1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/seq_buf.h>
66 #include <linux/uaccess.h>
68 #include <trace/events/vmscan.h>
70 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
71 EXPORT_SYMBOL(memory_cgrp_subsys);
73 struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #define MEM_CGROUP_RECLAIM_RETRIES 5
77 /* Socket memory accounting disabled? */
78 static bool cgroup_memory_nosocket;
80 /* Kernel memory accounting disabled? */
81 static bool cgroup_memory_nokmem;
83 /* Whether the swap controller is active */
84 #ifdef CONFIG_MEMCG_SWAP
85 int do_swap_account __read_mostly;
87 #define do_swap_account 0
90 /* Whether legacy memory+swap accounting is active */
91 static bool do_memsw_account(void)
93 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
96 static const char *const mem_cgroup_lru_names[] = {
104 #define THRESHOLDS_EVENTS_TARGET 128
105 #define SOFTLIMIT_EVENTS_TARGET 1024
106 #define NUMAINFO_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node {
114 struct rb_root rb_root;
115 struct rb_node *rb_rightmost;
119 struct mem_cgroup_tree {
120 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
126 struct mem_cgroup_eventfd_list {
127 struct list_head list;
128 struct eventfd_ctx *eventfd;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event {
136 * memcg which the event belongs to.
138 struct mem_cgroup *memcg;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx *eventfd;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event)(struct mem_cgroup *memcg,
153 struct eventfd_ctx *eventfd, const char *args);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t *wqh;
167 wait_queue_entry_t wait;
168 struct work_struct remove;
171 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
172 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct {
184 spinlock_t lock; /* for from, to */
185 struct mm_struct *mm;
186 struct mem_cgroup *from;
187 struct mem_cgroup *to;
189 unsigned long precharge;
190 unsigned long moved_charge;
191 unsigned long moved_swap;
192 struct task_struct *moving_task; /* a task moving charges */
193 wait_queue_head_t waitq; /* a waitq for other context */
195 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
196 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
200 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
207 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
208 MEM_CGROUP_CHARGE_TYPE_ANON,
209 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
210 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
214 /* for encoding cft->private value on file */
223 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
224 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
225 #define MEMFILE_ATTR(val) ((val) & 0xffff)
226 /* Used for OOM nofiier */
227 #define OOM_CONTROL (0)
230 * Iteration constructs for visiting all cgroups (under a tree). If
231 * loops are exited prematurely (break), mem_cgroup_iter_break() must
232 * be used for reference counting.
234 #define for_each_mem_cgroup_tree(iter, root) \
235 for (iter = mem_cgroup_iter(root, NULL, NULL); \
237 iter = mem_cgroup_iter(root, iter, NULL))
239 #define for_each_mem_cgroup(iter) \
240 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
242 iter = mem_cgroup_iter(NULL, iter, NULL))
244 static inline bool should_force_charge(void)
246 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
247 (current->flags & PF_EXITING);
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
254 memcg = root_mem_cgroup;
255 return &memcg->vmpressure;
258 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
260 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
263 #ifdef CONFIG_MEMCG_KMEM
265 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
266 * The main reason for not using cgroup id for this:
267 * this works better in sparse environments, where we have a lot of memcgs,
268 * but only a few kmem-limited. Or also, if we have, for instance, 200
269 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
270 * 200 entry array for that.
272 * The current size of the caches array is stored in memcg_nr_cache_ids. It
273 * will double each time we have to increase it.
275 static DEFINE_IDA(memcg_cache_ida);
276 int memcg_nr_cache_ids;
278 /* Protects memcg_nr_cache_ids */
279 static DECLARE_RWSEM(memcg_cache_ids_sem);
281 void memcg_get_cache_ids(void)
283 down_read(&memcg_cache_ids_sem);
286 void memcg_put_cache_ids(void)
288 up_read(&memcg_cache_ids_sem);
292 * MIN_SIZE is different than 1, because we would like to avoid going through
293 * the alloc/free process all the time. In a small machine, 4 kmem-limited
294 * cgroups is a reasonable guess. In the future, it could be a parameter or
295 * tunable, but that is strictly not necessary.
297 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
298 * this constant directly from cgroup, but it is understandable that this is
299 * better kept as an internal representation in cgroup.c. In any case, the
300 * cgrp_id space is not getting any smaller, and we don't have to necessarily
301 * increase ours as well if it increases.
303 #define MEMCG_CACHES_MIN_SIZE 4
304 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
307 * A lot of the calls to the cache allocation functions are expected to be
308 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
309 * conditional to this static branch, we'll have to allow modules that does
310 * kmem_cache_alloc and the such to see this symbol as well
312 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
313 EXPORT_SYMBOL(memcg_kmem_enabled_key);
315 struct workqueue_struct *memcg_kmem_cache_wq;
317 static int memcg_shrinker_map_size;
318 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
320 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
322 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
325 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
326 int size, int old_size)
328 struct memcg_shrinker_map *new, *old;
331 lockdep_assert_held(&memcg_shrinker_map_mutex);
334 old = rcu_dereference_protected(
335 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
336 /* Not yet online memcg */
340 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
344 /* Set all old bits, clear all new bits */
345 memset(new->map, (int)0xff, old_size);
346 memset((void *)new->map + old_size, 0, size - old_size);
348 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
349 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
355 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
357 struct mem_cgroup_per_node *pn;
358 struct memcg_shrinker_map *map;
361 if (mem_cgroup_is_root(memcg))
365 pn = mem_cgroup_nodeinfo(memcg, nid);
366 map = rcu_dereference_protected(pn->shrinker_map, true);
369 rcu_assign_pointer(pn->shrinker_map, NULL);
373 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
375 struct memcg_shrinker_map *map;
376 int nid, size, ret = 0;
378 if (mem_cgroup_is_root(memcg))
381 mutex_lock(&memcg_shrinker_map_mutex);
382 size = memcg_shrinker_map_size;
384 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
386 memcg_free_shrinker_maps(memcg);
390 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
392 mutex_unlock(&memcg_shrinker_map_mutex);
397 int memcg_expand_shrinker_maps(int new_id)
399 int size, old_size, ret = 0;
400 struct mem_cgroup *memcg;
402 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
403 old_size = memcg_shrinker_map_size;
404 if (size <= old_size)
407 mutex_lock(&memcg_shrinker_map_mutex);
408 if (!root_mem_cgroup)
411 for_each_mem_cgroup(memcg) {
412 if (mem_cgroup_is_root(memcg))
414 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
420 memcg_shrinker_map_size = size;
421 mutex_unlock(&memcg_shrinker_map_mutex);
425 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
427 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
428 struct memcg_shrinker_map *map;
431 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
432 /* Pairs with smp mb in shrink_slab() */
433 smp_mb__before_atomic();
434 set_bit(shrinker_id, map->map);
439 #else /* CONFIG_MEMCG_KMEM */
440 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
444 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
445 #endif /* CONFIG_MEMCG_KMEM */
448 * mem_cgroup_css_from_page - css of the memcg associated with a page
449 * @page: page of interest
451 * If memcg is bound to the default hierarchy, css of the memcg associated
452 * with @page is returned. The returned css remains associated with @page
453 * until it is released.
455 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
458 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
460 struct mem_cgroup *memcg;
462 memcg = page->mem_cgroup;
464 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
465 memcg = root_mem_cgroup;
471 * page_cgroup_ino - return inode number of the memcg a page is charged to
474 * Look up the closest online ancestor of the memory cgroup @page is charged to
475 * and return its inode number or 0 if @page is not charged to any cgroup. It
476 * is safe to call this function without holding a reference to @page.
478 * Note, this function is inherently racy, because there is nothing to prevent
479 * the cgroup inode from getting torn down and potentially reallocated a moment
480 * after page_cgroup_ino() returns, so it only should be used by callers that
481 * do not care (such as procfs interfaces).
