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_context: Print OOM information relevant to
1297 * memory controller.
1298 * @memcg: The memory cgroup that went over limit
1299 * @p: Task that is going to be killed
1301 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1304 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1309 pr_cont(",oom_memcg=");
1310 pr_cont_cgroup_path(memcg->css.cgroup);
1312 pr_cont(",global_oom");
1314 pr_cont(",task_memcg=");
1315 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1321 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1322 * memory controller.
1323 * @memcg: The memory cgroup that went over limit
1325 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1327 struct mem_cgroup *iter;
1330 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1331 K((u64)page_counter_read(&memcg->memory)),
1332 K((u64)memcg->memory.max), memcg->memory.failcnt);
1333 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1334 K((u64)page_counter_read(&memcg->memsw)),
1335 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1336 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1337 K((u64)page_counter_read(&memcg->kmem)),
1338 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1340 for_each_mem_cgroup_tree(iter, memcg) {
1341 pr_info("Memory cgroup stats for ");
1342 pr_cont_cgroup_path(iter->css.cgroup);
1345 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1346 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1348 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1349 K(memcg_page_state(iter, memcg1_stats[i])));
1352 for (i = 0; i < NR_LRU_LISTS; i++)
1353 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1354 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1361 * Return the memory (and swap, if configured) limit for a memcg.
1363 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1367 max = memcg->memory.max;
1368 if (mem_cgroup_swappiness(memcg)) {
1369 unsigned long memsw_max;
1370 unsigned long swap_max;
1372 memsw_max = memcg->memsw.max;
1373 swap_max = memcg->swap.max;
1374 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1375 max = min(max + swap_max, memsw_max);
1380 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1383 struct oom_control oc = {
1387 .gfp_mask = gfp_mask,
1392 mutex_lock(&oom_lock);
1393 ret = out_of_memory(&oc);
1394 mutex_unlock(&oom_lock);
1398 #if MAX_NUMNODES > 1
1401 * test_mem_cgroup_node_reclaimable
1402 * @memcg: the target memcg
1403 * @nid: the node ID to be checked.
1404 * @noswap : specify true here if the user wants flle only information.
1406 * This function returns whether the specified memcg contains any
1407 * reclaimable pages on a node. Returns true if there are any reclaimable
1408 * pages in the node.
1410 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1411 int nid, bool noswap)
1413 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1415 if (noswap || !total_swap_pages)
1417 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1424 * Always updating the nodemask is not very good - even if we have an empty
1425 * list or the wrong list here, we can start from some node and traverse all
1426 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1429 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1433 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1434 * pagein/pageout changes since the last update.
1436 if (!atomic_read(&memcg->numainfo_events))
1438 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1441 /* make a nodemask where this memcg uses memory from */
1442 memcg->scan_nodes = node_states[N_MEMORY];
1444 for_each_node_mask(nid, node_states[N_MEMORY]) {
1446 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1447 node_clear(nid, memcg->scan_nodes);
1450 atomic_set(&memcg->numainfo_events, 0);
1451 atomic_set(&memcg->numainfo_updating, 0);
1455 * Selecting a node where we start reclaim from. Because what we need is just
1456 * reducing usage counter, start from anywhere is O,K. Considering
1457 * memory reclaim from current node, there are pros. and cons.
1459 * Freeing memory from current node means freeing memory from a node which
1460 * we'll use or we've used. So, it may make LRU bad. And if several threads
1461 * hit limits, it will see a contention on a node. But freeing from remote
1462 * node means more costs for memory reclaim because of memory latency.
1464 * Now, we use round-robin. Better algorithm is welcomed.
1466 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1470 mem_cgroup_may_update_nodemask(memcg);
1471 node = memcg->last_scanned_node;
1473 node = next_node_in(node, memcg->scan_nodes);
1475 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1476 * last time it really checked all the LRUs due to rate limiting.
1477 * Fallback to the current node in that case for simplicity.
1479 if (unlikely(node == MAX_NUMNODES))
1480 node = numa_node_id();
1482 memcg->last_scanned_node = node;
1486 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1492 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1495 unsigned long *total_scanned)
1497 struct mem_cgroup *victim = NULL;
1500 unsigned long excess;
1501 unsigned long nr_scanned;
1502 struct mem_cgroup_reclaim_cookie reclaim = {
1507 excess = soft_limit_excess(root_memcg);
1510 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1515 * If we have not been able to reclaim
1516 * anything, it might because there are
1517 * no reclaimable pages under this hierarchy
1522 * We want to do more targeted reclaim.
1523 * excess >> 2 is not to excessive so as to
1524 * reclaim too much, nor too less that we keep
1525 * coming back to reclaim from this cgroup
1527 if (total >= (excess >> 2) ||
1528 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1533 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1534 pgdat, &nr_scanned);
1535 *total_scanned += nr_scanned;
1536 if (!soft_limit_excess(root_memcg))
1539 mem_cgroup_iter_break(root_memcg, victim);
1543 #ifdef CONFIG_LOCKDEP
1544 static struct lockdep_map memcg_oom_lock_dep_map = {
1545 .name = "memcg_oom_lock",
1549 static DEFINE_SPINLOCK(memcg_oom_lock);
1552 * Check OOM-Killer is already running under our hierarchy.
1553 * If someone is running, return false.
1555 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1557 struct mem_cgroup *iter, *failed = NULL;
1559 spin_lock(&memcg_oom_lock);
1561 for_each_mem_cgroup_tree(iter, memcg) {
1562 if (iter->oom_lock) {
1564 * this subtree of our hierarchy is already locked
1565 * so we cannot give a lock.
1568 mem_cgroup_iter_break(memcg, iter);
1571 iter->oom_lock = true;
1576 * OK, we failed to lock the whole subtree so we have
1577 * to clean up what we set up to the failing subtree
1579 for_each_mem_cgroup_tree(iter, memcg) {
1580 if (iter == failed) {
1581 mem_cgroup_iter_break(memcg, iter);
1584 iter->oom_lock = false;
1587 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1589 spin_unlock(&memcg_oom_lock);
1594 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1596 struct mem_cgroup *iter;
1598 spin_lock(&memcg_oom_lock);
1599 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1600 for_each_mem_cgroup_tree(iter, memcg)
1601 iter->oom_lock = false;
1602 spin_unlock(&memcg_oom_lock);
1605 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1607 struct mem_cgroup *iter;
1609 spin_lock(&memcg_oom_lock);
1610 for_each_mem_cgroup_tree(iter, memcg)
1612 spin_unlock(&memcg_oom_lock);
1615 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1617 struct mem_cgroup *iter;
1620 * When a new child is created while the hierarchy is under oom,
1621 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1623 spin_lock(&memcg_oom_lock);
1624 for_each_mem_cgroup_tree(iter, memcg)
1625 if (iter->under_oom > 0)
1627 spin_unlock(&memcg_oom_lock);
1630 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1632 struct oom_wait_info {
1633 struct mem_cgroup *memcg;
1634 wait_queue_entry_t wait;
1637 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1638 unsigned mode, int sync, void *arg)
1640 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1641 struct mem_cgroup *oom_wait_memcg;
1642 struct oom_wait_info *oom_wait_info;
1644 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1645 oom_wait_memcg = oom_wait_info->memcg;
1647 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1648 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1650 return autoremove_wake_function(wait, mode, sync, arg);
1653 static void memcg_oom_recover(struct mem_cgroup *memcg)
1656 * For the following lockless ->under_oom test, the only required
1657 * guarantee is that it must see the state asserted by an OOM when
1658 * this function is called as a result of userland actions
1659 * triggered by the notification of the OOM. This is trivially
1660 * achieved by invoking mem_cgroup_mark_under_oom() before
1661 * triggering notification.
1663 if (memcg && memcg->under_oom)
1664 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1674 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1676 if (order > PAGE_ALLOC_COSTLY_ORDER)
1679 memcg_memory_event(memcg, MEMCG_OOM);
1682 * We are in the middle of the charge context here, so we
1683 * don't want to block when potentially sitting on a callstack
1684 * that holds all kinds of filesystem and mm locks.
1686 * cgroup1 allows disabling the OOM killer and waiting for outside
1687 * handling until the charge can succeed; remember the context and put
1688 * the task to sleep at the end of the page fault when all locks are
1691 * On the other hand, in-kernel OOM killer allows for an async victim
1692 * memory reclaim (oom_reaper) and that means that we are not solely
1693 * relying on the oom victim to make a forward progress and we can
1694 * invoke the oom killer here.
1696 * Please note that mem_cgroup_out_of_memory might fail to find a
1697 * victim and then we have to bail out from the charge path.
1699 if (memcg->oom_kill_disable) {
1700 if (!current->in_user_fault)
1702 css_get(&memcg->css);
1703 current->memcg_in_oom = memcg;
1704 current->memcg_oom_gfp_mask = mask;
1705 current->memcg_oom_order = order;
1710 if (mem_cgroup_out_of_memory(memcg, mask, order))
1717 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1718 * @handle: actually kill/wait or just clean up the OOM state
1720 * This has to be called at the end of a page fault if the memcg OOM
1721 * handler was enabled.
1723 * Memcg supports userspace OOM handling where failed allocations must
1724 * sleep on a waitqueue until the userspace task resolves the
1725 * situation. Sleeping directly in the charge context with all kinds
1726 * of locks held is not a good idea, instead we remember an OOM state
1727 * in the task and mem_cgroup_oom_synchronize() has to be called at
1728 * the end of the page fault to complete the OOM handling.
1730 * Returns %true if an ongoing memcg OOM situation was detected and
1731 * completed, %false otherwise.
1733 bool mem_cgroup_oom_synchronize(bool handle)
1735 struct mem_cgroup *memcg = current->memcg_in_oom;
1736 struct oom_wait_info owait;
1739 /* OOM is global, do not handle */
1746 owait.memcg = memcg;
1747 owait.wait.flags = 0;
1748 owait.wait.func = memcg_oom_wake_function;
1749 owait.wait.private = current;
1750 INIT_LIST_HEAD(&owait.wait.entry);
1752 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1753 mem_cgroup_mark_under_oom(memcg);
1755 locked = mem_cgroup_oom_trylock(memcg);
1758 mem_cgroup_oom_notify(memcg);
1760 if (locked && !memcg->oom_kill_disable) {
1761 mem_cgroup_unmark_under_oom(memcg);
1762 finish_wait(&memcg_oom_waitq, &owait.wait);
1763 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1764 current->memcg_oom_order);
1767 mem_cgroup_unmark_under_oom(memcg);
1768 finish_wait(&memcg_oom_waitq, &owait.wait);
1772 mem_cgroup_oom_unlock(memcg);
1774 * There is no guarantee that an OOM-lock contender
1775 * sees the wakeups triggered by the OOM kill
1776 * uncharges. Wake any sleepers explicitely.
1778 memcg_oom_recover(memcg);
1781 current->memcg_in_oom = NULL;
1782 css_put(&memcg->css);
1787 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1788 * @victim: task to be killed by the OOM killer
1789 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1791 * Returns a pointer to a memory cgroup, which has to be cleaned up
1792 * by killing all belonging OOM-killable tasks.
1794 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1796 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1797 struct mem_cgroup *oom_domain)
1799 struct mem_cgroup *oom_group = NULL;
1800 struct mem_cgroup *memcg;
1802 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1806 oom_domain = root_mem_cgroup;
1810 memcg = mem_cgroup_from_task(victim);
1811 if (memcg == root_mem_cgroup)
1815 * Traverse the memory cgroup hierarchy from the victim task's
1816 * cgroup up to the OOMing cgroup (or root) to find the
1817 * highest-level memory cgroup with oom.group set.
1819 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1820 if (memcg->oom_group)
1823 if (memcg == oom_domain)
1828 css_get(&oom_group->css);
1835 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1837 pr_info("Tasks in ");
1838 pr_cont_cgroup_path(memcg->css.cgroup);
1839 pr_cont(" are going to be killed due to memory.oom.group set\n");
1843 * lock_page_memcg - lock a page->mem_cgroup binding
1846 * This function protects unlocked LRU pages from being moved to
1849 * It ensures lifetime of the returned memcg. Caller is responsible
1850 * for the lifetime of the page; __unlock_page_memcg() is available
1851 * when @page might get freed inside the locked section.
1853 struct mem_cgroup *lock_page_memcg(struct page *page)
1855 struct mem_cgroup *memcg;
1856 unsigned long flags;
1859 * The RCU lock is held throughout the transaction. The fast
1860 * path can get away without acquiring the memcg->move_lock
1861 * because page moving starts with an RCU grace period.
1863 * The RCU lock also protects the memcg from being freed when
1864 * the page state that is going to change is the only thing
1865 * preventing the page itself from being freed. E.g. writeback
1866 * doesn't hold a page reference and relies on PG_writeback to
1867 * keep off truncation, migration and so forth.
1871 if (mem_cgroup_disabled())
1874 memcg = page->mem_cgroup;
1875 if (unlikely(!memcg))
1878 if (atomic_read(&memcg->moving_account) <= 0)
1881 spin_lock_irqsave(&memcg->move_lock, flags);
1882 if (memcg != page->mem_cgroup) {
1883 spin_unlock_irqrestore(&memcg->move_lock, flags);
1888 * When charge migration first begins, we can have locked and
1889 * unlocked page stat updates happening concurrently. Track
1890 * the task who has the lock for unlock_page_memcg().
1892 memcg->move_lock_task = current;
1893 memcg->move_lock_flags = flags;
1897 EXPORT_SYMBOL(lock_page_memcg);
1900 * __unlock_page_memcg - unlock and unpin a memcg
1903 * Unlock and unpin a memcg returned by lock_page_memcg().
1905 void __unlock_page_memcg(struct mem_cgroup *memcg)
1907 if (memcg && memcg->move_lock_task == current) {
1908 unsigned long flags = memcg->move_lock_flags;
1910 memcg->move_lock_task = NULL;
1911 memcg->move_lock_flags = 0;
1913 spin_unlock_irqrestore(&memcg->move_lock, flags);
1920 * unlock_page_memcg - unlock a page->mem_cgroup binding
1923 void unlock_page_memcg(struct page *page)
1925 __unlock_page_memcg(page->mem_cgroup);
1927 EXPORT_SYMBOL(unlock_page_memcg);
1929 struct memcg_stock_pcp {
1930 struct mem_cgroup *cached; /* this never be root cgroup */
1931 unsigned int nr_pages;
1932 struct work_struct work;
1933 unsigned long flags;
1934 #define FLUSHING_CACHED_CHARGE 0
1936 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1937 static DEFINE_MUTEX(percpu_charge_mutex);
1940 * consume_stock: Try to consume stocked charge on this cpu.
1941 * @memcg: memcg to consume from.
1942 * @nr_pages: how many pages to charge.
1944 * The charges will only happen if @memcg matches the current cpu's memcg
1945 * stock, and at least @nr_pages are available in that stock. Failure to
1946 * service an allocation will refill the stock.
1948 * returns true if successful, false otherwise.
