1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/sched/mm.h>
39 #include <linux/shmem_fs.h>
40 #include <linux/hugetlb.h>
41 #include <linux/pagemap.h>
42 #include <linux/smp.h>
43 #include <linux/page-flags.h>
44 #include <linux/backing-dev.h>
45 #include <linux/bit_spinlock.h>
46 #include <linux/rcupdate.h>
47 #include <linux/limits.h>
48 #include <linux/export.h>
49 #include <linux/mutex.h>
50 #include <linux/rbtree.h>
51 #include <linux/slab.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/spinlock.h>
55 #include <linux/eventfd.h>
56 #include <linux/poll.h>
57 #include <linux/sort.h>
59 #include <linux/seq_file.h>
60 #include <linux/vmpressure.h>
61 #include <linux/mm_inline.h>
62 #include <linux/swap_cgroup.h>
63 #include <linux/cpu.h>
64 #include <linux/oom.h>
65 #include <linux/lockdep.h>
66 #include <linux/file.h>
67 #include <linux/tracehook.h>
73 #include <linux/uaccess.h>
75 #include <trace/events/vmscan.h>
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
78 EXPORT_SYMBOL(memory_cgrp_subsys);
80 struct mem_cgroup *root_mem_cgroup __read_mostly;
82 #define MEM_CGROUP_RECLAIM_RETRIES 5
84 /* Socket memory accounting disabled? */
85 static bool cgroup_memory_nosocket;
87 /* Kernel memory accounting disabled? */
88 static bool cgroup_memory_nokmem;
90 /* Whether the swap controller is active */
91 #ifdef CONFIG_MEMCG_SWAP
92 int do_swap_account __read_mostly;
94 #define do_swap_account 0
97 /* Whether legacy memory+swap accounting is active */
98 static bool do_memsw_account(void)
100 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
103 static const char *const mem_cgroup_lru_names[] = {
111 #define THRESHOLDS_EVENTS_TARGET 128
112 #define SOFTLIMIT_EVENTS_TARGET 1024
113 #define NUMAINFO_EVENTS_TARGET 1024
116 * Cgroups above their limits are maintained in a RB-Tree, independent of
117 * their hierarchy representation
120 struct mem_cgroup_tree_per_node {
121 struct rb_root rb_root;
122 struct rb_node *rb_rightmost;
126 struct mem_cgroup_tree {
127 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
130 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
133 struct mem_cgroup_eventfd_list {
134 struct list_head list;
135 struct eventfd_ctx *eventfd;
139 * cgroup_event represents events which userspace want to receive.
141 struct mem_cgroup_event {
143 * memcg which the event belongs to.
145 struct mem_cgroup *memcg;
147 * eventfd to signal userspace about the event.
149 struct eventfd_ctx *eventfd;
151 * Each of these stored in a list by the cgroup.
153 struct list_head list;
155 * register_event() callback will be used to add new userspace
156 * waiter for changes related to this event. Use eventfd_signal()
157 * on eventfd to send notification to userspace.
159 int (*register_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd, const char *args);
162 * unregister_event() callback will be called when userspace closes
163 * the eventfd or on cgroup removing. This callback must be set,
164 * if you want provide notification functionality.
166 void (*unregister_event)(struct mem_cgroup *memcg,
167 struct eventfd_ctx *eventfd);
169 * All fields below needed to unregister event when
170 * userspace closes eventfd.
173 wait_queue_head_t *wqh;
174 wait_queue_entry_t wait;
175 struct work_struct remove;
178 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
179 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
181 /* Stuffs for move charges at task migration. */
183 * Types of charges to be moved.
185 #define MOVE_ANON 0x1U
186 #define MOVE_FILE 0x2U
187 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
189 /* "mc" and its members are protected by cgroup_mutex */
190 static struct move_charge_struct {
191 spinlock_t lock; /* for from, to */
192 struct mm_struct *mm;
193 struct mem_cgroup *from;
194 struct mem_cgroup *to;
196 unsigned long precharge;
197 unsigned long moved_charge;
198 unsigned long moved_swap;
199 struct task_struct *moving_task; /* a task moving charges */
200 wait_queue_head_t waitq; /* a waitq for other context */
202 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
203 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
207 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
208 * limit reclaim to prevent infinite loops, if they ever occur.
210 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
211 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
214 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
215 MEM_CGROUP_CHARGE_TYPE_ANON,
216 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
217 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
221 /* for encoding cft->private value on file */
230 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
231 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
232 #define MEMFILE_ATTR(val) ((val) & 0xffff)
233 /* Used for OOM nofiier */
234 #define OOM_CONTROL (0)
237 * Iteration constructs for visiting all cgroups (under a tree). If
238 * loops are exited prematurely (break), mem_cgroup_iter_break() must
239 * be used for reference counting.
241 #define for_each_mem_cgroup_tree(iter, root) \
242 for (iter = mem_cgroup_iter(root, NULL, NULL); \
244 iter = mem_cgroup_iter(root, iter, NULL))
246 #define for_each_mem_cgroup(iter) \
247 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
249 iter = mem_cgroup_iter(NULL, iter, NULL))
251 static inline bool should_force_charge(void)
253 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
254 (current->flags & PF_EXITING);
257 /* Some nice accessors for the vmpressure. */
258 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
261 memcg = root_mem_cgroup;
262 return &memcg->vmpressure;
265 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
267 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
270 #ifdef CONFIG_MEMCG_KMEM
272 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
273 * The main reason for not using cgroup id for this:
274 * this works better in sparse environments, where we have a lot of memcgs,
275 * but only a few kmem-limited. Or also, if we have, for instance, 200
276 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
277 * 200 entry array for that.
279 * The current size of the caches array is stored in memcg_nr_cache_ids. It
280 * will double each time we have to increase it.
282 static DEFINE_IDA(memcg_cache_ida);
283 int memcg_nr_cache_ids;
285 /* Protects memcg_nr_cache_ids */
286 static DECLARE_RWSEM(memcg_cache_ids_sem);
288 void memcg_get_cache_ids(void)
290 down_read(&memcg_cache_ids_sem);
293 void memcg_put_cache_ids(void)
295 up_read(&memcg_cache_ids_sem);
299 * MIN_SIZE is different than 1, because we would like to avoid going through
300 * the alloc/free process all the time. In a small machine, 4 kmem-limited
301 * cgroups is a reasonable guess. In the future, it could be a parameter or
302 * tunable, but that is strictly not necessary.
304 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
305 * this constant directly from cgroup, but it is understandable that this is
306 * better kept as an internal representation in cgroup.c. In any case, the
307 * cgrp_id space is not getting any smaller, and we don't have to necessarily
308 * increase ours as well if it increases.
310 #define MEMCG_CACHES_MIN_SIZE 4
311 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
314 * A lot of the calls to the cache allocation functions are expected to be
315 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
316 * conditional to this static branch, we'll have to allow modules that does
317 * kmem_cache_alloc and the such to see this symbol as well
319 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
320 EXPORT_SYMBOL(memcg_kmem_enabled_key);
322 struct workqueue_struct *memcg_kmem_cache_wq;
324 static int memcg_shrinker_map_size;
325 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
327 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
329 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
332 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
333 int size, int old_size)
335 struct memcg_shrinker_map *new, *old;
338 lockdep_assert_held(&memcg_shrinker_map_mutex);
341 old = rcu_dereference_protected(
342 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
343 /* Not yet online memcg */
347 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
351 /* Set all old bits, clear all new bits */
352 memset(new->map, (int)0xff, old_size);
353 memset((void *)new->map + old_size, 0, size - old_size);
355 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
356 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
362 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
364 struct mem_cgroup_per_node *pn;
365 struct memcg_shrinker_map *map;
368 if (mem_cgroup_is_root(memcg))
372 pn = mem_cgroup_nodeinfo(memcg, nid);
373 map = rcu_dereference_protected(pn->shrinker_map, true);
376 rcu_assign_pointer(pn->shrinker_map, NULL);
380 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
382 struct memcg_shrinker_map *map;
383 int nid, size, ret = 0;
385 if (mem_cgroup_is_root(memcg))
388 mutex_lock(&memcg_shrinker_map_mutex);
389 size = memcg_shrinker_map_size;
391 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
393 memcg_free_shrinker_maps(memcg);
397 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
399 mutex_unlock(&memcg_shrinker_map_mutex);
404 int memcg_expand_shrinker_maps(int new_id)
406 int size, old_size, ret = 0;
407 struct mem_cgroup *memcg;
409 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
410 old_size = memcg_shrinker_map_size;
411 if (size <= old_size)
414 mutex_lock(&memcg_shrinker_map_mutex);
415 if (!root_mem_cgroup)
418 for_each_mem_cgroup(memcg) {
419 if (mem_cgroup_is_root(memcg))
421 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
427 memcg_shrinker_map_size = size;
428 mutex_unlock(&memcg_shrinker_map_mutex);
432 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
434 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
435 struct memcg_shrinker_map *map;
438 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
439 /* Pairs with smp mb in shrink_slab() */
440 smp_mb__before_atomic();
441 set_bit(shrinker_id, map->map);
446 #else /* CONFIG_MEMCG_KMEM */
447 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
451 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
452 #endif /* CONFIG_MEMCG_KMEM */
455 * mem_cgroup_css_from_page - css of the memcg associated with a page
456 * @page: page of interest
458 * If memcg is bound to the default hierarchy, css of the memcg associated
459 * with @page is returned. The returned css remains associated with @page
460 * until it is released.
462 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
465 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
467 struct mem_cgroup *memcg;
469 memcg = page->mem_cgroup;
471 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
472 memcg = root_mem_cgroup;
478 * page_cgroup_ino - return inode number of the memcg a page is charged to
481 * Look up the closest online ancestor of the memory cgroup @page is charged to
482 * and return its inode number or 0 if @page is not charged to any cgroup. It
483 * is safe to call this function without holding a reference to @page.
485 * Note, this function is inherently racy, because there is nothing to prevent
486 * the cgroup inode from getting torn down and potentially reallocated a moment
487 * after page_cgroup_ino() returns, so it only should be used by callers that
488 * do not care (such as procfs interfaces).
490 ino_t page_cgroup_ino(struct page *page)
492 struct mem_cgroup *memcg;
493 unsigned long ino = 0;
496 memcg = READ_ONCE(page->mem_cgroup);
497 while (memcg && !(memcg->css.flags & CSS_ONLINE))
498 memcg = parent_mem_cgroup(memcg);
500 ino = cgroup_ino(memcg->css.cgroup);
505 static struct mem_cgroup_per_node *
506 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
508 int nid = page_to_nid(page);
510 return memcg->nodeinfo[nid];
513 static struct mem_cgroup_tree_per_node *
514 soft_limit_tree_node(int nid)
516 return soft_limit_tree.rb_tree_per_node[nid];
519 static struct mem_cgroup_tree_per_node *
520 soft_limit_tree_from_page(struct page *page)
522 int nid = page_to_nid(page);
524 return soft_limit_tree.rb_tree_per_node[nid];
527 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
528 struct mem_cgroup_tree_per_node *mctz,
529 unsigned long new_usage_in_excess)
531 struct rb_node **p = &mctz->rb_root.rb_node;
532 struct rb_node *parent = NULL;
533 struct mem_cgroup_per_node *mz_node;
534 bool rightmost = true;
539 mz->usage_in_excess = new_usage_in_excess;
540 if (!mz->usage_in_excess)
544 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
546 if (mz->usage_in_excess < mz_node->usage_in_excess) {
552 * We can't avoid mem cgroups that are over their soft
553 * limit by the same amount
555 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
560 mctz->rb_rightmost = &mz->tree_node;
562 rb_link_node(&mz->tree_node, parent, p);
563 rb_insert_color(&mz->tree_node, &mctz->rb_root);
567 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
568 struct mem_cgroup_tree_per_node *mctz)
573 if (&mz->tree_node == mctz->rb_rightmost)
574 mctz->rb_rightmost = rb_prev(&mz->tree_node);
576 rb_erase(&mz->tree_node, &mctz->rb_root);
580 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
581 struct mem_cgroup_tree_per_node *mctz)
585 spin_lock_irqsave(&mctz->lock, flags);
586 __mem_cgroup_remove_exceeded(mz, mctz);
587 spin_unlock_irqrestore(&mctz->lock, flags);
590 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
592 unsigned long nr_pages = page_counter_read(&memcg->memory);
593 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
594 unsigned long excess = 0;
596 if (nr_pages > soft_limit)
597 excess = nr_pages - soft_limit;
602 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
604 unsigned long excess;
605 struct mem_cgroup_per_node *mz;
606 struct mem_cgroup_tree_per_node *mctz;
608 mctz = soft_limit_tree_from_page(page);
612 * Necessary to update all ancestors when hierarchy is used.
613 * because their event counter is not touched.
615 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
616 mz = mem_cgroup_page_nodeinfo(memcg, page);
617 excess = soft_limit_excess(memcg);
619 * We have to update the tree if mz is on RB-tree or
620 * mem is over its softlimit.
622 if (excess || mz->on_tree) {
625 spin_lock_irqsave(&mctz->lock, flags);
626 /* if on-tree, remove it */
628 __mem_cgroup_remove_exceeded(mz, mctz);
630 * Insert again. mz->usage_in_excess will be updated.
631 * If excess is 0, no tree ops.
633 __mem_cgroup_insert_exceeded(mz, mctz, excess);
634 spin_unlock_irqrestore(&mctz->lock, flags);
639 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
641 struct mem_cgroup_tree_per_node *mctz;
642 struct mem_cgroup_per_node *mz;
646 mz = mem_cgroup_nodeinfo(memcg, nid);
647 mctz = soft_limit_tree_node(nid);
649 mem_cgroup_remove_exceeded(mz, mctz);
653 static struct mem_cgroup_per_node *
654 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
656 struct mem_cgroup_per_node *mz;
660 if (!mctz->rb_rightmost)
661 goto done; /* Nothing to reclaim from */
663 mz = rb_entry(mctz->rb_rightmost,
664 struct mem_cgroup_per_node, tree_node);
666 * Remove the node now but someone else can add it back,
667 * we will to add it back at the end of reclaim to its correct
668 * position in the tree.
670 __mem_cgroup_remove_exceeded(mz, mctz);
671 if (!soft_limit_excess(mz->memcg) ||
672 !css_tryget_online(&mz->memcg->css))
678 static struct mem_cgroup_per_node *
679 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
681 struct mem_cgroup_per_node *mz;
683 spin_lock_irq(&mctz->lock);
684 mz = __mem_cgroup_largest_soft_limit_node(mctz);
685 spin_unlock_irq(&mctz->lock);
689 static unsigned long memcg_sum_events(struct mem_cgroup *memcg,
692 return atomic_long_read(&memcg->events[event]);
695 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
697 bool compound, int nr_pages)
700 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
701 * counted as CACHE even if it's on ANON LRU.
704 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
706 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
707 if (PageSwapBacked(page))
708 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
712 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
713 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
716 /* pagein of a big page is an event. So, ignore page size */
718 __count_memcg_events(memcg, PGPGIN, 1);
720 __count_memcg_events(memcg, PGPGOUT, 1);
721 nr_pages = -nr_pages; /* for event */
724 __this_cpu_add(memcg->stat_cpu->nr_page_events, nr_pages);
727 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
728 int nid, unsigned int lru_mask)
730 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
731 unsigned long nr = 0;
734 VM_BUG_ON((unsigned)nid >= nr_node_ids);
737 if (!(BIT(lru) & lru_mask))
739 nr += mem_cgroup_get_lru_size(lruvec, lru);
744 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
745 unsigned int lru_mask)
747 unsigned long nr = 0;
750 for_each_node_state(nid, N_MEMORY)
751 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
755 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
756 enum mem_cgroup_events_target target)
758 unsigned long val, next;
760 val = __this_cpu_read(memcg->stat_cpu->nr_page_events);
761 next = __this_cpu_read(memcg->stat_cpu->targets[target]);
762 /* from time_after() in jiffies.h */
763 if ((long)(next - val) < 0) {
765 case MEM_CGROUP_TARGET_THRESH:
766 next = val + THRESHOLDS_EVENTS_TARGET;
768 case MEM_CGROUP_TARGET_SOFTLIMIT:
769 next = val + SOFTLIMIT_EVENTS_TARGET;
771 case MEM_CGROUP_TARGET_NUMAINFO:
772 next = val + NUMAINFO_EVENTS_TARGET;
777 __this_cpu_write(memcg->stat_cpu->targets[target], next);
784 * Check events in order.
787 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
789 /* threshold event is triggered in finer grain than soft limit */
790 if (unlikely(mem_cgroup_event_ratelimit(memcg,
791 MEM_CGROUP_TARGET_THRESH))) {
793 bool do_numainfo __maybe_unused;
795 do_softlimit = mem_cgroup_event_ratelimit(memcg,
796 MEM_CGROUP_TARGET_SOFTLIMIT);
798 do_numainfo = mem_cgroup_event_ratelimit(memcg,
799 MEM_CGROUP_TARGET_NUMAINFO);
801 mem_cgroup_threshold(memcg);
802 if (unlikely(do_softlimit))
803 mem_cgroup_update_tree(memcg, page);
805 if (unlikely(do_numainfo))
806 atomic_inc(&memcg->numainfo_events);
811 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
814 * mm_update_next_owner() may clear mm->owner to NULL
815 * if it races with swapoff, page migration, etc.
816 * So this can be called with p == NULL.
821 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
823 EXPORT_SYMBOL(mem_cgroup_from_task);
826 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
827 * @mm: mm from which memcg should be extracted. It can be NULL.
829 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
830 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
833 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
835 struct mem_cgroup *memcg;
837 if (mem_cgroup_disabled())
843 * Page cache insertions can happen withou an
844 * actual mm context, e.g. during disk probing
845 * on boot, loopback IO, acct() writes etc.
848 memcg = root_mem_cgroup;
850 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
851 if (unlikely(!memcg))
852 memcg = root_mem_cgroup;
854 } while (!css_tryget_online(&memcg->css));
858 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
861 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
862 * @page: page from which memcg should be extracted.
864 * Obtain a reference on page->memcg and returns it if successful. Otherwise
865 * root_mem_cgroup is returned.
