT]);
5576 seq_printf(m, "pgmajfault %lu\n", acc.events[PGMAJFAULT]);
5578 seq_printf(m, "workingset_refault %lu\n",
5579 acc.stat[WORKINGSET_REFAULT]);
5580 seq_printf(m, "workingset_activate %lu\n",
5581 acc.stat[WORKINGSET_ACTIVATE]);
5582 seq_printf(m, "workingset_nodereclaim %lu\n",
5583 acc.stat[WORKINGSET_NODERECLAIM]);
5585 seq_printf(m, "pgrefill %lu\n", acc.events[PGREFILL]);
5586 seq_printf(m, "pgscan %lu\n", acc.events[PGSCAN_KSWAPD] +
5587 acc.events[PGSCAN_DIRECT]);
5588 seq_printf(m, "pgsteal %lu\n", acc.events[PGSTEAL_KSWAPD] +
5589 acc.events[PGSTEAL_DIRECT]);
5590 seq_printf(m, "pgactivate %lu\n", acc.events[PGACTIVATE]);
5591 seq_printf(m, "pgdeactivate %lu\n", acc.events[PGDEACTIVATE]);
5592 seq_printf(m, "pglazyfree %lu\n", acc.events[PGLAZYFREE]);
5593 seq_printf(m, "pglazyfreed %lu\n", acc.events[PGLAZYFREED]);
5598 static int memory_oom_group_show(struct seq_file *m, void *v)
5600 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5602 seq_printf(m, "%d\n", memcg->oom_group);
5607 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5608 char *buf, size_t nbytes, loff_t off)
5610 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5613 buf = strstrip(buf);
5617 ret = kstrtoint(buf, 0, &oom_group);
5621 if (oom_group != 0 && oom_group != 1)
5624 memcg->oom_group = oom_group;
5629 static struct cftype memory_files[] = {
5632 .flags = CFTYPE_NOT_ON_ROOT,
5633 .read_u64 = memory_current_read,
5637 .flags = CFTYPE_NOT_ON_ROOT,
5638 .seq_show = memory_min_show,
5639 .write = memory_min_write,
5643 .flags = CFTYPE_NOT_ON_ROOT,
5644 .seq_show = memory_low_show,
5645 .write = memory_low_write,
5649 .flags = CFTYPE_NOT_ON_ROOT,
5650 .seq_show = memory_high_show,
5651 .write = memory_high_write,
5655 .flags = CFTYPE_NOT_ON_ROOT,
5656 .seq_show = memory_max_show,
5657 .write = memory_max_write,
5661 .flags = CFTYPE_NOT_ON_ROOT,
5662 .file_offset = offsetof(struct mem_cgroup, events_file),
5663 .seq_show = memory_events_show,
5667 .flags = CFTYPE_NOT_ON_ROOT,
5668 .seq_show = memory_stat_show,
5671 .name = "oom.group",
5672 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5673 .seq_show = memory_oom_group_show,
5674 .write = memory_oom_group_write,
5679 struct cgroup_subsys memory_cgrp_subsys = {
5680 .css_alloc = mem_cgroup_css_alloc,
5681 .css_online = mem_cgroup_css_online,
5682 .css_offline = mem_cgroup_css_offline,
5683 .css_released = mem_cgroup_css_released,
5684 .css_free = mem_cgroup_css_free,
5685 .css_reset = mem_cgroup_css_reset,
5686 .can_attach = mem_cgroup_can_attach,
5687 .cancel_attach = mem_cgroup_cancel_attach,
5688 .post_attach = mem_cgroup_move_task,
5689 .bind = mem_cgroup_bind,
5690 .dfl_cftypes = memory_files,
5691 .legacy_cftypes = mem_cgroup_legacy_files,
5696 * mem_cgroup_protected - check if memory consumption is in the normal range
5697 * @root: the top ancestor of the sub-tree being checked
5698 * @memcg: the memory cgroup to check
5700 * WARNING: This function is not stateless! It can only be used as part
5701 * of a top-down tree iteration, not for isolated queries.