1950 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1952 struct memcg_stock_pcp *stock;
1953 unsigned long flags;
1956 if (nr_pages > MEMCG_CHARGE_BATCH)
1959 local_irq_save(flags);
1961 stock = this_cpu_ptr(&memcg_stock);
1962 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1963 stock->nr_pages -= nr_pages;
1967 local_irq_restore(flags);
1973 * Returns stocks cached in percpu and reset cached information.
1975 static void drain_stock(struct memcg_stock_pcp *stock)
1977 struct mem_cgroup *old = stock->cached;
1979 if (stock->nr_pages) {
1980 page_counter_uncharge(&old->memory, stock->nr_pages);
1981 if (do_memsw_account())
1982 page_counter_uncharge(&old->memsw, stock->nr_pages);
1983 css_put_many(&old->css, stock->nr_pages);
1984 stock->nr_pages = 0;
1986 stock->cached = NULL;
1989 static void drain_local_stock(struct work_struct *dummy)
1991 struct memcg_stock_pcp *stock;
1992 unsigned long flags;
1995 * The only protection from memory hotplug vs. drain_stock races is
1996 * that we always operate on local CPU stock here with IRQ disabled
1998 local_irq_save(flags);
2000 stock = this_cpu_ptr(&memcg_stock);
2002 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2004 local_irq_restore(flags);
2008 * Cache charges(val) to local per_cpu area.
2009 * This will be consumed by consume_stock() function, later.
2011 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2013 struct memcg_stock_pcp *stock;
2014 unsigned long flags;
2016 local_irq_save(flags);
2018 stock = this_cpu_ptr(&memcg_stock);
2019 if (stock->cached != memcg) { /* reset if necessary */
2021 stock->cached = memcg;
2023 stock->nr_pages += nr_pages;
2025 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2028 local_irq_restore(flags);
2032 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2033 * of the hierarchy under it.
2035 static void drain_all_stock(struct mem_cgroup *root_memcg)
2039 /* If someone's already draining, avoid adding running more workers. */
2040 if (!mutex_trylock(&percpu_charge_mutex))
2043 * Notify other cpus that system-wide "drain" is running
2044 * We do not care about races with the cpu hotplug because cpu down
2045 * as well as workers from this path always operate on the local
2046 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2049 for_each_online_cpu(cpu) {
2050 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2051 struct mem_cgroup *memcg;
2053 memcg = stock->cached;
2054 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2056 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2057 css_put(&memcg->css);
2060 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2062 drain_local_stock(&stock->work);
2064 schedule_work_on(cpu, &stock->work);
2066 css_put(&memcg->css);
2069 mutex_unlock(&percpu_charge_mutex);
2072 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2074 struct memcg_stock_pcp *stock;
2075 struct mem_cgroup *memcg;
2077 stock = &per_cpu(memcg_stock, cpu);
2080 for_each_mem_cgroup(memcg) {
2083 for (i = 0; i < MEMCG_NR_STAT; i++) {
2087 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
2089 atomic_long_add(x, &memcg->stat[i]);
2091 if (i >= NR_VM_NODE_STAT_ITEMS)
2094 for_each_node(nid) {
2095 struct mem_cgroup_per_node *pn;
2097 pn = mem_cgroup_nodeinfo(memcg, nid);
2098 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2100 atomic_long_add(x, &pn->lruvec_stat[i]);
2104 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2107 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
2109 atomic_long_add(x, &memcg->events[i]);
2116 static void reclaim_high(struct mem_cgroup *memcg,
2117 unsigned int nr_pages,
2121 if (page_counter_read(&memcg->memory) <= memcg->high)
2123 memcg_memory_event(memcg, MEMCG_HIGH);
2124 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2125 } while ((memcg = parent_mem_cgroup(memcg)));
2128 static void high_work_func(struct work_struct *work)
2130 struct mem_cgroup *memcg;
2132 memcg = container_of(work, struct mem_cgroup, high_work);
2133 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2137 * Scheduled by try_charge() to be executed from the userland return path
2138 * and reclaims memory over the high limit.
2140 void mem_cgroup_handle_over_high(void)
2142 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2143 struct mem_cgroup *memcg;
2145 if (likely(!nr_pages))
2148 memcg = get_mem_cgroup_from_mm(current->mm);
2149 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2150 css_put(&memcg->css);
2151 current->memcg_nr_pages_over_high = 0;
2154 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2155 unsigned int nr_pages)
2157 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2158 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2159 struct mem_cgroup *mem_over_limit;
2160 struct page_counter *counter;
2161 unsigned long nr_reclaimed;
2162 bool may_swap = true;
2163 bool drained = false;
2165 enum oom_status oom_status;
2167 if (mem_cgroup_is_root(memcg))
2170 if (consume_stock(memcg, nr_pages))
2173 if (!do_memsw_account() ||
2174 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2175 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2177 if (do_memsw_account())
2178 page_counter_uncharge(&memcg->memsw, batch);
2179 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2181 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2185 if (batch > nr_pages) {
2191 * Unlike in global OOM situations, memcg is not in a physical
2192 * memory shortage. Allow dying and OOM-killed tasks to
2193 * bypass the last charges so that they can exit quickly and
2194 * free their memory.
2196 if (unlikely(tsk_is_oom_victim(current) ||
2197 fatal_signal_pending(current) ||
2198 current->flags & PF_EXITING))
2202 * Prevent unbounded recursion when reclaim operations need to
2203 * allocate memory. This might exceed the limits temporarily,
2204 * but we prefer facilitating memory reclaim and getting back
2205 * under the limit over triggering OOM kills in these cases.
2207 if (unlikely(current->flags & PF_MEMALLOC))
2210 if (unlikely(task_in_memcg_oom(current)))
2213 if (!gfpflags_allow_blocking(gfp_mask))
2216 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2218 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2219 gfp_mask, may_swap);
2221 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2225 drain_all_stock(mem_over_limit);
2230 if (gfp_mask & __GFP_NORETRY)
2233 * Even though the limit is exceeded at this point, reclaim
2234 * may have been able to free some pages. Retry the charge
2235 * before killing the task.
2237 * Only for regular pages, though: huge pages are rather
2238 * unlikely to succeed so close to the limit, and we fall back
2239 * to regular pages anyway in case of failure.
2241 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2244 * At task move, charge accounts can be doubly counted. So, it's
2245 * better to wait until the end of task_move if something is going on.
2247 if (mem_cgroup_wait_acct_move(mem_over_limit))
2253 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2256 if (gfp_mask & __GFP_NOFAIL)
2259 if (fatal_signal_pending(current))
2263 * keep retrying as long as the memcg oom killer is able to make
2264 * a forward progress or bypass the charge if the oom killer
2265 * couldn't make any progress.
2267 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2268 get_order(nr_pages * PAGE_SIZE));
2269 switch (oom_status) {
2271 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2280 if (!(gfp_mask & __GFP_NOFAIL))
2284 * The allocation either can't fail or will lead to more memory
2285 * being freed very soon. Allow memory usage go over the limit
2286 * temporarily by force charging it.
2288 page_counter_charge(&memcg->memory, nr_pages);
2289 if (do_memsw_account())
2290 page_counter_charge(&memcg->memsw, nr_pages);
2291 css_get_many(&memcg->css, nr_pages);
2296 css_get_many(&memcg->css, batch);
2297 if (batch > nr_pages)
2298 refill_stock(memcg, batch - nr_pages);
2301 * If the hierarchy is above the normal consumption range, schedule
2302 * reclaim on returning to userland. We can perform reclaim here
2303 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2304 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2305 * not recorded as it most likely matches current's and won't
2306 * change in the meantime. As high limit is checked again before
2307 * reclaim, the cost of mismatch is negligible.
2310 if (page_counter_read(&memcg->memory) > memcg->high) {
2311 /* Don't bother a random interrupted task */
2312 if (in_interrupt()) {
2313 schedule_work(&memcg->high_work);
2316 current->memcg_nr_pages_over_high += batch;
2317 set_notify_resume(current);
2320 } while ((memcg = parent_mem_cgroup(memcg)));
2325 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2327 if (mem_cgroup_is_root(memcg))
2330 page_counter_uncharge(&memcg->memory, nr_pages);
2331 if (do_memsw_account())
2332 page_counter_uncharge(&memcg->memsw, nr_pages);
2334 css_put_many(&memcg->css, nr_pages);
2337 static void lock_page_lru(struct page *page, int *isolated)
2339 struct zone *zone = page_zone(page);
2341 spin_lock_irq(zone_lru_lock(zone));
2342 if (PageLRU(page)) {
2343 struct lruvec *lruvec;
2345 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2347 del_page_from_lru_list(page, lruvec, page_lru(page));
2353 static void unlock_page_lru(struct page *page, int isolated)
2355 struct zone *zone = page_zone(page);
2358 struct lruvec *lruvec;
2360 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2361 VM_BUG_ON_PAGE(PageLRU(page), page);
2363 add_page_to_lru_list(page, lruvec, page_lru(page));
2365 spin_unlock_irq(zone_lru_lock(zone));
2368 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2373 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2376 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2377 * may already be on some other mem_cgroup's LRU. Take care of it.
2380 lock_page_lru(page, &isolated);
2383 * Nobody should be changing or seriously looking at
2384 * page->mem_cgroup at this point:
2386 * - the page is uncharged
2388 * - the page is off-LRU
2390 * - an anonymous fault has exclusive page access, except for
2391 * a locked page table
2393 * - a page cache insertion, a swapin fault, or a migration
2394 * have the page locked
2396 page->mem_cgroup = memcg;
2399 unlock_page_lru(page, isolated);
2402 #ifdef CONFIG_MEMCG_KMEM
2403 static int memcg_alloc_cache_id(void)
2408 id = ida_simple_get(&memcg_cache_ida,
2409 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2413 if (id < memcg_nr_cache_ids)
2417 * There's no space for the new id in memcg_caches arrays,
2418 * so we have to grow them.
2420 down_write(&memcg_cache_ids_sem);
2422 size = 2 * (id + 1);
2423 if (size < MEMCG_CACHES_MIN_SIZE)
2424 size = MEMCG_CACHES_MIN_SIZE;
2425 else if (size > MEMCG_CACHES_MAX_SIZE)
2426 size = MEMCG_CACHES_MAX_SIZE;
2428 err = memcg_update_all_caches(size);
2430 err = memcg_update_all_list_lrus(size);
2432 memcg_nr_cache_ids = size;
2434 up_write(&memcg_cache_ids_sem);
2437 ida_simple_remove(&memcg_cache_ida, id);
2443 static void memcg_free_cache_id(int id)
2445 ida_simple_remove(&memcg_cache_ida, id);
2448 struct memcg_kmem_cache_create_work {
2449 struct mem_cgroup *memcg;
2450 struct kmem_cache *cachep;
2451 struct work_struct work;
2454 static void memcg_kmem_cache_create_func(struct work_struct *w)
2456 struct memcg_kmem_cache_create_work *cw =
2457 container_of(w, struct memcg_kmem_cache_create_work, work);
2458 struct mem_cgroup *memcg = cw->memcg;
2459 struct kmem_cache *cachep = cw->cachep;
2461 memcg_create_kmem_cache(memcg, cachep);
2463 css_put(&memcg->css);
2468 * Enqueue the creation of a per-memcg kmem_cache.
2470 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2471 struct kmem_cache *cachep)
2473 struct memcg_kmem_cache_create_work *cw;
2475 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2479 css_get(&memcg->css);
2482 cw->cachep = cachep;
2483 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2485 queue_work(memcg_kmem_cache_wq, &cw->work);
2488 static inline bool memcg_kmem_bypass(void)
2490 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2496 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2497 * @cachep: the original global kmem cache
2499 * Return the kmem_cache we're supposed to use for a slab allocation.
2500 * We try to use the current memcg's version of the cache.
2502 * If the cache does not exist yet, if we are the first user of it, we
2503 * create it asynchronously in a workqueue and let the current allocation
2504 * go through with the original cache.
2506 * This function takes a reference to the cache it returns to assure it
2507 * won't get destroyed while we are working with it. Once the caller is
2508 * done with it, memcg_kmem_put_cache() must be called to release the
2511 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2513 struct mem_cgroup *memcg;
2514 struct kmem_cache *memcg_cachep;
2517 VM_BUG_ON(!is_root_cache(cachep));
2519 if (memcg_kmem_bypass())
2522 memcg = get_mem_cgroup_from_current();
2523 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2527 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2528 if (likely(memcg_cachep))
2529 return memcg_cachep;
2532 * If we are in a safe context (can wait, and not in interrupt
2533 * context), we could be be predictable and return right away.
2534 * This would guarantee that the allocation being performed
2535 * already belongs in the new cache.
2537 * However, there are some clashes that can arrive from locking.
2538 * For instance, because we acquire the slab_mutex while doing
2539 * memcg_create_kmem_cache, this means no further allocation
2540 * could happen with the slab_mutex held. So it's better to
2543 memcg_schedule_kmem_cache_create(memcg, cachep);
2545 css_put(&memcg->css);
2550 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2551 * @cachep: the cache returned by memcg_kmem_get_cache
2553 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2555 if (!is_root_cache(cachep))
2556 css_put(&cachep->memcg_params.memcg->css);
2560 * memcg_kmem_charge_memcg: charge a kmem page
2561 * @page: page to charge
2562 * @gfp: reclaim mode
2563 * @order: allocation order
2564 * @memcg: memory cgroup to charge
2566 * Returns 0 on success, an error code on failure.
2568 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2569 struct mem_cgroup *memcg)
2571 unsigned int nr_pages = 1 << order;
2572 struct page_counter *counter;
2575 ret = try_charge(memcg, gfp, nr_pages);
2579 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2580 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2581 cancel_charge(memcg, nr_pages);
2585 page->mem_cgroup = memcg;
2591 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2592 * @page: page to charge
2593 * @gfp: reclaim mode
2594 * @order: allocation order
2596 * Returns 0 on success, an error code on failure.
2598 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2600 struct mem_cgroup *memcg;
2603 if (mem_cgroup_disabled() || memcg_kmem_bypass())
2606 memcg = get_mem_cgroup_from_current();
2607 if (!mem_cgroup_is_root(memcg)) {
2608 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2610 __SetPageKmemcg(page);
2612 css_put(&memcg->css);
2616 * memcg_kmem_uncharge: uncharge a kmem page
2617 * @page: page to uncharge
2618 * @order: allocation order
2620 void memcg_kmem_uncharge(struct page *page, int order)
2622 struct mem_cgroup *memcg = page->mem_cgroup;
2623 unsigned int nr_pages = 1 << order;
2628 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2630 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2631 page_counter_uncharge(&memcg->kmem, nr_pages);
2633 page_counter_uncharge(&memcg->memory, nr_pages);
2634 if (do_memsw_account())
2635 page_counter_uncharge(&memcg->memsw, nr_pages);
2637 page->mem_cgroup = NULL;
2639 /* slab pages do not have PageKmemcg flag set */
2640 if (PageKmemcg(page))
2641 __ClearPageKmemcg(page);
2643 css_put_many(&memcg->css, nr_pages);
2645 #endif /* CONFIG_MEMCG_KMEM */
2647 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2650 * Because tail pages are not marked as "used", set it. We're under
2651 * zone_lru_lock and migration entries setup in all page mappings.