867 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
869 struct mem_cgroup *memcg = page->mem_cgroup;
871 if (mem_cgroup_disabled())
875 if (!memcg || !css_tryget_online(&memcg->css))
876 memcg = root_mem_cgroup;
880 EXPORT_SYMBOL(get_mem_cgroup_from_page);
883 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
885 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
887 if (unlikely(current->active_memcg)) {
888 struct mem_cgroup *memcg = root_mem_cgroup;
891 if (css_tryget_online(¤t->active_memcg->css))
892 memcg = current->active_memcg;
896 return get_mem_cgroup_from_mm(current->mm);
900 * mem_cgroup_iter - iterate over memory cgroup hierarchy
901 * @root: hierarchy root
902 * @prev: previously returned memcg, NULL on first invocation
903 * @reclaim: cookie for shared reclaim walks, NULL for full walks
905 * Returns references to children of the hierarchy below @root, or
906 * @root itself, or %NULL after a full round-trip.
908 * Caller must pass the return value in @prev on subsequent
909 * invocations for reference counting, or use mem_cgroup_iter_break()
910 * to cancel a hierarchy walk before the round-trip is complete.
912 * Reclaimers can specify a node and a priority level in @reclaim to
913 * divide up the memcgs in the hierarchy among all concurrent
914 * reclaimers operating on the same node and priority.
916 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
917 struct mem_cgroup *prev,
918 struct mem_cgroup_reclaim_cookie *reclaim)
920 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
921 struct cgroup_subsys_state *css = NULL;
922 struct mem_cgroup *memcg = NULL;
923 struct mem_cgroup *pos = NULL;
925 if (mem_cgroup_disabled())
929 root = root_mem_cgroup;
931 if (prev && !reclaim)
934 if (!root->use_hierarchy && root != root_mem_cgroup) {
943 struct mem_cgroup_per_node *mz;
945 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
946 iter = &mz->iter[reclaim->priority];
948 if (prev && reclaim->generation != iter->generation)
952 pos = READ_ONCE(iter->position);
953 if (!pos || css_tryget(&pos->css))
956 * css reference reached zero, so iter->position will
957 * be cleared by ->css_released. However, we should not
958 * rely on this happening soon, because ->css_released
959 * is called from a work queue, and by busy-waiting we
960 * might block it. So we clear iter->position right
963 (void)cmpxchg(&iter->position, pos, NULL);
971 css = css_next_descendant_pre(css, &root->css);
974 * Reclaimers share the hierarchy walk, and a
975 * new one might jump in right at the end of
976 * the hierarchy - make sure they see at least
977 * one group and restart from the beginning.
985 * Verify the css and acquire a reference. The root
986 * is provided by the caller, so we know it's alive
987 * and kicking, and don't take an extra reference.
989 memcg = mem_cgroup_from_css(css);
991 if (css == &root->css)
1002 * The position could have already been updated by a competing
1003 * thread, so check that the value hasn't changed since we read
1004 * it to avoid reclaiming from the same cgroup twice.
1006 (void)cmpxchg(&iter->position, pos, memcg);
1014 reclaim->generation = iter->generation;
1020 if (prev && prev != root)
1021 css_put(&prev->css);
1027 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1028 * @root: hierarchy root
1029 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1031 void mem_cgroup_iter_break(struct mem_cgroup *root,
1032 struct mem_cgroup *prev)
1035 root = root_mem_cgroup;
1036 if (prev && prev != root)
1037 css_put(&prev->css);
1040 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1042 struct mem_cgroup *memcg = dead_memcg;
1043 struct mem_cgroup_reclaim_iter *iter;
1044 struct mem_cgroup_per_node *mz;
1048 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1049 for_each_node(nid) {
1050 mz = mem_cgroup_nodeinfo(memcg, nid);
1051 for (i = 0; i <= DEF_PRIORITY; i++) {
1052 iter = &mz->iter[i];
1053 cmpxchg(&iter->position,
1061 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1062 * @memcg: hierarchy root
1063 * @fn: function to call for each task
1064 * @arg: argument passed to @fn
1066 * This function iterates over tasks attached to @memcg or to any of its
1067 * descendants and calls @fn for each task. If @fn returns a non-zero
1068 * value, the function breaks the iteration loop and returns the value.
1069 * Otherwise, it will iterate over all tasks and return 0.
1071 * This function must not be called for the root memory cgroup.
1073 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1074 int (*fn)(struct task_struct *, void *), void *arg)
1076 struct mem_cgroup *iter;
1079 BUG_ON(memcg == root_mem_cgroup);
1081 for_each_mem_cgroup_tree(iter, memcg) {
1082 struct css_task_iter it;
1083 struct task_struct *task;
1085 css_task_iter_start(&iter->css, 0, &it);
1086 while (!ret && (task = css_task_iter_next(&it)))
1087 ret = fn(task, arg);
1088 css_task_iter_end(&it);
1090 mem_cgroup_iter_break(memcg, iter);
1098 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1100 * @pgdat: pgdat of the page
1102 * This function is only safe when following the LRU page isolation
1103 * and putback protocol: the LRU lock must be held, and the page must
1104 * either be PageLRU() or the caller must have isolated/allocated it.
1106 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1108 struct mem_cgroup_per_node *mz;
1109 struct mem_cgroup *memcg;
1110 struct lruvec *lruvec;
1112 if (mem_cgroup_disabled()) {
1113 lruvec = &pgdat->lruvec;
1117 memcg = page->mem_cgroup;
1119 * Swapcache readahead pages are added to the LRU - and
1120 * possibly migrated - before they are charged.
1123 memcg = root_mem_cgroup;
1125 mz = mem_cgroup_page_nodeinfo(memcg, page);
1126 lruvec = &mz->lruvec;
1129 * Since a node can be onlined after the mem_cgroup was created,
1130 * we have to be prepared to initialize lruvec->zone here;
1131 * and if offlined then reonlined, we need to reinitialize it.
1133 if (unlikely(lruvec->pgdat != pgdat))
1134 lruvec->pgdat = pgdat;
1139 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1140 * @lruvec: mem_cgroup per zone lru vector
1141 * @lru: index of lru list the page is sitting on
1142 * @zid: zone id of the accounted pages
1143 * @nr_pages: positive when adding or negative when removing
1145 * This function must be called under lru_lock, just before a page is added
1146 * to or just after a page is removed from an lru list (that ordering being
1147 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1149 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1150 int zid, int nr_pages)
1152 struct mem_cgroup_per_node *mz;
1153 unsigned long *lru_size;
1156 if (mem_cgroup_disabled())
1159 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1160 lru_size = &mz->lru_zone_size[zid][lru];
1163 *lru_size += nr_pages;
1166 if (WARN_ONCE(size < 0,
1167 "%s(%p, %d, %d): lru_size %ld\n",
1168 __func__, lruvec, lru, nr_pages, size)) {
1174 *lru_size += nr_pages;
1177 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1179 struct mem_cgroup *task_memcg;
1180 struct task_struct *p;
1183 p = find_lock_task_mm(task);
1185 task_memcg = get_mem_cgroup_from_mm(p->mm);
1189 * All threads may have already detached their mm's, but the oom
1190 * killer still needs to detect if they have already been oom
1191 * killed to prevent needlessly killing additional tasks.
1194 task_memcg = mem_cgroup_from_task(task);
1195 css_get(&task_memcg->css);
1198 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1199 css_put(&task_memcg->css);
1204 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1205 * @memcg: the memory cgroup
1207 * Returns the maximum amount of memory @mem can be charged with, in
1210 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1212 unsigned long margin = 0;
1213 unsigned long count;
1214 unsigned long limit;
1216 count = page_counter_read(&memcg->memory);
1217 limit = READ_ONCE(memcg->memory.max);
1219 margin = limit - count;
1221 if (do_memsw_account()) {
1222 count = page_counter_read(&memcg->memsw);
1223 limit = READ_ONCE(memcg->memsw.max);
1225 margin = min(margin, limit - count);
1234 * A routine for checking "mem" is under move_account() or not.
1236 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1237 * moving cgroups. This is for waiting at high-memory pressure
1240 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1242 struct mem_cgroup *from;
1243 struct mem_cgroup *to;
1246 * Unlike task_move routines, we access mc.to, mc.from not under
1247 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1249 spin_lock(&mc.lock);
1255 ret = mem_cgroup_is_descendant(from, memcg) ||
1256 mem_cgroup_is_descendant(to, memcg);
1258 spin_unlock(&mc.lock);
1262 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1264 if (mc.moving_task && current != mc.moving_task) {
1265 if (mem_cgroup_under_move(memcg)) {
1267 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1268 /* moving charge context might have finished. */
1271 finish_wait(&mc.waitq, &wait);
1278 static const unsigned int memcg1_stats[] = {
1289 static const char *const memcg1_stat_names[] = {
1300 #define K(x) ((x) << (PAGE_SHIFT-10))
1302 * mem_cgroup_print_oom_context: Print OOM information relevant to
1303 * memory controller.
1304 * @memcg: The memory cgroup that went over limit
1305 * @p: Task that is going to be killed
1307 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1310 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1315 pr_cont(",oom_memcg=");
1316 pr_cont_cgroup_path(memcg->css.cgroup);
1318 pr_cont(",global_oom");
1320 pr_cont(",task_memcg=");
1321 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1327 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1328 * memory controller.
1329 * @memcg: The memory cgroup that went over limit
1331 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1333 struct mem_cgroup *iter;
1336 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1337 K((u64)page_counter_read(&memcg->memory)),
1338 K((u64)memcg->memory.max), memcg->memory.failcnt);
1339 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1340 K((u64)page_counter_read(&memcg->memsw)),
1341 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1342 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1343 K((u64)page_counter_read(&memcg->kmem)),
1344 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1346 for_each_mem_cgroup_tree(iter, memcg) {
1347 pr_info("Memory cgroup stats for ");
1348 pr_cont_cgroup_path(iter->css.cgroup);
1351 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1352 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1354 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1355 K(memcg_page_state(iter, memcg1_stats[i])));
1358 for (i = 0; i < NR_LRU_LISTS; i++)
1359 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1360 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1367 * Return the memory (and swap, if configured) limit for a memcg.
1369 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1373 max = memcg->memory.max;
1374 if (mem_cgroup_swappiness(memcg)) {
1375 unsigned long memsw_max;
1376 unsigned long swap_max;
1378 memsw_max = memcg->memsw.max;
1379 swap_max = memcg->swap.max;
1380 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1381 max = min(max + swap_max, memsw_max);
1386 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1389 struct oom_control oc = {
1393 .gfp_mask = gfp_mask,
1398 if (mutex_lock_killable(&oom_lock))
1401 * A few threads which were not waiting at mutex_lock_killable() can
1402 * fail to bail out. Therefore, check again after holding oom_lock.
1404 ret = should_force_charge() || out_of_memory(&oc);
1405 mutex_unlock(&oom_lock);
1409 #if MAX_NUMNODES > 1
1412 * test_mem_cgroup_node_reclaimable
1413 * @memcg: the target memcg
1414 * @nid: the node ID to be checked.
1415 * @noswap : specify true here if the user wants flle only information.
1417 * This function returns whether the specified memcg contains any
1418 * reclaimable pages on a node. Returns true if there are any reclaimable
1419 * pages in the node.
1421 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1422 int nid, bool noswap)
1424 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1426 if (noswap || !total_swap_pages)
1428 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1435 * Always updating the nodemask is not very good - even if we have an empty
1436 * list or the wrong list here, we can start from some node and traverse all
1437 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1440 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1444 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1445 * pagein/pageout changes since the last update.
1447 if (!atomic_read(&memcg->numainfo_events))
1449 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1452 /* make a nodemask where this memcg uses memory from */
1453 memcg->scan_nodes = node_states[N_MEMORY];
1455 for_each_node_mask(nid, node_states[N_MEMORY]) {
1457 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1458 node_clear(nid, memcg->scan_nodes);
1461 atomic_set(&memcg->numainfo_events, 0);
1462 atomic_set(&memcg->numainfo_updating, 0);
1466 * Selecting a node where we start reclaim from. Because what we need is just
1467 * reducing usage counter, start from anywhere is O,K. Considering
1468 * memory reclaim from current node, there are pros. and cons.
1470 * Freeing memory from current node means freeing memory from a node which
1471 * we'll use or we've used. So, it may make LRU bad. And if several threads
1472 * hit limits, it will see a contention on a node. But freeing from remote
1473 * node means more costs for memory reclaim because of memory latency.
1475 * Now, we use round-robin. Better algorithm is welcomed.
1477 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1481 mem_cgroup_may_update_nodemask(memcg);
1482 node = memcg->last_scanned_node;
1484 node = next_node_in(node, memcg->scan_nodes);
1486 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1487 * last time it really checked all the LRUs due to rate limiting.
1488 * Fallback to the current node in that case for simplicity.
1490 if (unlikely(node == MAX_NUMNODES))
1491 node = numa_node_id();
1493 memcg->last_scanned_node = node;
1497 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1503 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1506 unsigned long *total_scanned)
1508 struct mem_cgroup *victim = NULL;
1511 unsigned long excess;
1512 unsigned long nr_scanned;
1513 struct mem_cgroup_reclaim_cookie reclaim = {
1518 excess = soft_limit_excess(root_memcg);
1521 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1526 * If we have not been able to reclaim
1527 * anything, it might because there are
1528 * no reclaimable pages under this hierarchy
1533 * We want to do more targeted reclaim.
1534 * excess >> 2 is not to excessive so as to
1535 * reclaim too much, nor too less that we keep
1536 * coming back to reclaim from this cgroup
1538 if (total >= (excess >> 2) ||
1539 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1544 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1545 pgdat, &nr_scanned);
1546 *total_scanned += nr_scanned;
1547 if (!soft_limit_excess(root_memcg))
1550 mem_cgroup_iter_break(root_memcg, victim);
1554 #ifdef CONFIG_LOCKDEP
1555 static struct lockdep_map memcg_oom_lock_dep_map = {
1556 .name = "memcg_oom_lock",
1560 static DEFINE_SPINLOCK(memcg_oom_lock);
1563 * Check OOM-Killer is already running under our hierarchy.
1564 * If someone is running, return false.
1566 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1568 struct mem_cgroup *iter, *failed = NULL;
1570 spin_lock(&memcg_oom_lock);
1572 for_each_mem_cgroup_tree(iter, memcg) {
1573 if (iter->oom_lock) {
1575 * this subtree of our hierarchy is already locked
1576 * so we cannot give a lock.
1579 mem_cgroup_iter_break(memcg, iter);
1582 iter->oom_lock = true;
1587 * OK, we failed to lock the whole subtree so we have
1588 * to clean up what we set up to the failing subtree
1590 for_each_mem_cgroup_tree(iter, memcg) {
1591 if (iter == failed) {
1592 mem_cgroup_iter_break(memcg, iter);
1595 iter->oom_lock = false;
1598 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1600 spin_unlock(&memcg_oom_lock);
1605 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1607 struct mem_cgroup *iter;
1609 spin_lock(&memcg_oom_lock);
1610 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1611 for_each_mem_cgroup_tree(iter, memcg)
1612 iter->oom_lock = false;
1613 spin_unlock(&memcg_oom_lock);
1616 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1618 struct mem_cgroup *iter;
1620 spin_lock(&memcg_oom_lock);
1621 for_each_mem_cgroup_tree(iter, memcg)
1623 spin_unlock(&memcg_oom_lock);
1626 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1628 struct mem_cgroup *iter;
1631 * When a new child is created while the hierarchy is under oom,
1632 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1634 spin_lock(&memcg_oom_lock);
1635 for_each_mem_cgroup_tree(iter, memcg)
1636 if (iter->under_oom > 0)
1638 spin_unlock(&memcg_oom_lock);
1641 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1643 struct oom_wait_info {
1644 struct mem_cgroup *memcg;
1645 wait_queue_entry_t wait;
1648 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1649 unsigned mode, int sync, void *arg)
1651 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1652 struct mem_cgroup *oom_wait_memcg;
1653 struct oom_wait_info *oom_wait_info;
1655 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1656 oom_wait_memcg = oom_wait_info->memcg;
1658 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1659 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1661 return autoremove_wake_function(wait, mode, sync, arg);
1664 static void memcg_oom_recover(struct mem_cgroup *memcg)
1667 * For the following lockless ->under_oom test, the only required
1668 * guarantee is that it must see the state asserted by an OOM when
1669 * this function is called as a result of userland actions
1670 * triggered by the notification of the OOM. This is trivially
1671 * achieved by invoking mem_cgroup_mark_under_oom() before
1672 * triggering notification.
1674 if (memcg && memcg->under_oom)
1675 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1685 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1687 enum oom_status ret;
1690 if (order > PAGE_ALLOC_COSTLY_ORDER)
1693 memcg_memory_event(memcg, MEMCG_OOM);
1696 * We are in the middle of the charge context here, so we
1697 * don't want to block when potentially sitting on a callstack
1698 * that holds all kinds of filesystem and mm locks.
1700 * cgroup1 allows disabling the OOM killer and waiting for outside
1701 * handling until the charge can succeed; remember the context and put
1702 * the task to sleep at the end of the page fault when all locks are
1705 * On the other hand, in-kernel OOM killer allows for an async victim
1706 * memory reclaim (oom_reaper) and that means that we are not solely
1707 * relying on the oom victim to make a forward progress and we can
1708 * invoke the oom killer here.
1710 * Please note that mem_cgroup_out_of_memory might fail to find a
1711 * victim and then we have to bail out from the charge path.
1713 if (memcg->oom_kill_disable) {
1714 if (!current->in_user_fault)
1716 css_get(&memcg->css);
1717 current->memcg_in_oom = memcg;
1718 current->memcg_oom_gfp_mask = mask;
1719 current->memcg_oom_order = order;
1724 mem_cgroup_mark_under_oom(memcg);
1726 locked = mem_cgroup_oom_trylock(memcg);
1729 mem_cgroup_oom_notify(memcg);
1731 mem_cgroup_unmark_under_oom(memcg);
1732 if (mem_cgroup_out_of_memory(memcg, mask, order))
1738 mem_cgroup_oom_unlock(memcg);
1744 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1745 * @handle: actually kill/wait or just clean up the OOM state
1747 * This has to be called at the end of a page fault if the memcg OOM
1748 * handler was enabled.
1750 * Memcg supports userspace OOM handling where failed allocations must
1751 * sleep on a waitqueue until the userspace task resolves the
1752 * situation. Sleeping directly in the charge context with all kinds
1753 * of locks held is not a good idea, instead we remember an OOM state
1754 * in the task and mem_cgroup_oom_synchronize() has to be called at
1755 * the end of the page fault to complete the OOM handling.