5703 * Returns one of the following:
5704 * MEMCG_PROT_NONE: cgroup memory is not protected
5705 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5706 * an unprotected supply of reclaimable memory from other cgroups.
5707 * MEMCG_PROT_MIN: cgroup memory is protected
5709 * @root is exclusive; it is never protected when looked at directly
5711 * To provide a proper hierarchical behavior, effective memory.min/low values
5712 * are used. Below is the description of how effective memory.low is calculated.
5713 * Effective memory.min values is calculated in the same way.
5715 * Effective memory.low is always equal or less than the original memory.low.
5716 * If there is no memory.low overcommittment (which is always true for
5717 * top-level memory cgroups), these two values are equal.
5718 * Otherwise, it's a part of parent's effective memory.low,
5719 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5720 * memory.low usages, where memory.low usage is the size of actually
5724 * elow = min( memory.low, parent->elow * ------------------ ),
5725 * siblings_low_usage
5727 * | memory.current, if memory.current < memory.low
5732 * Such definition of the effective memory.low provides the expected
5733 * hierarchical behavior: parent's memory.low value is limiting
5734 * children, unprotected memory is reclaimed first and cgroups,
5735 * which are not using their guarantee do not affect actual memory
5738 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5740 * A A/memory.low = 2G, A/memory.current = 6G
5742 * BC DE B/memory.low = 3G B/memory.current = 2G
5743 * C/memory.low = 1G C/memory.current = 2G
5744 * D/memory.low = 0 D/memory.current = 2G
5745 * E/memory.low = 10G E/memory.current = 0
5747 * and the memory pressure is applied, the following memory distribution
5748 * is expected (approximately):
5750 * A/memory.current = 2G
5752 * B/memory.current = 1.3G
5753 * C/memory.current = 0.6G
5754 * D/memory.current = 0
5755 * E/memory.current = 0
5757 * These calculations require constant tracking of the actual low usages
5758 * (see propagate_protected_usage()), as well as recursive calculation of
5759 * effective memory.low values. But as we do call mem_cgroup_protected()
5760 * path for each memory cgroup top-down from the reclaim,
5761 * it's possible to optimize this part, and save calculated elow
5762 * for next usage. This part is intentionally racy, but it's ok,
5763 * as memory.low is a best-effort mechanism.
5765 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5766 struct mem_cgroup *memcg)
5768 struct mem_cgroup *parent;
5769 unsigned long emin, parent_emin;
5770 unsigned long elow, parent_elow;
5771 unsigned long usage;
5773 if (mem_cgroup_disabled())
5774 return MEMCG_PROT_NONE;
5777 root = root_mem_cgroup;
5779 return MEMCG_PROT_NONE;
5781 usage = page_counter_read(&memcg->memory);
5783 return MEMCG_PROT_NONE;
5785 emin = memcg->memory.min;
5786 elow = memcg->memory.low;
5788 parent = parent_mem_cgroup(memcg);
5789 /* No parent means a non-hierarchical mode on v1 memcg */
5791 return MEMCG_PROT_NONE;
5796 parent_emin = READ_ONCE(parent->memory.emin);
5797 emin = min(emin, parent_emin);
5798 if (emin && parent_emin) {
5799 unsigned long min_usage, siblings_min_usage;
5801 min_usage = min(usage, memcg->memory.min);
5802 siblings_min_usage = atomic_long_read(
5803 &parent->memory.children_min_usage);
5805 if (min_usage && siblings_min_usage)
5806 emin = min(emin, parent_emin * min_usage /
5807 siblings_min_usage);
5810 parent_elow = READ_ONCE(parent->memory.elow);
5811 elow = min(elow, parent_elow);
5812 if (elow && parent_elow) {
5813 unsigned long low_usage, siblings_low_usage;
5815 low_usage = min(usage, memcg->memory.low);
5816 siblings_low_usage = atomic_long_read(
5817 &parent->memory.children_low_usage);
5819 if (low_usage && siblings_low_usage)
5820 elow = min(elow, parent_elow * low_usage /
5821 siblings_low_usage);
5825 memcg->memory.emin = emin;
5826 memcg->memory.elow = elow;
5829 return MEMCG_PROT_MIN;
5830 else if (usage <= elow)
5831 return MEMCG_PROT_LOW;
5833 return MEMCG_PROT_NONE;
5837 * mem_cgroup_try_charge - try charging a page
5838 * @page: page to charge
5839 * @mm: mm context of the victim
5840 * @gfp_mask: reclaim mode
5841 * @memcgp: charged memcg return
5842 * @compound: charge the page as compound or small page
5844 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5845 * pages according to @gfp_mask if necessary.