2653 void mem_cgroup_split_huge_fixup(struct page *head)
2657 if (mem_cgroup_disabled())
2660 for (i = 1; i < HPAGE_PMD_NR; i++)
2661 head[i].mem_cgroup = head->mem_cgroup;
2663 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2665 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2667 #ifdef CONFIG_MEMCG_SWAP
2669 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2670 * @entry: swap entry to be moved
2671 * @from: mem_cgroup which the entry is moved from
2672 * @to: mem_cgroup which the entry is moved to
2674 * It succeeds only when the swap_cgroup's record for this entry is the same
2675 * as the mem_cgroup's id of @from.
2677 * Returns 0 on success, -EINVAL on failure.
2679 * The caller must have charged to @to, IOW, called page_counter_charge() about
2680 * both res and memsw, and called css_get().
2682 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2683 struct mem_cgroup *from, struct mem_cgroup *to)
2685 unsigned short old_id, new_id;
2687 old_id = mem_cgroup_id(from);
2688 new_id = mem_cgroup_id(to);
2690 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2691 mod_memcg_state(from, MEMCG_SWAP, -1);
2692 mod_memcg_state(to, MEMCG_SWAP, 1);
2698 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2699 struct mem_cgroup *from, struct mem_cgroup *to)
2705 static DEFINE_MUTEX(memcg_max_mutex);
2707 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2708 unsigned long max, bool memsw)
2710 bool enlarge = false;
2711 bool drained = false;
2713 bool limits_invariant;
2714 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2717 if (signal_pending(current)) {
2722 mutex_lock(&memcg_max_mutex);
2724 * Make sure that the new limit (memsw or memory limit) doesn't
2725 * break our basic invariant rule memory.max <= memsw.max.
2727 limits_invariant = memsw ? max >= memcg->memory.max :
2728 max <= memcg->memsw.max;
2729 if (!limits_invariant) {
2730 mutex_unlock(&memcg_max_mutex);
2734 if (max > counter->max)
2736 ret = page_counter_set_max(counter, max);
2737 mutex_unlock(&memcg_max_mutex);
2743 drain_all_stock(memcg);
2748 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2749 GFP_KERNEL, !memsw)) {
2755 if (!ret && enlarge)
2756 memcg_oom_recover(memcg);
2761 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2763 unsigned long *total_scanned)
2765 unsigned long nr_reclaimed = 0;
2766 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2767 unsigned long reclaimed;
2769 struct mem_cgroup_tree_per_node *mctz;
2770 unsigned long excess;
2771 unsigned long nr_scanned;
2776 mctz = soft_limit_tree_node(pgdat->node_id);
2779 * Do not even bother to check the largest node if the root
2780 * is empty. Do it lockless to prevent lock bouncing. Races
2781 * are acceptable as soft limit is best effort anyway.
2783 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2787 * This loop can run a while, specially if mem_cgroup's continuously
2788 * keep exceeding their soft limit and putting the system under
2795 mz = mem_cgroup_largest_soft_limit_node(mctz);
2800 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2801 gfp_mask, &nr_scanned);
2802 nr_reclaimed += reclaimed;
2803 *total_scanned += nr_scanned;
2804 spin_lock_irq(&mctz->lock);
2805 __mem_cgroup_remove_exceeded(mz, mctz);
2808 * If we failed to reclaim anything from this memory cgroup
2809 * it is time to move on to the next cgroup
2813 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2815 excess = soft_limit_excess(mz->memcg);
2817 * One school of thought says that we should not add
2818 * back the node to the tree if reclaim returns 0.
2819 * But our reclaim could return 0, simply because due
2820 * to priority we are exposing a smaller subset of
2821 * memory to reclaim from. Consider this as a longer
2824 /* If excess == 0, no tree ops */
2825 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2826 spin_unlock_irq(&mctz->lock);
2827 css_put(&mz->memcg->css);
2830 * Could not reclaim anything and there are no more
2831 * mem cgroups to try or we seem to be looping without
2832 * reclaiming anything.
2834 if (!nr_reclaimed &&
2836 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2838 } while (!nr_reclaimed);
2840 css_put(&next_mz->memcg->css);
2841 return nr_reclaimed;
2845 * Test whether @memcg has children, dead or alive. Note that this
2846 * function doesn't care whether @memcg has use_hierarchy enabled and
2847 * returns %true if there are child csses according to the cgroup
2848 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2850 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2855 ret = css_next_child(NULL, &memcg->css);
2861 * Reclaims as many pages from the given memcg as possible.
2863 * Caller is responsible for holding css reference for memcg.
2865 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2867 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2869 /* we call try-to-free pages for make this cgroup empty */
2870 lru_add_drain_all();
2872 drain_all_stock(memcg);
2874 /* try to free all pages in this cgroup */
2875 while (nr_retries && page_counter_read(&memcg->memory)) {
2878 if (signal_pending(current))
2881 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2885 /* maybe some writeback is necessary */
2886 congestion_wait(BLK_RW_ASYNC, HZ/10);
2894 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2895 char *buf, size_t nbytes,
2898 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2900 if (mem_cgroup_is_root(memcg))
2902 return mem_cgroup_force_empty(memcg) ?: nbytes;
2905 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2908 return mem_cgroup_from_css(css)->use_hierarchy;
2911 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2912 struct cftype *cft, u64 val)
2915 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2916 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2918 if (memcg->use_hierarchy == val)
2922 * If parent's use_hierarchy is set, we can't make any modifications
2923 * in the child subtrees. If it is unset, then the change can
2924 * occur, provided the current cgroup has no children.
2926 * For the root cgroup, parent_mem is NULL, we allow value to be
2927 * set if there are no children.
2929 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2930 (val == 1 || val == 0)) {
2931 if (!memcg_has_children(memcg))
2932 memcg->use_hierarchy = val;
2941 struct accumulated_stats {
2942 unsigned long stat[MEMCG_NR_STAT];
2943 unsigned long events[NR_VM_EVENT_ITEMS];
2944 unsigned long lru_pages[NR_LRU_LISTS];
2945 const unsigned int *stats_array;
2946 const unsigned int *events_array;
2951 static void accumulate_memcg_tree(struct mem_cgroup *memcg,
2952 struct accumulated_stats *acc)
2954 struct mem_cgroup *mi;
2957 for_each_mem_cgroup_tree(mi, memcg) {
2958 for (i = 0; i < acc->stats_size; i++)
2959 acc->stat[i] += memcg_page_state(mi,
2960 acc->stats_array ? acc->stats_array[i] : i);
2962 for (i = 0; i < acc->events_size; i++)
2963 acc->events[i] += memcg_sum_events(mi,
2964 acc->events_array ? acc->events_array[i] : i);
2966 for (i = 0; i < NR_LRU_LISTS; i++)
2967 acc->lru_pages[i] +=
2968 mem_cgroup_nr_lru_pages(mi, BIT(i));
2972 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2974 unsigned long val = 0;
2976 if (mem_cgroup_is_root(memcg)) {
2977 struct mem_cgroup *iter;
2979 for_each_mem_cgroup_tree(iter, memcg) {
2980 val += memcg_page_state(iter, MEMCG_CACHE);
2981 val += memcg_page_state(iter, MEMCG_RSS);
2983 val += memcg_page_state(iter, MEMCG_SWAP);
2987 val = page_counter_read(&memcg->memory);
2989 val = page_counter_read(&memcg->memsw);
3002 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3005 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3006 struct page_counter *counter;
3008 switch (MEMFILE_TYPE(cft->private)) {
3010 counter = &memcg->memory;
3013 counter = &memcg->memsw;
3016 counter = &memcg->kmem;
3019 counter = &memcg->tcpmem;
3025 switch (MEMFILE_ATTR(cft->private)) {
3027 if (counter == &memcg->memory)
3028 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3029 if (counter == &memcg->memsw)
3030 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3031 return (u64)page_counter_read(counter) * PAGE_SIZE;
3033 return (u64)counter->max * PAGE_SIZE;
3035 return (u64)counter->watermark * PAGE_SIZE;
3037 return counter->failcnt;
3038 case RES_SOFT_LIMIT:
3039 return (u64)memcg->soft_limit * PAGE_SIZE;
3045 #ifdef CONFIG_MEMCG_KMEM
3046 static int memcg_online_kmem(struct mem_cgroup *memcg)
3050 if (cgroup_memory_nokmem)
3053 BUG_ON(memcg->kmemcg_id >= 0);
3054 BUG_ON(memcg->kmem_state);
3056 memcg_id = memcg_alloc_cache_id();
3060 static_branch_inc(&memcg_kmem_enabled_key);
3062 * A memory cgroup is considered kmem-online as soon as it gets
3063 * kmemcg_id. Setting the id after enabling static branching will
3064 * guarantee no one starts accounting before all call sites are
3067 memcg->kmemcg_id = memcg_id;
3068 memcg->kmem_state = KMEM_ONLINE;
3069 INIT_LIST_HEAD(&memcg->kmem_caches);
3074 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3076 struct cgroup_subsys_state *css;
3077 struct mem_cgroup *parent, *child;
3080 if (memcg->kmem_state != KMEM_ONLINE)
3083 * Clear the online state before clearing memcg_caches array
3084 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3085 * guarantees that no cache will be created for this cgroup
3086 * after we are done (see memcg_create_kmem_cache()).
3088 memcg->kmem_state = KMEM_ALLOCATED;
3090 memcg_deactivate_kmem_caches(memcg);
3092 kmemcg_id = memcg->kmemcg_id;
3093 BUG_ON(kmemcg_id < 0);
3095 parent = parent_mem_cgroup(memcg);
3097 parent = root_mem_cgroup;
3100 * Change kmemcg_id of this cgroup and all its descendants to the
3101 * parent's id, and then move all entries from this cgroup's list_lrus
3102 * to ones of the parent. After we have finished, all list_lrus
3103 * corresponding to this cgroup are guaranteed to remain empty. The
3104 * ordering is imposed by list_lru_node->lock taken by
3105 * memcg_drain_all_list_lrus().
3107 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3108 css_for_each_descendant_pre(css, &memcg->css) {
3109 child = mem_cgroup_from_css(css);
3110 BUG_ON(child->kmemcg_id != kmemcg_id);
3111 child->kmemcg_id = parent->kmemcg_id;
3112 if (!memcg->use_hierarchy)
3117 memcg_drain_all_list_lrus(kmemcg_id, parent);
3119 memcg_free_cache_id(kmemcg_id);
3122 static void memcg_free_kmem(struct mem_cgroup *memcg)
3124 /* css_alloc() failed, offlining didn't happen */
3125 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3126 memcg_offline_kmem(memcg);
3128 if (memcg->kmem_state == KMEM_ALLOCATED) {
3129 memcg_destroy_kmem_caches(memcg);
3130 static_branch_dec(&memcg_kmem_enabled_key);
3131 WARN_ON(page_counter_read(&memcg->kmem));
3135 static int memcg_online_kmem(struct mem_cgroup *memcg)
3139 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3142 static void memcg_free_kmem(struct mem_cgroup *memcg)
3145 #endif /* CONFIG_MEMCG_KMEM */
3147 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3152 mutex_lock(&memcg_max_mutex);
3153 ret = page_counter_set_max(&memcg->kmem, max);
3154 mutex_unlock(&memcg_max_mutex);
3158 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3162 mutex_lock(&memcg_max_mutex);
3164 ret = page_counter_set_max(&memcg->tcpmem, max);
3168 if (!memcg->tcpmem_active) {
3170 * The active flag needs to be written after the static_key
3171 * update. This is what guarantees that the socket activation
3172 * function is the last one to run. See mem_cgroup_sk_alloc()
3173 * for details, and note that we don't mark any socket as
3174 * belonging to this memcg until that flag is up.
3176 * We need to do this, because static_keys will span multiple
3177 * sites, but we can't control their order. If we mark a socket
3178 * as accounted, but the accounting functions are not patched in
3179 * yet, we'll lose accounting.
3181 * We never race with the readers in mem_cgroup_sk_alloc(),
3182 * because when this value change, the code to process it is not
3185 static_branch_inc(&memcg_sockets_enabled_key);
3186 memcg->tcpmem_active = true;
3189 mutex_unlock(&memcg_max_mutex);
3194 * The user of this function is...
3197 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3198 char *buf, size_t nbytes, loff_t off)
3200 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3201 unsigned long nr_pages;
3204 buf = strstrip(buf);
3205 ret = page_counter_memparse(buf, "-1", &nr_pages);
3209 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3211 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3215 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3217 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3220 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3223 ret = memcg_update_kmem_max(memcg, nr_pages);
3226 ret = memcg_update_tcp_max(memcg, nr_pages);
3230 case RES_SOFT_LIMIT:
3231 memcg->soft_limit = nr_pages;
3235 return ret ?: nbytes;
3238 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3239 size_t nbytes, loff_t off)
3241 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3242 struct page_counter *counter;
3244 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3246 counter = &memcg->memory;
3249 counter = &memcg->memsw;
3252 counter = &memcg->kmem;
3255 counter = &memcg->tcpmem;
3261 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3263 page_counter_reset_watermark(counter);
3266 counter->failcnt = 0;
3275 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3278 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3282 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3283 struct cftype *cft, u64 val)
3285 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3287 if (val & ~MOVE_MASK)
3291 * No kind of locking is needed in here, because ->can_attach() will
3292 * check this value once in the beginning of the process, and then carry
3293 * on with stale data. This means that changes to this value will only
3294 * affect task migrations starting after the change.