1757 * Returns %true if an ongoing memcg OOM situation was detected and
1758 * completed, %false otherwise.
1760 bool mem_cgroup_oom_synchronize(bool handle)
1762 struct mem_cgroup *memcg = current->memcg_in_oom;
1763 struct oom_wait_info owait;
1766 /* OOM is global, do not handle */
1773 owait.memcg = memcg;
1774 owait.wait.flags = 0;
1775 owait.wait.func = memcg_oom_wake_function;
1776 owait.wait.private = current;
1777 INIT_LIST_HEAD(&owait.wait.entry);
1779 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1780 mem_cgroup_mark_under_oom(memcg);
1782 locked = mem_cgroup_oom_trylock(memcg);
1785 mem_cgroup_oom_notify(memcg);
1787 if (locked && !memcg->oom_kill_disable) {
1788 mem_cgroup_unmark_under_oom(memcg);
1789 finish_wait(&memcg_oom_waitq, &owait.wait);
1790 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1791 current->memcg_oom_order);
1794 mem_cgroup_unmark_under_oom(memcg);
1795 finish_wait(&memcg_oom_waitq, &owait.wait);
1799 mem_cgroup_oom_unlock(memcg);
1801 * There is no guarantee that an OOM-lock contender
1802 * sees the wakeups triggered by the OOM kill
1803 * uncharges. Wake any sleepers explicitely.
1805 memcg_oom_recover(memcg);
1808 current->memcg_in_oom = NULL;
1809 css_put(&memcg->css);
1814 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1815 * @victim: task to be killed by the OOM killer
1816 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1818 * Returns a pointer to a memory cgroup, which has to be cleaned up
1819 * by killing all belonging OOM-killable tasks.
1821 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1823 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1824 struct mem_cgroup *oom_domain)
1826 struct mem_cgroup *oom_group = NULL;
1827 struct mem_cgroup *memcg;
1829 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1833 oom_domain = root_mem_cgroup;
1837 memcg = mem_cgroup_from_task(victim);
1838 if (memcg == root_mem_cgroup)
1842 * Traverse the memory cgroup hierarchy from the victim task's
1843 * cgroup up to the OOMing cgroup (or root) to find the
1844 * highest-level memory cgroup with oom.group set.
1846 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1847 if (memcg->oom_group)
1850 if (memcg == oom_domain)
1855 css_get(&oom_group->css);
1862 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1864 pr_info("Tasks in ");
1865 pr_cont_cgroup_path(memcg->css.cgroup);
1866 pr_cont(" are going to be killed due to memory.oom.group set\n");
1870 * lock_page_memcg - lock a page->mem_cgroup binding
1873 * This function protects unlocked LRU pages from being moved to
1876 * It ensures lifetime of the returned memcg. Caller is responsible
1877 * for the lifetime of the page; __unlock_page_memcg() is available
1878 * when @page might get freed inside the locked section.
1880 struct mem_cgroup *lock_page_memcg(struct page *page)
1882 struct mem_cgroup *memcg;
1883 unsigned long flags;
1886 * The RCU lock is held throughout the transaction. The fast
1887 * path can get away without acquiring the memcg->move_lock
1888 * because page moving starts with an RCU grace period.
1890 * The RCU lock also protects the memcg from being freed when
1891 * the page state that is going to change is the only thing
1892 * preventing the page itself from being freed. E.g. writeback
1893 * doesn't hold a page reference and relies on PG_writeback to
1894 * keep off truncation, migration and so forth.
1898 if (mem_cgroup_disabled())
1901 memcg = page->mem_cgroup;
1902 if (unlikely(!memcg))
1905 if (atomic_read(&memcg->moving_account) <= 0)
1908 spin_lock_irqsave(&memcg->move_lock, flags);
1909 if (memcg != page->mem_cgroup) {
1910 spin_unlock_irqrestore(&memcg->move_lock, flags);
1915 * When charge migration first begins, we can have locked and
1916 * unlocked page stat updates happening concurrently. Track
1917 * the task who has the lock for unlock_page_memcg().
1919 memcg->move_lock_task = current;
1920 memcg->move_lock_flags = flags;
1924 EXPORT_SYMBOL(lock_page_memcg);
1927 * __unlock_page_memcg - unlock and unpin a memcg
1930 * Unlock and unpin a memcg returned by lock_page_memcg().
1932 void __unlock_page_memcg(struct mem_cgroup *memcg)
1934 if (memcg && memcg->move_lock_task == current) {
1935 unsigned long flags = memcg->move_lock_flags;
1937 memcg->move_lock_task = NULL;
1938 memcg->move_lock_flags = 0;
1940 spin_unlock_irqrestore(&memcg->move_lock, flags);
1947 * unlock_page_memcg - unlock a page->mem_cgroup binding
1950 void unlock_page_memcg(struct page *page)
1952 __unlock_page_memcg(page->mem_cgroup);
1954 EXPORT_SYMBOL(unlock_page_memcg);
1956 struct memcg_stock_pcp {
1957 struct mem_cgroup *cached; /* this never be root cgroup */
1958 unsigned int nr_pages;
1959 struct work_struct work;
1960 unsigned long flags;
1961 #define FLUSHING_CACHED_CHARGE 0
1963 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1964 static DEFINE_MUTEX(percpu_charge_mutex);
1967 * consume_stock: Try to consume stocked charge on this cpu.
1968 * @memcg: memcg to consume from.
1969 * @nr_pages: how many pages to charge.
1971 * The charges will only happen if @memcg matches the current cpu's memcg
1972 * stock, and at least @nr_pages are available in that stock. Failure to
1973 * service an allocation will refill the stock.
1975 * returns true if successful, false otherwise.
1977 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1979 struct memcg_stock_pcp *stock;
1980 unsigned long flags;
1983 if (nr_pages > MEMCG_CHARGE_BATCH)
1986 local_irq_save(flags);
1988 stock = this_cpu_ptr(&memcg_stock);
1989 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1990 stock->nr_pages -= nr_pages;
1994 local_irq_restore(flags);
2000 * Returns stocks cached in percpu and reset cached information.
2002 static void drain_stock(struct memcg_stock_pcp *stock)
2004 struct mem_cgroup *old = stock->cached;
2006 if (stock->nr_pages) {
2007 page_counter_uncharge(&old->memory, stock->nr_pages);
2008 if (do_memsw_account())
2009 page_counter_uncharge(&old->memsw, stock->nr_pages);
2010 css_put_many(&old->css, stock->nr_pages);
2011 stock->nr_pages = 0;
2013 stock->cached = NULL;
2016 static void drain_local_stock(struct work_struct *dummy)
2018 struct memcg_stock_pcp *stock;
2019 unsigned long flags;
2022 * The only protection from memory hotplug vs. drain_stock races is
2023 * that we always operate on local CPU stock here with IRQ disabled
2025 local_irq_save(flags);
2027 stock = this_cpu_ptr(&memcg_stock);
2029 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2031 local_irq_restore(flags);
2035 * Cache charges(val) to local per_cpu area.
2036 * This will be consumed by consume_stock() function, later.
2038 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2040 struct memcg_stock_pcp *stock;
2041 unsigned long flags;
2043 local_irq_save(flags);
2045 stock = this_cpu_ptr(&memcg_stock);
2046 if (stock->cached != memcg) { /* reset if necessary */
2048 stock->cached = memcg;
2050 stock->nr_pages += nr_pages;
2052 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2055 local_irq_restore(flags);
2059 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2060 * of the hierarchy under it.
2062 static void drain_all_stock(struct mem_cgroup *root_memcg)
2066 /* If someone's already draining, avoid adding running more workers. */
2067 if (!mutex_trylock(&percpu_charge_mutex))
2070 * Notify other cpus that system-wide "drain" is running
2071 * We do not care about races with the cpu hotplug because cpu down
2072 * as well as workers from this path always operate on the local
2073 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2076 for_each_online_cpu(cpu) {
2077 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2078 struct mem_cgroup *memcg;
2080 memcg = stock->cached;
2081 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2083 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2084 css_put(&memcg->css);
2087 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2089 drain_local_stock(&stock->work);
2091 schedule_work_on(cpu, &stock->work);
2093 css_put(&memcg->css);
2096 mutex_unlock(&percpu_charge_mutex);
2099 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2101 struct memcg_stock_pcp *stock;
2102 struct mem_cgroup *memcg;
2104 stock = &per_cpu(memcg_stock, cpu);
2107 for_each_mem_cgroup(memcg) {
2110 for (i = 0; i < MEMCG_NR_STAT; i++) {
2114 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
2116 atomic_long_add(x, &memcg->stat[i]);
2118 if (i >= NR_VM_NODE_STAT_ITEMS)
2121 for_each_node(nid) {
2122 struct mem_cgroup_per_node *pn;
2124 pn = mem_cgroup_nodeinfo(memcg, nid);
2125 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2127 atomic_long_add(x, &pn->lruvec_stat[i]);
2131 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2134 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
2136 atomic_long_add(x, &memcg->events[i]);
2143 static void reclaim_high(struct mem_cgroup *memcg,
2144 unsigned int nr_pages,
2148 if (page_counter_read(&memcg->memory) <= memcg->high)
2150 memcg_memory_event(memcg, MEMCG_HIGH);
2151 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2152 } while ((memcg = parent_mem_cgroup(memcg)));
2155 static void high_work_func(struct work_struct *work)
2157 struct mem_cgroup *memcg;
2159 memcg = container_of(work, struct mem_cgroup, high_work);
2160 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2164 * Scheduled by try_charge() to be executed from the userland return path
2165 * and reclaims memory over the high limit.
2167 void mem_cgroup_handle_over_high(void)
2169 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2170 struct mem_cgroup *memcg;
2172 if (likely(!nr_pages))
2175 memcg = get_mem_cgroup_from_mm(current->mm);
2176 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2177 css_put(&memcg->css);
2178 current->memcg_nr_pages_over_high = 0;
2181 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2182 unsigned int nr_pages)
2184 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2185 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2186 struct mem_cgroup *mem_over_limit;
2187 struct page_counter *counter;
2188 unsigned long nr_reclaimed;
2189 bool may_swap = true;
2190 bool drained = false;
2192 enum oom_status oom_status;
2194 if (mem_cgroup_is_root(memcg))
2197 if (consume_stock(memcg, nr_pages))
2200 if (!do_memsw_account() ||
2201 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2202 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2204 if (do_memsw_account())
2205 page_counter_uncharge(&memcg->memsw, batch);
2206 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2208 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2212 if (batch > nr_pages) {
2218 * Unlike in global OOM situations, memcg is not in a physical
2219 * memory shortage. Allow dying and OOM-killed tasks to
2220 * bypass the last charges so that they can exit quickly and
2221 * free their memory.
2223 if (unlikely(should_force_charge()))
2227 * Prevent unbounded recursion when reclaim operations need to
2228 * allocate memory. This might exceed the limits temporarily,
2229 * but we prefer facilitating memory reclaim and getting back
2230 * under the limit over triggering OOM kills in these cases.
2232 if (unlikely(current->flags & PF_MEMALLOC))
2235 if (unlikely(task_in_memcg_oom(current)))
2238 if (!gfpflags_allow_blocking(gfp_mask))
2241 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2243 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2244 gfp_mask, may_swap);
2246 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2250 drain_all_stock(mem_over_limit);
2255 if (gfp_mask & __GFP_NORETRY)
2258 * Even though the limit is exceeded at this point, reclaim
2259 * may have been able to free some pages. Retry the charge
2260 * before killing the task.
2262 * Only for regular pages, though: huge pages are rather
2263 * unlikely to succeed so close to the limit, and we fall back
2264 * to regular pages anyway in case of failure.
2266 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2269 * At task move, charge accounts can be doubly counted. So, it's
2270 * better to wait until the end of task_move if something is going on.
2272 if (mem_cgroup_wait_acct_move(mem_over_limit))
2278 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2281 if (gfp_mask & __GFP_NOFAIL)
2284 if (fatal_signal_pending(current))
2288 * keep retrying as long as the memcg oom killer is able to make
2289 * a forward progress or bypass the charge if the oom killer
2290 * couldn't make any progress.
2292 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2293 get_order(nr_pages * PAGE_SIZE));
2294 switch (oom_status) {
2296 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2305 if (!(gfp_mask & __GFP_NOFAIL))
2309 * The allocation either can't fail or will lead to more memory
2310 * being freed very soon. Allow memory usage go over the limit
2311 * temporarily by force charging it.
2313 page_counter_charge(&memcg->memory, nr_pages);
2314 if (do_memsw_account())
2315 page_counter_charge(&memcg->memsw, nr_pages);
2316 css_get_many(&memcg->css, nr_pages);
2321 css_get_many(&memcg->css, batch);
2322 if (batch > nr_pages)
2323 refill_stock(memcg, batch - nr_pages);
2326 * If the hierarchy is above the normal consumption range, schedule
2327 * reclaim on returning to userland. We can perform reclaim here
2328 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2329 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2330 * not recorded as it most likely matches current's and won't
2331 * change in the meantime. As high limit is checked again before
2332 * reclaim, the cost of mismatch is negligible.
2335 if (page_counter_read(&memcg->memory) > memcg->high) {
2336 /* Don't bother a random interrupted task */
2337 if (in_interrupt()) {
2338 schedule_work(&memcg->high_work);
2341 current->memcg_nr_pages_over_high += batch;
2342 set_notify_resume(current);
2345 } while ((memcg = parent_mem_cgroup(memcg)));
2350 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2352 if (mem_cgroup_is_root(memcg))
2355 page_counter_uncharge(&memcg->memory, nr_pages);
2356 if (do_memsw_account())
2357 page_counter_uncharge(&memcg->memsw, nr_pages);
2359 css_put_many(&memcg->css, nr_pages);
2362 static void lock_page_lru(struct page *page, int *isolated)
2364 struct zone *zone = page_zone(page);
2366 spin_lock_irq(zone_lru_lock(zone));
2367 if (PageLRU(page)) {
2368 struct lruvec *lruvec;
2370 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2372 del_page_from_lru_list(page, lruvec, page_lru(page));
2378 static void unlock_page_lru(struct page *page, int isolated)
2380 struct zone *zone = page_zone(page);
2383 struct lruvec *lruvec;
2385 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2386 VM_BUG_ON_PAGE(PageLRU(page), page);
2388 add_page_to_lru_list(page, lruvec, page_lru(page));
2390 spin_unlock_irq(zone_lru_lock(zone));
2393 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2398 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2401 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2402 * may already be on some other mem_cgroup's LRU. Take care of it.
2405 lock_page_lru(page, &isolated);
2408 * Nobody should be changing or seriously looking at
2409 * page->mem_cgroup at this point:
2411 * - the page is uncharged
2413 * - the page is off-LRU
2415 * - an anonymous fault has exclusive page access, except for
2416 * a locked page table
2418 * - a page cache insertion, a swapin fault, or a migration
2419 * have the page locked
2421 page->mem_cgroup = memcg;
2424 unlock_page_lru(page, isolated);
2427 #ifdef CONFIG_MEMCG_KMEM
2428 static int memcg_alloc_cache_id(void)
2433 id = ida_simple_get(&memcg_cache_ida,
2434 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2438 if (id < memcg_nr_cache_ids)
2442 * There's no space for the new id in memcg_caches arrays,
2443 * so we have to grow them.
2445 down_write(&memcg_cache_ids_sem);
2447 size = 2 * (id + 1);
2448 if (size < MEMCG_CACHES_MIN_SIZE)
2449 size = MEMCG_CACHES_MIN_SIZE;
2450 else if (size > MEMCG_CACHES_MAX_SIZE)
2451 size = MEMCG_CACHES_MAX_SIZE;
2453 err = memcg_update_all_caches(size);
2455 err = memcg_update_all_list_lrus(size);
2457 memcg_nr_cache_ids = size;
2459 up_write(&memcg_cache_ids_sem);
2462 ida_simple_remove(&memcg_cache_ida, id);
2468 static void memcg_free_cache_id(int id)
2470 ida_simple_remove(&memcg_cache_ida, id);
2473 struct memcg_kmem_cache_create_work {
2474 struct mem_cgroup *memcg;
2475 struct kmem_cache *cachep;
2476 struct work_struct work;
2479 static void memcg_kmem_cache_create_func(struct work_struct *w)
2481 struct memcg_kmem_cache_create_work *cw =
2482 container_of(w, struct memcg_kmem_cache_create_work, work);
2483 struct mem_cgroup *memcg = cw->memcg;
2484 struct kmem_cache *cachep = cw->cachep;
2486 memcg_create_kmem_cache(memcg, cachep);
2488 css_put(&memcg->css);
2493 * Enqueue the creation of a per-memcg kmem_cache.
2495 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2496 struct kmem_cache *cachep)
2498 struct memcg_kmem_cache_create_work *cw;
2500 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2504 css_get(&memcg->css);
2507 cw->cachep = cachep;
2508 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2510 queue_work(memcg_kmem_cache_wq, &cw->work);
2513 static inline bool memcg_kmem_bypass(void)
2515 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2521 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2522 * @cachep: the original global kmem cache
2524 * Return the kmem_cache we're supposed to use for a slab allocation.
2525 * We try to use the current memcg's version of the cache.
2527 * If the cache does not exist yet, if we are the first user of it, we
2528 * create it asynchronously in a workqueue and let the current allocation
2529 * go through with the original cache.
2531 * This function takes a reference to the cache it returns to assure it
2532 * won't get destroyed while we are working with it. Once the caller is
2533 * done with it, memcg_kmem_put_cache() must be called to release the
2536 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2538 struct mem_cgroup *memcg;
2539 struct kmem_cache *memcg_cachep;
2542 VM_BUG_ON(!is_root_cache(cachep));
2544 if (memcg_kmem_bypass())
2547 memcg = get_mem_cgroup_from_current();
2548 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2552 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2553 if (likely(memcg_cachep))
2554 return memcg_cachep;
2557 * If we are in a safe context (can wait, and not in interrupt
2558 * context), we could be be predictable and return right away.
2559 * This would guarantee that the allocation being performed
2560 * already belongs in the new cache.
2562 * However, there are some clashes that can arrive from locking.