483 ino_t page_cgroup_ino(struct page *page)
485 struct mem_cgroup *memcg;
486 unsigned long ino = 0;
489 if (PageHead(page) && PageSlab(page))
490 memcg = memcg_from_slab_page(page);
492 memcg = READ_ONCE(page->mem_cgroup);
493 while (memcg && !(memcg->css.flags & CSS_ONLINE))
494 memcg = parent_mem_cgroup(memcg);
496 ino = cgroup_ino(memcg->css.cgroup);
501 static struct mem_cgroup_per_node *
502 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
504 int nid = page_to_nid(page);
506 return memcg->nodeinfo[nid];
509 static struct mem_cgroup_tree_per_node *
510 soft_limit_tree_node(int nid)
512 return soft_limit_tree.rb_tree_per_node[nid];
515 static struct mem_cgroup_tree_per_node *
516 soft_limit_tree_from_page(struct page *page)
518 int nid = page_to_nid(page);
520 return soft_limit_tree.rb_tree_per_node[nid];
523 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
524 struct mem_cgroup_tree_per_node *mctz,
525 unsigned long new_usage_in_excess)
527 struct rb_node **p = &mctz->rb_root.rb_node;
528 struct rb_node *parent = NULL;
529 struct mem_cgroup_per_node *mz_node;
530 bool rightmost = true;
535 mz->usage_in_excess = new_usage_in_excess;
536 if (!mz->usage_in_excess)
540 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
542 if (mz->usage_in_excess < mz_node->usage_in_excess) {
548 * We can't avoid mem cgroups that are over their soft
549 * limit by the same amount
551 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
556 mctz->rb_rightmost = &mz->tree_node;
558 rb_link_node(&mz->tree_node, parent, p);
559 rb_insert_color(&mz->tree_node, &mctz->rb_root);
563 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
564 struct mem_cgroup_tree_per_node *mctz)
569 if (&mz->tree_node == mctz->rb_rightmost)
570 mctz->rb_rightmost = rb_prev(&mz->tree_node);
572 rb_erase(&mz->tree_node, &mctz->rb_root);
576 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
577 struct mem_cgroup_tree_per_node *mctz)
581 spin_lock_irqsave(&mctz->lock, flags);
582 __mem_cgroup_remove_exceeded(mz, mctz);
583 spin_unlock_irqrestore(&mctz->lock, flags);
586 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
588 unsigned long nr_pages = page_counter_read(&memcg->memory);
589 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
590 unsigned long excess = 0;
592 if (nr_pages > soft_limit)
593 excess = nr_pages - soft_limit;
598 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
600 unsigned long excess;
601 struct mem_cgroup_per_node *mz;
602 struct mem_cgroup_tree_per_node *mctz;
604 mctz = soft_limit_tree_from_page(page);
608 * Necessary to update all ancestors when hierarchy is used.
609 * because their event counter is not touched.
611 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
612 mz = mem_cgroup_page_nodeinfo(memcg, page);
613 excess = soft_limit_excess(memcg);
615 * We have to update the tree if mz is on RB-tree or
616 * mem is over its softlimit.
618 if (excess || mz->on_tree) {
621 spin_lock_irqsave(&mctz->lock, flags);
622 /* if on-tree, remove it */
624 __mem_cgroup_remove_exceeded(mz, mctz);
626 * Insert again. mz->usage_in_excess will be updated.
627 * If excess is 0, no tree ops.
629 __mem_cgroup_insert_exceeded(mz, mctz, excess);
630 spin_unlock_irqrestore(&mctz->lock, flags);
635 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
637 struct mem_cgroup_tree_per_node *mctz;
638 struct mem_cgroup_per_node *mz;
642 mz = mem_cgroup_nodeinfo(memcg, nid);
643 mctz = soft_limit_tree_node(nid);
645 mem_cgroup_remove_exceeded(mz, mctz);
649 static struct mem_cgroup_per_node *
650 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
652 struct mem_cgroup_per_node *mz;
656 if (!mctz->rb_rightmost)
657 goto done; /* Nothing to reclaim from */
659 mz = rb_entry(mctz->rb_rightmost,
660 struct mem_cgroup_per_node, tree_node);
662 * Remove the node now but someone else can add it back,
663 * we will to add it back at the end of reclaim to its correct
664 * position in the tree.
666 __mem_cgroup_remove_exceeded(mz, mctz);
667 if (!soft_limit_excess(mz->memcg) ||
668 !css_tryget_online(&mz->memcg->css))
674 static struct mem_cgroup_per_node *
675 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
677 struct mem_cgroup_per_node *mz;
679 spin_lock_irq(&mctz->lock);
680 mz = __mem_cgroup_largest_soft_limit_node(mctz);
681 spin_unlock_irq(&mctz->lock);
686 * __mod_memcg_state - update cgroup memory statistics
687 * @memcg: the memory cgroup
688 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
689 * @val: delta to add to the counter, can be negative
691 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
695 if (mem_cgroup_disabled())
698 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
699 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
700 struct mem_cgroup *mi;
703 * Batch local counters to keep them in sync with
704 * the hierarchical ones.
706 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
707 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
708 atomic_long_add(x, &mi->vmstats[idx]);
711 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
714 static struct mem_cgroup_per_node *
715 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
717 struct mem_cgroup *parent;
719 parent = parent_mem_cgroup(pn->memcg);
722 return mem_cgroup_nodeinfo(parent, nid);
726 * __mod_lruvec_state - update lruvec memory statistics
727 * @lruvec: the lruvec
728 * @idx: the stat item
729 * @val: delta to add to the counter, can be negative
731 * The lruvec is the intersection of the NUMA node and a cgroup. This
732 * function updates the all three counters that are affected by a
733 * change of state at this level: per-node, per-cgroup, per-lruvec.
735 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
738 pg_data_t *pgdat = lruvec_pgdat(lruvec);
739 struct mem_cgroup_per_node *pn;
740 struct mem_cgroup *memcg;
744 __mod_node_page_state(pgdat, idx, val);
746 if (mem_cgroup_disabled())
749 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
753 __mod_memcg_state(memcg, idx, val);
755 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
756 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
757 struct mem_cgroup_per_node *pi;
760 * Batch local counters to keep them in sync with
761 * the hierarchical ones.
763 __this_cpu_add(pn->lruvec_stat_local->count[idx], x);
764 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
765 atomic_long_add(x, &pi->lruvec_stat[idx]);
768 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
771 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
773 struct page *page = virt_to_head_page(p);
774 pg_data_t *pgdat = page_pgdat(page);
775 struct mem_cgroup *memcg;
776 struct lruvec *lruvec;
779 memcg = memcg_from_slab_page(page);
781 /* Untracked pages have no memcg, no lruvec. Update only the node */
782 if (!memcg || memcg == root_mem_cgroup) {
783 __mod_node_page_state(pgdat, idx, val);
785 lruvec = mem_cgroup_lruvec(pgdat, memcg);
786 __mod_lruvec_state(lruvec, idx, val);
792 * __count_memcg_events - account VM events in a cgroup
793 * @memcg: the memory cgroup
794 * @idx: the event item
795 * @count: the number of events that occured
797 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
802 if (mem_cgroup_disabled())
805 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
806 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
807 struct mem_cgroup *mi;
810 * Batch local counters to keep them in sync with
811 * the hierarchical ones.
813 __this_cpu_add(memcg->vmstats_local->events[idx], x);
814 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
815 atomic_long_add(x, &mi->vmevents[idx]);
818 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
821 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
823 return atomic_long_read(&memcg->vmevents[event]);
826 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
831 for_each_possible_cpu(cpu)
832 x += per_cpu(memcg->vmstats_local->events[event], cpu);
836 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
838 bool compound, int nr_pages)
841 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
842 * counted as CACHE even if it's on ANON LRU.
845 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
847 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
848 if (PageSwapBacked(page))
849 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
853 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
854 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
857 /* pagein of a big page is an event. So, ignore page size */
859 __count_memcg_events(memcg, PGPGIN, 1);
861 __count_memcg_events(memcg, PGPGOUT, 1);
862 nr_pages = -nr_pages; /* for event */
865 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
868 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
869 enum mem_cgroup_events_target target)
871 unsigned long val, next;
873 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
874 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
875 /* from time_after() in jiffies.h */
876 if ((long)(next - val) < 0) {
878 case MEM_CGROUP_TARGET_THRESH:
879 next = val + THRESHOLDS_EVENTS_TARGET;
881 case MEM_CGROUP_TARGET_SOFTLIMIT:
882 next = val + SOFTLIMIT_EVENTS_TARGET;
884 case MEM_CGROUP_TARGET_NUMAINFO:
885 next = val + NUMAINFO_EVENTS_TARGET;
890 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
897 * Check events in order.
900 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
902 /* threshold event is triggered in finer grain than soft limit */
903 if (unlikely(mem_cgroup_event_ratelimit(memcg,
904 MEM_CGROUP_TARGET_THRESH))) {
906 bool do_numainfo __maybe_unused;
908 do_softlimit = mem_cgroup_event_ratelimit(memcg,
909 MEM_CGROUP_TARGET_SOFTLIMIT);
911 do_numainfo = mem_cgroup_event_ratelimit(memcg,
912 MEM_CGROUP_TARGET_NUMAINFO);
914 mem_cgroup_threshold(memcg);
915 if (unlikely(do_softlimit))
916 mem_cgroup_update_tree(memcg, page);
918 if (unlikely(do_numainfo))
919 atomic_inc(&memcg->numainfo_events);
924 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
927 * mm_update_next_owner() may clear mm->owner to NULL
928 * if it races with swapoff, page migration, etc.
929 * So this can be called with p == NULL.
934 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
936 EXPORT_SYMBOL(mem_cgroup_from_task);
939 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
940 * @mm: mm from which memcg should be extracted. It can be NULL.
942 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
943 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
946 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
948 struct mem_cgroup *memcg;
950 if (mem_cgroup_disabled())
956 * Page cache insertions can happen withou an
957 * actual mm context, e.g. during disk probing
958 * on boot, loopback IO, acct() writes etc.
961 memcg = root_mem_cgroup;
963 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
964 if (unlikely(!memcg))
965 memcg = root_mem_cgroup;
967 } while (!css_tryget_online(&memcg->css));
971 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
974 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
975 * @page: page from which memcg should be extracted.