3296 memcg->move_charge_at_immigrate = val;
3300 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3301 struct cftype *cft, u64 val)
3308 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3312 unsigned int lru_mask;
3315 static const struct numa_stat stats[] = {
3316 { "total", LRU_ALL },
3317 { "file", LRU_ALL_FILE },
3318 { "anon", LRU_ALL_ANON },
3319 { "unevictable", BIT(LRU_UNEVICTABLE) },
3321 const struct numa_stat *stat;
3324 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3326 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3327 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3328 seq_printf(m, "%s=%lu", stat->name, nr);
3329 for_each_node_state(nid, N_MEMORY) {
3330 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3332 seq_printf(m, " N%d=%lu", nid, nr);
3337 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3338 struct mem_cgroup *iter;
3341 for_each_mem_cgroup_tree(iter, memcg)
3342 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3343 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3344 for_each_node_state(nid, N_MEMORY) {
3346 for_each_mem_cgroup_tree(iter, memcg)
3347 nr += mem_cgroup_node_nr_lru_pages(
3348 iter, nid, stat->lru_mask);
3349 seq_printf(m, " N%d=%lu", nid, nr);
3356 #endif /* CONFIG_NUMA */
3358 /* Universal VM events cgroup1 shows, original sort order */
3359 static const unsigned int memcg1_events[] = {
3366 static const char *const memcg1_event_names[] = {
3373 static int memcg_stat_show(struct seq_file *m, void *v)
3375 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3376 unsigned long memory, memsw;
3377 struct mem_cgroup *mi;
3379 struct accumulated_stats acc;
3381 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3382 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3384 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3385 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3387 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3388 memcg_page_state(memcg, memcg1_stats[i]) *
3392 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3393 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3394 memcg_sum_events(memcg, memcg1_events[i]));
3396 for (i = 0; i < NR_LRU_LISTS; i++)
3397 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3398 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3400 /* Hierarchical information */
3401 memory = memsw = PAGE_COUNTER_MAX;
3402 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3403 memory = min(memory, mi->memory.max);
3404 memsw = min(memsw, mi->memsw.max);
3406 seq_printf(m, "hierarchical_memory_limit %llu\n",
3407 (u64)memory * PAGE_SIZE);
3408 if (do_memsw_account())
3409 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3410 (u64)memsw * PAGE_SIZE);
3412 memset(&acc, 0, sizeof(acc));
3413 acc.stats_size = ARRAY_SIZE(memcg1_stats);
3414 acc.stats_array = memcg1_stats;
3415 acc.events_size = ARRAY_SIZE(memcg1_events);
3416 acc.events_array = memcg1_events;
3417 accumulate_memcg_tree(memcg, &acc);
3419 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3420 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3422 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3423 (u64)acc.stat[i] * PAGE_SIZE);
3426 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3427 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3428 (u64)acc.events[i]);
3430 for (i = 0; i < NR_LRU_LISTS; i++)
3431 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3432 (u64)acc.lru_pages[i] * PAGE_SIZE);
3434 #ifdef CONFIG_DEBUG_VM
3437 struct mem_cgroup_per_node *mz;
3438 struct zone_reclaim_stat *rstat;
3439 unsigned long recent_rotated[2] = {0, 0};
3440 unsigned long recent_scanned[2] = {0, 0};
3442 for_each_online_pgdat(pgdat) {
3443 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3444 rstat = &mz->lruvec.reclaim_stat;
3446 recent_rotated[0] += rstat->recent_rotated[0];
3447 recent_rotated[1] += rstat->recent_rotated[1];
3448 recent_scanned[0] += rstat->recent_scanned[0];
3449 recent_scanned[1] += rstat->recent_scanned[1];
3451 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3452 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3453 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3454 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3461 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3464 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3466 return mem_cgroup_swappiness(memcg);
3469 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3470 struct cftype *cft, u64 val)
3472 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3478 memcg->swappiness = val;
3480 vm_swappiness = val;
3485 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3487 struct mem_cgroup_threshold_ary *t;
3488 unsigned long usage;
3493 t = rcu_dereference(memcg->thresholds.primary);
3495 t = rcu_dereference(memcg->memsw_thresholds.primary);
3500 usage = mem_cgroup_usage(memcg, swap);
3503 * current_threshold points to threshold just below or equal to usage.
3504 * If it's not true, a threshold was crossed after last
3505 * call of __mem_cgroup_threshold().
3507 i = t->current_threshold;
3510 * Iterate backward over array of thresholds starting from
3511 * current_threshold and check if a threshold is crossed.
3512 * If none of thresholds below usage is crossed, we read
3513 * only one element of the array here.
3515 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3516 eventfd_signal(t->entries[i].eventfd, 1);
3518 /* i = current_threshold + 1 */
3522 * Iterate forward over array of thresholds starting from
3523 * current_threshold+1 and check if a threshold is crossed.
3524 * If none of thresholds above usage is crossed, we read
3525 * only one element of the array here.
3527 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3528 eventfd_signal(t->entries[i].eventfd, 1);
3530 /* Update current_threshold */
3531 t->current_threshold = i - 1;
3536 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3539 __mem_cgroup_threshold(memcg, false);
3540 if (do_memsw_account())
3541 __mem_cgroup_threshold(memcg, true);
3543 memcg = parent_mem_cgroup(memcg);
3547 static int compare_thresholds(const void *a, const void *b)
3549 const struct mem_cgroup_threshold *_a = a;
3550 const struct mem_cgroup_threshold *_b = b;
3552 if (_a->threshold > _b->threshold)
3555 if (_a->threshold < _b->threshold)
3561 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3563 struct mem_cgroup_eventfd_list *ev;
3565 spin_lock(&memcg_oom_lock);
3567 list_for_each_entry(ev, &memcg->oom_notify, list)
3568 eventfd_signal(ev->eventfd, 1);
3570 spin_unlock(&memcg_oom_lock);
3574 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3576 struct mem_cgroup *iter;
3578 for_each_mem_cgroup_tree(iter, memcg)
3579 mem_cgroup_oom_notify_cb(iter);
3582 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3583 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3585 struct mem_cgroup_thresholds *thresholds;
3586 struct mem_cgroup_threshold_ary *new;
3587 unsigned long threshold;
3588 unsigned long usage;
3591 ret = page_counter_memparse(args, "-1", &threshold);
3595 mutex_lock(&memcg->thresholds_lock);
3598 thresholds = &memcg->thresholds;
3599 usage = mem_cgroup_usage(memcg, false);
3600 } else if (type == _MEMSWAP) {
3601 thresholds = &memcg->memsw_thresholds;
3602 usage = mem_cgroup_usage(memcg, true);
3606 /* Check if a threshold crossed before adding a new one */
3607 if (thresholds->primary)
3608 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3610 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3612 /* Allocate memory for new array of thresholds */
3613 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3621 /* Copy thresholds (if any) to new array */
3622 if (thresholds->primary) {
3623 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3624 sizeof(struct mem_cgroup_threshold));
3627 /* Add new threshold */
3628 new->entries[size - 1].eventfd = eventfd;
3629 new->entries[size - 1].threshold = threshold;
3631 /* Sort thresholds. Registering of new threshold isn't time-critical */
3632 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3633 compare_thresholds, NULL);
3635 /* Find current threshold */
3636 new->current_threshold = -1;
3637 for (i = 0; i < size; i++) {
3638 if (new->entries[i].threshold <= usage) {
3640 * new->current_threshold will not be used until
3641 * rcu_assign_pointer(), so it's safe to increment
3644 ++new->current_threshold;
3649 /* Free old spare buffer and save old primary buffer as spare */
3650 kfree(thresholds->spare);
3651 thresholds->spare = thresholds->primary;
3653 rcu_assign_pointer(thresholds->primary, new);
3655 /* To be sure that nobody uses thresholds */
3659 mutex_unlock(&memcg->thresholds_lock);
3664 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3665 struct eventfd_ctx *eventfd, const char *args)
3667 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3670 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3671 struct eventfd_ctx *eventfd, const char *args)
3673 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3676 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3677 struct eventfd_ctx *eventfd, enum res_type type)
3679 struct mem_cgroup_thresholds *thresholds;
3680 struct mem_cgroup_threshold_ary *new;
3681 unsigned long usage;
3684 mutex_lock(&memcg->thresholds_lock);
3687 thresholds = &memcg->thresholds;
3688 usage = mem_cgroup_usage(memcg, false);
3689 } else if (type == _MEMSWAP) {
3690 thresholds = &memcg->memsw_thresholds;
3691 usage = mem_cgroup_usage(memcg, true);
3695 if (!thresholds->primary)
3698 /* Check if a threshold crossed before removing */
3699 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3701 /* Calculate new number of threshold */
3703 for (i = 0; i < thresholds->primary->size; i++) {
3704 if (thresholds->primary->entries[i].eventfd != eventfd)
3708 new = thresholds->spare;
3710 /* Set thresholds array to NULL if we don't have thresholds */
3719 /* Copy thresholds and find current threshold */
3720 new->current_threshold = -1;
3721 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3722 if (thresholds->primary->entries[i].eventfd == eventfd)
3725 new->entries[j] = thresholds->primary->entries[i];
3726 if (new->entries[j].threshold <= usage) {
3728 * new->current_threshold will not be used
3729 * until rcu_assign_pointer(), so it's safe to increment
3732 ++new->current_threshold;
3738 /* Swap primary and spare array */
3739 thresholds->spare = thresholds->primary;
3741 rcu_assign_pointer(thresholds->primary, new);
3743 /* To be sure that nobody uses thresholds */
3746 /* If all events are unregistered, free the spare array */
3748 kfree(thresholds->spare);
3749 thresholds->spare = NULL;
3752 mutex_unlock(&memcg->thresholds_lock);
3755 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3756 struct eventfd_ctx *eventfd)
3758 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3761 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3762 struct eventfd_ctx *eventfd)
3764 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3767 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3768 struct eventfd_ctx *eventfd, const char *args)
3770 struct mem_cgroup_eventfd_list *event;
3772 event = kmalloc(sizeof(*event), GFP_KERNEL);
3776 spin_lock(&memcg_oom_lock);
3778 event->eventfd = eventfd;
3779 list_add(&event->list, &memcg->oom_notify);
3781 /* already in OOM ? */
3782 if (memcg->under_oom)
3783 eventfd_signal(eventfd, 1);
3784 spin_unlock(&memcg_oom_lock);
3789 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3790 struct eventfd_ctx *eventfd)
3792 struct mem_cgroup_eventfd_list *ev, *tmp;
3794 spin_lock(&memcg_oom_lock);
3796 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3797 if (ev->eventfd == eventfd) {
3798 list_del(&ev->list);
3803 spin_unlock(&memcg_oom_lock);
3806 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3808 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3810 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3811 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3812 seq_printf(sf, "oom_kill %lu\n",
3813 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3817 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3818 struct cftype *cft, u64 val)
3820 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3822 /* cannot set to root cgroup and only 0 and 1 are allowed */
3823 if (!css->parent || !((val == 0) || (val == 1)))
3826 memcg->oom_kill_disable = val;
3828 memcg_oom_recover(memcg);
3833 #ifdef CONFIG_CGROUP_WRITEBACK
3835 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3837 return wb_domain_init(&memcg->cgwb_domain, gfp);
3840 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3842 wb_domain_exit(&memcg->cgwb_domain);
3845 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3847 wb_domain_size_changed(&memcg->cgwb_domain);
3850 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3852 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3854 if (!memcg->css.parent)
3857 return &memcg->cgwb_domain;
3861 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3862 * @wb: bdi_writeback in question
3863 * @pfilepages: out parameter for number of file pages
3864 * @pheadroom: out parameter for number of allocatable pages according to memcg
3865 * @pdirty: out parameter for number of dirty pages
3866 * @pwriteback: out parameter for number of pages under writeback
3868 * Determine the numbers of file, headroom, dirty, and writeback pages in
3869 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3870 * is a bit more involved.
3872 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3873 * headroom is calculated as the lowest headroom of itself and the
3874 * ancestors. Note that this doesn't consider the actual amount of
3875 * available memory in the system. The caller should further cap
3876 * *@pheadroom accordingly.
3878 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3879 unsigned long *pheadroom, unsigned long *pdirty,
3880 unsigned long *pwriteback)
3882 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3883 struct mem_cgroup *parent;
3885 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
3887 /* this should eventually include NR_UNSTABLE_NFS */
3888 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
3889 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3890 (1 << LRU_ACTIVE_FILE));
3891 *pheadroom = PAGE_COUNTER_MAX;
3893 while ((parent = parent_mem_cgroup(memcg))) {
3894 unsigned long ceiling = min(memcg->memory.max, memcg->high);
3895 unsigned long used = page_counter_read(&memcg->memory);
3897 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3902 #else /* CONFIG_CGROUP_WRITEBACK */
3904 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3909 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3913 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3917 #endif /* CONFIG_CGROUP_WRITEBACK */
3920 * DO NOT USE IN NEW FILES.
3922 * "cgroup.event_control" implementation.
3924 * This is way over-engineered. It tries to support fully configurable
3925 * events for each user. Such level of flexibility is completely
3926 * unnecessary especially in the light of the planned unified hierarchy.
3928 * Please deprecate this and replace with something simpler if at all
3933 * Unregister event and free resources.
3935 * Gets called from workqueue.
3937 static void memcg_event_remove(struct work_struct *work)
3939 struct mem_cgroup_event *event =
3940 container_of(work, struct mem_cgroup_event, remove);
3941 struct mem_cgroup *memcg = event->memcg;
3943 remove_wait_queue(event->wqh, &event->wait);
3945 event->unregister_event(memcg, event->eventfd);
3947 /* Notify userspace the event is going away. */
3948 eventfd_signal(event->eventfd, 1);
3950 eventfd_ctx_put(event->eventfd);
3952 css_put(&memcg->css);
3956 * Gets called on EPOLLHUP on eventfd when user closes it.
3958 * Called with wqh->lock held and interrupts disabled.
3960 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
3961 int sync, void *key)
3963 struct mem_cgroup_event *event =
3964 container_of(wait, struct mem_cgroup_event, wait);
3965 struct mem_cgroup *memcg = event->memcg;
3966 __poll_t flags = key_to_poll(key);
3968 if (flags & EPOLLHUP) {
3970 * If the event has been detached at cgroup removal, we
3971 * can simply return knowing the other side will cleanup
3974 * We can't race against event freeing since the other
3975 * side will require wqh->lock via remove_wait_queue(),
3978 spin_lock(&memcg->event_list_lock);
3979 if (!list_empty(&event->list)) {
3980 list_del_init(&event->list);
3982 * We are in atomic context, but cgroup_event_remove()
3983 * may sleep, so we have to call it in workqueue.
3985 schedule_work(&event->remove);
3987 spin_unlock(&memcg->event_list_lock);
3993 static void memcg_event_ptable_queue_proc(struct file *file,
3994 wait_queue_head_t *wqh, poll_table *pt)
3996 struct mem_cgroup_event *event =
3997 container_of(pt, struct mem_cgroup_event, pt);
4000 add_wait_queue(wqh, &event->wait);
4004 * DO NOT USE IN NEW FILES.
4006 * Parse input and register new cgroup event handler.
4008 * Input must be in format '<event_fd> <control_fd> <args>'.
4009 * Interpretation of args is defined by control file implementation.
4011 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4012 char *buf, size_t nbytes, loff_t off)
4014 struct cgroup_subsys_state *css = of_css(of);
4015 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4016 struct mem_cgroup_event *event;
4017 struct cgroup_subsys_state *cfile_css;
4018 unsigned int efd, cfd;
4025 buf = strstrip(buf);
4027 efd = simple_strtoul(buf, &endp, 10);
4032 cfd = simple_strtoul(buf, &endp, 10);
4033 if ((*endp != ' ') && (*endp != '\0'))
4037 event = kzalloc(sizeof(*event), GFP_KERNEL);
4041 event->memcg = memcg;
4042 INIT_LIST_HEAD(&event->list);
4043 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4044 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4045 INIT_WORK(&event->remove, memcg_event_remove);
4053 event->eventfd = eventfd_ctx_fileget(efile.file);
4054 if (IS_ERR(event->eventfd)) {
4055 ret = PTR_ERR(event->eventfd);
4062 goto out_put_eventfd;
4065 /* the process need read permission on control file */
4066 /* AV: shouldn't we check that it's been opened for read instead? */
4067 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4072 * Determine the event callbacks and set them in @event. This used
4073 * to be done via struct cftype but cgroup core no longer knows
4074 * about these events. The following is crude but the whole thing
4075 * is for compatibility anyway.
4077 * DO NOT ADD NEW FILES.