2563 * For instance, because we acquire the slab_mutex while doing
2564 * memcg_create_kmem_cache, this means no further allocation
2565 * could happen with the slab_mutex held. So it's better to
2568 memcg_schedule_kmem_cache_create(memcg, cachep);
2570 css_put(&memcg->css);
2575 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2576 * @cachep: the cache returned by memcg_kmem_get_cache
2578 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2580 if (!is_root_cache(cachep))
2581 css_put(&cachep->memcg_params.memcg->css);
2585 * __memcg_kmem_charge_memcg: charge a kmem page
2586 * @page: page to charge
2587 * @gfp: reclaim mode
2588 * @order: allocation order
2589 * @memcg: memory cgroup to charge
2591 * Returns 0 on success, an error code on failure.
2593 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2594 struct mem_cgroup *memcg)
2596 unsigned int nr_pages = 1 << order;
2597 struct page_counter *counter;
2600 ret = try_charge(memcg, gfp, nr_pages);
2604 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2605 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2606 cancel_charge(memcg, nr_pages);
2610 page->mem_cgroup = memcg;
2616 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2617 * @page: page to charge
2618 * @gfp: reclaim mode
2619 * @order: allocation order
2621 * Returns 0 on success, an error code on failure.
2623 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2625 struct mem_cgroup *memcg;
2628 if (memcg_kmem_bypass())
2631 memcg = get_mem_cgroup_from_current();
2632 if (!mem_cgroup_is_root(memcg)) {
2633 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2635 __SetPageKmemcg(page);
2637 css_put(&memcg->css);
2641 * __memcg_kmem_uncharge: uncharge a kmem page
2642 * @page: page to uncharge
2643 * @order: allocation order
2645 void __memcg_kmem_uncharge(struct page *page, int order)
2647 struct mem_cgroup *memcg = page->mem_cgroup;
2648 unsigned int nr_pages = 1 << order;
2653 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2655 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2656 page_counter_uncharge(&memcg->kmem, nr_pages);
2658 page_counter_uncharge(&memcg->memory, nr_pages);
2659 if (do_memsw_account())
2660 page_counter_uncharge(&memcg->memsw, nr_pages);
2662 page->mem_cgroup = NULL;
2664 /* slab pages do not have PageKmemcg flag set */
2665 if (PageKmemcg(page))
2666 __ClearPageKmemcg(page);
2668 css_put_many(&memcg->css, nr_pages);
2670 #endif /* CONFIG_MEMCG_KMEM */
2672 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2675 * Because tail pages are not marked as "used", set it. We're under
2676 * zone_lru_lock and migration entries setup in all page mappings.
2678 void mem_cgroup_split_huge_fixup(struct page *head)
2682 if (mem_cgroup_disabled())
2685 for (i = 1; i < HPAGE_PMD_NR; i++)
2686 head[i].mem_cgroup = head->mem_cgroup;
2688 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2690 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2692 #ifdef CONFIG_MEMCG_SWAP
2694 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2695 * @entry: swap entry to be moved
2696 * @from: mem_cgroup which the entry is moved from
2697 * @to: mem_cgroup which the entry is moved to
2699 * It succeeds only when the swap_cgroup's record for this entry is the same
2700 * as the mem_cgroup's id of @from.
2702 * Returns 0 on success, -EINVAL on failure.
2704 * The caller must have charged to @to, IOW, called page_counter_charge() about
2705 * both res and memsw, and called css_get().
2707 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2708 struct mem_cgroup *from, struct mem_cgroup *to)
2710 unsigned short old_id, new_id;
2712 old_id = mem_cgroup_id(from);
2713 new_id = mem_cgroup_id(to);
2715 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2716 mod_memcg_state(from, MEMCG_SWAP, -1);
2717 mod_memcg_state(to, MEMCG_SWAP, 1);
2723 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2724 struct mem_cgroup *from, struct mem_cgroup *to)
2730 static DEFINE_MUTEX(memcg_max_mutex);
2732 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2733 unsigned long max, bool memsw)
2735 bool enlarge = false;
2736 bool drained = false;
2738 bool limits_invariant;
2739 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2742 if (signal_pending(current)) {
2747 mutex_lock(&memcg_max_mutex);
2749 * Make sure that the new limit (memsw or memory limit) doesn't
2750 * break our basic invariant rule memory.max <= memsw.max.
2752 limits_invariant = memsw ? max >= memcg->memory.max :
2753 max <= memcg->memsw.max;
2754 if (!limits_invariant) {
2755 mutex_unlock(&memcg_max_mutex);
2759 if (max > counter->max)
2761 ret = page_counter_set_max(counter, max);
2762 mutex_unlock(&memcg_max_mutex);
2768 drain_all_stock(memcg);
2773 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2774 GFP_KERNEL, !memsw)) {
2780 if (!ret && enlarge)
2781 memcg_oom_recover(memcg);
2786 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2788 unsigned long *total_scanned)
2790 unsigned long nr_reclaimed = 0;
2791 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2792 unsigned long reclaimed;
2794 struct mem_cgroup_tree_per_node *mctz;
2795 unsigned long excess;
2796 unsigned long nr_scanned;
2801 mctz = soft_limit_tree_node(pgdat->node_id);
2804 * Do not even bother to check the largest node if the root
2805 * is empty. Do it lockless to prevent lock bouncing. Races
2806 * are acceptable as soft limit is best effort anyway.
2808 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2812 * This loop can run a while, specially if mem_cgroup's continuously
2813 * keep exceeding their soft limit and putting the system under
2820 mz = mem_cgroup_largest_soft_limit_node(mctz);
2825 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2826 gfp_mask, &nr_scanned);
2827 nr_reclaimed += reclaimed;
2828 *total_scanned += nr_scanned;
2829 spin_lock_irq(&mctz->lock);
2830 __mem_cgroup_remove_exceeded(mz, mctz);
2833 * If we failed to reclaim anything from this memory cgroup
2834 * it is time to move on to the next cgroup
2838 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2840 excess = soft_limit_excess(mz->memcg);
2842 * One school of thought says that we should not add
2843 * back the node to the tree if reclaim returns 0.
2844 * But our reclaim could return 0, simply because due
2845 * to priority we are exposing a smaller subset of
2846 * memory to reclaim from. Consider this as a longer
2849 /* If excess == 0, no tree ops */
2850 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2851 spin_unlock_irq(&mctz->lock);
2852 css_put(&mz->memcg->css);
2855 * Could not reclaim anything and there are no more
2856 * mem cgroups to try or we seem to be looping without
2857 * reclaiming anything.
2859 if (!nr_reclaimed &&
2861 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2863 } while (!nr_reclaimed);
2865 css_put(&next_mz->memcg->css);
2866 return nr_reclaimed;
2870 * Test whether @memcg has children, dead or alive. Note that this
2871 * function doesn't care whether @memcg has use_hierarchy enabled and
2872 * returns %true if there are child csses according to the cgroup
2873 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2875 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2880 ret = css_next_child(NULL, &memcg->css);
2886 * Reclaims as many pages from the given memcg as possible.
2888 * Caller is responsible for holding css reference for memcg.
2890 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2892 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2894 /* we call try-to-free pages for make this cgroup empty */
2895 lru_add_drain_all();
2897 drain_all_stock(memcg);
2899 /* try to free all pages in this cgroup */
2900 while (nr_retries && page_counter_read(&memcg->memory)) {
2903 if (signal_pending(current))
2906 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2910 /* maybe some writeback is necessary */
2911 congestion_wait(BLK_RW_ASYNC, HZ/10);
2919 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2920 char *buf, size_t nbytes,
2923 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2925 if (mem_cgroup_is_root(memcg))
2927 return mem_cgroup_force_empty(memcg) ?: nbytes;
2930 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2933 return mem_cgroup_from_css(css)->use_hierarchy;
2936 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2937 struct cftype *cft, u64 val)
2940 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2941 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2943 if (memcg->use_hierarchy == val)
2947 * If parent's use_hierarchy is set, we can't make any modifications
2948 * in the child subtrees. If it is unset, then the change can
2949 * occur, provided the current cgroup has no children.
2951 * For the root cgroup, parent_mem is NULL, we allow value to be
2952 * set if there are no children.
2954 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2955 (val == 1 || val == 0)) {
2956 if (!memcg_has_children(memcg))
2957 memcg->use_hierarchy = val;
2966 struct accumulated_stats {
2967 unsigned long stat[MEMCG_NR_STAT];
2968 unsigned long events[NR_VM_EVENT_ITEMS];
2969 unsigned long lru_pages[NR_LRU_LISTS];
2970 const unsigned int *stats_array;
2971 const unsigned int *events_array;
2976 static void accumulate_memcg_tree(struct mem_cgroup *memcg,
2977 struct accumulated_stats *acc)
2979 struct mem_cgroup *mi;
2982 for_each_mem_cgroup_tree(mi, memcg) {
2983 for (i = 0; i < acc->stats_size; i++)
2984 acc->stat[i] += memcg_page_state(mi,
2985 acc->stats_array ? acc->stats_array[i] : i);
2987 for (i = 0; i < acc->events_size; i++)
2988 acc->events[i] += memcg_sum_events(mi,
2989 acc->events_array ? acc->events_array[i] : i);
2991 for (i = 0; i < NR_LRU_LISTS; i++)
2992 acc->lru_pages[i] +=
2993 mem_cgroup_nr_lru_pages(mi, BIT(i));
2997 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2999 unsigned long val = 0;
3001 if (mem_cgroup_is_root(memcg)) {
3002 struct mem_cgroup *iter;
3004 for_each_mem_cgroup_tree(iter, memcg) {
3005 val += memcg_page_state(iter, MEMCG_CACHE);
3006 val += memcg_page_state(iter, MEMCG_RSS);
3008 val += memcg_page_state(iter, MEMCG_SWAP);
3012 val = page_counter_read(&memcg->memory);
3014 val = page_counter_read(&memcg->memsw);
3027 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3030 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3031 struct page_counter *counter;
3033 switch (MEMFILE_TYPE(cft->private)) {
3035 counter = &memcg->memory;
3038 counter = &memcg->memsw;
3041 counter = &memcg->kmem;
3044 counter = &memcg->tcpmem;
3050 switch (MEMFILE_ATTR(cft->private)) {
3052 if (counter == &memcg->memory)
3053 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3054 if (counter == &memcg->memsw)
3055 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3056 return (u64)page_counter_read(counter) * PAGE_SIZE;
3058 return (u64)counter->max * PAGE_SIZE;
3060 return (u64)counter->watermark * PAGE_SIZE;
3062 return counter->failcnt;
3063 case RES_SOFT_LIMIT:
3064 return (u64)memcg->soft_limit * PAGE_SIZE;
3070 #ifdef CONFIG_MEMCG_KMEM
3071 static int memcg_online_kmem(struct mem_cgroup *memcg)
3075 if (cgroup_memory_nokmem)
3078 BUG_ON(memcg->kmemcg_id >= 0);
3079 BUG_ON(memcg->kmem_state);
3081 memcg_id = memcg_alloc_cache_id();
3085 static_branch_inc(&memcg_kmem_enabled_key);
3087 * A memory cgroup is considered kmem-online as soon as it gets
3088 * kmemcg_id. Setting the id after enabling static branching will
3089 * guarantee no one starts accounting before all call sites are
3092 memcg->kmemcg_id = memcg_id;
3093 memcg->kmem_state = KMEM_ONLINE;
3094 INIT_LIST_HEAD(&memcg->kmem_caches);
3099 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3101 struct cgroup_subsys_state *css;
3102 struct mem_cgroup *parent, *child;
3105 if (memcg->kmem_state != KMEM_ONLINE)
3108 * Clear the online state before clearing memcg_caches array
3109 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3110 * guarantees that no cache will be created for this cgroup
3111 * after we are done (see memcg_create_kmem_cache()).
3113 memcg->kmem_state = KMEM_ALLOCATED;
3115 memcg_deactivate_kmem_caches(memcg);
3117 kmemcg_id = memcg->kmemcg_id;
3118 BUG_ON(kmemcg_id < 0);
3120 parent = parent_mem_cgroup(memcg);
3122 parent = root_mem_cgroup;
3125 * Change kmemcg_id of this cgroup and all its descendants to the
3126 * parent's id, and then move all entries from this cgroup's list_lrus
3127 * to ones of the parent. After we have finished, all list_lrus
3128 * corresponding to this cgroup are guaranteed to remain empty. The
3129 * ordering is imposed by list_lru_node->lock taken by
3130 * memcg_drain_all_list_lrus().
3132 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3133 css_for_each_descendant_pre(css, &memcg->css) {
3134 child = mem_cgroup_from_css(css);
3135 BUG_ON(child->kmemcg_id != kmemcg_id);
3136 child->kmemcg_id = parent->kmemcg_id;
3137 if (!memcg->use_hierarchy)
3142 memcg_drain_all_list_lrus(kmemcg_id, parent);
3144 memcg_free_cache_id(kmemcg_id);
3147 static void memcg_free_kmem(struct mem_cgroup *memcg)
3149 /* css_alloc() failed, offlining didn't happen */
3150 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3151 memcg_offline_kmem(memcg);
3153 if (memcg->kmem_state == KMEM_ALLOCATED) {
3154 memcg_destroy_kmem_caches(memcg);
3155 static_branch_dec(&memcg_kmem_enabled_key);
3156 WARN_ON(page_counter_read(&memcg->kmem));
3160 static int memcg_online_kmem(struct mem_cgroup *memcg)
3164 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3167 static void memcg_free_kmem(struct mem_cgroup *memcg)
3170 #endif /* CONFIG_MEMCG_KMEM */
3172 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3177 mutex_lock(&memcg_max_mutex);
3178 ret = page_counter_set_max(&memcg->kmem, max);
3179 mutex_unlock(&memcg_max_mutex);
3183 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3187 mutex_lock(&memcg_max_mutex);
3189 ret = page_counter_set_max(&memcg->tcpmem, max);
3193 if (!memcg->tcpmem_active) {
3195 * The active flag needs to be written after the static_key
3196 * update. This is what guarantees that the socket activation
3197 * function is the last one to run. See mem_cgroup_sk_alloc()
3198 * for details, and note that we don't mark any socket as
3199 * belonging to this memcg until that flag is up.
3201 * We need to do this, because static_keys will span multiple
3202 * sites, but we can't control their order. If we mark a socket
3203 * as accounted, but the accounting functions are not patched in
3204 * yet, we'll lose accounting.
3206 * We never race with the readers in mem_cgroup_sk_alloc(),
3207 * because when this value change, the code to process it is not
3210 static_branch_inc(&memcg_sockets_enabled_key);
3211 memcg->tcpmem_active = true;
3214 mutex_unlock(&memcg_max_mutex);
3219 * The user of this function is...
3222 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3223 char *buf, size_t nbytes, loff_t off)
3225 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3226 unsigned long nr_pages;
3229 buf = strstrip(buf);
3230 ret = page_counter_memparse(buf, "-1", &nr_pages);
3234 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3236 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3240 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3242 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3245 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3248 ret = memcg_update_kmem_max(memcg, nr_pages);
3251 ret = memcg_update_tcp_max(memcg, nr_pages);
3255 case RES_SOFT_LIMIT:
3256 memcg->soft_limit = nr_pages;
3260 return ret ?: nbytes;
3263 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3264 size_t nbytes, loff_t off)
3266 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3267 struct page_counter *counter;
3269 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3271 counter = &memcg->memory;
3274 counter = &memcg->memsw;
3277 counter = &memcg->kmem;
3280 counter = &memcg->tcpmem;
3286 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3288 page_counter_reset_watermark(counter);
3291 counter->failcnt = 0;
3300 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3303 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3307 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3308 struct cftype *cft, u64 val)
3310 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3312 if (val & ~MOVE_MASK)
3316 * No kind of locking is needed in here, because ->can_attach() will
3317 * check this value once in the beginning of the process, and then carry
3318 * on with stale data. This means that changes to this value will only
3319 * affect task migrations starting after the change.
3321 memcg->move_charge_at_immigrate = val;
3325 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3326 struct cftype *cft, u64 val)
3333 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3337 unsigned int lru_mask;
3340 static const struct numa_stat stats[] = {
3341 { "total", LRU_ALL },
3342 { "file", LRU_ALL_FILE },
3343 { "anon", LRU_ALL_ANON },
3344 { "unevictable", BIT(LRU_UNEVICTABLE) },
3346 const struct numa_stat *stat;
3349 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3351 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3352 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3353 seq_printf(m, "%s=%lu", stat->name, nr);
3354 for_each_node_state(nid, N_MEMORY) {
3355 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3357 seq_printf(m, " N%d=%lu", nid, nr);
3362 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3363 struct mem_cgroup *iter;
3366 for_each_mem_cgroup_tree(iter, memcg)
3367 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3368 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3369 for_each_node_state(nid, N_MEMORY) {
3371 for_each_mem_cgroup_tree(iter, memcg)
3372 nr += mem_cgroup_node_nr_lru_pages(
3373 iter, nid, stat->lru_mask);
3374 seq_printf(m, " N%d=%lu", nid, nr);
3381 #endif /* CONFIG_NUMA */
3383 /* Universal VM events cgroup1 shows, original sort order */
3384 static const unsigned int memcg1_events[] = {
3391 static const char *const memcg1_event_names[] = {
3398 static int memcg_stat_show(struct seq_file *m, void *v)
3400 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3401 unsigned long memory, memsw;
3402 struct mem_cgroup *mi;
3404 struct accumulated_stats acc;
3406 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3407 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3409 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3410 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3412 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3413 memcg_page_state(memcg, memcg1_stats[i]) *
3417 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3418 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3419 memcg_sum_events(memcg, memcg1_events[i]));
3421 for (i = 0; i < NR_LRU_LISTS; i++)
3422 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3423 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3425 /* Hierarchical information */
3426 memory = memsw = PAGE_COUNTER_MAX;
3427 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3428 memory = min(memory, mi->memory.max);
3429 memsw = min(memsw, mi->memsw.max);
3431 seq_printf(m, "hierarchical_memory_limit %llu\n",
3432 (u64)memory * PAGE_SIZE);
3433 if (do_memsw_account())
3434 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3435 (u64)memsw * PAGE_SIZE);
3437 memset(&acc, 0, sizeof(acc));
3438 acc.stats_size = ARRAY_SIZE(memcg1_stats);
3439 acc.stats_array = memcg1_stats;
3440 acc.events_size = ARRAY_SIZE(memcg1_events);
3441 acc.events_array = memcg1_events;
3442 accumulate_memcg_tree(memcg, &acc);
3444 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3445 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3447 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3448 (u64)acc.stat[i] * PAGE_SIZE);
3451 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3452 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3453 (u64)acc.events[i]);
3455 for (i = 0; i < NR_LRU_LISTS; i++)
3456 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3457 (u64)acc.lru_pages[i] * PAGE_SIZE);
3459 #ifdef CONFIG_DEBUG_VM
3462 struct mem_cgroup_per_node *mz;
3463 struct zone_reclaim_stat *rstat;
3464 unsigned long recent_rotated[2] = {0, 0};
3465 unsigned long recent_scanned[2] = {0, 0};
3467 for_each_online_pgdat(pgdat) {
3468 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3469 rstat = &mz->lruvec.reclaim_stat;
3471 recent_rotated[0] += rstat->recent_rotated[0];
3472 recent_rotated[1] += rstat->recent_rotated[1];
3473 recent_scanned[0] += rstat->recent_scanned[0];
3474 recent_scanned[1] += rstat->recent_scanned[1];
3476 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3477 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3478 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3479 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3486 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3489 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3491 return mem_cgroup_swappiness(memcg);
3494 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3495 struct cftype *cft, u64 val)
3497 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3503 memcg->swappiness = val;
3505 vm_swappiness = val;
3510 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3512 struct mem_cgroup_threshold_ary *t;
3513 unsigned long usage;
3518 t = rcu_dereference(memcg->thresholds.primary);
3520 t = rcu_dereference(memcg->memsw_thresholds.primary);
3525 usage = mem_cgroup_usage(memcg, swap);
3528 * current_threshold points to threshold just below or equal to usage.