5847 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5848 * Otherwise, an error code is returned.
5850 * After page->mapping has been set up, the caller must finalize the
5851 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5852 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5854 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5855 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5858 struct mem_cgroup *memcg = NULL;
5859 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5862 if (mem_cgroup_disabled())
5865 if (PageSwapCache(page)) {
5867 * Every swap fault against a single page tries to charge the
5868 * page, bail as early as possible. shmem_unuse() encounters
5869 * already charged pages, too. The USED bit is protected by
5870 * the page lock, which serializes swap cache removal, which
5871 * in turn serializes uncharging.
5873 VM_BUG_ON_PAGE(!PageLocked(page), page);
5874 if (compound_head(page)->mem_cgroup)
5877 if (do_swap_account) {
5878 swp_entry_t ent = { .val = page_private(page), };
5879 unsigned short id = lookup_swap_cgroup_id(ent);
5882 memcg = mem_cgroup_from_id(id);
5883 if (memcg && !css_tryget_online(&memcg->css))
5890 memcg = get_mem_cgroup_from_mm(mm);
5892 ret = try_charge(memcg, gfp_mask, nr_pages);
5894 css_put(&memcg->css);
5900 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
5901 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5904 struct mem_cgroup *memcg;
5907 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
5909 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
5914 * mem_cgroup_commit_charge - commit a page charge
5915 * @page: page to charge
5916 * @memcg: memcg to charge the page to
5917 * @lrucare: page might be on LRU already
5918 * @compound: charge the page as compound or small page
5920 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5921 * after page->mapping has been set up. This must happen atomically
5922 * as part of the page instantiation, i.e. under the page table lock
5923 * for anonymous pages, under the page lock for page and swap cache.
5925 * In addition, the page must not be on the LRU during the commit, to
5926 * prevent racing with task migration. If it might be, use @lrucare.
5928 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5930 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5931 bool lrucare, bool compound)
5933 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5935 VM_BUG_ON_PAGE(!page->mapping, page);
5936 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5938 if (mem_cgroup_disabled())
5941 * Swap faults will attempt to charge the same page multiple
5942 * times. But reuse_swap_page() might have removed the page
5943 * from swapcache already, so we can't check PageSwapCache().
5948 commit_charge(page, memcg, lrucare);
5950 local_irq_disable();
5951 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5952 memcg_check_events(memcg, page);
5955 if (do_memsw_account() && PageSwapCache(page)) {
5956 swp_entry_t entry = { .val = page_private(page) };
5958 * The swap entry might not get freed for a long time,
5959 * let's not wait for it. The page already received a
5960 * memory+swap charge, drop the swap entry duplicate.
5962 mem_cgroup_uncharge_swap(entry, nr_pages);
5967 * mem_cgroup_cancel_charge - cancel a page charge
5968 * @page: page to charge
5969 * @memcg: memcg to charge the page to
5970 * @compound: charge the page as compound or small page
5972 * Cancel a charge transaction started by mem_cgroup_try_charge().