4079 name = cfile.file->f_path.dentry->d_name.name;
4081 if (!strcmp(name, "memory.usage_in_bytes")) {
4082 event->register_event = mem_cgroup_usage_register_event;
4083 event->unregister_event = mem_cgroup_usage_unregister_event;
4084 } else if (!strcmp(name, "memory.oom_control")) {
4085 event->register_event = mem_cgroup_oom_register_event;
4086 event->unregister_event = mem_cgroup_oom_unregister_event;
4087 } else if (!strcmp(name, "memory.pressure_level")) {
4088 event->register_event = vmpressure_register_event;
4089 event->unregister_event = vmpressure_unregister_event;
4090 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4091 event->register_event = memsw_cgroup_usage_register_event;
4092 event->unregister_event = memsw_cgroup_usage_unregister_event;
4099 * Verify @cfile should belong to @css. Also, remaining events are
4100 * automatically removed on cgroup destruction but the removal is
4101 * asynchronous, so take an extra ref on @css.
4103 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4104 &memory_cgrp_subsys);
4106 if (IS_ERR(cfile_css))
4108 if (cfile_css != css) {
4113 ret = event->register_event(memcg, event->eventfd, buf);
4117 vfs_poll(efile.file, &event->pt);
4119 spin_lock(&memcg->event_list_lock);
4120 list_add(&event->list, &memcg->event_list);
4121 spin_unlock(&memcg->event_list_lock);
4133 eventfd_ctx_put(event->eventfd);
4142 static struct cftype mem_cgroup_legacy_files[] = {
4144 .name = "usage_in_bytes",
4145 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4146 .read_u64 = mem_cgroup_read_u64,
4149 .name = "max_usage_in_bytes",
4150 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4151 .write = mem_cgroup_reset,
4152 .read_u64 = mem_cgroup_read_u64,
4155 .name = "limit_in_bytes",
4156 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4157 .write = mem_cgroup_write,
4158 .read_u64 = mem_cgroup_read_u64,
4161 .name = "soft_limit_in_bytes",
4162 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4163 .write = mem_cgroup_write,
4164 .read_u64 = mem_cgroup_read_u64,
4168 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4169 .write = mem_cgroup_reset,
4170 .read_u64 = mem_cgroup_read_u64,
4174 .seq_show = memcg_stat_show,
4177 .name = "force_empty",
4178 .write = mem_cgroup_force_empty_write,
4181 .name = "use_hierarchy",
4182 .write_u64 = mem_cgroup_hierarchy_write,
4183 .read_u64 = mem_cgroup_hierarchy_read,
4186 .name = "cgroup.event_control", /* XXX: for compat */
4187 .write = memcg_write_event_control,
4188 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4191 .name = "swappiness",
4192 .read_u64 = mem_cgroup_swappiness_read,
4193 .write_u64 = mem_cgroup_swappiness_write,
4196 .name = "move_charge_at_immigrate",
4197 .read_u64 = mem_cgroup_move_charge_read,
4198 .write_u64 = mem_cgroup_move_charge_write,
4201 .name = "oom_control",
4202 .seq_show = mem_cgroup_oom_control_read,
4203 .write_u64 = mem_cgroup_oom_control_write,
4204 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4207 .name = "pressure_level",
4211 .name = "numa_stat",
4212 .seq_show = memcg_numa_stat_show,
4216 .name = "kmem.limit_in_bytes",
4217 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4218 .write = mem_cgroup_write,
4219 .read_u64 = mem_cgroup_read_u64,
4222 .name = "kmem.usage_in_bytes",
4223 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4224 .read_u64 = mem_cgroup_read_u64,
4227 .name = "kmem.failcnt",
4228 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4229 .write = mem_cgroup_reset,
4230 .read_u64 = mem_cgroup_read_u64,
4233 .name = "kmem.max_usage_in_bytes",
4234 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4235 .write = mem_cgroup_reset,
4236 .read_u64 = mem_cgroup_read_u64,
4238 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4240 .name = "kmem.slabinfo",
4241 .seq_start = memcg_slab_start,
4242 .seq_next = memcg_slab_next,
4243 .seq_stop = memcg_slab_stop,
4244 .seq_show = memcg_slab_show,
4248 .name = "kmem.tcp.limit_in_bytes",
4249 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4250 .write = mem_cgroup_write,
4251 .read_u64 = mem_cgroup_read_u64,
4254 .name = "kmem.tcp.usage_in_bytes",
4255 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4256 .read_u64 = mem_cgroup_read_u64,
4259 .name = "kmem.tcp.failcnt",
4260 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4261 .write = mem_cgroup_reset,
4262 .read_u64 = mem_cgroup_read_u64,
4265 .name = "kmem.tcp.max_usage_in_bytes",
4266 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4267 .write = mem_cgroup_reset,
4268 .read_u64 = mem_cgroup_read_u64,
4270 { }, /* terminate */
4274 * Private memory cgroup IDR
4276 * Swap-out records and page cache shadow entries need to store memcg
4277 * references in constrained space, so we maintain an ID space that is
4278 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4279 * memory-controlled cgroups to 64k.
4281 * However, there usually are many references to the oflline CSS after
4282 * the cgroup has been destroyed, such as page cache or reclaimable
4283 * slab objects, that don't need to hang on to the ID. We want to keep
4284 * those dead CSS from occupying IDs, or we might quickly exhaust the
4285 * relatively small ID space and prevent the creation of new cgroups
4286 * even when there are much fewer than 64k cgroups - possibly none.
4288 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4289 * be freed and recycled when it's no longer needed, which is usually
4290 * when the CSS is offlined.
4292 * The only exception to that are records of swapped out tmpfs/shmem
4293 * pages that need to be attributed to live ancestors on swapin. But
4294 * those references are manageable from userspace.
4297 static DEFINE_IDR(mem_cgroup_idr);
4299 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4301 if (memcg->id.id > 0) {
4302 idr_remove(&mem_cgroup_idr, memcg->id.id);
4307 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4309 refcount_add(n, &memcg->id.ref);
4312 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4314 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4315 mem_cgroup_id_remove(memcg);
4317 /* Memcg ID pins CSS */
4318 css_put(&memcg->css);
4322 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4324 mem_cgroup_id_get_many(memcg, 1);
4327 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4329 mem_cgroup_id_put_many(memcg, 1);
4333 * mem_cgroup_from_id - look up a memcg from a memcg id
4334 * @id: the memcg id to look up
4336 * Caller must hold rcu_read_lock().
4338 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4340 WARN_ON_ONCE(!rcu_read_lock_held());
4341 return idr_find(&mem_cgroup_idr, id);
4344 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4346 struct mem_cgroup_per_node *pn;
4349 * This routine is called against possible nodes.
4350 * But it's BUG to call kmalloc() against offline node.
4352 * TODO: this routine can waste much memory for nodes which will
4353 * never be onlined. It's better to use memory hotplug callback
4356 if (!node_state(node, N_NORMAL_MEMORY))
4358 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4362 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4363 if (!pn->lruvec_stat_cpu) {
4368 lruvec_init(&pn->lruvec);
4369 pn->usage_in_excess = 0;
4370 pn->on_tree = false;
4373 memcg->nodeinfo[node] = pn;
4377 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4379 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4384 free_percpu(pn->lruvec_stat_cpu);
4388 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4393 free_mem_cgroup_per_node_info(memcg, node);
4394 free_percpu(memcg->stat_cpu);
4398 static void mem_cgroup_free(struct mem_cgroup *memcg)
4400 memcg_wb_domain_exit(memcg);
4401 __mem_cgroup_free(memcg);
4404 static struct mem_cgroup *mem_cgroup_alloc(void)
4406 struct mem_cgroup *memcg;
4410 size = sizeof(struct mem_cgroup);
4411 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4413 memcg = kzalloc(size, GFP_KERNEL);
4417 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4418 1, MEM_CGROUP_ID_MAX,
4420 if (memcg->id.id < 0)
4423 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4424 if (!memcg->stat_cpu)
4428 if (alloc_mem_cgroup_per_node_info(memcg, node))
4431 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4434 INIT_WORK(&memcg->high_work, high_work_func);
4435 memcg->last_scanned_node = MAX_NUMNODES;
4436 INIT_LIST_HEAD(&memcg->oom_notify);
4437 mutex_init(&memcg->thresholds_lock);
4438 spin_lock_init(&memcg->move_lock);
4439 vmpressure_init(&memcg->vmpressure);
4440 INIT_LIST_HEAD(&memcg->event_list);
4441 spin_lock_init(&memcg->event_list_lock);
4442 memcg->socket_pressure = jiffies;
4443 #ifdef CONFIG_MEMCG_KMEM
4444 memcg->kmemcg_id = -1;
4446 #ifdef CONFIG_CGROUP_WRITEBACK
4447 INIT_LIST_HEAD(&memcg->cgwb_list);
4449 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4452 mem_cgroup_id_remove(memcg);
4453 __mem_cgroup_free(memcg);
4457 static struct cgroup_subsys_state * __ref
4458 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4460 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4461 struct mem_cgroup *memcg;
4462 long error = -ENOMEM;
4464 memcg = mem_cgroup_alloc();
4466 return ERR_PTR(error);
4468 memcg->high = PAGE_COUNTER_MAX;
4469 memcg->soft_limit = PAGE_COUNTER_MAX;
4471 memcg->swappiness = mem_cgroup_swappiness(parent);
4472 memcg->oom_kill_disable = parent->oom_kill_disable;
4474 if (parent && parent->use_hierarchy) {
4475 memcg->use_hierarchy = true;
4476 page_counter_init(&memcg->memory, &parent->memory);
4477 page_counter_init(&memcg->swap, &parent->swap);
4478 page_counter_init(&memcg->memsw, &parent->memsw);
4479 page_counter_init(&memcg->kmem, &parent->kmem);
4480 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4482 page_counter_init(&memcg->memory, NULL);
4483 page_counter_init(&memcg->swap, NULL);
4484 page_counter_init(&memcg->memsw, NULL);
4485 page_counter_init(&memcg->kmem, NULL);
4486 page_counter_init(&memcg->tcpmem, NULL);
4488 * Deeper hierachy with use_hierarchy == false doesn't make
4489 * much sense so let cgroup subsystem know about this
4490 * unfortunate state in our controller.
4492 if (parent != root_mem_cgroup)
4493 memory_cgrp_subsys.broken_hierarchy = true;
4496 /* The following stuff does not apply to the root */
4498 root_mem_cgroup = memcg;
4502 error = memcg_online_kmem(memcg);
4506 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4507 static_branch_inc(&memcg_sockets_enabled_key);
4511 mem_cgroup_id_remove(memcg);
4512 mem_cgroup_free(memcg);
4513 return ERR_PTR(-ENOMEM);
4516 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4518 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4521 * A memcg must be visible for memcg_expand_shrinker_maps()
4522 * by the time the maps are allocated. So, we allocate maps
4523 * here, when for_each_mem_cgroup() can't skip it.
4525 if (memcg_alloc_shrinker_maps(memcg)) {
4526 mem_cgroup_id_remove(memcg);
4530 /* Online state pins memcg ID, memcg ID pins CSS */
4531 refcount_set(&memcg->id.ref, 1);
4536 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4538 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4539 struct mem_cgroup_event *event, *tmp;
4542 * Unregister events and notify userspace.
4543 * Notify userspace about cgroup removing only after rmdir of cgroup
4544 * directory to avoid race between userspace and kernelspace.
4546 spin_lock(&memcg->event_list_lock);
4547 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4548 list_del_init(&event->list);
4549 schedule_work(&event->remove);
4551 spin_unlock(&memcg->event_list_lock);
4553 page_counter_set_min(&memcg->memory, 0);
4554 page_counter_set_low(&memcg->memory, 0);
4556 memcg_offline_kmem(memcg);
4557 wb_memcg_offline(memcg);
4559 drain_all_stock(memcg);
4561 mem_cgroup_id_put(memcg);
4564 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4566 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4568 invalidate_reclaim_iterators(memcg);
4571 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4573 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4575 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4576 static_branch_dec(&memcg_sockets_enabled_key);
4578 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4579 static_branch_dec(&memcg_sockets_enabled_key);
4581 vmpressure_cleanup(&memcg->vmpressure);
4582 cancel_work_sync(&memcg->high_work);
4583 mem_cgroup_remove_from_trees(memcg);
4584 memcg_free_shrinker_maps(memcg);
4585 memcg_free_kmem(memcg);
4586 mem_cgroup_free(memcg);
4590 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4591 * @css: the target css
4593 * Reset the states of the mem_cgroup associated with @css. This is
4594 * invoked when the userland requests disabling on the default hierarchy
4595 * but the memcg is pinned through dependency. The memcg should stop
4596 * applying policies and should revert to the vanilla state as it may be
4597 * made visible again.
4599 * The current implementation only resets the essential configurations.
4600 * This needs to be expanded to cover all the visible parts.
4602 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4604 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4606 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4607 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4608 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4609 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4610 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4611 page_counter_set_min(&memcg->memory, 0);
4612 page_counter_set_low(&memcg->memory, 0);
4613 memcg->high = PAGE_COUNTER_MAX;
4614 memcg->soft_limit = PAGE_COUNTER_MAX;
4615 memcg_wb_domain_size_changed(memcg);
4619 /* Handlers for move charge at task migration. */
4620 static int mem_cgroup_do_precharge(unsigned long count)
4624 /* Try a single bulk charge without reclaim first, kswapd may wake */
4625 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4627 mc.precharge += count;
4631 /* Try charges one by one with reclaim, but do not retry */
4633 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4647 enum mc_target_type {
4654 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4655 unsigned long addr, pte_t ptent)
4657 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4659 if (!page || !page_mapped(page))
4661 if (PageAnon(page)) {
4662 if (!(mc.flags & MOVE_ANON))
4665 if (!(mc.flags & MOVE_FILE))
4668 if (!get_page_unless_zero(page))
4674 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4675 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4676 pte_t ptent, swp_entry_t *entry)
4678 struct page *page = NULL;
4679 swp_entry_t ent = pte_to_swp_entry(ptent);
4681 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4685 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4686 * a device and because they are not accessible by CPU they are store
4687 * as special swap entry in the CPU page table.
4689 if (is_device_private_entry(ent)) {
4690 page = device_private_entry_to_page(ent);
4692 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4693 * a refcount of 1 when free (unlike normal page)
4695 if (!page_ref_add_unless(page, 1, 1))
4701 * Because lookup_swap_cache() updates some statistics counter,
4702 * we call find_get_page() with swapper_space directly.
4704 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4705 if (do_memsw_account())
4706 entry->val = ent.val;
4711 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4712 pte_t ptent, swp_entry_t *entry)
4718 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4719 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4721 struct page *page = NULL;
4722 struct address_space *mapping;
4725 if (!vma->vm_file) /* anonymous vma */
4727 if (!(mc.flags & MOVE_FILE))
4730 mapping = vma->vm_file->f_mapping;
4731 pgoff = linear_page_index(vma, addr);
4733 /* page is moved even if it's not RSS of this task(page-faulted). */
4735 /* shmem/tmpfs may report page out on swap: account for that too. */
4736 if (shmem_mapping(mapping)) {
4737 page = find_get_entry(mapping, pgoff);
4738 if (xa_is_value(page)) {
4739 swp_entry_t swp = radix_to_swp_entry(page);
4740 if (do_memsw_account())
4742 page = find_get_page(swap_address_space(swp),
4746 page = find_get_page(mapping, pgoff);
4748 page = find_get_page(mapping, pgoff);
4754 * mem_cgroup_move_account - move account of the page
4756 * @compound: charge the page as compound or small page
4757 * @from: mem_cgroup which the page is moved from.