977 * Obtain a reference on page->memcg and returns it if successful. Otherwise
978 * root_mem_cgroup is returned.
980 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
982 struct mem_cgroup *memcg = page->mem_cgroup;
984 if (mem_cgroup_disabled())
988 if (!memcg || !css_tryget_online(&memcg->css))
989 memcg = root_mem_cgroup;
993 EXPORT_SYMBOL(get_mem_cgroup_from_page);
996 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
998 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1000 if (unlikely(current->active_memcg)) {
1001 struct mem_cgroup *memcg = root_mem_cgroup;
1004 if (css_tryget_online(¤t->active_memcg->css))
1005 memcg = current->active_memcg;
1009 return get_mem_cgroup_from_mm(current->mm);
1013 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1014 * @root: hierarchy root
1015 * @prev: previously returned memcg, NULL on first invocation
1016 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1018 * Returns references to children of the hierarchy below @root, or
1019 * @root itself, or %NULL after a full round-trip.
1021 * Caller must pass the return value in @prev on subsequent
1022 * invocations for reference counting, or use mem_cgroup_iter_break()
1023 * to cancel a hierarchy walk before the round-trip is complete.
1025 * Reclaimers can specify a node and a priority level in @reclaim to
1026 * divide up the memcgs in the hierarchy among all concurrent
1027 * reclaimers operating on the same node and priority.
1029 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1030 struct mem_cgroup *prev,
1031 struct mem_cgroup_reclaim_cookie *reclaim)
1033 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1034 struct cgroup_subsys_state *css = NULL;
1035 struct mem_cgroup *memcg = NULL;
1036 struct mem_cgroup *pos = NULL;
1038 if (mem_cgroup_disabled())
1042 root = root_mem_cgroup;
1044 if (prev && !reclaim)
1047 if (!root->use_hierarchy && root != root_mem_cgroup) {
1056 struct mem_cgroup_per_node *mz;
1058 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1059 iter = &mz->iter[reclaim->priority];
1061 if (prev && reclaim->generation != iter->generation)
1065 pos = READ_ONCE(iter->position);
1066 if (!pos || css_tryget(&pos->css))
1069 * css reference reached zero, so iter->position will
1070 * be cleared by ->css_released. However, we should not
1071 * rely on this happening soon, because ->css_released
1072 * is called from a work queue, and by busy-waiting we
1073 * might block it. So we clear iter->position right
1076 (void)cmpxchg(&iter->position, pos, NULL);
1084 css = css_next_descendant_pre(css, &root->css);
1087 * Reclaimers share the hierarchy walk, and a
1088 * new one might jump in right at the end of
1089 * the hierarchy - make sure they see at least
1090 * one group and restart from the beginning.
1098 * Verify the css and acquire a reference. The root
1099 * is provided by the caller, so we know it's alive
1100 * and kicking, and don't take an extra reference.
1102 memcg = mem_cgroup_from_css(css);
1104 if (css == &root->css)
1107 if (css_tryget(css))
1115 * The position could have already been updated by a competing
1116 * thread, so check that the value hasn't changed since we read
1117 * it to avoid reclaiming from the same cgroup twice.
1119 (void)cmpxchg(&iter->position, pos, memcg);
1127 reclaim->generation = iter->generation;
1133 if (prev && prev != root)
1134 css_put(&prev->css);
1140 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1141 * @root: hierarchy root
1142 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1144 void mem_cgroup_iter_break(struct mem_cgroup *root,
1145 struct mem_cgroup *prev)
1148 root = root_mem_cgroup;
1149 if (prev && prev != root)
1150 css_put(&prev->css);
1153 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1154 struct mem_cgroup *dead_memcg)
1156 struct mem_cgroup_reclaim_iter *iter;
1157 struct mem_cgroup_per_node *mz;
1161 for_each_node(nid) {
1162 mz = mem_cgroup_nodeinfo(from, nid);
1163 for (i = 0; i <= DEF_PRIORITY; i++) {
1164 iter = &mz->iter[i];
1165 cmpxchg(&iter->position,
1171 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1173 struct mem_cgroup *memcg = dead_memcg;
1174 struct mem_cgroup *last;
1177 __invalidate_reclaim_iterators(memcg, dead_memcg);
1179 } while ((memcg = parent_mem_cgroup(memcg)));
1182 * When cgruop1 non-hierarchy mode is used,
1183 * parent_mem_cgroup() does not walk all the way up to the
1184 * cgroup root (root_mem_cgroup). So we have to handle
1185 * dead_memcg from cgroup root separately.
1187 if (last != root_mem_cgroup)
1188 __invalidate_reclaim_iterators(root_mem_cgroup,
1193 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1194 * @memcg: hierarchy root
1195 * @fn: function to call for each task
1196 * @arg: argument passed to @fn
1198 * This function iterates over tasks attached to @memcg or to any of its
1199 * descendants and calls @fn for each task. If @fn returns a non-zero
1200 * value, the function breaks the iteration loop and returns the value.
1201 * Otherwise, it will iterate over all tasks and return 0.
1203 * This function must not be called for the root memory cgroup.
1205 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1206 int (*fn)(struct task_struct *, void *), void *arg)
1208 struct mem_cgroup *iter;
1211 BUG_ON(memcg == root_mem_cgroup);
1213 for_each_mem_cgroup_tree(iter, memcg) {
1214 struct css_task_iter it;
1215 struct task_struct *task;
1217 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1218 while (!ret && (task = css_task_iter_next(&it)))
1219 ret = fn(task, arg);
1220 css_task_iter_end(&it);
1222 mem_cgroup_iter_break(memcg, iter);
1230 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1232 * @pgdat: pgdat of the page
1234 * This function is only safe when following the LRU page isolation
1235 * and putback protocol: the LRU lock must be held, and the page must
1236 * either be PageLRU() or the caller must have isolated/allocated it.
1238 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1240 struct mem_cgroup_per_node *mz;
1241 struct mem_cgroup *memcg;
1242 struct lruvec *lruvec;
1244 if (mem_cgroup_disabled()) {
1245 lruvec = &pgdat->lruvec;
1249 memcg = page->mem_cgroup;
1251 * Swapcache readahead pages are added to the LRU - and
1252 * possibly migrated - before they are charged.
1255 memcg = root_mem_cgroup;
1257 mz = mem_cgroup_page_nodeinfo(memcg, page);
1258 lruvec = &mz->lruvec;
1261 * Since a node can be onlined after the mem_cgroup was created,
1262 * we have to be prepared to initialize lruvec->zone here;
1263 * and if offlined then reonlined, we need to reinitialize it.
1265 if (unlikely(lruvec->pgdat != pgdat))
1266 lruvec->pgdat = pgdat;
1271 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1272 * @lruvec: mem_cgroup per zone lru vector
1273 * @lru: index of lru list the page is sitting on
1274 * @zid: zone id of the accounted pages
1275 * @nr_pages: positive when adding or negative when removing
1277 * This function must be called under lru_lock, just before a page is added
1278 * to or just after a page is removed from an lru list (that ordering being
1279 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1281 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1282 int zid, int nr_pages)
1284 struct mem_cgroup_per_node *mz;
1285 unsigned long *lru_size;
1288 if (mem_cgroup_disabled())
1291 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1292 lru_size = &mz->lru_zone_size[zid][lru];
1295 *lru_size += nr_pages;
1298 if (WARN_ONCE(size < 0,
1299 "%s(%p, %d, %d): lru_size %ld\n",
1300 __func__, lruvec, lru, nr_pages, size)) {
1306 *lru_size += nr_pages;
1310 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1311 * @memcg: the memory cgroup
1313 * Returns the maximum amount of memory @mem can be charged with, in
1316 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1318 unsigned long margin = 0;
1319 unsigned long count;
1320 unsigned long limit;
1322 count = page_counter_read(&memcg->memory);
1323 limit = READ_ONCE(memcg->memory.max);
1325 margin = limit - count;
1327 if (do_memsw_account()) {
1328 count = page_counter_read(&memcg->memsw);
1329 limit = READ_ONCE(memcg->memsw.max);
1331 margin = min(margin, limit - count);
1340 * A routine for checking "mem" is under move_account() or not.
1342 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1343 * moving cgroups. This is for waiting at high-memory pressure
1346 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1348 struct mem_cgroup *from;
1349 struct mem_cgroup *to;
1352 * Unlike task_move routines, we access mc.to, mc.from not under
1353 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1355 spin_lock(&mc.lock);
1361 ret = mem_cgroup_is_descendant(from, memcg) ||
1362 mem_cgroup_is_descendant(to, memcg);
1364 spin_unlock(&mc.lock);
1368 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1370 if (mc.moving_task && current != mc.moving_task) {
1371 if (mem_cgroup_under_move(memcg)) {
1373 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1374 /* moving charge context might have finished. */
1377 finish_wait(&mc.waitq, &wait);
1384 static char *memory_stat_format(struct mem_cgroup *memcg)
1389 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1394 * Provide statistics on the state of the memory subsystem as
1395 * well as cumulative event counters that show past behavior.