5974 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5977 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5979 if (mem_cgroup_disabled())
5982 * Swap faults will attempt to charge the same page multiple
5983 * times. But reuse_swap_page() might have removed the page
5984 * from swapcache already, so we can't check PageSwapCache().
5989 cancel_charge(memcg, nr_pages);
5992 struct uncharge_gather {
5993 struct mem_cgroup *memcg;
5994 unsigned long pgpgout;
5995 unsigned long nr_anon;
5996 unsigned long nr_file;
5997 unsigned long nr_kmem;
5998 unsigned long nr_huge;
5999 unsigned long nr_shmem;
6000 struct page *dummy_page;
6003 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6005 memset(ug, 0, sizeof(*ug));
6008 static void uncharge_batch(const struct uncharge_gather *ug)
6010 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6011 unsigned long flags;
6013 if (!mem_cgroup_is_root(ug->memcg)) {
6014 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6015 if (do_memsw_account())
6016 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6017 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6018 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6019 memcg_oom_recover(ug->memcg);
6022 local_irq_save(flags);
6023 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6024 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6025 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6026 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6027 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6028 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
6029 memcg_check_events(ug->memcg, ug->dummy_page);
6030 local_irq_restore(flags);
6032 if (!mem_cgroup_is_root(ug->memcg))
6033 css_put_many(&ug->memcg->css, nr_pages);
6036 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6038 VM_BUG_ON_PAGE(PageLRU(page), page);
6039 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6040 !PageHWPoison(page) , page);
6042 if (!page->mem_cgroup)
6046 * Nobody should be changing or seriously looking at
6047 * page->mem_cgroup at this point, we have fully
6048 * exclusive access to the page.
6051 if (ug->memcg != page->mem_cgroup) {
6054 uncharge_gather_clear(ug);
6056 ug->memcg = page->mem_cgroup;
6059 if (!PageKmemcg(page)) {
6060 unsigned int nr_pages = 1;
6062 if (PageTransHuge(page)) {
6063 nr_pages <<= compound_order(page);
6064 ug->nr_huge += nr_pages;
6067 ug->nr_anon += nr_pages;
6069 ug->nr_file += nr_pages;
6070 if (PageSwapBacked(page))
6071 ug->nr_shmem += nr_pages;
6075 ug->nr_kmem += 1 << compound_order(page);
6076 __ClearPageKmemcg(page);
6079 ug->dummy_page = page;
6080 page->mem_cgroup = NULL;
6083 static void uncharge_list(struct list_head *page_list)
6085 struct uncharge_gather ug;
6086 struct list_head *next;
6088 uncharge_gather_clear(&ug);
6091 * Note that the list can be a single page->lru; hence the
6092 * do-while loop instead of a simple list_for_each_entry().
6094 next = page_list->next;
6098 page = list_entry(next, struct page, lru);
6099 next = page->lru.next;
6101 uncharge_page(page, &ug);
6102 } while (next != page_list);
6105 uncharge_batch(&ug);
6109 * mem_cgroup_uncharge - uncharge a page
6110 * @page: page to uncharge
6112 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6113 * mem_cgroup_commit_charge().
6115 void mem_cgroup_uncharge(struct page *page)
6117 struct uncharge_gather ug;
6119 if (mem_cgroup_disabled())
6122 /* Don't touch page->lru of any random page, pre-check: */
6123 if (!page->mem_cgroup)
6126 uncharge_gather_clear(&ug);
6127 uncharge_page(page, &ug);
6128 uncharge_batch(&ug);
6132 * mem_cgroup_uncharge_list - uncharge a list of page
6133 * @page_list: list of pages to uncharge
6135 * Uncharge a list of pages previously charged with
6136 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6138 void mem_cgroup_uncharge_list(struct list_head *page_list)
6140 if (mem_cgroup_disabled())
6143 if (!list_empty(page_list))
6144 uncharge_list(page_list);
6148 * mem_cgroup_migrate - charge a page's replacement
6149 * @oldpage: currently circulating page
6150 * @newpage: replacement page
6152 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6153 * be uncharged upon free.