3529 * If it's not true, a threshold was crossed after last
3530 * call of __mem_cgroup_threshold().
3532 i = t->current_threshold;
3535 * Iterate backward over array of thresholds starting from
3536 * current_threshold and check if a threshold is crossed.
3537 * If none of thresholds below usage is crossed, we read
3538 * only one element of the array here.
3540 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3541 eventfd_signal(t->entries[i].eventfd, 1);
3543 /* i = current_threshold + 1 */
3547 * Iterate forward over array of thresholds starting from
3548 * current_threshold+1 and check if a threshold is crossed.
3549 * If none of thresholds above usage is crossed, we read
3550 * only one element of the array here.
3552 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3553 eventfd_signal(t->entries[i].eventfd, 1);
3555 /* Update current_threshold */
3556 t->current_threshold = i - 1;
3561 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3564 __mem_cgroup_threshold(memcg, false);
3565 if (do_memsw_account())
3566 __mem_cgroup_threshold(memcg, true);
3568 memcg = parent_mem_cgroup(memcg);
3572 static int compare_thresholds(const void *a, const void *b)
3574 const struct mem_cgroup_threshold *_a = a;
3575 const struct mem_cgroup_threshold *_b = b;
3577 if (_a->threshold > _b->threshold)
3580 if (_a->threshold < _b->threshold)
3586 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3588 struct mem_cgroup_eventfd_list *ev;
3590 spin_lock(&memcg_oom_lock);
3592 list_for_each_entry(ev, &memcg->oom_notify, list)
3593 eventfd_signal(ev->eventfd, 1);
3595 spin_unlock(&memcg_oom_lock);
3599 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3601 struct mem_cgroup *iter;
3603 for_each_mem_cgroup_tree(iter, memcg)
3604 mem_cgroup_oom_notify_cb(iter);
3607 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3608 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3610 struct mem_cgroup_thresholds *thresholds;
3611 struct mem_cgroup_threshold_ary *new;
3612 unsigned long threshold;
3613 unsigned long usage;
3616 ret = page_counter_memparse(args, "-1", &threshold);
3620 mutex_lock(&memcg->thresholds_lock);
3623 thresholds = &memcg->thresholds;
3624 usage = mem_cgroup_usage(memcg, false);
3625 } else if (type == _MEMSWAP) {
3626 thresholds = &memcg->memsw_thresholds;
3627 usage = mem_cgroup_usage(memcg, true);
3631 /* Check if a threshold crossed before adding a new one */
3632 if (thresholds->primary)
3633 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3635 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3637 /* Allocate memory for new array of thresholds */
3638 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3645 /* Copy thresholds (if any) to new array */
3646 if (thresholds->primary) {
3647 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3648 sizeof(struct mem_cgroup_threshold));
3651 /* Add new threshold */
3652 new->entries[size - 1].eventfd = eventfd;
3653 new->entries[size - 1].threshold = threshold;
3655 /* Sort thresholds. Registering of new threshold isn't time-critical */
3656 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3657 compare_thresholds, NULL);
3659 /* Find current threshold */
3660 new->current_threshold = -1;
3661 for (i = 0; i < size; i++) {
3662 if (new->entries[i].threshold <= usage) {
3664 * new->current_threshold will not be used until
3665 * rcu_assign_pointer(), so it's safe to increment
3668 ++new->current_threshold;
3673 /* Free old spare buffer and save old primary buffer as spare */
3674 kfree(thresholds->spare);
3675 thresholds->spare = thresholds->primary;
3677 rcu_assign_pointer(thresholds->primary, new);
3679 /* To be sure that nobody uses thresholds */
3683 mutex_unlock(&memcg->thresholds_lock);
3688 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3689 struct eventfd_ctx *eventfd, const char *args)
3691 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3694 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3695 struct eventfd_ctx *eventfd, const char *args)
3697 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3700 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3701 struct eventfd_ctx *eventfd, enum res_type type)
3703 struct mem_cgroup_thresholds *thresholds;
3704 struct mem_cgroup_threshold_ary *new;
3705 unsigned long usage;
3708 mutex_lock(&memcg->thresholds_lock);
3711 thresholds = &memcg->thresholds;
3712 usage = mem_cgroup_usage(memcg, false);
3713 } else if (type == _MEMSWAP) {
3714 thresholds = &memcg->memsw_thresholds;
3715 usage = mem_cgroup_usage(memcg, true);
3719 if (!thresholds->primary)
3722 /* Check if a threshold crossed before removing */
3723 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3725 /* Calculate new number of threshold */
3727 for (i = 0; i < thresholds->primary->size; i++) {
3728 if (thresholds->primary->entries[i].eventfd != eventfd)
3732 new = thresholds->spare;
3734 /* Set thresholds array to NULL if we don't have thresholds */
3743 /* Copy thresholds and find current threshold */
3744 new->current_threshold = -1;
3745 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3746 if (thresholds->primary->entries[i].eventfd == eventfd)
3749 new->entries[j] = thresholds->primary->entries[i];
3750 if (new->entries[j].threshold <= usage) {
3752 * new->current_threshold will not be used
3753 * until rcu_assign_pointer(), so it's safe to increment
3756 ++new->current_threshold;
3762 /* Swap primary and spare array */
3763 thresholds->spare = thresholds->primary;
3765 rcu_assign_pointer(thresholds->primary, new);
3767 /* To be sure that nobody uses thresholds */
3770 /* If all events are unregistered, free the spare array */
3772 kfree(thresholds->spare);
3773 thresholds->spare = NULL;
3776 mutex_unlock(&memcg->thresholds_lock);
3779 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3780 struct eventfd_ctx *eventfd)
3782 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3785 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3786 struct eventfd_ctx *eventfd)
3788 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3791 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3792 struct eventfd_ctx *eventfd, const char *args)
3794 struct mem_cgroup_eventfd_list *event;
3796 event = kmalloc(sizeof(*event), GFP_KERNEL);
3800 spin_lock(&memcg_oom_lock);
3802 event->eventfd = eventfd;
3803 list_add(&event->list, &memcg->oom_notify);
3805 /* already in OOM ? */
3806 if (memcg->under_oom)
3807 eventfd_signal(eventfd, 1);
3808 spin_unlock(&memcg_oom_lock);
3813 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3814 struct eventfd_ctx *eventfd)
3816 struct mem_cgroup_eventfd_list *ev, *tmp;
3818 spin_lock(&memcg_oom_lock);
3820 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3821 if (ev->eventfd == eventfd) {
3822 list_del(&ev->list);
3827 spin_unlock(&memcg_oom_lock);
3830 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3832 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
3834 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3835 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3836 seq_printf(sf, "oom_kill %lu\n",
3837 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3841 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3842 struct cftype *cft, u64 val)
3844 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3846 /* cannot set to root cgroup and only 0 and 1 are allowed */
3847 if (!css->parent || !((val == 0) || (val == 1)))
3850 memcg->oom_kill_disable = val;
3852 memcg_oom_recover(memcg);
3857 #ifdef CONFIG_CGROUP_WRITEBACK
3859 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3861 return wb_domain_init(&memcg->cgwb_domain, gfp);
3864 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3866 wb_domain_exit(&memcg->cgwb_domain);
3869 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3871 wb_domain_size_changed(&memcg->cgwb_domain);
3874 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3876 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3878 if (!memcg->css.parent)
3881 return &memcg->cgwb_domain;
3885 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3886 * @wb: bdi_writeback in question
3887 * @pfilepages: out parameter for number of file pages
3888 * @pheadroom: out parameter for number of allocatable pages according to memcg
3889 * @pdirty: out parameter for number of dirty pages
3890 * @pwriteback: out parameter for number of pages under writeback
3892 * Determine the numbers of file, headroom, dirty, and writeback pages in
3893 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3894 * is a bit more involved.
3896 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3897 * headroom is calculated as the lowest headroom of itself and the
3898 * ancestors. Note that this doesn't consider the actual amount of
3899 * available memory in the system. The caller should further cap
3900 * *@pheadroom accordingly.
3902 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3903 unsigned long *pheadroom, unsigned long *pdirty,
3904 unsigned long *pwriteback)
3906 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3907 struct mem_cgroup *parent;
3909 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
3911 /* this should eventually include NR_UNSTABLE_NFS */
3912 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
3913 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3914 (1 << LRU_ACTIVE_FILE));
3915 *pheadroom = PAGE_COUNTER_MAX;
3917 while ((parent = parent_mem_cgroup(memcg))) {
3918 unsigned long ceiling = min(memcg->memory.max, memcg->high);
3919 unsigned long used = page_counter_read(&memcg->memory);
3921 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3926 #else /* CONFIG_CGROUP_WRITEBACK */
3928 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3933 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3937 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3941 #endif /* CONFIG_CGROUP_WRITEBACK */
3944 * DO NOT USE IN NEW FILES.
3946 * "cgroup.event_control" implementation.
3948 * This is way over-engineered. It tries to support fully configurable
3949 * events for each user. Such level of flexibility is completely
3950 * unnecessary especially in the light of the planned unified hierarchy.
3952 * Please deprecate this and replace with something simpler if at all
3957 * Unregister event and free resources.
3959 * Gets called from workqueue.
3961 static void memcg_event_remove(struct work_struct *work)
3963 struct mem_cgroup_event *event =
3964 container_of(work, struct mem_cgroup_event, remove);
3965 struct mem_cgroup *memcg = event->memcg;
3967 remove_wait_queue(event->wqh, &event->wait);
3969 event->unregister_event(memcg, event->eventfd);
3971 /* Notify userspace the event is going away. */
3972 eventfd_signal(event->eventfd, 1);
3974 eventfd_ctx_put(event->eventfd);
3976 css_put(&memcg->css);
3980 * Gets called on EPOLLHUP on eventfd when user closes it.
3982 * Called with wqh->lock held and interrupts disabled.
3984 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
3985 int sync, void *key)
3987 struct mem_cgroup_event *event =
3988 container_of(wait, struct mem_cgroup_event, wait);
3989 struct mem_cgroup *memcg = event->memcg;
3990 __poll_t flags = key_to_poll(key);
3992 if (flags & EPOLLHUP) {
3994 * If the event has been detached at cgroup removal, we
3995 * can simply return knowing the other side will cleanup
3998 * We can't race against event freeing since the other
3999 * side will require wqh->lock via remove_wait_queue(),
4002 spin_lock(&memcg->event_list_lock);
4003 if (!list_empty(&event->list)) {
4004 list_del_init(&event->list);
4006 * We are in atomic context, but cgroup_event_remove()
4007 * may sleep, so we have to call it in workqueue.
4009 schedule_work(&event->remove);
4011 spin_unlock(&memcg->event_list_lock);
4017 static void memcg_event_ptable_queue_proc(struct file *file,
4018 wait_queue_head_t *wqh, poll_table *pt)
4020 struct mem_cgroup_event *event =
4021 container_of(pt, struct mem_cgroup_event, pt);
4024 add_wait_queue(wqh, &event->wait);
4028 * DO NOT USE IN NEW FILES.
4030 * Parse input and register new cgroup event handler.
4032 * Input must be in format '<event_fd> <control_fd> <args>'.
4033 * Interpretation of args is defined by control file implementation.
4035 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4036 char *buf, size_t nbytes, loff_t off)
4038 struct cgroup_subsys_state *css = of_css(of);
4039 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4040 struct mem_cgroup_event *event;
4041 struct cgroup_subsys_state *cfile_css;
4042 unsigned int efd, cfd;
4049 buf = strstrip(buf);
4051 efd = simple_strtoul(buf, &endp, 10);
4056 cfd = simple_strtoul(buf, &endp, 10);
4057 if ((*endp != ' ') && (*endp != '\0'))
4061 event = kzalloc(sizeof(*event), GFP_KERNEL);
4065 event->memcg = memcg;
4066 INIT_LIST_HEAD(&event->list);
4067 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4068 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4069 INIT_WORK(&event->remove, memcg_event_remove);
4077 event->eventfd = eventfd_ctx_fileget(efile.file);
4078 if (IS_ERR(event->eventfd)) {
4079 ret = PTR_ERR(event->eventfd);
4086 goto out_put_eventfd;
4089 /* the process need read permission on control file */
4090 /* AV: shouldn't we check that it's been opened for read instead? */
4091 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4096 * Determine the event callbacks and set them in @event. This used
4097 * to be done via struct cftype but cgroup core no longer knows
4098 * about these events. The following is crude but the whole thing
4099 * is for compatibility anyway.
4101 * DO NOT ADD NEW FILES.
4103 name = cfile.file->f_path.dentry->d_name.name;
4105 if (!strcmp(name, "memory.usage_in_bytes")) {
4106 event->register_event = mem_cgroup_usage_register_event;
4107 event->unregister_event = mem_cgroup_usage_unregister_event;
4108 } else if (!strcmp(name, "memory.oom_control")) {
4109 event->register_event = mem_cgroup_oom_register_event;
4110 event->unregister_event = mem_cgroup_oom_unregister_event;
4111 } else if (!strcmp(name, "memory.pressure_level")) {
4112 event->register_event = vmpressure_register_event;
4113 event->unregister_event = vmpressure_unregister_event;
4114 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4115 event->register_event = memsw_cgroup_usage_register_event;
4116 event->unregister_event = memsw_cgroup_usage_unregister_event;
4123 * Verify @cfile should belong to @css. Also, remaining events are
4124 * automatically removed on cgroup destruction but the removal is
4125 * asynchronous, so take an extra ref on @css.
4127 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4128 &memory_cgrp_subsys);
4130 if (IS_ERR(cfile_css))
4132 if (cfile_css != css) {
4137 ret = event->register_event(memcg, event->eventfd, buf);
4141 vfs_poll(efile.file, &event->pt);
4143 spin_lock(&memcg->event_list_lock);
4144 list_add(&event->list, &memcg->event_list);
4145 spin_unlock(&memcg->event_list_lock);
4157 eventfd_ctx_put(event->eventfd);
4166 static struct cftype mem_cgroup_legacy_files[] = {
4168 .name = "usage_in_bytes",
4169 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4170 .read_u64 = mem_cgroup_read_u64,
4173 .name = "max_usage_in_bytes",
4174 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4175 .write = mem_cgroup_reset,
4176 .read_u64 = mem_cgroup_read_u64,
4179 .name = "limit_in_bytes",
4180 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4181 .write = mem_cgroup_write,
4182 .read_u64 = mem_cgroup_read_u64,
4185 .name = "soft_limit_in_bytes",
4186 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4187 .write = mem_cgroup_write,
4188 .read_u64 = mem_cgroup_read_u64,
4192 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4193 .write = mem_cgroup_reset,
4194 .read_u64 = mem_cgroup_read_u64,
4198 .seq_show = memcg_stat_show,
4201 .name = "force_empty",
4202 .write = mem_cgroup_force_empty_write,
4205 .name = "use_hierarchy",
4206 .write_u64 = mem_cgroup_hierarchy_write,
4207 .read_u64 = mem_cgroup_hierarchy_read,
4210 .name = "cgroup.event_control", /* XXX: for compat */
4211 .write = memcg_write_event_control,
4212 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4215 .name = "swappiness",
4216 .read_u64 = mem_cgroup_swappiness_read,
4217 .write_u64 = mem_cgroup_swappiness_write,
4220 .name = "move_charge_at_immigrate",
4221 .read_u64 = mem_cgroup_move_charge_read,
4222 .write_u64 = mem_cgroup_move_charge_write,
4225 .name = "oom_control",
4226 .seq_show = mem_cgroup_oom_control_read,
4227 .write_u64 = mem_cgroup_oom_control_write,
4228 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4231 .name = "pressure_level",
4235 .name = "numa_stat",
4236 .seq_show = memcg_numa_stat_show,
4240 .name = "kmem.limit_in_bytes",
4241 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4242 .write = mem_cgroup_write,
4243 .read_u64 = mem_cgroup_read_u64,
4246 .name = "kmem.usage_in_bytes",
4247 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4248 .read_u64 = mem_cgroup_read_u64,
4251 .name = "kmem.failcnt",
4252 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4253 .write = mem_cgroup_reset,
4254 .read_u64 = mem_cgroup_read_u64,
4257 .name = "kmem.max_usage_in_bytes",
4258 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4259 .write = mem_cgroup_reset,
4260 .read_u64 = mem_cgroup_read_u64,
4262 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4264 .name = "kmem.slabinfo",
4265 .seq_start = memcg_slab_start,
4266 .seq_next = memcg_slab_next,
4267 .seq_stop = memcg_slab_stop,
4268 .seq_show = memcg_slab_show,
4272 .name = "kmem.tcp.limit_in_bytes",
4273 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4274 .write = mem_cgroup_write,
4275 .read_u64 = mem_cgroup_read_u64,
4278 .name = "kmem.tcp.usage_in_bytes",
4279 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4280 .read_u64 = mem_cgroup_read_u64,
4283 .name = "kmem.tcp.failcnt",
4284 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4285 .write = mem_cgroup_reset,
4286 .read_u64 = mem_cgroup_read_u64,
4289 .name = "kmem.tcp.max_usage_in_bytes",
4290 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4291 .write = mem_cgroup_reset,
4292 .read_u64 = mem_cgroup_read_u64,
4294 { }, /* terminate */
4298 * Private memory cgroup IDR
4300 * Swap-out records and page cache shadow entries need to store memcg
4301 * references in constrained space, so we maintain an ID space that is
4302 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4303 * memory-controlled cgroups to 64k.