1397 * This list is ordered following a combination of these gradients:
1398 * 1) generic big picture -> specifics and details
1399 * 2) reflecting userspace activity -> reflecting kernel heuristics
1401 * Current memory state:
1404 seq_buf_printf(&s, "anon %llu\n",
1405 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1407 seq_buf_printf(&s, "file %llu\n",
1408 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1410 seq_buf_printf(&s, "kernel_stack %llu\n",
1411 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1413 seq_buf_printf(&s, "slab %llu\n",
1414 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1415 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1417 seq_buf_printf(&s, "sock %llu\n",
1418 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1421 seq_buf_printf(&s, "shmem %llu\n",
1422 (u64)memcg_page_state(memcg, NR_SHMEM) *
1424 seq_buf_printf(&s, "file_mapped %llu\n",
1425 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1427 seq_buf_printf(&s, "file_dirty %llu\n",
1428 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1430 seq_buf_printf(&s, "file_writeback %llu\n",
1431 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1435 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1436 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1437 * arse because it requires migrating the work out of rmap to a place
1438 * where the page->mem_cgroup is set up and stable.
1440 seq_buf_printf(&s, "anon_thp %llu\n",
1441 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1444 for (i = 0; i < NR_LRU_LISTS; i++)
1445 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1446 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1449 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1450 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1452 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1453 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1456 /* Accumulated memory events */
1458 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1459 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1461 seq_buf_printf(&s, "workingset_refault %lu\n",
1462 memcg_page_state(memcg, WORKINGSET_REFAULT));
1463 seq_buf_printf(&s, "workingset_activate %lu\n",
1464 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1465 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1466 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1468 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1469 seq_buf_printf(&s, "pgscan %lu\n",
1470 memcg_events(memcg, PGSCAN_KSWAPD) +
1471 memcg_events(memcg, PGSCAN_DIRECT));
1472 seq_buf_printf(&s, "pgsteal %lu\n",
1473 memcg_events(memcg, PGSTEAL_KSWAPD) +
1474 memcg_events(memcg, PGSTEAL_DIRECT));
1475 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1476 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1477 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1478 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1480 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1481 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1482 memcg_events(memcg, THP_FAULT_ALLOC));
1483 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1484 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1485 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1487 /* The above should easily fit into one page */
1488 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1493 #define K(x) ((x) << (PAGE_SHIFT-10))
1495 * mem_cgroup_print_oom_context: Print OOM information relevant to
1496 * memory controller.
1497 * @memcg: The memory cgroup that went over limit
1498 * @p: Task that is going to be killed
1500 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1503 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1508 pr_cont(",oom_memcg=");
1509 pr_cont_cgroup_path(memcg->css.cgroup);
1511 pr_cont(",global_oom");
1513 pr_cont(",task_memcg=");
1514 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1520 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1521 * memory controller.
1522 * @memcg: The memory cgroup that went over limit
1524 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1528 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1529 K((u64)page_counter_read(&memcg->memory)),
1530 K((u64)memcg->memory.max), memcg->memory.failcnt);
1531 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1532 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1533 K((u64)page_counter_read(&memcg->swap)),
1534 K((u64)memcg->swap.max), memcg->swap.failcnt);
1536 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1537 K((u64)page_counter_read(&memcg->memsw)),
1538 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1539 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1540 K((u64)page_counter_read(&memcg->kmem)),
1541 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1544 pr_info("Memory cgroup stats for ");
1545 pr_cont_cgroup_path(memcg->css.cgroup);
1547 buf = memory_stat_format(memcg);
1555 * Return the memory (and swap, if configured) limit for a memcg.
1557 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1561 max = memcg->memory.max;
1562 if (mem_cgroup_swappiness(memcg)) {
1563 unsigned long memsw_max;
1564 unsigned long swap_max;
1566 memsw_max = memcg->memsw.max;
1567 swap_max = memcg->swap.max;
1568 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1569 max = min(max + swap_max, memsw_max);
1574 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1577 struct oom_control oc = {
1581 .gfp_mask = gfp_mask,
1586 if (mutex_lock_killable(&oom_lock))
1589 * A few threads which were not waiting at mutex_lock_killable() can
1590 * fail to bail out. Therefore, check again after holding oom_lock.
1592 ret = should_force_charge() || out_of_memory(&oc);
1593 mutex_unlock(&oom_lock);
1597 #if MAX_NUMNODES > 1
1600 * test_mem_cgroup_node_reclaimable
1601 * @memcg: the target memcg
1602 * @nid: the node ID to be checked.
1603 * @noswap : specify true here if the user wants flle only information.
1605 * This function returns whether the specified memcg contains any
1606 * reclaimable pages on a node. Returns true if there are any reclaimable
1607 * pages in the node.
1609 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1610 int nid, bool noswap)
1612 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1614 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1615 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1617 if (noswap || !total_swap_pages)
1619 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1620 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1627 * Always updating the nodemask is not very good - even if we have an empty
1628 * list or the wrong list here, we can start from some node and traverse all
1629 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1632 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1636 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1637 * pagein/pageout changes since the last update.
1639 if (!atomic_read(&memcg->numainfo_events))
1641 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1644 /* make a nodemask where this memcg uses memory from */
1645 memcg->scan_nodes = node_states[N_MEMORY];
1647 for_each_node_mask(nid, node_states[N_MEMORY]) {
1649 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1650 node_clear(nid, memcg->scan_nodes);
1653 atomic_set(&memcg->numainfo_events, 0);
1654 atomic_set(&memcg->numainfo_updating, 0);
1658 * Selecting a node where we start reclaim from. Because what we need is just
1659 * reducing usage counter, start from anywhere is O,K. Considering
1660 * memory reclaim from current node, there are pros. and cons.
1662 * Freeing memory from current node means freeing memory from a node which
1663 * we'll use or we've used. So, it may make LRU bad. And if several threads
1664 * hit limits, it will see a contention on a node. But freeing from remote
1665 * node means more costs for memory reclaim because of memory latency.
1667 * Now, we use round-robin. Better algorithm is welcomed.
1669 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1673 mem_cgroup_may_update_nodemask(memcg);
1674 node = memcg->last_scanned_node;
1676 node = next_node_in(node, memcg->scan_nodes);
1678 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1679 * last time it really checked all the LRUs due to rate limiting.
1680 * Fallback to the current node in that case for simplicity.
1682 if (unlikely(node == MAX_NUMNODES))
1683 node = numa_node_id();
1685 memcg->last_scanned_node = node;
1689 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1695 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1698 unsigned long *total_scanned)
1700 struct mem_cgroup *victim = NULL;
1703 unsigned long excess;
1704 unsigned long nr_scanned;
1705 struct mem_cgroup_reclaim_cookie reclaim = {
1710 excess = soft_limit_excess(root_memcg);
1713 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1718 * If we have not been able to reclaim
1719 * anything, it might because there are
1720 * no reclaimable pages under this hierarchy
1725 * We want to do more targeted reclaim.
1726 * excess >> 2 is not to excessive so as to
1727 * reclaim too much, nor too less that we keep
1728 * coming back to reclaim from this cgroup
1730 if (total >= (excess >> 2) ||
1731 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1736 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1737 pgdat, &nr_scanned);
1738 *total_scanned += nr_scanned;
1739 if (!soft_limit_excess(root_memcg))
1742 mem_cgroup_iter_break(root_memcg, victim);
1746 #ifdef CONFIG_LOCKDEP
1747 static struct lockdep_map memcg_oom_lock_dep_map = {
1748 .name = "memcg_oom_lock",
1752 static DEFINE_SPINLOCK(memcg_oom_lock);
1755 * Check OOM-Killer is already running under our hierarchy.
1756 * If someone is running, return false.
1758 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1760 struct mem_cgroup *iter, *failed = NULL;
1762 spin_lock(&memcg_oom_lock);
1764 for_each_mem_cgroup_tree(iter, memcg) {
1765 if (iter->oom_lock) {
1767 * this subtree of our hierarchy is already locked
1768 * so we cannot give a lock.