4758 * @to: mem_cgroup which the page is moved to. @from != @to.
4760 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4762 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4765 static int mem_cgroup_move_account(struct page *page,
4767 struct mem_cgroup *from,
4768 struct mem_cgroup *to)
4770 unsigned long flags;
4771 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4775 VM_BUG_ON(from == to);
4776 VM_BUG_ON_PAGE(PageLRU(page), page);
4777 VM_BUG_ON(compound && !PageTransHuge(page));
4780 * Prevent mem_cgroup_migrate() from looking at
4781 * page->mem_cgroup of its source page while we change it.
4784 if (!trylock_page(page))
4788 if (page->mem_cgroup != from)
4791 anon = PageAnon(page);
4793 spin_lock_irqsave(&from->move_lock, flags);
4795 if (!anon && page_mapped(page)) {
4796 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4797 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4801 * move_lock grabbed above and caller set from->moving_account, so
4802 * mod_memcg_page_state will serialize updates to PageDirty.
4803 * So mapping should be stable for dirty pages.
4805 if (!anon && PageDirty(page)) {
4806 struct address_space *mapping = page_mapping(page);
4808 if (mapping_cap_account_dirty(mapping)) {
4809 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4810 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4814 if (PageWriteback(page)) {
4815 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4816 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4820 * It is safe to change page->mem_cgroup here because the page
4821 * is referenced, charged, and isolated - we can't race with
4822 * uncharging, charging, migration, or LRU putback.
4825 /* caller should have done css_get */
4826 page->mem_cgroup = to;
4827 spin_unlock_irqrestore(&from->move_lock, flags);
4831 local_irq_disable();
4832 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4833 memcg_check_events(to, page);
4834 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4835 memcg_check_events(from, page);
4844 * get_mctgt_type - get target type of moving charge
4845 * @vma: the vma the pte to be checked belongs
4846 * @addr: the address corresponding to the pte to be checked
4847 * @ptent: the pte to be checked
4848 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4851 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4852 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4853 * move charge. if @target is not NULL, the page is stored in target->page
4854 * with extra refcnt got(Callers should handle it).
4855 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4856 * target for charge migration. if @target is not NULL, the entry is stored
4858 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4859 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4860 * For now we such page is charge like a regular page would be as for all
4861 * intent and purposes it is just special memory taking the place of a
4864 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4866 * Called with pte lock held.
4869 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4870 unsigned long addr, pte_t ptent, union mc_target *target)
4872 struct page *page = NULL;
4873 enum mc_target_type ret = MC_TARGET_NONE;
4874 swp_entry_t ent = { .val = 0 };
4876 if (pte_present(ptent))
4877 page = mc_handle_present_pte(vma, addr, ptent);
4878 else if (is_swap_pte(ptent))
4879 page = mc_handle_swap_pte(vma, ptent, &ent);
4880 else if (pte_none(ptent))
4881 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4883 if (!page && !ent.val)
4887 * Do only loose check w/o serialization.
4888 * mem_cgroup_move_account() checks the page is valid or
4889 * not under LRU exclusion.
4891 if (page->mem_cgroup == mc.from) {
4892 ret = MC_TARGET_PAGE;
4893 if (is_device_private_page(page) ||
4894 is_device_public_page(page))
4895 ret = MC_TARGET_DEVICE;
4897 target->page = page;
4899 if (!ret || !target)
4903 * There is a swap entry and a page doesn't exist or isn't charged.
4904 * But we cannot move a tail-page in a THP.
4906 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
4907 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4908 ret = MC_TARGET_SWAP;
4915 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4917 * We don't consider PMD mapped swapping or file mapped pages because THP does
4918 * not support them for now.
4919 * Caller should make sure that pmd_trans_huge(pmd) is true.
4921 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4922 unsigned long addr, pmd_t pmd, union mc_target *target)
4924 struct page *page = NULL;
4925 enum mc_target_type ret = MC_TARGET_NONE;
4927 if (unlikely(is_swap_pmd(pmd))) {
4928 VM_BUG_ON(thp_migration_supported() &&
4929 !is_pmd_migration_entry(pmd));
4932 page = pmd_page(pmd);
4933 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4934 if (!(mc.flags & MOVE_ANON))
4936 if (page->mem_cgroup == mc.from) {
4937 ret = MC_TARGET_PAGE;
4940 target->page = page;
4946 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4947 unsigned long addr, pmd_t pmd, union mc_target *target)
4949 return MC_TARGET_NONE;
4953 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4954 unsigned long addr, unsigned long end,
4955 struct mm_walk *walk)
4957 struct vm_area_struct *vma = walk->vma;
4961 ptl = pmd_trans_huge_lock(pmd, vma);
4964 * Note their can not be MC_TARGET_DEVICE for now as we do not
4965 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
4966 * MEMORY_DEVICE_PRIVATE but this might change.
4968 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4969 mc.precharge += HPAGE_PMD_NR;
4974 if (pmd_trans_unstable(pmd))
4976 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4977 for (; addr != end; pte++, addr += PAGE_SIZE)
4978 if (get_mctgt_type(vma, addr, *pte, NULL))
4979 mc.precharge++; /* increment precharge temporarily */
4980 pte_unmap_unlock(pte - 1, ptl);
4986 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4988 unsigned long precharge;
4990 struct mm_walk mem_cgroup_count_precharge_walk = {
4991 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4994 down_read(&mm->mmap_sem);
4995 walk_page_range(0, mm->highest_vm_end,
4996 &mem_cgroup_count_precharge_walk);
4997 up_read(&mm->mmap_sem);
4999 precharge = mc.precharge;
5005 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5007 unsigned long precharge = mem_cgroup_count_precharge(mm);
5009 VM_BUG_ON(mc.moving_task);
5010 mc.moving_task = current;
5011 return mem_cgroup_do_precharge(precharge);
5014 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5015 static void __mem_cgroup_clear_mc(void)
5017 struct mem_cgroup *from = mc.from;
5018 struct mem_cgroup *to = mc.to;
5020 /* we must uncharge all the leftover precharges from mc.to */
5022 cancel_charge(mc.to, mc.precharge);
5026 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5027 * we must uncharge here.
5029 if (mc.moved_charge) {
5030 cancel_charge(mc.from, mc.moved_charge);
5031 mc.moved_charge = 0;
5033 /* we must fixup refcnts and charges */
5034 if (mc.moved_swap) {
5035 /* uncharge swap account from the old cgroup */
5036 if (!mem_cgroup_is_root(mc.from))
5037 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5039 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5042 * we charged both to->memory and to->memsw, so we
5043 * should uncharge to->memory.
5045 if (!mem_cgroup_is_root(mc.to))
5046 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5048 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5049 css_put_many(&mc.to->css, mc.moved_swap);
5053 memcg_oom_recover(from);
5054 memcg_oom_recover(to);
5055 wake_up_all(&mc.waitq);
5058 static void mem_cgroup_clear_mc(void)
5060 struct mm_struct *mm = mc.mm;
5063 * we must clear moving_task before waking up waiters at the end of
5066 mc.moving_task = NULL;
5067 __mem_cgroup_clear_mc();
5068 spin_lock(&mc.lock);
5072 spin_unlock(&mc.lock);
5077 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5079 struct cgroup_subsys_state *css;
5080 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5081 struct mem_cgroup *from;
5082 struct task_struct *leader, *p;
5083 struct mm_struct *mm;
5084 unsigned long move_flags;
5087 /* charge immigration isn't supported on the default hierarchy */
5088 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5092 * Multi-process migrations only happen on the default hierarchy
5093 * where charge immigration is not used. Perform charge
5094 * immigration if @tset contains a leader and whine if there are
5098 cgroup_taskset_for_each_leader(leader, css, tset) {
5101 memcg = mem_cgroup_from_css(css);
5107 * We are now commited to this value whatever it is. Changes in this
5108 * tunable will only affect upcoming migrations, not the current one.
5109 * So we need to save it, and keep it going.
5111 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5115 from = mem_cgroup_from_task(p);
5117 VM_BUG_ON(from == memcg);
5119 mm = get_task_mm(p);
5122 /* We move charges only when we move a owner of the mm */
5123 if (mm->owner == p) {
5126 VM_BUG_ON(mc.precharge);
5127 VM_BUG_ON(mc.moved_charge);
5128 VM_BUG_ON(mc.moved_swap);
5130 spin_lock(&mc.lock);
5134 mc.flags = move_flags;
5135 spin_unlock(&mc.lock);
5136 /* We set mc.moving_task later */
5138 ret = mem_cgroup_precharge_mc(mm);
5140 mem_cgroup_clear_mc();
5147 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5150 mem_cgroup_clear_mc();
5153 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5154 unsigned long addr, unsigned long end,
5155 struct mm_walk *walk)
5158 struct vm_area_struct *vma = walk->vma;
5161 enum mc_target_type target_type;
5162 union mc_target target;
5165 ptl = pmd_trans_huge_lock(pmd, vma);
5167 if (mc.precharge < HPAGE_PMD_NR) {
5171 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5172 if (target_type == MC_TARGET_PAGE) {
5174 if (!isolate_lru_page(page)) {
5175 if (!mem_cgroup_move_account(page, true,
5177 mc.precharge -= HPAGE_PMD_NR;
5178 mc.moved_charge += HPAGE_PMD_NR;
5180 putback_lru_page(page);
5183 } else if (target_type == MC_TARGET_DEVICE) {
5185 if (!mem_cgroup_move_account(page, true,
5187 mc.precharge -= HPAGE_PMD_NR;
5188 mc.moved_charge += HPAGE_PMD_NR;
5196 if (pmd_trans_unstable(pmd))
5199 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5200 for (; addr != end; addr += PAGE_SIZE) {
5201 pte_t ptent = *(pte++);
5202 bool device = false;
5208 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5209 case MC_TARGET_DEVICE:
5212 case MC_TARGET_PAGE:
5215 * We can have a part of the split pmd here. Moving it
5216 * can be done but it would be too convoluted so simply
5217 * ignore such a partial THP and keep it in original
5218 * memcg. There should be somebody mapping the head.
5220 if (PageTransCompound(page))
5222 if (!device && isolate_lru_page(page))
5224 if (!mem_cgroup_move_account(page, false,
5227 /* we uncharge from mc.from later. */
5231 putback_lru_page(page);
5232 put: /* get_mctgt_type() gets the page */
5235 case MC_TARGET_SWAP:
5237 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5239 /* we fixup refcnts and charges later. */
5247 pte_unmap_unlock(pte - 1, ptl);
5252 * We have consumed all precharges we got in can_attach().
5253 * We try charge one by one, but don't do any additional
5254 * charges to mc.to if we have failed in charge once in attach()
5257 ret = mem_cgroup_do_precharge(1);
5265 static void mem_cgroup_move_charge(void)
5267 struct mm_walk mem_cgroup_move_charge_walk = {
5268 .pmd_entry = mem_cgroup_move_charge_pte_range,
5272 lru_add_drain_all();
5274 * Signal lock_page_memcg() to take the memcg's move_lock
5275 * while we're moving its pages to another memcg. Then wait
5276 * for already started RCU-only updates to finish.
5278 atomic_inc(&mc.from->moving_account);
5281 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5283 * Someone who are holding the mmap_sem might be waiting in
5284 * waitq. So we cancel all extra charges, wake up all waiters,
5285 * and retry. Because we cancel precharges, we might not be able
5286 * to move enough charges, but moving charge is a best-effort
5287 * feature anyway, so it wouldn't be a big problem.
5289 __mem_cgroup_clear_mc();
5294 * When we have consumed all precharges and failed in doing
5295 * additional charge, the page walk just aborts.
5297 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5299 up_read(&mc.mm->mmap_sem);
5300 atomic_dec(&mc.from->moving_account);
5303 static void mem_cgroup_move_task(void)
5306 mem_cgroup_move_charge();
5307 mem_cgroup_clear_mc();
5310 #else /* !CONFIG_MMU */
5311 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5315 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5318 static void mem_cgroup_move_task(void)
5324 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5325 * to verify whether we're attached to the default hierarchy on each mount
5328 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5331 * use_hierarchy is forced on the default hierarchy. cgroup core
5332 * guarantees that @root doesn't have any children, so turning it
5333 * on for the root memcg is enough.
5335 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5336 root_mem_cgroup->use_hierarchy = true;
5338 root_mem_cgroup->use_hierarchy = false;
5341 static u64 memory_current_read(struct cgroup_subsys_state *css,
5344 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5346 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5349 static int memory_min_show(struct seq_file *m, void *v)
5351 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5352 unsigned long min = READ_ONCE(memcg->memory.min);
5354 if (min == PAGE_COUNTER_MAX)
5355 seq_puts(m, "max\n");
5357 seq_printf(m, "%llu\n", (u64)min * PAGE_SIZE);
5362 static ssize_t memory_min_write(struct kernfs_open_file *of,
5363 char *buf, size_t nbytes, loff_t off)
5365 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5369 buf = strstrip(buf);
5370 err = page_counter_memparse(buf, "max", &min);
5374 page_counter_set_min(&memcg->memory, min);
5379 static int memory_low_show(struct seq_file *m, void *v)
5381 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5382 unsigned long low = READ_ONCE(memcg->memory.low);
5384 if (low == PAGE_COUNTER_MAX)
5385 seq_puts(m, "max\n");
5387 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5392 static ssize_t memory_low_write(struct kernfs_open_file *of,
5393 char *buf, size_t nbytes, loff_t off)
5395 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5399 buf = strstrip(buf);
5400 err = page_counter_memparse(buf, "max", &low);
5404 page_counter_set_low(&memcg->memory, low);
5409 static int memory_high_show(struct seq_file *m, void *v)
5411 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5412 unsigned long high = READ_ONCE(memcg->high);
5414 if (high == PAGE_COUNTER_MAX)
5415 seq_puts(m, "max\n");
5417 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5422 static ssize_t memory_high_write(struct kernfs_open_file *of,
5423 char *buf, size_t nbytes, loff_t off)
5425 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5426 unsigned long nr_pages;
5430 buf = strstrip(buf);
5431 err = page_counter_memparse(buf, "max", &high);
5437 nr_pages = page_counter_read(&memcg->memory);
5438 if (nr_pages > high)
5439 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5442 memcg_wb_domain_size_changed(memcg);
5446 static int memory_max_show(struct seq_file *m, void *v)
5448 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5449 unsigned long max = READ_ONCE(memcg->memory.max);
5451 if (max == PAGE_COUNTER_MAX)
5452 seq_puts(m, "max\n");
5454 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5459 static ssize_t memory_max_write(struct kernfs_open_file *of,
5460 char *buf, size_t nbytes, loff_t off)
5462 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5463 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5464 bool drained = false;
5468 buf = strstrip(buf);
5469 err = page_counter_memparse(buf, "max", &max);
5473 xchg(&memcg->memory.max, max);
5476 unsigned long nr_pages = page_counter_read(&memcg->memory);
5478 if (nr_pages <= max)
5481 if (signal_pending(current)) {
5487 drain_all_stock(memcg);
5493 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5499 memcg_memory_event(memcg, MEMCG_OOM);
5500 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5504 memcg_wb_domain_size_changed(memcg);
5508 static int memory_events_show(struct seq_file *m, void *v)
5510 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5512 seq_printf(m, "low %lu\n",
5513 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5514 seq_printf(m, "high %lu\n",
5515 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5516 seq_printf(m, "max %lu\n",
5517 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5518 seq_printf(m, "oom %lu\n",
5519 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5520 seq_printf(m, "oom_kill %lu\n",
5521 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5526 static int memory_stat_show(struct seq_file *m, void *v)
5528 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5529 struct accumulated_stats acc;
5533 * Provide statistics on the state of the memory subsystem as
5534 * well as cumulative event counters that show past behavior.