4305 * However, there usually are many references to the oflline CSS after
4306 * the cgroup has been destroyed, such as page cache or reclaimable
4307 * slab objects, that don't need to hang on to the ID. We want to keep
4308 * those dead CSS from occupying IDs, or we might quickly exhaust the
4309 * relatively small ID space and prevent the creation of new cgroups
4310 * even when there are much fewer than 64k cgroups - possibly none.
4312 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4313 * be freed and recycled when it's no longer needed, which is usually
4314 * when the CSS is offlined.
4316 * The only exception to that are records of swapped out tmpfs/shmem
4317 * pages that need to be attributed to live ancestors on swapin. But
4318 * those references are manageable from userspace.
4321 static DEFINE_IDR(mem_cgroup_idr);
4323 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4325 if (memcg->id.id > 0) {
4326 idr_remove(&mem_cgroup_idr, memcg->id.id);
4331 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4333 refcount_add(n, &memcg->id.ref);
4336 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4338 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4339 mem_cgroup_id_remove(memcg);
4341 /* Memcg ID pins CSS */
4342 css_put(&memcg->css);
4346 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4348 mem_cgroup_id_get_many(memcg, 1);
4351 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4353 mem_cgroup_id_put_many(memcg, 1);
4357 * mem_cgroup_from_id - look up a memcg from a memcg id
4358 * @id: the memcg id to look up
4360 * Caller must hold rcu_read_lock().
4362 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4364 WARN_ON_ONCE(!rcu_read_lock_held());
4365 return idr_find(&mem_cgroup_idr, id);
4368 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4370 struct mem_cgroup_per_node *pn;
4373 * This routine is called against possible nodes.
4374 * But it's BUG to call kmalloc() against offline node.
4376 * TODO: this routine can waste much memory for nodes which will
4377 * never be onlined. It's better to use memory hotplug callback
4380 if (!node_state(node, N_NORMAL_MEMORY))
4382 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4386 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4387 if (!pn->lruvec_stat_cpu) {
4392 lruvec_init(&pn->lruvec);
4393 pn->usage_in_excess = 0;
4394 pn->on_tree = false;
4397 memcg->nodeinfo[node] = pn;
4401 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4403 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4408 free_percpu(pn->lruvec_stat_cpu);
4412 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4417 free_mem_cgroup_per_node_info(memcg, node);
4418 free_percpu(memcg->stat_cpu);
4422 static void mem_cgroup_free(struct mem_cgroup *memcg)
4424 memcg_wb_domain_exit(memcg);
4425 __mem_cgroup_free(memcg);
4428 static struct mem_cgroup *mem_cgroup_alloc(void)
4430 struct mem_cgroup *memcg;
4434 size = sizeof(struct mem_cgroup);
4435 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4437 memcg = kzalloc(size, GFP_KERNEL);
4441 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4442 1, MEM_CGROUP_ID_MAX,
4444 if (memcg->id.id < 0)
4447 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4448 if (!memcg->stat_cpu)
4452 if (alloc_mem_cgroup_per_node_info(memcg, node))
4455 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4458 INIT_WORK(&memcg->high_work, high_work_func);
4459 memcg->last_scanned_node = MAX_NUMNODES;
4460 INIT_LIST_HEAD(&memcg->oom_notify);
4461 mutex_init(&memcg->thresholds_lock);
4462 spin_lock_init(&memcg->move_lock);
4463 vmpressure_init(&memcg->vmpressure);
4464 INIT_LIST_HEAD(&memcg->event_list);
4465 spin_lock_init(&memcg->event_list_lock);
4466 memcg->socket_pressure = jiffies;
4467 #ifdef CONFIG_MEMCG_KMEM
4468 memcg->kmemcg_id = -1;
4470 #ifdef CONFIG_CGROUP_WRITEBACK
4471 INIT_LIST_HEAD(&memcg->cgwb_list);
4473 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4476 mem_cgroup_id_remove(memcg);
4477 __mem_cgroup_free(memcg);
4481 static struct cgroup_subsys_state * __ref
4482 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4484 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4485 struct mem_cgroup *memcg;
4486 long error = -ENOMEM;
4488 memcg = mem_cgroup_alloc();
4490 return ERR_PTR(error);
4492 memcg->high = PAGE_COUNTER_MAX;
4493 memcg->soft_limit = PAGE_COUNTER_MAX;
4495 memcg->swappiness = mem_cgroup_swappiness(parent);
4496 memcg->oom_kill_disable = parent->oom_kill_disable;
4498 if (parent && parent->use_hierarchy) {
4499 memcg->use_hierarchy = true;
4500 page_counter_init(&memcg->memory, &parent->memory);
4501 page_counter_init(&memcg->swap, &parent->swap);
4502 page_counter_init(&memcg->memsw, &parent->memsw);
4503 page_counter_init(&memcg->kmem, &parent->kmem);
4504 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4506 page_counter_init(&memcg->memory, NULL);
4507 page_counter_init(&memcg->swap, NULL);
4508 page_counter_init(&memcg->memsw, NULL);
4509 page_counter_init(&memcg->kmem, NULL);
4510 page_counter_init(&memcg->tcpmem, NULL);
4512 * Deeper hierachy with use_hierarchy == false doesn't make
4513 * much sense so let cgroup subsystem know about this
4514 * unfortunate state in our controller.
4516 if (parent != root_mem_cgroup)
4517 memory_cgrp_subsys.broken_hierarchy = true;
4520 /* The following stuff does not apply to the root */
4522 root_mem_cgroup = memcg;
4526 error = memcg_online_kmem(memcg);
4530 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4531 static_branch_inc(&memcg_sockets_enabled_key);
4535 mem_cgroup_id_remove(memcg);
4536 mem_cgroup_free(memcg);
4537 return ERR_PTR(-ENOMEM);
4540 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4542 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4545 * A memcg must be visible for memcg_expand_shrinker_maps()
4546 * by the time the maps are allocated. So, we allocate maps
4547 * here, when for_each_mem_cgroup() can't skip it.
4549 if (memcg_alloc_shrinker_maps(memcg)) {
4550 mem_cgroup_id_remove(memcg);
4554 /* Online state pins memcg ID, memcg ID pins CSS */
4555 refcount_set(&memcg->id.ref, 1);
4560 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4562 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4563 struct mem_cgroup_event *event, *tmp;
4566 * Unregister events and notify userspace.
4567 * Notify userspace about cgroup removing only after rmdir of cgroup
4568 * directory to avoid race between userspace and kernelspace.
4570 spin_lock(&memcg->event_list_lock);
4571 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4572 list_del_init(&event->list);
4573 schedule_work(&event->remove);
4575 spin_unlock(&memcg->event_list_lock);
4577 page_counter_set_min(&memcg->memory, 0);
4578 page_counter_set_low(&memcg->memory, 0);
4580 memcg_offline_kmem(memcg);
4581 wb_memcg_offline(memcg);
4583 drain_all_stock(memcg);
4585 mem_cgroup_id_put(memcg);
4588 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4590 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4592 invalidate_reclaim_iterators(memcg);
4595 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4597 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4599 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4600 static_branch_dec(&memcg_sockets_enabled_key);
4602 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4603 static_branch_dec(&memcg_sockets_enabled_key);
4605 vmpressure_cleanup(&memcg->vmpressure);
4606 cancel_work_sync(&memcg->high_work);
4607 mem_cgroup_remove_from_trees(memcg);
4608 memcg_free_shrinker_maps(memcg);
4609 memcg_free_kmem(memcg);
4610 mem_cgroup_free(memcg);
4614 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4615 * @css: the target css
4617 * Reset the states of the mem_cgroup associated with @css. This is
4618 * invoked when the userland requests disabling on the default hierarchy
4619 * but the memcg is pinned through dependency. The memcg should stop
4620 * applying policies and should revert to the vanilla state as it may be
4621 * made visible again.
4623 * The current implementation only resets the essential configurations.
4624 * This needs to be expanded to cover all the visible parts.
4626 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4628 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4630 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4631 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4632 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4633 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4634 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4635 page_counter_set_min(&memcg->memory, 0);
4636 page_counter_set_low(&memcg->memory, 0);
4637 memcg->high = PAGE_COUNTER_MAX;
4638 memcg->soft_limit = PAGE_COUNTER_MAX;
4639 memcg_wb_domain_size_changed(memcg);
4643 /* Handlers for move charge at task migration. */
4644 static int mem_cgroup_do_precharge(unsigned long count)
4648 /* Try a single bulk charge without reclaim first, kswapd may wake */
4649 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4651 mc.precharge += count;
4655 /* Try charges one by one with reclaim, but do not retry */
4657 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4671 enum mc_target_type {
4678 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4679 unsigned long addr, pte_t ptent)
4681 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4683 if (!page || !page_mapped(page))
4685 if (PageAnon(page)) {
4686 if (!(mc.flags & MOVE_ANON))
4689 if (!(mc.flags & MOVE_FILE))
4692 if (!get_page_unless_zero(page))
4698 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4699 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4700 pte_t ptent, swp_entry_t *entry)
4702 struct page *page = NULL;
4703 swp_entry_t ent = pte_to_swp_entry(ptent);
4705 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4709 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4710 * a device and because they are not accessible by CPU they are store
4711 * as special swap entry in the CPU page table.
4713 if (is_device_private_entry(ent)) {
4714 page = device_private_entry_to_page(ent);
4716 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4717 * a refcount of 1 when free (unlike normal page)
4719 if (!page_ref_add_unless(page, 1, 1))
4725 * Because lookup_swap_cache() updates some statistics counter,
4726 * we call find_get_page() with swapper_space directly.
4728 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4729 if (do_memsw_account())
4730 entry->val = ent.val;
4735 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4736 pte_t ptent, swp_entry_t *entry)
4742 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4743 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4745 struct page *page = NULL;
4746 struct address_space *mapping;
4749 if (!vma->vm_file) /* anonymous vma */
4751 if (!(mc.flags & MOVE_FILE))
4754 mapping = vma->vm_file->f_mapping;
4755 pgoff = linear_page_index(vma, addr);
4757 /* page is moved even if it's not RSS of this task(page-faulted). */
4759 /* shmem/tmpfs may report page out on swap: account for that too. */
4760 if (shmem_mapping(mapping)) {
4761 page = find_get_entry(mapping, pgoff);
4762 if (xa_is_value(page)) {
4763 swp_entry_t swp = radix_to_swp_entry(page);
4764 if (do_memsw_account())
4766 page = find_get_page(swap_address_space(swp),
4770 page = find_get_page(mapping, pgoff);
4772 page = find_get_page(mapping, pgoff);
4778 * mem_cgroup_move_account - move account of the page
4780 * @compound: charge the page as compound or small page
4781 * @from: mem_cgroup which the page is moved from.
4782 * @to: mem_cgroup which the page is moved to. @from != @to.
4784 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4786 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4789 static int mem_cgroup_move_account(struct page *page,
4791 struct mem_cgroup *from,
4792 struct mem_cgroup *to)
4794 unsigned long flags;
4795 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4799 VM_BUG_ON(from == to);
4800 VM_BUG_ON_PAGE(PageLRU(page), page);
4801 VM_BUG_ON(compound && !PageTransHuge(page));
4804 * Prevent mem_cgroup_migrate() from looking at
4805 * page->mem_cgroup of its source page while we change it.
4808 if (!trylock_page(page))
4812 if (page->mem_cgroup != from)
4815 anon = PageAnon(page);
4817 spin_lock_irqsave(&from->move_lock, flags);
4819 if (!anon && page_mapped(page)) {
4820 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4821 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4825 * move_lock grabbed above and caller set from->moving_account, so
4826 * mod_memcg_page_state will serialize updates to PageDirty.
4827 * So mapping should be stable for dirty pages.
4829 if (!anon && PageDirty(page)) {
4830 struct address_space *mapping = page_mapping(page);
4832 if (mapping_cap_account_dirty(mapping)) {
4833 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4834 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4838 if (PageWriteback(page)) {
4839 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4840 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4844 * It is safe to change page->mem_cgroup here because the page
4845 * is referenced, charged, and isolated - we can't race with
4846 * uncharging, charging, migration, or LRU putback.
4849 /* caller should have done css_get */
4850 page->mem_cgroup = to;
4851 spin_unlock_irqrestore(&from->move_lock, flags);
4855 local_irq_disable();
4856 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4857 memcg_check_events(to, page);
4858 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4859 memcg_check_events(from, page);
4868 * get_mctgt_type - get target type of moving charge
4869 * @vma: the vma the pte to be checked belongs
4870 * @addr: the address corresponding to the pte to be checked
4871 * @ptent: the pte to be checked
4872 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4875 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4876 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4877 * move charge. if @target is not NULL, the page is stored in target->page
4878 * with extra refcnt got(Callers should handle it).
4879 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4880 * target for charge migration. if @target is not NULL, the entry is stored
4882 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4883 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4884 * For now we such page is charge like a regular page would be as for all
4885 * intent and purposes it is just special memory taking the place of a
4888 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4890 * Called with pte lock held.
4893 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4894 unsigned long addr, pte_t ptent, union mc_target *target)
4896 struct page *page = NULL;
4897 enum mc_target_type ret = MC_TARGET_NONE;
4898 swp_entry_t ent = { .val = 0 };
4900 if (pte_present(ptent))
4901 page = mc_handle_present_pte(vma, addr, ptent);
4902 else if (is_swap_pte(ptent))
4903 page = mc_handle_swap_pte(vma, ptent, &ent);
4904 else if (pte_none(ptent))
4905 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4907 if (!page && !ent.val)
4911 * Do only loose check w/o serialization.
4912 * mem_cgroup_move_account() checks the page is valid or
4913 * not under LRU exclusion.
4915 if (page->mem_cgroup == mc.from) {
4916 ret = MC_TARGET_PAGE;
4917 if (is_device_private_page(page) ||
4918 is_device_public_page(page))
4919 ret = MC_TARGET_DEVICE;
4921 target->page = page;
4923 if (!ret || !target)
4927 * There is a swap entry and a page doesn't exist or isn't charged.
4928 * But we cannot move a tail-page in a THP.
4930 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
4931 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4932 ret = MC_TARGET_SWAP;
4939 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4941 * We don't consider PMD mapped swapping or file mapped pages because THP does
4942 * not support them for now.
4943 * Caller should make sure that pmd_trans_huge(pmd) is true.
4945 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4946 unsigned long addr, pmd_t pmd, union mc_target *target)
4948 struct page *page = NULL;
4949 enum mc_target_type ret = MC_TARGET_NONE;
4951 if (unlikely(is_swap_pmd(pmd))) {
4952 VM_BUG_ON(thp_migration_supported() &&
4953 !is_pmd_migration_entry(pmd));
4956 page = pmd_page(pmd);
4957 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4958 if (!(mc.flags & MOVE_ANON))
4960 if (page->mem_cgroup == mc.from) {
4961 ret = MC_TARGET_PAGE;
4964 target->page = page;
4970 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4971 unsigned long addr, pmd_t pmd, union mc_target *target)
4973 return MC_TARGET_NONE;
4977 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4978 unsigned long addr, unsigned long end,
4979 struct mm_walk *walk)
4981 struct vm_area_struct *vma = walk->vma;
4985 ptl = pmd_trans_huge_lock(pmd, vma);
4988 * Note their can not be MC_TARGET_DEVICE for now as we do not
4989 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
4990 * MEMORY_DEVICE_PRIVATE but this might change.
4992 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4993 mc.precharge += HPAGE_PMD_NR;
4998 if (pmd_trans_unstable(pmd))
5000 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5001 for (; addr != end; pte++, addr += PAGE_SIZE)
5002 if (get_mctgt_type(vma, addr, *pte, NULL))
5003 mc.precharge++; /* increment precharge temporarily */
5004 pte_unmap_unlock(pte - 1, ptl);
5010 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5012 unsigned long precharge;
5014 struct mm_walk mem_cgroup_count_precharge_walk = {
5015 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5018 down_read(&mm->mmap_sem);
5019 walk_page_range(0, mm->highest_vm_end,
5020 &mem_cgroup_count_precharge_walk);
5021 up_read(&mm->mmap_sem);
5023 precharge = mc.precharge;
5029 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5031 unsigned long precharge = mem_cgroup_count_precharge(mm);
5033 VM_BUG_ON(mc.moving_task);
5034 mc.moving_task = current;
5035 return mem_cgroup_do_precharge(precharge);
5038 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5039 static void __mem_cgroup_clear_mc(void)
5041 struct mem_cgroup *from = mc.from;
5042 struct mem_cgroup *to = mc.to;
5044 /* we must uncharge all the leftover precharges from mc.to */
5046 cancel_charge(mc.to, mc.precharge);
5050 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5051 * we must uncharge here.
5053 if (mc.moved_charge) {
5054 cancel_charge(mc.from, mc.moved_charge);
5055 mc.moved_charge = 0;
5057 /* we must fixup refcnts and charges */
5058 if (mc.moved_swap) {
5059 /* uncharge swap account from the old cgroup */
5060 if (!mem_cgroup_is_root(mc.from))
5061 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5063 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5066 * we charged both to->memory and to->memsw, so we
5067 * should uncharge to->memory.
5069 if (!mem_cgroup_is_root(mc.to))
5070 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5072 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5073 css_put_many(&mc.to->css, mc.moved_swap);
5077 memcg_oom_recover(from);
5078 memcg_oom_recover(to);
5079 wake_up_all(&mc.waitq);
5082 static void mem_cgroup_clear_mc(void)
5084 struct mm_struct *mm = mc.mm;
5087 * we must clear moving_task before waking up waiters at the end of
5090 mc.moving_task = NULL;
5091 __mem_cgroup_clear_mc();
5092 spin_lock(&mc.lock);
5096 spin_unlock(&mc.lock);
5101 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5103 struct cgroup_subsys_state *css;
5104 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5105 struct mem_cgroup *from;
5106 struct task_struct *leader, *p;
5107 struct mm_struct *mm;
5108 unsigned long move_flags;
5111 /* charge immigration isn't supported on the default hierarchy */
5112 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5116 * Multi-process migrations only happen on the default hierarchy
5117 * where charge immigration is not used. Perform charge
5118 * immigration if @tset contains a leader and whine if there are
5122 cgroup_taskset_for_each_leader(leader, css, tset) {
5125 memcg = mem_cgroup_from_css(css);
5131 * We are now commited to this value whatever it is. Changes in this
5132 * tunable will only affect upcoming migrations, not the current one.