1771 mem_cgroup_iter_break(memcg, iter);
1774 iter->oom_lock = true;
1779 * OK, we failed to lock the whole subtree so we have
1780 * to clean up what we set up to the failing subtree
1782 for_each_mem_cgroup_tree(iter, memcg) {
1783 if (iter == failed) {
1784 mem_cgroup_iter_break(memcg, iter);
1787 iter->oom_lock = false;
1790 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1792 spin_unlock(&memcg_oom_lock);
1797 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1799 struct mem_cgroup *iter;
1801 spin_lock(&memcg_oom_lock);
1802 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1803 for_each_mem_cgroup_tree(iter, memcg)
1804 iter->oom_lock = false;
1805 spin_unlock(&memcg_oom_lock);
1808 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1810 struct mem_cgroup *iter;
1812 spin_lock(&memcg_oom_lock);
1813 for_each_mem_cgroup_tree(iter, memcg)
1815 spin_unlock(&memcg_oom_lock);
1818 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1820 struct mem_cgroup *iter;
1823 * When a new child is created while the hierarchy is under oom,
1824 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1826 spin_lock(&memcg_oom_lock);
1827 for_each_mem_cgroup_tree(iter, memcg)
1828 if (iter->under_oom > 0)
1830 spin_unlock(&memcg_oom_lock);
1833 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1835 struct oom_wait_info {
1836 struct mem_cgroup *memcg;
1837 wait_queue_entry_t wait;
1840 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1841 unsigned mode, int sync, void *arg)
1843 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1844 struct mem_cgroup *oom_wait_memcg;
1845 struct oom_wait_info *oom_wait_info;
1847 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1848 oom_wait_memcg = oom_wait_info->memcg;
1850 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1851 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1853 return autoremove_wake_function(wait, mode, sync, arg);
1856 static void memcg_oom_recover(struct mem_cgroup *memcg)
1859 * For the following lockless ->under_oom test, the only required
1860 * guarantee is that it must see the state asserted by an OOM when
1861 * this function is called as a result of userland actions
1862 * triggered by the notification of the OOM. This is trivially
1863 * achieved by invoking mem_cgroup_mark_under_oom() before
1864 * triggering notification.
1866 if (memcg && memcg->under_oom)
1867 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1877 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1879 enum oom_status ret;
1882 if (order > PAGE_ALLOC_COSTLY_ORDER)
1885 memcg_memory_event(memcg, MEMCG_OOM);
1888 * We are in the middle of the charge context here, so we
1889 * don't want to block when potentially sitting on a callstack
1890 * that holds all kinds of filesystem and mm locks.
1892 * cgroup1 allows disabling the OOM killer and waiting for outside
1893 * handling until the charge can succeed; remember the context and put
1894 * the task to sleep at the end of the page fault when all locks are
1897 * On the other hand, in-kernel OOM killer allows for an async victim
1898 * memory reclaim (oom_reaper) and that means that we are not solely
1899 * relying on the oom victim to make a forward progress and we can
1900 * invoke the oom killer here.
1902 * Please note that mem_cgroup_out_of_memory might fail to find a
1903 * victim and then we have to bail out from the charge path.
1905 if (memcg->oom_kill_disable) {
1906 if (!current->in_user_fault)
1908 css_get(&memcg->css);
1909 current->memcg_in_oom = memcg;
1910 current->memcg_oom_gfp_mask = mask;
1911 current->memcg_oom_order = order;
1916 mem_cgroup_mark_under_oom(memcg);
1918 locked = mem_cgroup_oom_trylock(memcg);
1921 mem_cgroup_oom_notify(memcg);
1923 mem_cgroup_unmark_under_oom(memcg);
1924 if (mem_cgroup_out_of_memory(memcg, mask, order))
1930 mem_cgroup_oom_unlock(memcg);
1936 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1937 * @handle: actually kill/wait or just clean up the OOM state
1939 * This has to be called at the end of a page fault if the memcg OOM
1940 * handler was enabled.
1942 * Memcg supports userspace OOM handling where failed allocations must
1943 * sleep on a waitqueue until the userspace task resolves the
1944 * situation. Sleeping directly in the charge context with all kinds
1945 * of locks held is not a good idea, instead we remember an OOM state
1946 * in the task and mem_cgroup_oom_synchronize() has to be called at
1947 * the end of the page fault to complete the OOM handling.
1949 * Returns %true if an ongoing memcg OOM situation was detected and
1950 * completed, %false otherwise.
1952 bool mem_cgroup_oom_synchronize(bool handle)
1954 struct mem_cgroup *memcg = current->memcg_in_oom;
1955 struct oom_wait_info owait;
1958 /* OOM is global, do not handle */
1965 owait.memcg = memcg;
1966 owait.wait.flags = 0;
1967 owait.wait.func = memcg_oom_wake_function;
1968 owait.wait.private = current;
1969 INIT_LIST_HEAD(&owait.wait.entry);
1971 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1972 mem_cgroup_mark_under_oom(memcg);
1974 locked = mem_cgroup_oom_trylock(memcg);
1977 mem_cgroup_oom_notify(memcg);
1979 if (locked && !memcg->oom_kill_disable) {
1980 mem_cgroup_unmark_under_oom(memcg);
1981 finish_wait(&memcg_oom_waitq, &owait.wait);
1982 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1983 current->memcg_oom_order);
1986 mem_cgroup_unmark_under_oom(memcg);
1987 finish_wait(&memcg_oom_waitq, &owait.wait);
1991 mem_cgroup_oom_unlock(memcg);
1993 * There is no guarantee that an OOM-lock contender
1994 * sees the wakeups triggered by the OOM kill
1995 * uncharges. Wake any sleepers explicitely.
1997 memcg_oom_recover(memcg);
2000 current->memcg_in_oom = NULL;
2001 css_put(&memcg->css);
2006 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2007 * @victim: task to be killed by the OOM killer
2008 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2010 * Returns a pointer to a memory cgroup, which has to be cleaned up
2011 * by killing all belonging OOM-killable tasks.
2013 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2015 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2016 struct mem_cgroup *oom_domain)
2018 struct mem_cgroup *oom_group = NULL;
2019 struct mem_cgroup *memcg;
2021 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2025 oom_domain = root_mem_cgroup;
2029 memcg = mem_cgroup_from_task(victim);
2030 if (memcg == root_mem_cgroup)
2034 * Traverse the memory cgroup hierarchy from the victim task's
2035 * cgroup up to the OOMing cgroup (or root) to find the
2036 * highest-level memory cgroup with oom.group set.
2038 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2039 if (memcg->oom_group)
2042 if (memcg == oom_domain)
2047 css_get(&oom_group->css);
2054 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2056 pr_info("Tasks in ");
2057 pr_cont_cgroup_path(memcg->css.cgroup);
2058 pr_cont(" are going to be killed due to memory.oom.group set\n");
2062 * lock_page_memcg - lock a page->mem_cgroup binding
2065 * This function protects unlocked LRU pages from being moved to
2068 * It ensures lifetime of the returned memcg. Caller is responsible
2069 * for the lifetime of the page; __unlock_page_memcg() is available
2070 * when @page might get freed inside the locked section.
2072 struct mem_cgroup *lock_page_memcg(struct page *page)
2074 struct mem_cgroup *memcg;
2075 unsigned long flags;
2078 * The RCU lock is held throughout the transaction. The fast
2079 * path can get away without acquiring the memcg->move_lock
2080 * because page moving starts with an RCU grace period.
2082 * The RCU lock also protects the memcg from being freed when
2083 * the page state that is going to change is the only thing
2084 * preventing the page itself from being freed. E.g. writeback
2085 * doesn't hold a page reference and relies on PG_writeback to
2086 * keep off truncation, migration and so forth.
2090 if (mem_cgroup_disabled())
2093 memcg = page->mem_cgroup;
2094 if (unlikely(!memcg))
2097 if (atomic_read(&memcg->moving_account) <= 0)
2100 spin_lock_irqsave(&memcg->move_lock, flags);
2101 if (memcg != page->mem_cgroup) {
2102 spin_unlock_irqrestore(&memcg->move_lock, flags);
2107 * When charge migration first begins, we can have locked and
2108 * unlocked page stat updates happening concurrently. Track
2109 * the task who has the lock for unlock_page_memcg().
2111 memcg->move_lock_task = current;
2112 memcg->move_lock_flags = flags;
2116 EXPORT_SYMBOL(lock_page_memcg);
2119 * __unlock_page_memcg - unlock and unpin a memcg
2122 * Unlock and unpin a memcg returned by lock_page_memcg().
2124 void __unlock_page_memcg(struct mem_cgroup *memcg)
2126 if (memcg && memcg->move_lock_task == current) {
2127 unsigned long flags = memcg->move_lock_flags;
2129 memcg->move_lock_task = NULL;
2130 memcg->move_lock_flags = 0;
2132 spin_unlock_irqrestore(&memcg->move_lock, flags);
2139 * unlock_page_memcg - unlock a page->mem_cgroup binding
2142 void unlock_page_memcg(struct page *page)
2144 __unlock_page_memcg(page->mem_cgroup);
2146 EXPORT_SYMBOL(unlock_page_memcg);
2148 struct memcg_stock_pcp {
2149 struct mem_cgroup *cached; /* this never be root cgroup */
2150 unsigned int nr_pages;
2151 struct work_struct work;
2152 unsigned long flags;
2153 #define FLUSHING_CACHED_CHARGE 0
2155 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2156 static DEFINE_MUTEX(percpu_charge_mutex);
2159 * consume_stock: Try to consume stocked charge on this cpu.
2160 * @memcg: memcg to consume from.
2161 * @nr_pages: how many pages to charge.
2163 * The charges will only happen if @memcg matches the current cpu's memcg
2164 * stock, and at least @nr_pages are available in that stock. Failure to
2165 * service an allocation will refill the stock.
2167 * returns true if successful, false otherwise.