6155 * Both pages must be locked, @newpage->mapping must be set up.
6157 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6159 struct mem_cgroup *memcg;
6160 unsigned int nr_pages;
6162 unsigned long flags;
6164 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6165 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6166 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6167 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6170 if (mem_cgroup_disabled())
6173 /* Page cache replacement: new page already charged? */
6174 if (newpage->mem_cgroup)
6177 /* Swapcache readahead pages can get replaced before being charged */
6178 memcg = oldpage->mem_cgroup;
6182 /* Force-charge the new page. The old one will be freed soon */
6183 compound = PageTransHuge(newpage);
6184 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6186 page_counter_charge(&memcg->memory, nr_pages);
6187 if (do_memsw_account())
6188 page_counter_charge(&memcg->memsw, nr_pages);
6189 css_get_many(&memcg->css, nr_pages);
6191 commit_charge(newpage, memcg, false);
6193 local_irq_save(flags);
6194 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6195 memcg_check_events(memcg, newpage);
6196 local_irq_restore(flags);
6199 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6200 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6202 void mem_cgroup_sk_alloc(struct sock *sk)
6204 struct mem_cgroup *memcg;
6206 if (!mem_cgroup_sockets_enabled)
6210 * Socket cloning can throw us here with sk_memcg already
6211 * filled. It won't however, necessarily happen from
6212 * process context. So the test for root memcg given
6213 * the current task's memcg won't help us in this case.
6215 * Respecting the original socket's memcg is a better
6216 * decision in this case.
6219 css_get(&sk->sk_memcg->css);
6224 memcg = mem_cgroup_from_task(current);
6225 if (memcg == root_mem_cgroup)
6227 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6229 if (css_tryget_online(&memcg->css))
6230 sk->sk_memcg = memcg;
6235 void mem_cgroup_sk_free(struct sock *sk)
6238 css_put(&sk->sk_memcg->css);
6242 * mem_cgroup_charge_skmem - charge socket memory
6243 * @memcg: memcg to charge
6244 * @nr_pages: number of pages to charge
6246 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6247 * @memcg's configured limit, %false if the charge had to be forced.
6249 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6251 gfp_t gfp_mask = GFP_KERNEL;
6253 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6254 struct page_counter *fail;
6256 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6257 memcg->tcpmem_pressure = 0;
6260 page_counter_charge(&memcg->tcpmem, nr_pages);
6261 memcg->tcpmem_pressure = 1;
6265 /* Don't block in the packet receive path */
6267 gfp_mask = GFP_NOWAIT;
6269 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6271 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6274 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6279 * mem_cgroup_uncharge_skmem - uncharge socket memory
6280 * @memcg: memcg to uncharge
6281 * @nr_pages: number of pages to uncharge
6283 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6285 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6286 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6290 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6292 refill_stock(memcg, nr_pages);
6295 static int __init cgroup_memory(char *s)
6299 while ((token = strsep(&s, ",")) != NULL) {
6302 if (!strcmp(token, "nosocket"))
6303 cgroup_memory_nosocket = true;
6304 if (!strcmp(token, "nokmem"))
6305 cgroup_memory_nokmem = true;
6309 __setup("cgroup.memory=", cgroup_memory);
6312 * subsys_initcall() for memory controller.
6314 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6315 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6316 * basically everything that doesn't depend on a specific mem_cgroup structure
6317 * should be initialized from here.
6319 static int __init mem_cgroup_init(void)
6323 #ifdef CONFIG_MEMCG_KMEM
6325 * Kmem cache creation is mostly done with the slab_mutex held,
6326 * so use a workqueue with limited concurrency to avoid stalling
6327 * all worker threads in case lots of cgroups are created and
6328 * destroyed simultaneously.