5133 * So we need to save it, and keep it going.
5135 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5139 from = mem_cgroup_from_task(p);
5141 VM_BUG_ON(from == memcg);
5143 mm = get_task_mm(p);
5146 /* We move charges only when we move a owner of the mm */
5147 if (mm->owner == p) {
5150 VM_BUG_ON(mc.precharge);
5151 VM_BUG_ON(mc.moved_charge);
5152 VM_BUG_ON(mc.moved_swap);
5154 spin_lock(&mc.lock);
5158 mc.flags = move_flags;
5159 spin_unlock(&mc.lock);
5160 /* We set mc.moving_task later */
5162 ret = mem_cgroup_precharge_mc(mm);
5164 mem_cgroup_clear_mc();
5171 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5174 mem_cgroup_clear_mc();
5177 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5178 unsigned long addr, unsigned long end,
5179 struct mm_walk *walk)
5182 struct vm_area_struct *vma = walk->vma;
5185 enum mc_target_type target_type;
5186 union mc_target target;
5189 ptl = pmd_trans_huge_lock(pmd, vma);
5191 if (mc.precharge < HPAGE_PMD_NR) {
5195 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5196 if (target_type == MC_TARGET_PAGE) {
5198 if (!isolate_lru_page(page)) {
5199 if (!mem_cgroup_move_account(page, true,
5201 mc.precharge -= HPAGE_PMD_NR;
5202 mc.moved_charge += HPAGE_PMD_NR;
5204 putback_lru_page(page);
5207 } else if (target_type == MC_TARGET_DEVICE) {
5209 if (!mem_cgroup_move_account(page, true,
5211 mc.precharge -= HPAGE_PMD_NR;
5212 mc.moved_charge += HPAGE_PMD_NR;
5220 if (pmd_trans_unstable(pmd))
5223 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5224 for (; addr != end; addr += PAGE_SIZE) {
5225 pte_t ptent = *(pte++);
5226 bool device = false;
5232 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5233 case MC_TARGET_DEVICE:
5236 case MC_TARGET_PAGE:
5239 * We can have a part of the split pmd here. Moving it
5240 * can be done but it would be too convoluted so simply
5241 * ignore such a partial THP and keep it in original
5242 * memcg. There should be somebody mapping the head.
5244 if (PageTransCompound(page))
5246 if (!device && isolate_lru_page(page))
5248 if (!mem_cgroup_move_account(page, false,
5251 /* we uncharge from mc.from later. */
5255 putback_lru_page(page);
5256 put: /* get_mctgt_type() gets the page */
5259 case MC_TARGET_SWAP:
5261 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5263 /* we fixup refcnts and charges later. */
5271 pte_unmap_unlock(pte - 1, ptl);
5276 * We have consumed all precharges we got in can_attach().
5277 * We try charge one by one, but don't do any additional
5278 * charges to mc.to if we have failed in charge once in attach()
5281 ret = mem_cgroup_do_precharge(1);
5289 static void mem_cgroup_move_charge(void)
5291 struct mm_walk mem_cgroup_move_charge_walk = {
5292 .pmd_entry = mem_cgroup_move_charge_pte_range,
5296 lru_add_drain_all();
5298 * Signal lock_page_memcg() to take the memcg's move_lock
5299 * while we're moving its pages to another memcg. Then wait
5300 * for already started RCU-only updates to finish.
5302 atomic_inc(&mc.from->moving_account);
5305 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5307 * Someone who are holding the mmap_sem might be waiting in
5308 * waitq. So we cancel all extra charges, wake up all waiters,
5309 * and retry. Because we cancel precharges, we might not be able
5310 * to move enough charges, but moving charge is a best-effort
5311 * feature anyway, so it wouldn't be a big problem.
5313 __mem_cgroup_clear_mc();
5318 * When we have consumed all precharges and failed in doing
5319 * additional charge, the page walk just aborts.
5321 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5323 up_read(&mc.mm->mmap_sem);
5324 atomic_dec(&mc.from->moving_account);
5327 static void mem_cgroup_move_task(void)
5330 mem_cgroup_move_charge();
5331 mem_cgroup_clear_mc();
5334 #else /* !CONFIG_MMU */
5335 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5339 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5342 static void mem_cgroup_move_task(void)
5348 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5349 * to verify whether we're attached to the default hierarchy on each mount
5352 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5355 * use_hierarchy is forced on the default hierarchy. cgroup core
5356 * guarantees that @root doesn't have any children, so turning it
5357 * on for the root memcg is enough.
5359 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5360 root_mem_cgroup->use_hierarchy = true;
5362 root_mem_cgroup->use_hierarchy = false;
5365 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5367 if (value == PAGE_COUNTER_MAX)
5368 seq_puts(m, "max\n");
5370 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5375 static u64 memory_current_read(struct cgroup_subsys_state *css,
5378 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5380 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5383 static int memory_min_show(struct seq_file *m, void *v)
5385 return seq_puts_memcg_tunable(m,
5386 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5389 static ssize_t memory_min_write(struct kernfs_open_file *of,
5390 char *buf, size_t nbytes, loff_t off)
5392 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5396 buf = strstrip(buf);
5397 err = page_counter_memparse(buf, "max", &min);
5401 page_counter_set_min(&memcg->memory, min);
5406 static int memory_low_show(struct seq_file *m, void *v)
5408 return seq_puts_memcg_tunable(m,
5409 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5412 static ssize_t memory_low_write(struct kernfs_open_file *of,
5413 char *buf, size_t nbytes, loff_t off)
5415 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5419 buf = strstrip(buf);
5420 err = page_counter_memparse(buf, "max", &low);
5424 page_counter_set_low(&memcg->memory, low);
5429 static int memory_high_show(struct seq_file *m, void *v)
5431 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5434 static ssize_t memory_high_write(struct kernfs_open_file *of,
5435 char *buf, size_t nbytes, loff_t off)
5437 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5438 unsigned long nr_pages;
5442 buf = strstrip(buf);
5443 err = page_counter_memparse(buf, "max", &high);
5449 nr_pages = page_counter_read(&memcg->memory);
5450 if (nr_pages > high)
5451 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5454 memcg_wb_domain_size_changed(memcg);
5458 static int memory_max_show(struct seq_file *m, void *v)
5460 return seq_puts_memcg_tunable(m,
5461 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5464 static ssize_t memory_max_write(struct kernfs_open_file *of,
5465 char *buf, size_t nbytes, loff_t off)
5467 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5468 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5469 bool drained = false;
5473 buf = strstrip(buf);
5474 err = page_counter_memparse(buf, "max", &max);
5478 xchg(&memcg->memory.max, max);
5481 unsigned long nr_pages = page_counter_read(&memcg->memory);
5483 if (nr_pages <= max)
5486 if (signal_pending(current)) {
5492 drain_all_stock(memcg);
5498 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5504 memcg_memory_event(memcg, MEMCG_OOM);
5505 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5509 memcg_wb_domain_size_changed(memcg);
5513 static int memory_events_show(struct seq_file *m, void *v)
5515 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5517 seq_printf(m, "low %lu\n",
5518 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5519 seq_printf(m, "high %lu\n",
5520 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5521 seq_printf(m, "max %lu\n",
5522 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5523 seq_printf(m, "oom %lu\n",
5524 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5525 seq_printf(m, "oom_kill %lu\n",
5526 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5531 static int memory_stat_show(struct seq_file *m, void *v)
5533 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5534 struct accumulated_stats acc;
5538 * Provide statistics on the state of the memory subsystem as
5539 * well as cumulative event counters that show past behavior.
5541 * This list is ordered following a combination of these gradients:
5542 * 1) generic big picture -> specifics and details
5543 * 2) reflecting userspace activity -> reflecting kernel heuristics
5545 * Current memory state:
5548 memset(&acc, 0, sizeof(acc));
5549 acc.stats_size = MEMCG_NR_STAT;
5550 acc.events_size = NR_VM_EVENT_ITEMS;
5551 accumulate_memcg_tree(memcg, &acc);
5553 seq_printf(m, "anon %llu\n",
5554 (u64)acc.stat[MEMCG_RSS] * PAGE_SIZE);
5555 seq_printf(m, "file %llu\n",
5556 (u64)acc.stat[MEMCG_CACHE] * PAGE_SIZE);
5557 seq_printf(m, "kernel_stack %llu\n",
5558 (u64)acc.stat[MEMCG_KERNEL_STACK_KB] * 1024);
5559 seq_printf(m, "slab %llu\n",
5560 (u64)(acc.stat[NR_SLAB_RECLAIMABLE] +
5561 acc.stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5562 seq_printf(m, "sock %llu\n",
5563 (u64)acc.stat[MEMCG_SOCK] * PAGE_SIZE);
5565 seq_printf(m, "shmem %llu\n",
5566 (u64)acc.stat[NR_SHMEM] * PAGE_SIZE);
5567 seq_printf(m, "file_mapped %llu\n",
5568 (u64)acc.stat[NR_FILE_MAPPED] * PAGE_SIZE);
5569 seq_printf(m, "file_dirty %llu\n",
5570 (u64)acc.stat[NR_FILE_DIRTY] * PAGE_SIZE);
5571 seq_printf(m, "file_writeback %llu\n",
5572 (u64)acc.stat[NR_WRITEBACK] * PAGE_SIZE);
5574 for (i = 0; i < NR_LRU_LISTS; i++)
5575 seq_printf(m, "%s %llu\n", mem_cgroup_lru_names[i],
5576 (u64)acc.lru_pages[i] * PAGE_SIZE);
5578 seq_printf(m, "slab_reclaimable %llu\n",
5579 (u64)acc.stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5580 seq_printf(m, "slab_unreclaimable %llu\n",
5581 (u64)acc.stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5583 /* Accumulated memory events */
5585 seq_printf(m, "pgfault %lu\n", acc.events[PGFAULT]);
5586 seq_printf(m, "pgmajfault %lu\n", acc.events[PGMAJFAULT]);
5588 seq_printf(m, "workingset_refault %lu\n",
5589 acc.stat[WORKINGSET_REFAULT]);
5590 seq_printf(m, "workingset_activate %lu\n",
5591 acc.stat[WORKINGSET_ACTIVATE]);
5592 seq_printf(m, "workingset_nodereclaim %lu\n",
5593 acc.stat[WORKINGSET_NODERECLAIM]);
5595 seq_printf(m, "pgrefill %lu\n", acc.events[PGREFILL]);
5596 seq_printf(m, "pgscan %lu\n", acc.events[PGSCAN_KSWAPD] +
5597 acc.events[PGSCAN_DIRECT]);
5598 seq_printf(m, "pgsteal %lu\n", acc.events[PGSTEAL_KSWAPD] +
5599 acc.events[PGSTEAL_DIRECT]);
5600 seq_printf(m, "pgactivate %lu\n", acc.events[PGACTIVATE]);
5601 seq_printf(m, "pgdeactivate %lu\n", acc.events[PGDEACTIVATE]);
5602 seq_printf(m, "pglazyfree %lu\n", acc.events[PGLAZYFREE]);
5603 seq_printf(m, "pglazyfreed %lu\n", acc.events[PGLAZYFREED]);
5608 static int memory_oom_group_show(struct seq_file *m, void *v)
5610 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5612 seq_printf(m, "%d\n", memcg->oom_group);
5617 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5618 char *buf, size_t nbytes, loff_t off)
5620 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5623 buf = strstrip(buf);
5627 ret = kstrtoint(buf, 0, &oom_group);
5631 if (oom_group != 0 && oom_group != 1)
5634 memcg->oom_group = oom_group;
5639 static struct cftype memory_files[] = {
5642 .flags = CFTYPE_NOT_ON_ROOT,
5643 .read_u64 = memory_current_read,
5647 .flags = CFTYPE_NOT_ON_ROOT,
5648 .seq_show = memory_min_show,
5649 .write = memory_min_write,
5653 .flags = CFTYPE_NOT_ON_ROOT,
5654 .seq_show = memory_low_show,
5655 .write = memory_low_write,
5659 .flags = CFTYPE_NOT_ON_ROOT,
5660 .seq_show = memory_high_show,
5661 .write = memory_high_write,
5665 .flags = CFTYPE_NOT_ON_ROOT,
5666 .seq_show = memory_max_show,
5667 .write = memory_max_write,
5671 .flags = CFTYPE_NOT_ON_ROOT,
5672 .file_offset = offsetof(struct mem_cgroup, events_file),
5673 .seq_show = memory_events_show,
5677 .flags = CFTYPE_NOT_ON_ROOT,
5678 .seq_show = memory_stat_show,
5681 .name = "oom.group",
5682 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5683 .seq_show = memory_oom_group_show,
5684 .write = memory_oom_group_write,
5689 struct cgroup_subsys memory_cgrp_subsys = {
5690 .css_alloc = mem_cgroup_css_alloc,
5691 .css_online = mem_cgroup_css_online,
5692 .css_offline = mem_cgroup_css_offline,
5693 .css_released = mem_cgroup_css_released,
5694 .css_free = mem_cgroup_css_free,
5695 .css_reset = mem_cgroup_css_reset,
5696 .can_attach = mem_cgroup_can_attach,
5697 .cancel_attach = mem_cgroup_cancel_attach,
5698 .post_attach = mem_cgroup_move_task,
5699 .bind = mem_cgroup_bind,
5700 .dfl_cftypes = memory_files,
5701 .legacy_cftypes = mem_cgroup_legacy_files,
5706 * mem_cgroup_protected - check if memory consumption is in the normal range
5707 * @root: the top ancestor of the sub-tree being checked
5708 * @memcg: the memory cgroup to check
5710 * WARNING: This function is not stateless! It can only be used as part
5711 * of a top-down tree iteration, not for isolated queries.
5713 * Returns one of the following:
5714 * MEMCG_PROT_NONE: cgroup memory is not protected
5715 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5716 * an unprotected supply of reclaimable memory from other cgroups.
5717 * MEMCG_PROT_MIN: cgroup memory is protected
5719 * @root is exclusive; it is never protected when looked at directly
5721 * To provide a proper hierarchical behavior, effective memory.min/low values
5722 * are used. Below is the description of how effective memory.low is calculated.
5723 * Effective memory.min values is calculated in the same way.
5725 * Effective memory.low is always equal or less than the original memory.low.
5726 * If there is no memory.low overcommittment (which is always true for
5727 * top-level memory cgroups), these two values are equal.
5728 * Otherwise, it's a part of parent's effective memory.low,
5729 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5730 * memory.low usages, where memory.low usage is the size of actually
5734 * elow = min( memory.low, parent->elow * ------------------ ),
5735 * siblings_low_usage
5737 * | memory.current, if memory.current < memory.low
5742 * Such definition of the effective memory.low provides the expected
5743 * hierarchical behavior: parent's memory.low value is limiting
5744 * children, unprotected memory is reclaimed first and cgroups,
5745 * which are not using their guarantee do not affect actual memory
5748 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5750 * A A/memory.low = 2G, A/memory.current = 6G
5752 * BC DE B/memory.low = 3G B/memory.current = 2G
5753 * C/memory.low = 1G C/memory.current = 2G
5754 * D/memory.low = 0 D/memory.current = 2G
5755 * E/memory.low = 10G E/memory.current = 0
5757 * and the memory pressure is applied, the following memory distribution
5758 * is expected (approximately):
5760 * A/memory.current = 2G
5762 * B/memory.current = 1.3G
5763 * C/memory.current = 0.6G
5764 * D/memory.current = 0
5765 * E/memory.current = 0
5767 * These calculations require constant tracking of the actual low usages
5768 * (see propagate_protected_usage()), as well as recursive calculation of
5769 * effective memory.low values. But as we do call mem_cgroup_protected()
5770 * path for each memory cgroup top-down from the reclaim,
5771 * it's possible to optimize this part, and save calculated elow
5772 * for next usage. This part is intentionally racy, but it's ok,
5773 * as memory.low is a best-effort mechanism.
5775 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5776 struct mem_cgroup *memcg)
5778 struct mem_cgroup *parent;
5779 unsigned long emin, parent_emin;
5780 unsigned long elow, parent_elow;
5781 unsigned long usage;
5783 if (mem_cgroup_disabled())
5784 return MEMCG_PROT_NONE;
5787 root = root_mem_cgroup;
5789 return MEMCG_PROT_NONE;
5791 usage = page_counter_read(&memcg->memory);
5793 return MEMCG_PROT_NONE;
5795 emin = memcg->memory.min;
5796 elow = memcg->memory.low;
5798 parent = parent_mem_cgroup(memcg);
5799 /* No parent means a non-hierarchical mode on v1 memcg */
5801 return MEMCG_PROT_NONE;
5806 parent_emin = READ_ONCE(parent->memory.emin);
5807 emin = min(emin, parent_emin);
5808 if (emin && parent_emin) {
5809 unsigned long min_usage, siblings_min_usage;
5811 min_usage = min(usage, memcg->memory.min);
5812 siblings_min_usage = atomic_long_read(
5813 &parent->memory.children_min_usage);
5815 if (min_usage && siblings_min_usage)
5816 emin = min(emin, parent_emin * min_usage /
5817 siblings_min_usage);
5820 parent_elow = READ_ONCE(parent->memory.elow);
5821 elow = min(elow, parent_elow);
5822 if (elow && parent_elow) {
5823 unsigned long low_usage, siblings_low_usage;
5825 low_usage = min(usage, memcg->memory.low);
5826 siblings_low_usage = atomic_long_read(
5827 &parent->memory.children_low_usage);
5829 if (low_usage && siblings_low_usage)
5830 elow = min(elow, parent_elow * low_usage /
5831 siblings_low_usage);
5835 memcg->memory.emin = emin;
5836 memcg->memory.elow = elow;
5839 return MEMCG_PROT_MIN;
5840 else if (usage <= elow)
5841 return MEMCG_PROT_LOW;
5843 return MEMCG_PROT_NONE;
5847 * mem_cgroup_try_charge - try charging a page
5848 * @page: page to charge
5849 * @mm: mm context of the victim
5850 * @gfp_mask: reclaim mode
5851 * @memcgp: charged memcg return
5852 * @compound: charge the page as compound or small page
5854 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5855 * pages according to @gfp_mask if necessary.