2169 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2171 struct memcg_stock_pcp *stock;
2172 unsigned long flags;
2175 if (nr_pages > MEMCG_CHARGE_BATCH)
2178 local_irq_save(flags);
2180 stock = this_cpu_ptr(&memcg_stock);
2181 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2182 stock->nr_pages -= nr_pages;
2186 local_irq_restore(flags);
2192 * Returns stocks cached in percpu and reset cached information.
2194 static void drain_stock(struct memcg_stock_pcp *stock)
2196 struct mem_cgroup *old = stock->cached;
2198 if (stock->nr_pages) {
2199 page_counter_uncharge(&old->memory, stock->nr_pages);
2200 if (do_memsw_account())
2201 page_counter_uncharge(&old->memsw, stock->nr_pages);
2202 css_put_many(&old->css, stock->nr_pages);
2203 stock->nr_pages = 0;
2205 stock->cached = NULL;
2208 static void drain_local_stock(struct work_struct *dummy)
2210 struct memcg_stock_pcp *stock;
2211 unsigned long flags;
2214 * The only protection from memory hotplug vs. drain_stock races is
2215 * that we always operate on local CPU stock here with IRQ disabled
2217 local_irq_save(flags);
2219 stock = this_cpu_ptr(&memcg_stock);
2221 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2223 local_irq_restore(flags);
2227 * Cache charges(val) to local per_cpu area.
2228 * This will be consumed by consume_stock() function, later.
2230 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2232 struct memcg_stock_pcp *stock;
2233 unsigned long flags;
2235 local_irq_save(flags);
2237 stock = this_cpu_ptr(&memcg_stock);
2238 if (stock->cached != memcg) { /* reset if necessary */
2240 stock->cached = memcg;
2242 stock->nr_pages += nr_pages;
2244 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2247 local_irq_restore(flags);
2251 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2252 * of the hierarchy under it.
2254 static void drain_all_stock(struct mem_cgroup *root_memcg)
2258 /* If someone's already draining, avoid adding running more workers. */
2259 if (!mutex_trylock(&percpu_charge_mutex))
2262 * Notify other cpus that system-wide "drain" is running
2263 * We do not care about races with the cpu hotplug because cpu down
2264 * as well as workers from this path always operate on the local
2265 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2268 for_each_online_cpu(cpu) {
2269 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2270 struct mem_cgroup *memcg;
2272 memcg = stock->cached;
2273 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2275 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2276 css_put(&memcg->css);
2279 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2281 drain_local_stock(&stock->work);
2283 schedule_work_on(cpu, &stock->work);
2285 css_put(&memcg->css);
2288 mutex_unlock(&percpu_charge_mutex);
2291 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2293 struct memcg_stock_pcp *stock;
2294 struct mem_cgroup *memcg, *mi;
2296 stock = &per_cpu(memcg_stock, cpu);
2299 for_each_mem_cgroup(memcg) {
2302 for (i = 0; i < MEMCG_NR_STAT; i++) {
2306 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2308 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2309 atomic_long_add(x, &memcg->vmstats[i]);
2311 if (i >= NR_VM_NODE_STAT_ITEMS)
2314 for_each_node(nid) {
2315 struct mem_cgroup_per_node *pn;
2317 pn = mem_cgroup_nodeinfo(memcg, nid);
2318 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2321 atomic_long_add(x, &pn->lruvec_stat[i]);
2322 } while ((pn = parent_nodeinfo(pn, nid)));
2326 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2329 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2331 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2332 atomic_long_add(x, &memcg->vmevents[i]);
2339 static void reclaim_high(struct mem_cgroup *memcg,
2340 unsigned int nr_pages,
2344 if (page_counter_read(&memcg->memory) <= memcg->high)
2346 memcg_memory_event(memcg, MEMCG_HIGH);
2347 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2348 } while ((memcg = parent_mem_cgroup(memcg)));
2351 static void high_work_func(struct work_struct *work)
2353 struct mem_cgroup *memcg;
2355 memcg = container_of(work, struct mem_cgroup, high_work);
2356 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2360 * Scheduled by try_charge() to be executed from the userland return path
2361 * and reclaims memory over the high limit.
2363 void mem_cgroup_handle_over_high(void)
2365 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2366 struct mem_cgroup *memcg;
2368 if (likely(!nr_pages))
2371 memcg = get_mem_cgroup_from_mm(current->mm);
2372 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2373 css_put(&memcg->css);
2374 current->memcg_nr_pages_over_high = 0;
2377 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2378 unsigned int nr_pages)
2380 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2381 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2382 struct mem_cgroup *mem_over_limit;
2383 struct page_counter *counter;
2384 unsigned long nr_reclaimed;
2385 bool may_swap = true;
2386 bool drained = false;
2387 enum oom_status oom_status;
2389 if (mem_cgroup_is_root(memcg))
2392 if (consume_stock(memcg, nr_pages))
2395 if (!do_memsw_account() ||
2396 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2397 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2399 if (do_memsw_account())
2400 page_counter_uncharge(&memcg->memsw, batch);
2401 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2403 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2407 if (batch > nr_pages) {
2413 * Unlike in global OOM situations, memcg is not in a physical
2414 * memory shortage. Allow dying and OOM-killed tasks to
2415 * bypass the last charges so that they can exit quickly and
2416 * free their memory.
2418 if (unlikely(should_force_charge()))
2422 * Prevent unbounded recursion when reclaim operations need to
2423 * allocate memory. This might exceed the limits temporarily,
2424 * but we prefer facilitating memory reclaim and getting back
2425 * under the limit over triggering OOM kills in these cases.
2427 if (unlikely(current->flags & PF_MEMALLOC))
2430 if (unlikely(task_in_memcg_oom(current)))
2433 if (!gfpflags_allow_blocking(gfp_mask))
2436 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2438 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2439 gfp_mask, may_swap);
2441 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2445 drain_all_stock(mem_over_limit);
2450 if (gfp_mask & __GFP_NORETRY)
2453 * Even though the limit is exceeded at this point, reclaim
2454 * may have been able to free some pages. Retry the charge
2455 * before killing the task.
2457 * Only for regular pages, though: huge pages are rather
2458 * unlikely to succeed so close to the limit, and we fall back
2459 * to regular pages anyway in case of failure.
2461 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2464 * At task move, charge accounts can be doubly counted. So, it's
2465 * better to wait until the end of task_move if something is going on.
2467 if (mem_cgroup_wait_acct_move(mem_over_limit))
2473 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2476 if (gfp_mask & __GFP_NOFAIL)
2479 if (fatal_signal_pending(current))
2483 * keep retrying as long as the memcg oom killer is able to make
2484 * a forward progress or bypass the charge if the oom killer
2485 * couldn't make any progress.
2487 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2488 get_order(nr_pages * PAGE_SIZE));
2489 switch (oom_status) {
2491 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2499 if (!(gfp_mask & __GFP_NOFAIL))
2503 * The allocation either can't fail or will lead to more memory
2504 * being freed very soon. Allow memory usage go over the limit
2505 * temporarily by force charging it.
2507 page_counter_charge(&memcg->memory, nr_pages);
2508 if (do_memsw_account())
2509 page_counter_charge(&memcg->memsw, nr_pages);
2510 css_get_many(&memcg->css, nr_pages);
2515 css_get_many(&memcg->css, batch);
2516 if (batch > nr_pages)
2517 refill_stock(memcg, batch - nr_pages);
2520 * If the hierarchy is above the normal consumption range, schedule
2521 * reclaim on returning to userland. We can perform reclaim here
2522 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2523 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2524 * not recorded as it most likely matches current's and won't
2525 * change in the meantime. As high limit is checked again before
2526 * reclaim, the cost of mismatch is negligible.
2529 if (page_counter_read(&memcg->memory) > memcg->high) {
2530 /* Don't bother a random interrupted task */
2531 if (in_interrupt()) {
2532 schedule_work(&memcg->high_work);
2535 current->memcg_nr_pages_over_high += batch;
2536 set_notify_resume(current);
2539 } while ((memcg = parent_mem_cgroup(memcg)));
2544 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2546 if (mem_cgroup_is_root(memcg))
2549 page_counter_uncharge(&memcg->memory, nr_pages);
2550 if (do_memsw_account())
2551 page_counter_uncharge(&memcg->memsw, nr_pages);
2553 css_put_many(&memcg->css, nr_pages);
2556 static void lock_page_lru(struct page *page, int *isolated)
2558 pg_data_t *pgdat = page_pgdat(page);
2560 spin_lock_irq(&pgdat->lru_lock);
2561 if (PageLRU(page)) {
2562 struct lruvec *lruvec;
2564 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2566 del_page_from_lru_list(page, lruvec, page_lru(page));
2572 static void unlock_page_lru(struct page *page, int isolated)
2574 pg_data_t *pgdat = page_pgdat(page);
2577 struct lruvec *lruvec;
2579 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2580 VM_BUG_ON_PAGE(PageLRU(page), page);
2582 add_page_to_lru_list(page, lruvec, page_lru(page));
2584 spin_unlock_irq(&pgdat->lru_lock);
2587 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2592 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2595 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2596 * may already be on some other mem_cgroup's LRU. Take care of it.