5536 * This list is ordered following a combination of these gradients:
5537 * 1) generic big picture -> specifics and details
5538 * 2) reflecting userspace activity -> reflecting kernel heuristics
5540 * Current memory state:
5543 memset(&acc, 0, sizeof(acc));
5544 acc.stats_size = MEMCG_NR_STAT;
5545 acc.events_size = NR_VM_EVENT_ITEMS;
5546 accumulate_memcg_tree(memcg, &acc);
5548 seq_printf(m, "anon %llu\n",
5549 (u64)acc.stat[MEMCG_RSS] * PAGE_SIZE);
5550 seq_printf(m, "file %llu\n",
5551 (u64)acc.stat[MEMCG_CACHE] * PAGE_SIZE);
5552 seq_printf(m, "kernel_stack %llu\n",
5553 (u64)acc.stat[MEMCG_KERNEL_STACK_KB] * 1024);
5554 seq_printf(m, "slab %llu\n",
5555 (u64)(acc.stat[NR_SLAB_RECLAIMABLE] +
5556 acc.stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5557 seq_printf(m, "sock %llu\n",
5558 (u64)acc.stat[MEMCG_SOCK] * PAGE_SIZE);
5560 seq_printf(m, "shmem %llu\n",
5561 (u64)acc.stat[NR_SHMEM] * PAGE_SIZE);
5562 seq_printf(m, "file_mapped %llu\n",
5563 (u64)acc.stat[NR_FILE_MAPPED] * PAGE_SIZE);
5564 seq_printf(m, "file_dirty %llu\n",
5565 (u64)acc.stat[NR_FILE_DIRTY] * PAGE_SIZE);
5566 seq_printf(m, "file_writeback %llu\n",
5567 (u64)acc.stat[NR_WRITEBACK] * PAGE_SIZE);
5569 for (i = 0; i < NR_LRU_LISTS; i++)
5570 seq_printf(m, "%s %llu\n", mem_cgroup_lru_names[i],
5571 (u64)acc.lru_pages[i] * PAGE_SIZE);
5573 seq_printf(m, "slab_reclaimable %llu\n",
5574 (u64)acc.stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5575 seq_printf(m, "slab_unreclaimable %llu\n",
5576 (u64)acc.stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5578 /* Accumulated memory events */
5580 seq_printf(m, "pgfault %lu\n", acc.events[PGFAULT]);
5581 seq_printf(m, "pgmajfault %lu\n", acc.events[PGMAJFAULT]);
5583 seq_printf(m, "workingset_refault %lu\n",
5584 acc.stat[WORKINGSET_REFAULT]);
5585 seq_printf(m, "workingset_activate %lu\n",
5586 acc.stat[WORKINGSET_ACTIVATE]);
5587 seq_printf(m, "workingset_nodereclaim %lu\n",
5588 acc.stat[WORKINGSET_NODERECLAIM]);
5590 seq_printf(m, "pgrefill %lu\n", acc.events[PGREFILL]);
5591 seq_printf(m, "pgscan %lu\n", acc.events[PGSCAN_KSWAPD] +
5592 acc.events[PGSCAN_DIRECT]);
5593 seq_printf(m, "pgsteal %lu\n", acc.events[PGSTEAL_KSWAPD] +
5594 acc.events[PGSTEAL_DIRECT]);
5595 seq_printf(m, "pgactivate %lu\n", acc.events[PGACTIVATE]);
5596 seq_printf(m, "pgdeactivate %lu\n", acc.events[PGDEACTIVATE]);
5597 seq_printf(m, "pglazyfree %lu\n", acc.events[PGLAZYFREE]);
5598 seq_printf(m, "pglazyfreed %lu\n", acc.events[PGLAZYFREED]);
5603 static int memory_oom_group_show(struct seq_file *m, void *v)
5605 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5607 seq_printf(m, "%d\n", memcg->oom_group);
5612 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5613 char *buf, size_t nbytes, loff_t off)
5615 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5618 buf = strstrip(buf);
5622 ret = kstrtoint(buf, 0, &oom_group);
5626 if (oom_group != 0 && oom_group != 1)
5629 memcg->oom_group = oom_group;
5634 static struct cftype memory_files[] = {
5637 .flags = CFTYPE_NOT_ON_ROOT,
5638 .read_u64 = memory_current_read,
5642 .flags = CFTYPE_NOT_ON_ROOT,
5643 .seq_show = memory_min_show,
5644 .write = memory_min_write,
5648 .flags = CFTYPE_NOT_ON_ROOT,
5649 .seq_show = memory_low_show,
5650 .write = memory_low_write,
5654 .flags = CFTYPE_NOT_ON_ROOT,
5655 .seq_show = memory_high_show,
5656 .write = memory_high_write,
5660 .flags = CFTYPE_NOT_ON_ROOT,
5661 .seq_show = memory_max_show,
5662 .write = memory_max_write,
5666 .flags = CFTYPE_NOT_ON_ROOT,
5667 .file_offset = offsetof(struct mem_cgroup, events_file),
5668 .seq_show = memory_events_show,
5672 .flags = CFTYPE_NOT_ON_ROOT,
5673 .seq_show = memory_stat_show,
5676 .name = "oom.group",
5677 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5678 .seq_show = memory_oom_group_show,
5679 .write = memory_oom_group_write,
5684 struct cgroup_subsys memory_cgrp_subsys = {
5685 .css_alloc = mem_cgroup_css_alloc,
5686 .css_online = mem_cgroup_css_online,
5687 .css_offline = mem_cgroup_css_offline,
5688 .css_released = mem_cgroup_css_released,
5689 .css_free = mem_cgroup_css_free,
5690 .css_reset = mem_cgroup_css_reset,
5691 .can_attach = mem_cgroup_can_attach,
5692 .cancel_attach = mem_cgroup_cancel_attach,
5693 .post_attach = mem_cgroup_move_task,
5694 .bind = mem_cgroup_bind,
5695 .dfl_cftypes = memory_files,
5696 .legacy_cftypes = mem_cgroup_legacy_files,
5701 * mem_cgroup_protected - check if memory consumption is in the normal range
5702 * @root: the top ancestor of the sub-tree being checked
5703 * @memcg: the memory cgroup to check
5705 * WARNING: This function is not stateless! It can only be used as part
5706 * of a top-down tree iteration, not for isolated queries.
5708 * Returns one of the following:
5709 * MEMCG_PROT_NONE: cgroup memory is not protected
5710 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5711 * an unprotected supply of reclaimable memory from other cgroups.
5712 * MEMCG_PROT_MIN: cgroup memory is protected
5714 * @root is exclusive; it is never protected when looked at directly
5716 * To provide a proper hierarchical behavior, effective memory.min/low values
5717 * are used. Below is the description of how effective memory.low is calculated.
5718 * Effective memory.min values is calculated in the same way.
5720 * Effective memory.low is always equal or less than the original memory.low.
5721 * If there is no memory.low overcommittment (which is always true for
5722 * top-level memory cgroups), these two values are equal.
5723 * Otherwise, it's a part of parent's effective memory.low,
5724 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5725 * memory.low usages, where memory.low usage is the size of actually
5729 * elow = min( memory.low, parent->elow * ------------------ ),
5730 * siblings_low_usage
5732 * | memory.current, if memory.current < memory.low
5737 * Such definition of the effective memory.low provides the expected
5738 * hierarchical behavior: parent's memory.low value is limiting
5739 * children, unprotected memory is reclaimed first and cgroups,
5740 * which are not using their guarantee do not affect actual memory
5743 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5745 * A A/memory.low = 2G, A/memory.current = 6G
5747 * BC DE B/memory.low = 3G B/memory.current = 2G
5748 * C/memory.low = 1G C/memory.current = 2G
5749 * D/memory.low = 0 D/memory.current = 2G
5750 * E/memory.low = 10G E/memory.current = 0
5752 * and the memory pressure is applied, the following memory distribution
5753 * is expected (approximately):
5755 * A/memory.current = 2G
5757 * B/memory.current = 1.3G
5758 * C/memory.current = 0.6G
5759 * D/memory.current = 0
5760 * E/memory.current = 0
5762 * These calculations require constant tracking of the actual low usages
5763 * (see propagate_protected_usage()), as well as recursive calculation of
5764 * effective memory.low values. But as we do call mem_cgroup_protected()
5765 * path for each memory cgroup top-down from the reclaim,
5766 * it's possible to optimize this part, and save calculated elow
5767 * for next usage. This part is intentionally racy, but it's ok,
5768 * as memory.low is a best-effort mechanism.
5770 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5771 struct mem_cgroup *memcg)
5773 struct mem_cgroup *parent;
5774 unsigned long emin, parent_emin;
5775 unsigned long elow, parent_elow;
5776 unsigned long usage;
5778 if (mem_cgroup_disabled())
5779 return MEMCG_PROT_NONE;
5782 root = root_mem_cgroup;
5784 return MEMCG_PROT_NONE;
5786 usage = page_counter_read(&memcg->memory);
5788 return MEMCG_PROT_NONE;
5790 emin = memcg->memory.min;
5791 elow = memcg->memory.low;
5793 parent = parent_mem_cgroup(memcg);
5794 /* No parent means a non-hierarchical mode on v1 memcg */
5796 return MEMCG_PROT_NONE;
5801 parent_emin = READ_ONCE(parent->memory.emin);
5802 emin = min(emin, parent_emin);
5803 if (emin && parent_emin) {
5804 unsigned long min_usage, siblings_min_usage;
5806 min_usage = min(usage, memcg->memory.min);
5807 siblings_min_usage = atomic_long_read(
5808 &parent->memory.children_min_usage);
5810 if (min_usage && siblings_min_usage)
5811 emin = min(emin, parent_emin * min_usage /
5812 siblings_min_usage);
5815 parent_elow = READ_ONCE(parent->memory.elow);
5816 elow = min(elow, parent_elow);
5817 if (elow && parent_elow) {
5818 unsigned long low_usage, siblings_low_usage;
5820 low_usage = min(usage, memcg->memory.low);
5821 siblings_low_usage = atomic_long_read(
5822 &parent->memory.children_low_usage);
5824 if (low_usage && siblings_low_usage)
5825 elow = min(elow, parent_elow * low_usage /
5826 siblings_low_usage);
5830 memcg->memory.emin = emin;
5831 memcg->memory.elow = elow;
5834 return MEMCG_PROT_MIN;
5835 else if (usage <= elow)
5836 return MEMCG_PROT_LOW;
5838 return MEMCG_PROT_NONE;
5842 * mem_cgroup_try_charge - try charging a page
5843 * @page: page to charge
5844 * @mm: mm context of the victim
5845 * @gfp_mask: reclaim mode
5846 * @memcgp: charged memcg return
5847 * @compound: charge the page as compound or small page
5849 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5850 * pages according to @gfp_mask if necessary.
5852 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5853 * Otherwise, an error code is returned.
5855 * After page->mapping has been set up, the caller must finalize the
5856 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5857 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5859 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5860 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5863 struct mem_cgroup *memcg = NULL;
5864 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5867 if (mem_cgroup_disabled())
5870 if (PageSwapCache(page)) {
5872 * Every swap fault against a single page tries to charge the
5873 * page, bail as early as possible. shmem_unuse() encounters
5874 * already charged pages, too. The USED bit is protected by
5875 * the page lock, which serializes swap cache removal, which
5876 * in turn serializes uncharging.
5878 VM_BUG_ON_PAGE(!PageLocked(page), page);
5879 if (compound_head(page)->mem_cgroup)
5882 if (do_swap_account) {
5883 swp_entry_t ent = { .val = page_private(page), };
5884 unsigned short id = lookup_swap_cgroup_id(ent);
5887 memcg = mem_cgroup_from_id(id);
5888 if (memcg && !css_tryget_online(&memcg->css))
5895 memcg = get_mem_cgroup_from_mm(mm);
5897 ret = try_charge(memcg, gfp_mask, nr_pages);
5899 css_put(&memcg->css);
5905 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
5906 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5909 struct mem_cgroup *memcg;
5912 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
5914 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
5919 * mem_cgroup_commit_charge - commit a page charge
5920 * @page: page to charge
5921 * @memcg: memcg to charge the page to
5922 * @lrucare: page might be on LRU already
5923 * @compound: charge the page as compound or small page
5925 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5926 * after page->mapping has been set up. This must happen atomically
5927 * as part of the page instantiation, i.e. under the page table lock
5928 * for anonymous pages, under the page lock for page and swap cache.
5930 * In addition, the page must not be on the LRU during the commit, to
5931 * prevent racing with task migration. If it might be, use @lrucare.
5933 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5935 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5936 bool lrucare, bool compound)
5938 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5940 VM_BUG_ON_PAGE(!page->mapping, page);
5941 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5943 if (mem_cgroup_disabled())
5946 * Swap faults will attempt to charge the same page multiple
5947 * times. But reuse_swap_page() might have removed the page
5948 * from swapcache already, so we can't check PageSwapCache().
5953 commit_charge(page, memcg, lrucare);
5955 local_irq_disable();
5956 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5957 memcg_check_events(memcg, page);
5960 if (do_memsw_account() && PageSwapCache(page)) {
5961 swp_entry_t entry = { .val = page_private(page) };
5963 * The swap entry might not get freed for a long time,
5964 * let's not wait for it. The page already received a
5965 * memory+swap charge, drop the swap entry duplicate.
5967 mem_cgroup_uncharge_swap(entry, nr_pages);
5972 * mem_cgroup_cancel_charge - cancel a page charge
5973 * @page: page to charge
5974 * @memcg: memcg to charge the page to
5975 * @compound: charge the page as compound or small page
5977 * Cancel a charge transaction started by mem_cgroup_try_charge().
5979 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5982 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5984 if (mem_cgroup_disabled())
5987 * Swap faults will attempt to charge the same page multiple
5988 * times. But reuse_swap_page() might have removed the page
5989 * from swapcache already, so we can't check PageSwapCache().