6330 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6331 BUG_ON(!memcg_kmem_cache_wq);
6334 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6335 memcg_hotplug_cpu_dead);
6337 for_each_possible_cpu(cpu)
6338 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6341 for_each_node(node) {
6342 struct mem_cgroup_tree_per_node *rtpn;
6344 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6345 node_online(node) ? node : NUMA_NO_NODE);
6347 rtpn->rb_root = RB_ROOT;
6348 rtpn->rb_rightmost = NULL;
6349 spin_lock_init(&rtpn->lock);
6350 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6355 subsys_initcall(mem_cgroup_init);
6357 #ifdef CONFIG_MEMCG_SWAP
6358 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6360 while (!atomic_inc_not_zero(&memcg->id.ref)) {
6362 * The root cgroup cannot be destroyed, so it's refcount must
6365 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6369 memcg = parent_mem_cgroup(memcg);
6371 memcg = root_mem_cgroup;
6377 * mem_cgroup_swapout - transfer a memsw charge to swap
6378 * @page: page whose memsw charge to transfer
6379 * @entry: swap entry to move the charge to
6381 * Transfer the memsw charge of @page to @entry.
6383 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6385 struct mem_cgroup *memcg, *swap_memcg;
6386 unsigned int nr_entries;
6387 unsigned short oldid;
6389 VM_BUG_ON_PAGE(PageLRU(page), page);
6390 VM_BUG_ON_PAGE(page_count(page), page);
6392 if (!do_memsw_account())
6395 memcg = page->mem_cgroup;
6397 /* Readahead page, never charged */
6402 * In case the memcg owning these pages has been offlined and doesn't
6403 * have an ID allocated to it anymore, charge the closest online
6404 * ancestor for the swap instead and transfer the memory+swap charge.
6406 swap_memcg = mem_cgroup_id_get_online(memcg);
6407 nr_entries = hpage_nr_pages(page);
6408 /* Get references for the tail pages, too */
6410 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6411 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6413 VM_BUG_ON_PAGE(oldid, page);
6414 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6416 page->mem_cgroup = NULL;
6418 if (!mem_cgroup_is_root(memcg))
6419 page_counter_uncharge(&memcg->memory, nr_entries);
6421 if (memcg != swap_memcg) {
6422 if (!mem_cgroup_is_root(swap_memcg))
6423 page_counter_charge(&swap_memcg->memsw, nr_entries);
6424 page_counter_uncharge(&memcg->memsw, nr_entries);
6428 * Interrupts should be disabled here because the caller holds the
6429 * i_pages lock which is taken with interrupts-off. It is
6430 * important here to have the interrupts disabled because it is the
6431 * only synchronisation we have for updating the per-CPU variables.
6433 VM_BUG_ON(!irqs_disabled());
6434 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6436 memcg_check_events(memcg, page);
6438 if (!mem_cgroup_is_root(memcg))
6439 css_put_many(&memcg->css, nr_entries);
6443 * mem_cgroup_try_charge_swap - try charging swap space for a page
6444 * @page: page being added to swap
6445 * @entry: swap entry to charge
6447 * Try to charge @page's memcg for the swap space at @entry.
6449 * Returns 0 on success, -ENOMEM on failure.