2599 lock_page_lru(page, &isolated);
2602 * Nobody should be changing or seriously looking at
2603 * page->mem_cgroup at this point:
2605 * - the page is uncharged
2607 * - the page is off-LRU
2609 * - an anonymous fault has exclusive page access, except for
2610 * a locked page table
2612 * - a page cache insertion, a swapin fault, or a migration
2613 * have the page locked
2615 page->mem_cgroup = memcg;
2618 unlock_page_lru(page, isolated);
2621 #ifdef CONFIG_MEMCG_KMEM
2622 static int memcg_alloc_cache_id(void)
2627 id = ida_simple_get(&memcg_cache_ida,
2628 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2632 if (id < memcg_nr_cache_ids)
2636 * There's no space for the new id in memcg_caches arrays,
2637 * so we have to grow them.
2639 down_write(&memcg_cache_ids_sem);
2641 size = 2 * (id + 1);
2642 if (size < MEMCG_CACHES_MIN_SIZE)
2643 size = MEMCG_CACHES_MIN_SIZE;
2644 else if (size > MEMCG_CACHES_MAX_SIZE)
2645 size = MEMCG_CACHES_MAX_SIZE;
2647 err = memcg_update_all_caches(size);
2649 err = memcg_update_all_list_lrus(size);
2651 memcg_nr_cache_ids = size;
2653 up_write(&memcg_cache_ids_sem);
2656 ida_simple_remove(&memcg_cache_ida, id);
2662 static void memcg_free_cache_id(int id)
2664 ida_simple_remove(&memcg_cache_ida, id);
2667 struct memcg_kmem_cache_create_work {
2668 struct mem_cgroup *memcg;
2669 struct kmem_cache *cachep;
2670 struct work_struct work;
2673 static void memcg_kmem_cache_create_func(struct work_struct *w)
2675 struct memcg_kmem_cache_create_work *cw =
2676 container_of(w, struct memcg_kmem_cache_create_work, work);
2677 struct mem_cgroup *memcg = cw->memcg;
2678 struct kmem_cache *cachep = cw->cachep;
2680 memcg_create_kmem_cache(memcg, cachep);
2682 css_put(&memcg->css);
2687 * Enqueue the creation of a per-memcg kmem_cache.
2689 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2690 struct kmem_cache *cachep)
2692 struct memcg_kmem_cache_create_work *cw;
2694 if (!css_tryget_online(&memcg->css))
2697 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2702 cw->cachep = cachep;
2703 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2705 queue_work(memcg_kmem_cache_wq, &cw->work);
2708 static inline bool memcg_kmem_bypass(void)
2710 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2716 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2717 * @cachep: the original global kmem cache
2719 * Return the kmem_cache we're supposed to use for a slab allocation.
2720 * We try to use the current memcg's version of the cache.
2722 * If the cache does not exist yet, if we are the first user of it, we
2723 * create it asynchronously in a workqueue and let the current allocation
2724 * go through with the original cache.
2726 * This function takes a reference to the cache it returns to assure it
2727 * won't get destroyed while we are working with it. Once the caller is
2728 * done with it, memcg_kmem_put_cache() must be called to release the
2731 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2733 struct mem_cgroup *memcg;
2734 struct kmem_cache *memcg_cachep;
2735 struct memcg_cache_array *arr;
2738 VM_BUG_ON(!is_root_cache(cachep));
2740 if (memcg_kmem_bypass())
2745 if (unlikely(current->active_memcg))
2746 memcg = current->active_memcg;
2748 memcg = mem_cgroup_from_task(current);
2750 if (!memcg || memcg == root_mem_cgroup)
2753 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2757 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2760 * Make sure we will access the up-to-date value. The code updating
2761 * memcg_caches issues a write barrier to match the data dependency
2762 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2764 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2767 * If we are in a safe context (can wait, and not in interrupt
2768 * context), we could be be predictable and return right away.
2769 * This would guarantee that the allocation being performed
2770 * already belongs in the new cache.
2772 * However, there are some clashes that can arrive from locking.
2773 * For instance, because we acquire the slab_mutex while doing
2774 * memcg_create_kmem_cache, this means no further allocation
2775 * could happen with the slab_mutex held. So it's better to
2778 * If the memcg is dying or memcg_cache is about to be released,
2779 * don't bother creating new kmem_caches. Because memcg_cachep
2780 * is ZEROed as the fist step of kmem offlining, we don't need
2781 * percpu_ref_tryget_live() here. css_tryget_online() check in
2782 * memcg_schedule_kmem_cache_create() will prevent us from
2783 * creation of a new kmem_cache.
2785 if (unlikely(!memcg_cachep))
2786 memcg_schedule_kmem_cache_create(memcg, cachep);
2787 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2788 cachep = memcg_cachep;
2795 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2796 * @cachep: the cache returned by memcg_kmem_get_cache
2798 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2800 if (!is_root_cache(cachep))
2801 percpu_ref_put(&cachep->memcg_params.refcnt);
2805 * __memcg_kmem_charge_memcg: charge a kmem page
2806 * @page: page to charge
2807 * @gfp: reclaim mode
2808 * @order: allocation order
2809 * @memcg: memory cgroup to charge
2811 * Returns 0 on success, an error code on failure.
2813 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2814 struct mem_cgroup *memcg)
2816 unsigned int nr_pages = 1 << order;
2817 struct page_counter *counter;
2820 ret = try_charge(memcg, gfp, nr_pages);
2824 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2825 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2826 cancel_charge(memcg, nr_pages);
2833 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2834 * @page: page to charge
2835 * @gfp: reclaim mode
2836 * @order: allocation order
2838 * Returns 0 on success, an error code on failure.
2840 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2842 struct mem_cgroup *memcg;
2845 if (memcg_kmem_bypass())
2848 memcg = get_mem_cgroup_from_current();
2849 if (!mem_cgroup_is_root(memcg)) {
2850 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2852 page->mem_cgroup = memcg;
2853 __SetPageKmemcg(page);
2856 css_put(&memcg->css);
2861 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2862 * @memcg: memcg to uncharge
2863 * @nr_pages: number of pages to uncharge
2865 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2866 unsigned int nr_pages)
2868 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2869 page_counter_uncharge(&memcg->kmem, nr_pages);
2871 page_counter_uncharge(&memcg->memory, nr_pages);
2872 if (do_memsw_account())
2873 page_counter_uncharge(&memcg->memsw, nr_pages);
2876 * __memcg_kmem_uncharge: uncharge a kmem page
2877 * @page: page to uncharge
2878 * @order: allocation order
2880 void __memcg_kmem_uncharge(struct page *page, int order)
2882 struct mem_cgroup *memcg = page->mem_cgroup;
2883 unsigned int nr_pages = 1 << order;
2888 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2889 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
2890 page->mem_cgroup = NULL;
2892 /* slab pages do not have PageKmemcg flag set */
2893 if (PageKmemcg(page))
2894 __ClearPageKmemcg(page);
2896 css_put_many(&memcg->css, nr_pages);
2898 #endif /* CONFIG_MEMCG_KMEM */
2900 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2903 * Because tail pages are not marked as "used", set it. We're under
2904 * pgdat->lru_lock and migration entries setup in all page mappings.
2906 void mem_cgroup_split_huge_fixup(struct page *head)
2910 if (mem_cgroup_disabled())
2913 for (i = 1; i < HPAGE_PMD_NR; i++)
2914 head[i].mem_cgroup = head->mem_cgroup;
2916 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2918 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2920 #ifdef CONFIG_MEMCG_SWAP
2922 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2923 * @entry: swap entry to be moved
2924 * @from: mem_cgroup which the entry is moved from
2925 * @to: mem_cgroup which the entry is moved to
2927 * It succeeds only when the swap_cgroup's record for this entry is the same
2928 * as the mem_cgroup's id of @from.
2930 * Returns 0 on success, -EINVAL on failure.
2932 * The caller must have charged to @to, IOW, called page_counter_charge() about
2933 * both res and memsw, and called css_get().
2935 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2936 struct mem_cgroup *from, struct mem_cgroup *to)
2938 unsigned short old_id, new_id;
2940 old_id = mem_cgroup_id(from);
2941 new_id = mem_cgroup_id(to);
2943 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2944 mod_memcg_state(from, MEMCG_SWAP, -1);
2945 mod_memcg_state(to, MEMCG_SWAP, 1);
2951 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2952 struct mem_cgroup *from, struct mem_cgroup *to)
2958 static DEFINE_MUTEX(memcg_max_mutex);
2960 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2961 unsigned long max, bool memsw)
2963 bool enlarge = false;
2964 bool drained = false;
2966 bool limits_invariant;
2967 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2970 if (signal_pending(current)) {
2975 mutex_lock(&memcg_max_mutex);
2977 * Make sure that the new limit (memsw or memory limit) doesn't
2978 * break our basic invariant rule memory.max <= memsw.max.