5994 cancel_charge(memcg, nr_pages);
5997 struct uncharge_gather {
5998 struct mem_cgroup *memcg;
5999 unsigned long pgpgout;
6000 unsigned long nr_anon;
6001 unsigned long nr_file;
6002 unsigned long nr_kmem;
6003 unsigned long nr_huge;
6004 unsigned long nr_shmem;
6005 struct page *dummy_page;
6008 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6010 memset(ug, 0, sizeof(*ug));
6013 static void uncharge_batch(const struct uncharge_gather *ug)
6015 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6016 unsigned long flags;
6018 if (!mem_cgroup_is_root(ug->memcg)) {
6019 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6020 if (do_memsw_account())
6021 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6022 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6023 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6024 memcg_oom_recover(ug->memcg);
6027 local_irq_save(flags);
6028 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6029 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6030 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6031 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6032 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6033 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
6034 memcg_check_events(ug->memcg, ug->dummy_page);
6035 local_irq_restore(flags);
6037 if (!mem_cgroup_is_root(ug->memcg))
6038 css_put_many(&ug->memcg->css, nr_pages);
6041 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6043 VM_BUG_ON_PAGE(PageLRU(page), page);
6044 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6045 !PageHWPoison(page) , page);
6047 if (!page->mem_cgroup)
6051 * Nobody should be changing or seriously looking at
6052 * page->mem_cgroup at this point, we have fully
6053 * exclusive access to the page.
6056 if (ug->memcg != page->mem_cgroup) {
6059 uncharge_gather_clear(ug);
6061 ug->memcg = page->mem_cgroup;
6064 if (!PageKmemcg(page)) {
6065 unsigned int nr_pages = 1;
6067 if (PageTransHuge(page)) {
6068 nr_pages <<= compound_order(page);
6069 ug->nr_huge += nr_pages;
6072 ug->nr_anon += nr_pages;
6074 ug->nr_file += nr_pages;
6075 if (PageSwapBacked(page))
6076 ug->nr_shmem += nr_pages;
6080 ug->nr_kmem += 1 << compound_order(page);
6081 __ClearPageKmemcg(page);
6084 ug->dummy_page = page;
6085 page->mem_cgroup = NULL;
6088 static void uncharge_list(struct list_head *page_list)
6090 struct uncharge_gather ug;
6091 struct list_head *next;
6093 uncharge_gather_clear(&ug);
6096 * Note that the list can be a single page->lru; hence the
6097 * do-while loop instead of a simple list_for_each_entry().
6099 next = page_list->next;
6103 page = list_entry(next, struct page, lru);
6104 next = page->lru.next;
6106 uncharge_page(page, &ug);
6107 } while (next != page_list);
6110 uncharge_batch(&ug);
6114 * mem_cgroup_uncharge - uncharge a page
6115 * @page: page to uncharge
6117 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6118 * mem_cgroup_commit_charge().
6120 void mem_cgroup_uncharge(struct page *page)
6122 struct uncharge_gather ug;
6124 if (mem_cgroup_disabled())
6127 /* Don't touch page->lru of any random page, pre-check: */
6128 if (!page->mem_cgroup)
6131 uncharge_gather_clear(&ug);
6132 uncharge_page(page, &ug);
6133 uncharge_batch(&ug);
6137 * mem_cgroup_uncharge_list - uncharge a list of page
6138 * @page_list: list of pages to uncharge
6140 * Uncharge a list of pages previously charged with
6141 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6143 void mem_cgroup_uncharge_list(struct list_head *page_list)
6145 if (mem_cgroup_disabled())
6148 if (!list_empty(page_list))
6149 uncharge_list(page_list);
6153 * mem_cgroup_migrate - charge a page's replacement
6154 * @oldpage: currently circulating page
6155 * @newpage: replacement page
6157 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6158 * be uncharged upon free.
6160 * Both pages must be locked, @newpage->mapping must be set up.
6162 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6164 struct mem_cgroup *memcg;
6165 unsigned int nr_pages;
6167 unsigned long flags;
6169 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6170 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6171 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6172 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6175 if (mem_cgroup_disabled())
6178 /* Page cache replacement: new page already charged? */
6179 if (newpage->mem_cgroup)
6182 /* Swapcache readahead pages can get replaced before being charged */
6183 memcg = oldpage->mem_cgroup;
6187 /* Force-charge the new page. The old one will be freed soon */
6188 compound = PageTransHuge(newpage);
6189 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6191 page_counter_charge(&memcg->memory, nr_pages);
6192 if (do_memsw_account())
6193 page_counter_charge(&memcg->memsw, nr_pages);
6194 css_get_many(&memcg->css, nr_pages);
6196 commit_charge(newpage, memcg, false);
6198 local_irq_save(flags);
6199 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6200 memcg_check_events(memcg, newpage);
6201 local_irq_restore(flags);
6204 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6205 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6207 void mem_cgroup_sk_alloc(struct sock *sk)
6209 struct mem_cgroup *memcg;
6211 if (!mem_cgroup_sockets_enabled)
6215 * Socket cloning can throw us here with sk_memcg already
6216 * filled. It won't however, necessarily happen from
6217 * process context. So the test for root memcg given
6218 * the current task's memcg won't help us in this case.
6220 * Respecting the original socket's memcg is a better
6221 * decision in this case.
6224 css_get(&sk->sk_memcg->css);
6229 memcg = mem_cgroup_from_task(current);
6230 if (memcg == root_mem_cgroup)
6232 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6234 if (css_tryget_online(&memcg->css))
6235 sk->sk_memcg = memcg;
6240 void mem_cgroup_sk_free(struct sock *sk)
6243 css_put(&sk->sk_memcg->css);
6247 * mem_cgroup_charge_skmem - charge socket memory
6248 * @memcg: memcg to charge
6249 * @nr_pages: number of pages to charge
6251 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6252 * @memcg's configured limit, %false if the charge had to be forced.
6254 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6256 gfp_t gfp_mask = GFP_KERNEL;
6258 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6259 struct page_counter *fail;
6261 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6262 memcg->tcpmem_pressure = 0;
6265 page_counter_charge(&memcg->tcpmem, nr_pages);
6266 memcg->tcpmem_pressure = 1;
6270 /* Don't block in the packet receive path */
6272 gfp_mask = GFP_NOWAIT;
6274 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6276 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6279 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6284 * mem_cgroup_uncharge_skmem - uncharge socket memory
6285 * @memcg: memcg to uncharge
6286 * @nr_pages: number of pages to uncharge
6288 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6290 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6291 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6295 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6297 refill_stock(memcg, nr_pages);
6300 static int __init cgroup_memory(char *s)
6304 while ((token = strsep(&s, ",")) != NULL) {
6307 if (!strcmp(token, "nosocket"))
6308 cgroup_memory_nosocket = true;
6309 if (!strcmp(token, "nokmem"))
6310 cgroup_memory_nokmem = true;
6314 __setup("cgroup.memory=", cgroup_memory);
6317 * subsys_initcall() for memory controller.
6319 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6320 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6321 * basically everything that doesn't depend on a specific mem_cgroup structure
6322 * should be initialized from here.
6324 static int __init mem_cgroup_init(void)
6328 #ifdef CONFIG_MEMCG_KMEM
6330 * Kmem cache creation is mostly done with the slab_mutex held,
6331 * so use a workqueue with limited concurrency to avoid stalling
6332 * all worker threads in case lots of cgroups are created and
6333 * destroyed simultaneously.
6335 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6336 BUG_ON(!memcg_kmem_cache_wq);
6339 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6340 memcg_hotplug_cpu_dead);
6342 for_each_possible_cpu(cpu)
6343 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6346 for_each_node(node) {
6347 struct mem_cgroup_tree_per_node *rtpn;
6349 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6350 node_online(node) ? node : NUMA_NO_NODE);
6352 rtpn->rb_root = RB_ROOT;
6353 rtpn->rb_rightmost = NULL;
6354 spin_lock_init(&rtpn->lock);
6355 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6360 subsys_initcall(mem_cgroup_init);
6362 #ifdef CONFIG_MEMCG_SWAP
6363 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6365 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6367 * The root cgroup cannot be destroyed, so it's refcount must
6370 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6374 memcg = parent_mem_cgroup(memcg);
6376 memcg = root_mem_cgroup;
6382 * mem_cgroup_swapout - transfer a memsw charge to swap
6383 * @page: page whose memsw charge to transfer
6384 * @entry: swap entry to move the charge to
6386 * Transfer the memsw charge of @page to @entry.
6388 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6390 struct mem_cgroup *memcg, *swap_memcg;
6391 unsigned int nr_entries;
6392 unsigned short oldid;
6394 VM_BUG_ON_PAGE(PageLRU(page), page);
6395 VM_BUG_ON_PAGE(page_count(page), page);
6397 if (!do_memsw_account())
6400 memcg = page->mem_cgroup;
6402 /* Readahead page, never charged */
6407 * In case the memcg owning these pages has been offlined and doesn't
6408 * have an ID allocated to it anymore, charge the closest online
6409 * ancestor for the swap instead and transfer the memory+swap charge.
6411 swap_memcg = mem_cgroup_id_get_online(memcg);
6412 nr_entries = hpage_nr_pages(page);
6413 /* Get references for the tail pages, too */
6415 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6416 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6418 VM_BUG_ON_PAGE(oldid, page);
6419 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6421 page->mem_cgroup = NULL;
6423 if (!mem_cgroup_is_root(memcg))
6424 page_counter_uncharge(&memcg->memory, nr_entries);
6426 if (memcg != swap_memcg) {
6427 if (!mem_cgroup_is_root(swap_memcg))
6428 page_counter_charge(&swap_memcg->memsw, nr_entries);
6429 page_counter_uncharge(&memcg->memsw, nr_entries);
6433 * Interrupts should be disabled here because the caller holds the
6434 * i_pages lock which is taken with interrupts-off. It is
6435 * important here to have the interrupts disabled because it is the
6436 * only synchronisation we have for updating the per-CPU variables.
6438 VM_BUG_ON(!irqs_disabled());
6439 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6441 memcg_check_events(memcg, page);
6443 if (!mem_cgroup_is_root(memcg))
6444 css_put_many(&memcg->css, nr_entries);
6448 * mem_cgroup_try_charge_swap - try charging swap space for a page
6449 * @page: page being added to swap
6450 * @entry: swap entry to charge
6452 * Try to charge @page's memcg for the swap space at @entry.
6454 * Returns 0 on success, -ENOMEM on failure.
6456 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6458 unsigned int nr_pages = hpage_nr_pages(page);
6459 struct page_counter *counter;
6460 struct mem_cgroup *memcg;
6461 unsigned short oldid;
6463 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6466 memcg = page->mem_cgroup;
6468 /* Readahead page, never charged */
6473 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6477 memcg = mem_cgroup_id_get_online(memcg);
6479 if (!mem_cgroup_is_root(memcg) &&
6480 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6481 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6482 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6483 mem_cgroup_id_put(memcg);
6487 /* Get references for the tail pages, too */
6489 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6490 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6491 VM_BUG_ON_PAGE(oldid, page);
6492 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6498 * mem_cgroup_uncharge_swap - uncharge swap space
6499 * @entry: swap entry to uncharge
6500 * @nr_pages: the amount of swap space to uncharge
6502 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6504 struct mem_cgroup *memcg;
6507 if (!do_swap_account)
6510 id = swap_cgroup_record(entry, 0, nr_pages);
6512 memcg = mem_cgroup_from_id(id);
6514 if (!mem_cgroup_is_root(memcg)) {
6515 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6516 page_counter_uncharge(&memcg->swap, nr_pages);
6518 page_counter_uncharge(&memcg->memsw, nr_pages);
6520 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6521 mem_cgroup_id_put_many(memcg, nr_pages);
6526 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6528 long nr_swap_pages = get_nr_swap_pages();
6530 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6531 return nr_swap_pages;
6532 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6533 nr_swap_pages = min_t(long, nr_swap_pages,
6534 READ_ONCE(memcg->swap.max) -
6535 page_counter_read(&memcg->swap));
6536 return nr_swap_pages;
6539 bool mem_cgroup_swap_full(struct page *page)
6541 struct mem_cgroup *memcg;
6543 VM_BUG_ON_PAGE(!PageLocked(page), page);
6547 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6550 memcg = page->mem_cgroup;
6554 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6555 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6561 /* for remember boot option*/
6562 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6563 static int really_do_swap_account __initdata = 1;
6565 static int really_do_swap_account __initdata;
6568 static int __init enable_swap_account(char *s)
6570 if (!strcmp(s, "1"))
6571 really_do_swap_account = 1;
6572 else if (!strcmp(s, "0"))
6573 really_do_swap_account = 0;
6576 __setup("swapaccount=", enable_swap_account);
6578 static u64 swap_current_read(struct cgroup_subsys_state *css,
6581 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6583 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6586 static int swap_max_show(struct seq_file *m, void *v)
6588 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6589 unsigned long max = READ_ONCE(memcg->swap.max);
6591 if (max == PAGE_COUNTER_MAX)
6592 seq_puts(m, "max\n");
6594 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6599 static ssize_t swap_max_write(struct kernfs_open_file *of,
6600 char *buf, size_t nbytes, loff_t off)
6602 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6606 buf = strstrip(buf);
6607 err = page_counter_memparse(buf, "max", &max);
6611 xchg(&memcg->swap.max, max);
6616 static int swap_events_show(struct seq_file *m, void *v)
6618 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6620 seq_printf(m, "max %lu\n",
6621 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6622 seq_printf(m, "fail %lu\n",
6623 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6628 static struct cftype swap_files[] = {
6630 .name = "swap.current",
6631 .flags = CFTYPE_NOT_ON_ROOT,
6632 .read_u64 = swap_current_read,
6636 .flags = CFTYPE_NOT_ON_ROOT,
6637 .seq_show = swap_max_show,
6638 .write = swap_max_write,
6641 .name = "swap.events",
6642 .flags = CFTYPE_NOT_ON_ROOT,
6643 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6644 .seq_show = swap_events_show,
6649 static struct cftype memsw_cgroup_files[] = {
6651 .name = "memsw.usage_in_bytes",
6652 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6653 .read_u64 = mem_cgroup_read_u64,
6656 .name = "memsw.max_usage_in_bytes",
6657 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6658 .write = mem_cgroup_reset,
6659 .read_u64 = mem_cgroup_read_u64,
6662 .name = "memsw.limit_in_bytes",
6663 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6664 .write = mem_cgroup_write,
6665 .read_u64 = mem_cgroup_read_u64,
6668 .name = "memsw.failcnt",
6669 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6670 .write = mem_cgroup_reset,
6671 .read_u64 = mem_cgroup_read_u64,
6673 { }, /* terminate */
6676 static int __init mem_cgroup_swap_init(void)
6678 if (!mem_cgroup_disabled() && really_do_swap_account) {
6679 do_swap_account = 1;
6680 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6682 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6683 memsw_cgroup_files));
6687 subsys_initcall(mem_cgroup_swap_init);
6689 #endif /* CONFIG_MEMCG_SWAP */