6451 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6453 unsigned int nr_pages = hpage_nr_pages(page);
6454 struct page_counter *counter;
6455 struct mem_cgroup *memcg;
6456 unsigned short oldid;
6458 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6461 memcg = page->mem_cgroup;
6463 /* Readahead page, never charged */
6468 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6472 memcg = mem_cgroup_id_get_online(memcg);
6474 if (!mem_cgroup_is_root(memcg) &&
6475 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6476 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6477 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6478 mem_cgroup_id_put(memcg);
6482 /* Get references for the tail pages, too */
6484 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6485 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6486 VM_BUG_ON_PAGE(oldid, page);
6487 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6493 * mem_cgroup_uncharge_swap - uncharge swap space
6494 * @entry: swap entry to uncharge
6495 * @nr_pages: the amount of swap space to uncharge
6497 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6499 struct mem_cgroup *memcg;
6502 if (!do_swap_account)
6505 id = swap_cgroup_record(entry, 0, nr_pages);
6507 memcg = mem_cgroup_from_id(id);
6509 if (!mem_cgroup_is_root(memcg)) {
6510 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6511 page_counter_uncharge(&memcg->swap, nr_pages);
6513 page_counter_uncharge(&memcg->memsw, nr_pages);
6515 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6516 mem_cgroup_id_put_many(memcg, nr_pages);
6521 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6523 long nr_swap_pages = get_nr_swap_pages();
6525 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6526 return nr_swap_pages;
6527 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6528 nr_swap_pages = min_t(long, nr_swap_pages,
6529 READ_ONCE(memcg->swap.max) -
6530 page_counter_read(&memcg->swap));
6531 return nr_swap_pages;
6534 bool mem_cgroup_swap_full(struct page *page)
6536 struct mem_cgroup *memcg;
6538 VM_BUG_ON_PAGE(!PageLocked(page), page);
6542 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6545 memcg = page->mem_cgroup;
6549 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6550 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6556 /* for remember boot option*/
6557 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6558 static int really_do_swap_account __initdata = 1;
6560 static int really_do_swap_account __initdata;
6563 static int __init enable_swap_account(char *s)
6565 if (!strcmp(s, "1"))
6566 really_do_swap_account = 1;
6567 else if (!strcmp(s, "0"))
6568 really_do_swap_account = 0;
6571 __setup("swapaccount=", enable_swap_account);
6573 static u64 swap_current_read(struct cgroup_subsys_state *css,
6576 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6578 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6581 static int swap_max_show(struct seq_file *m, void *v)
6583 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6584 unsigned long max = READ_ONCE(memcg->swap.max);
6586 if (max == PAGE_COUNTER_MAX)
6587 seq_puts(m, "max\n");
6589 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6594 static ssize_t swap_max_write(struct kernfs_open_file *of,
6595 char *buf, size_t nbytes, loff_t off)
6597 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6601 buf = strstrip(buf);
6602 err = page_counter_memparse(buf, "max", &max);
6606 xchg(&memcg->swap.max, max);
6611 static int swap_events_show(struct seq_file *m, void *v)
6613 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6615 seq_printf(m, "max %lu\n",
6616 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6617 seq_printf(m, "fail %lu\n",
6618 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6623 static struct cftype swap_files[] = {
6625 .name = "swap.current",
6626 .flags = CFTYPE_NOT_ON_ROOT,
6627 .read_u64 = swap_current_read,
6631 .flags = CFTYPE_NOT_ON_ROOT,
6632 .seq_show = swap_max_show,
6633 .write = swap_max_write,
6636 .name = "swap.events",
6637 .flags = CFTYPE_NOT_ON_ROOT,
6638 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6639 .seq_show = swap_events_show,
6644 static struct cftype memsw_cgroup_files[] = {
6646 .name = "memsw.usage_in_bytes",
6647 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6648 .read_u64 = mem_cgroup_read_u64,
6651 .name = "memsw.max_usage_in_bytes",
6652 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6653 .write = mem_cgroup_reset,
6654 .read_u64 = mem_cgroup_read_u64,
6657 .name = "memsw.limit_in_bytes",
6658 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6659 .write = mem_cgroup_write,
6660 .read_u64 = mem_cgroup_read_u64,
6663 .name = "memsw.failcnt",
6664 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6665 .write = mem_cgroup_reset,
6666 .read_u64 = mem_cgroup_read_u64,
6668 { }, /* terminate */
6671 static int __init mem_cgroup_swap_init(void)
6673 if (!mem_cgroup_disabled() && really_do_swap_account) {
6674 do_swap_account = 1;
6675 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6677 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6678 memsw_cgroup_files));
6682 subsys_initcall(mem_cgroup_swap_init);
6684 #endif /* CONFIG_MEMCG_SWAP */