5857 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5858 * Otherwise, an error code is returned.
5860 * After page->mapping has been set up, the caller must finalize the
5861 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5862 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5864 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5865 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5868 struct mem_cgroup *memcg = NULL;
5869 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5872 if (mem_cgroup_disabled())
5875 if (PageSwapCache(page)) {
5877 * Every swap fault against a single page tries to charge the
5878 * page, bail as early as possible. shmem_unuse() encounters
5879 * already charged pages, too. The USED bit is protected by
5880 * the page lock, which serializes swap cache removal, which
5881 * in turn serializes uncharging.
5883 VM_BUG_ON_PAGE(!PageLocked(page), page);
5884 if (compound_head(page)->mem_cgroup)
5887 if (do_swap_account) {
5888 swp_entry_t ent = { .val = page_private(page), };
5889 unsigned short id = lookup_swap_cgroup_id(ent);
5892 memcg = mem_cgroup_from_id(id);
5893 if (memcg && !css_tryget_online(&memcg->css))
5900 memcg = get_mem_cgroup_from_mm(mm);
5902 ret = try_charge(memcg, gfp_mask, nr_pages);
5904 css_put(&memcg->css);
5910 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
5911 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5914 struct mem_cgroup *memcg;
5917 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
5919 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
5924 * mem_cgroup_commit_charge - commit a page charge
5925 * @page: page to charge
5926 * @memcg: memcg to charge the page to
5927 * @lrucare: page might be on LRU already
5928 * @compound: charge the page as compound or small page
5930 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5931 * after page->mapping has been set up. This must happen atomically
5932 * as part of the page instantiation, i.e. under the page table lock
5933 * for anonymous pages, under the page lock for page and swap cache.
5935 * In addition, the page must not be on the LRU during the commit, to
5936 * prevent racing with task migration. If it might be, use @lrucare.
5938 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5940 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5941 bool lrucare, bool compound)
5943 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5945 VM_BUG_ON_PAGE(!page->mapping, page);
5946 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5948 if (mem_cgroup_disabled())
5951 * Swap faults will attempt to charge the same page multiple
5952 * times. But reuse_swap_page() might have removed the page
5953 * from swapcache already, so we can't check PageSwapCache().
5958 commit_charge(page, memcg, lrucare);
5960 local_irq_disable();
5961 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5962 memcg_check_events(memcg, page);
5965 if (do_memsw_account() && PageSwapCache(page)) {
5966 swp_entry_t entry = { .val = page_private(page) };
5968 * The swap entry might not get freed for a long time,
5969 * let's not wait for it. The page already received a
5970 * memory+swap charge, drop the swap entry duplicate.
5972 mem_cgroup_uncharge_swap(entry, nr_pages);
5977 * mem_cgroup_cancel_charge - cancel a page charge
5978 * @page: page to charge
5979 * @memcg: memcg to charge the page to
5980 * @compound: charge the page as compound or small page
5982 * Cancel a charge transaction started by mem_cgroup_try_charge().
5984 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5987 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5989 if (mem_cgroup_disabled())
5992 * Swap faults will attempt to charge the same page multiple
5993 * times. But reuse_swap_page() might have removed the page
5994 * from swapcache already, so we can't check PageSwapCache().
5999 cancel_charge(memcg, nr_pages);
6002 struct uncharge_gather {
6003 struct mem_cgroup *memcg;
6004 unsigned long pgpgout;
6005 unsigned long nr_anon;
6006 unsigned long nr_file;
6007 unsigned long nr_kmem;
6008 unsigned long nr_huge;
6009 unsigned long nr_shmem;
6010 struct page *dummy_page;
6013 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6015 memset(ug, 0, sizeof(*ug));
6018 static void uncharge_batch(const struct uncharge_gather *ug)
6020 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6021 unsigned long flags;
6023 if (!mem_cgroup_is_root(ug->memcg)) {
6024 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6025 if (do_memsw_account())
6026 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6027 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6028 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6029 memcg_oom_recover(ug->memcg);
6032 local_irq_save(flags);
6033 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6034 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6035 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6036 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6037 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6038 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
6039 memcg_check_events(ug->memcg, ug->dummy_page);
6040 local_irq_restore(flags);
6042 if (!mem_cgroup_is_root(ug->memcg))
6043 css_put_many(&ug->memcg->css, nr_pages);
6046 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6048 VM_BUG_ON_PAGE(PageLRU(page), page);
6049 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6050 !PageHWPoison(page) , page);
6052 if (!page->mem_cgroup)
6056 * Nobody should be changing or seriously looking at
6057 * page->mem_cgroup at this point, we have fully
6058 * exclusive access to the page.
6061 if (ug->memcg != page->mem_cgroup) {
6064 uncharge_gather_clear(ug);
6066 ug->memcg = page->mem_cgroup;
6069 if (!PageKmemcg(page)) {
6070 unsigned int nr_pages = 1;
6072 if (PageTransHuge(page)) {
6073 nr_pages <<= compound_order(page);
6074 ug->nr_huge += nr_pages;
6077 ug->nr_anon += nr_pages;
6079 ug->nr_file += nr_pages;
6080 if (PageSwapBacked(page))
6081 ug->nr_shmem += nr_pages;
6085 ug->nr_kmem += 1 << compound_order(page);
6086 __ClearPageKmemcg(page);
6089 ug->dummy_page = page;
6090 page->mem_cgroup = NULL;
6093 static void uncharge_list(struct list_head *page_list)
6095 struct uncharge_gather ug;
6096 struct list_head *next;
6098 uncharge_gather_clear(&ug);
6101 * Note that the list can be a single page->lru; hence the
6102 * do-while loop instead of a simple list_for_each_entry().
6104 next = page_list->next;
6108 page = list_entry(next, struct page, lru);
6109 next = page->lru.next;
6111 uncharge_page(page, &ug);
6112 } while (next != page_list);
6115 uncharge_batch(&ug);
6119 * mem_cgroup_uncharge - uncharge a page
6120 * @page: page to uncharge
6122 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6123 * mem_cgroup_commit_charge().
6125 void mem_cgroup_uncharge(struct page *page)
6127 struct uncharge_gather ug;
6129 if (mem_cgroup_disabled())
6132 /* Don't touch page->lru of any random page, pre-check: */
6133 if (!page->mem_cgroup)
6136 uncharge_gather_clear(&ug);
6137 uncharge_page(page, &ug);
6138 uncharge_batch(&ug);
6142 * mem_cgroup_uncharge_list - uncharge a list of page
6143 * @page_list: list of pages to uncharge
6145 * Uncharge a list of pages previously charged with
6146 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6148 void mem_cgroup_uncharge_list(struct list_head *page_list)
6150 if (mem_cgroup_disabled())
6153 if (!list_empty(page_list))
6154 uncharge_list(page_list);
6158 * mem_cgroup_migrate - charge a page's replacement
6159 * @oldpage: currently circulating page
6160 * @newpage: replacement page
6162 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6163 * be uncharged upon free.
6165 * Both pages must be locked, @newpage->mapping must be set up.
6167 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6169 struct mem_cgroup *memcg;
6170 unsigned int nr_pages;
6172 unsigned long flags;
6174 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6175 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6176 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6177 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6180 if (mem_cgroup_disabled())
6183 /* Page cache replacement: new page already charged? */
6184 if (newpage->mem_cgroup)
6187 /* Swapcache readahead pages can get replaced before being charged */
6188 memcg = oldpage->mem_cgroup;
6192 /* Force-charge the new page. The old one will be freed soon */
6193 compound = PageTransHuge(newpage);
6194 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6196 page_counter_charge(&memcg->memory, nr_pages);
6197 if (do_memsw_account())
6198 page_counter_charge(&memcg->memsw, nr_pages);
6199 css_get_many(&memcg->css, nr_pages);
6201 commit_charge(newpage, memcg, false);
6203 local_irq_save(flags);
6204 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6205 memcg_check_events(memcg, newpage);
6206 local_irq_restore(flags);
6209 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6210 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6212 void mem_cgroup_sk_alloc(struct sock *sk)
6214 struct mem_cgroup *memcg;
6216 if (!mem_cgroup_sockets_enabled)
6220 * Socket cloning can throw us here with sk_memcg already
6221 * filled. It won't however, necessarily happen from
6222 * process context. So the test for root memcg given
6223 * the current task's memcg won't help us in this case.
6225 * Respecting the original socket's memcg is a better
6226 * decision in this case.
6229 css_get(&sk->sk_memcg->css);
6234 memcg = mem_cgroup_from_task(current);
6235 if (memcg == root_mem_cgroup)
6237 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6239 if (css_tryget_online(&memcg->css))
6240 sk->sk_memcg = memcg;
6245 void mem_cgroup_sk_free(struct sock *sk)
6248 css_put(&sk->sk_memcg->css);
6252 * mem_cgroup_charge_skmem - charge socket memory
6253 * @memcg: memcg to charge
6254 * @nr_pages: number of pages to charge
6256 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6257 * @memcg's configured limit, %false if the charge had to be forced.
6259 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6261 gfp_t gfp_mask = GFP_KERNEL;
6263 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6264 struct page_counter *fail;
6266 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6267 memcg->tcpmem_pressure = 0;
6270 page_counter_charge(&memcg->tcpmem, nr_pages);
6271 memcg->tcpmem_pressure = 1;
6275 /* Don't block in the packet receive path */
6277 gfp_mask = GFP_NOWAIT;
6279 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6281 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6284 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6289 * mem_cgroup_uncharge_skmem - uncharge socket memory
6290 * @memcg: memcg to uncharge
6291 * @nr_pages: number of pages to uncharge
6293 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6295 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6296 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6300 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6302 refill_stock(memcg, nr_pages);
6305 static int __init cgroup_memory(char *s)
6309 while ((token = strsep(&s, ",")) != NULL) {
6312 if (!strcmp(token, "nosocket"))
6313 cgroup_memory_nosocket = true;
6314 if (!strcmp(token, "nokmem"))
6315 cgroup_memory_nokmem = true;
6319 __setup("cgroup.memory=", cgroup_memory);
6322 * subsys_initcall() for memory controller.
6324 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6325 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6326 * basically everything that doesn't depend on a specific mem_cgroup structure
6327 * should be initialized from here.
6329 static int __init mem_cgroup_init(void)
6333 #ifdef CONFIG_MEMCG_KMEM
6335 * Kmem cache creation is mostly done with the slab_mutex held,
6336 * so use a workqueue with limited concurrency to avoid stalling
6337 * all worker threads in case lots of cgroups are created and
6338 * destroyed simultaneously.
6340 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6341 BUG_ON(!memcg_kmem_cache_wq);
6344 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6345 memcg_hotplug_cpu_dead);
6347 for_each_possible_cpu(cpu)
6348 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6351 for_each_node(node) {
6352 struct mem_cgroup_tree_per_node *rtpn;
6354 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6355 node_online(node) ? node : NUMA_NO_NODE);
6357 rtpn->rb_root = RB_ROOT;
6358 rtpn->rb_rightmost = NULL;
6359 spin_lock_init(&rtpn->lock);
6360 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6365 subsys_initcall(mem_cgroup_init);
6367 #ifdef CONFIG_MEMCG_SWAP
6368 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6370 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6372 * The root cgroup cannot be destroyed, so it's refcount must
6375 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6379 memcg = parent_mem_cgroup(memcg);
6381 memcg = root_mem_cgroup;
6387 * mem_cgroup_swapout - transfer a memsw charge to swap
6388 * @page: page whose memsw charge to transfer
6389 * @entry: swap entry to move the charge to
6391 * Transfer the memsw charge of @page to @entry.
6393 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6395 struct mem_cgroup *memcg, *swap_memcg;
6396 unsigned int nr_entries;
6397 unsigned short oldid;
6399 VM_BUG_ON_PAGE(PageLRU(page), page);
6400 VM_BUG_ON_PAGE(page_count(page), page);
6402 if (!do_memsw_account())
6405 memcg = page->mem_cgroup;
6407 /* Readahead page, never charged */
6412 * In case the memcg owning these pages has been offlined and doesn't
6413 * have an ID allocated to it anymore, charge the closest online
6414 * ancestor for the swap instead and transfer the memory+swap charge.
6416 swap_memcg = mem_cgroup_id_get_online(memcg);
6417 nr_entries = hpage_nr_pages(page);
6418 /* Get references for the tail pages, too */
6420 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6421 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6423 VM_BUG_ON_PAGE(oldid, page);
6424 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6426 page->mem_cgroup = NULL;
6428 if (!mem_cgroup_is_root(memcg))
6429 page_counter_uncharge(&memcg->memory, nr_entries);
6431 if (memcg != swap_memcg) {
6432 if (!mem_cgroup_is_root(swap_memcg))
6433 page_counter_charge(&swap_memcg->memsw, nr_entries);
6434 page_counter_uncharge(&memcg->memsw, nr_entries);
6438 * Interrupts should be disabled here because the caller holds the
6439 * i_pages lock which is taken with interrupts-off. It is
6440 * important here to have the interrupts disabled because it is the
6441 * only synchronisation we have for updating the per-CPU variables.
6443 VM_BUG_ON(!irqs_disabled());
6444 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6446 memcg_check_events(memcg, page);
6448 if (!mem_cgroup_is_root(memcg))
6449 css_put_many(&memcg->css, nr_entries);
6453 * mem_cgroup_try_charge_swap - try charging swap space for a page
6454 * @page: page being added to swap
6455 * @entry: swap entry to charge
6457 * Try to charge @page's memcg for the swap space at @entry.
6459 * Returns 0 on success, -ENOMEM on failure.
6461 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6463 unsigned int nr_pages = hpage_nr_pages(page);
6464 struct page_counter *counter;
6465 struct mem_cgroup *memcg;
6466 unsigned short oldid;
6468 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6471 memcg = page->mem_cgroup;
6473 /* Readahead page, never charged */
6478 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6482 memcg = mem_cgroup_id_get_online(memcg);
6484 if (!mem_cgroup_is_root(memcg) &&
6485 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6486 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6487 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6488 mem_cgroup_id_put(memcg);
6492 /* Get references for the tail pages, too */
6494 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6495 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6496 VM_BUG_ON_PAGE(oldid, page);
6497 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6503 * mem_cgroup_uncharge_swap - uncharge swap space
6504 * @entry: swap entry to uncharge
6505 * @nr_pages: the amount of swap space to uncharge
6507 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6509 struct mem_cgroup *memcg;
6512 if (!do_swap_account)
6515 id = swap_cgroup_record(entry, 0, nr_pages);
6517 memcg = mem_cgroup_from_id(id);
6519 if (!mem_cgroup_is_root(memcg)) {
6520 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6521 page_counter_uncharge(&memcg->swap, nr_pages);
6523 page_counter_uncharge(&memcg->memsw, nr_pages);
6525 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6526 mem_cgroup_id_put_many(memcg, nr_pages);
6531 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6533 long nr_swap_pages = get_nr_swap_pages();
6535 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6536 return nr_swap_pages;
6537 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6538 nr_swap_pages = min_t(long, nr_swap_pages,
6539 READ_ONCE(memcg->swap.max) -
6540 page_counter_read(&memcg->swap));
6541 return nr_swap_pages;
6544 bool mem_cgroup_swap_full(struct page *page)
6546 struct mem_cgroup *memcg;
6548 VM_BUG_ON_PAGE(!PageLocked(page), page);
6552 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6555 memcg = page->mem_cgroup;
6559 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6560 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6566 /* for remember boot option*/
6567 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6568 static int really_do_swap_account __initdata = 1;
6570 static int really_do_swap_account __initdata;
6573 static int __init enable_swap_account(char *s)
6575 if (!strcmp(s, "1"))
6576 really_do_swap_account = 1;
6577 else if (!strcmp(s, "0"))
6578 really_do_swap_account = 0;
6581 __setup("swapaccount=", enable_swap_account);
6583 static u64 swap_current_read(struct cgroup_subsys_state *css,
6586 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6588 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6591 static int swap_max_show(struct seq_file *m, void *v)
6593 return seq_puts_memcg_tunable(m,
6594 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6597 static ssize_t swap_max_write(struct kernfs_open_file *of,
6598 char *buf, size_t nbytes, loff_t off)
6600 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6604 buf = strstrip(buf);
6605 err = page_counter_memparse(buf, "max", &max);
6609 xchg(&memcg->swap.max, max);
6614 static int swap_events_show(struct seq_file *m, void *v)
6616 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6618 seq_printf(m, "max %lu\n",
6619 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6620 seq_printf(m, "fail %lu\n",
6621 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6626 static struct cftype swap_files[] = {
6628 .name = "swap.current",
6629 .flags = CFTYPE_NOT_ON_ROOT,
6630 .read_u64 = swap_current_read,
6634 .flags = CFTYPE_NOT_ON_ROOT,
6635 .seq_show = swap_max_show,
6636 .write = swap_max_write,
6639 .name = "swap.events",
6640 .flags = CFTYPE_NOT_ON_ROOT,
6641 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6642 .seq_show = swap_events_show,
6647 static struct cftype memsw_cgroup_files[] = {
6649 .name = "memsw.usage_in_bytes",
6650 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6651 .read_u64 = mem_cgroup_read_u64,
6654 .name = "memsw.max_usage_in_bytes",
6655 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6656 .write = mem_cgroup_reset,
6657 .read_u64 = mem_cgroup_read_u64,
6660 .name = "memsw.limit_in_bytes",
6661 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6662 .write = mem_cgroup_write,
6663 .read_u64 = mem_cgroup_read_u64,
6666 .name = "memsw.failcnt",
6667 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6668 .write = mem_cgroup_reset,
6669 .read_u64 = mem_cgroup_read_u64,
6671 { }, /* terminate */
6674 static int __init mem_cgroup_swap_init(void)
6676 if (!mem_cgroup_disabled() && really_do_swap_account) {
6677 do_swap_account = 1;
6678 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6680 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6681 memsw_cgroup_files));
6685 subsys_initcall(mem_cgroup_swap_init);
6687 #endif /* CONFIG_MEMCG_SWAP */