2980 limits_invariant = memsw ? max >= memcg->memory.max :
2981 max <= memcg->memsw.max;
2982 if (!limits_invariant) {
2983 mutex_unlock(&memcg_max_mutex);
2987 if (max > counter->max)
2989 ret = page_counter_set_max(counter, max);
2990 mutex_unlock(&memcg_max_mutex);
2996 drain_all_stock(memcg);
3001 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3002 GFP_KERNEL, !memsw)) {
3008 if (!ret && enlarge)
3009 memcg_oom_recover(memcg);
3014 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3016 unsigned long *total_scanned)
3018 unsigned long nr_reclaimed = 0;
3019 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3020 unsigned long reclaimed;
3022 struct mem_cgroup_tree_per_node *mctz;
3023 unsigned long excess;
3024 unsigned long nr_scanned;
3029 mctz = soft_limit_tree_node(pgdat->node_id);
3032 * Do not even bother to check the largest node if the root
3033 * is empty. Do it lockless to prevent lock bouncing. Races
3034 * are acceptable as soft limit is best effort anyway.
3036 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3040 * This loop can run a while, specially if mem_cgroup's continuously
3041 * keep exceeding their soft limit and putting the system under
3048 mz = mem_cgroup_largest_soft_limit_node(mctz);
3053 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3054 gfp_mask, &nr_scanned);
3055 nr_reclaimed += reclaimed;
3056 *total_scanned += nr_scanned;
3057 spin_lock_irq(&mctz->lock);
3058 __mem_cgroup_remove_exceeded(mz, mctz);
3061 * If we failed to reclaim anything from this memory cgroup
3062 * it is time to move on to the next cgroup
3066 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3068 excess = soft_limit_excess(mz->memcg);
3070 * One school of thought says that we should not add
3071 * back the node to the tree if reclaim returns 0.
3072 * But our reclaim could return 0, simply because due
3073 * to priority we are exposing a smaller subset of
3074 * memory to reclaim from. Consider this as a longer
3077 /* If excess == 0, no tree ops */
3078 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3079 spin_unlock_irq(&mctz->lock);
3080 css_put(&mz->memcg->css);
3083 * Could not reclaim anything and there are no more
3084 * mem cgroups to try or we seem to be looping without
3085 * reclaiming anything.
3087 if (!nr_reclaimed &&
3089 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3091 } while (!nr_reclaimed);
3093 css_put(&next_mz->memcg->css);
3094 return nr_reclaimed;
3098 * Test whether @memcg has children, dead or alive. Note that this
3099 * function doesn't care whether @memcg has use_hierarchy enabled and
3100 * returns %true if there are child csses according to the cgroup
3101 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3103 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3108 ret = css_next_child(NULL, &memcg->css);
3114 * Reclaims as many pages from the given memcg as possible.
3116 * Caller is responsible for holding css reference for memcg.
3118 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3120 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3122 /* we call try-to-free pages for make this cgroup empty */
3123 lru_add_drain_all();
3125 drain_all_stock(memcg);
3127 /* try to free all pages in this cgroup */
3128 while (nr_retries && page_counter_read(&memcg->memory)) {
3131 if (signal_pending(current))
3134 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3138 /* maybe some writeback is necessary */
3139 congestion_wait(BLK_RW_ASYNC, HZ/10);
3147 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3148 char *buf, size_t nbytes,
3151 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3153 if (mem_cgroup_is_root(memcg))
3155 return mem_cgroup_force_empty(memcg) ?: nbytes;
3158 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3161 return mem_cgroup_from_css(css)->use_hierarchy;
3164 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3165 struct cftype *cft, u64 val)
3168 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3169 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3171 if (memcg->use_hierarchy == val)
3175 * If parent's use_hierarchy is set, we can't make any modifications
3176 * in the child subtrees. If it is unset, then the change can
3177 * occur, provided the current cgroup has no children.
3179 * For the root cgroup, parent_mem is NULL, we allow value to be
3180 * set if there are no children.
3182 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3183 (val == 1 || val == 0)) {
3184 if (!memcg_has_children(memcg))
3185 memcg->use_hierarchy = val;
3194 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3198 if (mem_cgroup_is_root(memcg)) {
3199 val = memcg_page_state(memcg, MEMCG_CACHE) +
3200 memcg_page_state(memcg, MEMCG_RSS);
3202 val += memcg_page_state(memcg, MEMCG_SWAP);
3205 val = page_counter_read(&memcg->memory);
3207 val = page_counter_read(&memcg->memsw);
3220 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3223 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3224 struct page_counter *counter;
3226 switch (MEMFILE_TYPE(cft->private)) {
3228 counter = &memcg->memory;
3231 counter = &memcg->memsw;
3234 counter = &memcg->kmem;
3237 counter = &memcg->tcpmem;
3243 switch (MEMFILE_ATTR(cft->private)) {
3245 if (counter == &memcg->memory)
3246 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3247 if (counter == &memcg->memsw)
3248 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3249 return (u64)page_counter_read(counter) * PAGE_SIZE;
3251 return (u64)counter->max * PAGE_SIZE;
3253 return (u64)counter->watermark * PAGE_SIZE;
3255 return counter->failcnt;
3256 case RES_SOFT_LIMIT:
3257 return (u64)memcg->soft_limit * PAGE_SIZE;
3263 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3265 unsigned long stat[MEMCG_NR_STAT];
3266 struct mem_cgroup *mi;
3269 for (i = 0; i < MEMCG_NR_STAT; i++)
3272 for_each_online_cpu(cpu)
3273 for (i = 0; i < MEMCG_NR_STAT; i++)
3274 stat[i] += raw_cpu_read(memcg->vmstats_percpu->stat[i]);
3276 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3277 for (i = 0; i < MEMCG_NR_STAT; i++)
3278 atomic_long_add(stat[i], &mi->vmstats[i]);
3280 for_each_node(node) {
3281 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3282 struct mem_cgroup_per_node *pi;
3284 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3287 for_each_online_cpu(cpu)
3288 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3289 stat[i] += raw_cpu_read(
3290 pn->lruvec_stat_cpu->count[i]);
3292 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3293 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3294 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3298 #ifdef CONFIG_MEMCG_KMEM
3299 static int memcg_online_kmem(struct mem_cgroup *memcg)
3303 if (cgroup_memory_nokmem)
3306 BUG_ON(memcg->kmemcg_id >= 0);
3307 BUG_ON(memcg->kmem_state);
3309 memcg_id = memcg_alloc_cache_id();
3313 static_branch_inc(&memcg_kmem_enabled_key);
3315 * A memory cgroup is considered kmem-online as soon as it gets
3316 * kmemcg_id. Setting the id after enabling static branching will
3317 * guarantee no one starts accounting before all call sites are
3320 memcg->kmemcg_id = memcg_id;
3321 memcg->kmem_state = KMEM_ONLINE;
3322 INIT_LIST_HEAD(&memcg->kmem_caches);
3327 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3329 struct cgroup_subsys_state *css;
3330 struct mem_cgroup *parent, *child;
3333 if (memcg->kmem_state != KMEM_ONLINE)
3336 * Clear the online state before clearing memcg_caches array
3337 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3338 * guarantees that no cache will be created for this cgroup
3339 * after we are done (see memcg_create_kmem_cache()).
3341 memcg->kmem_state = KMEM_ALLOCATED;
3343 parent = parent_mem_cgroup(memcg);
3345 parent = root_mem_cgroup;
3347 memcg_deactivate_kmem_caches(memcg, parent);
3349 kmemcg_id = memcg->kmemcg_id;
3350 BUG_ON(kmemcg_id < 0);
3353 * Change kmemcg_id of this cgroup and all its descendants to the
3354 * parent's id, and then move all entries from this cgroup's list_lrus
3355 * to ones of the parent. After we have finished, all list_lrus
3356 * corresponding to this cgroup are guaranteed to remain empty. The
3357 * ordering is imposed by list_lru_node->lock taken by
3358 * memcg_drain_all_list_lrus().
3360 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3361 css_for_each_descendant_pre(css, &memcg->css) {
3362 child = mem_cgroup_from_css(css);
3363 BUG_ON(child->kmemcg_id != kmemcg_id);
3364 child->kmemcg_id = parent->kmemcg_id;
3365 if (!memcg->use_hierarchy)
3370 memcg_drain_all_list_lrus(kmemcg_id, parent);
3372 memcg_free_cache_id(kmemcg_id);
3375 static void memcg_free_kmem(struct mem_cgroup *memcg)
3377 /* css_alloc() failed, offlining didn't happen */
3378 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3379 memcg_offline_kmem(memcg);
3381 if (memcg->kmem_state == KMEM_ALLOCATED) {
3382 WARN_ON(!list_empty(&memcg->kmem_caches));
3383 static_branch_dec(&memcg_kmem_enabled_key);
3387 static int memcg_online_kmem(struct mem_cgroup *memcg)
3391 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3394 static void memcg_free_kmem(struct mem_cgroup *memcg)
3397 #endif /* CONFIG_MEMCG_KMEM */
3399 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3404 mutex_lock(&memcg_max_mutex);
3405 ret = page_counter_set_max(&memcg->kmem, max);
3406 mutex_unlock(&memcg_max_mutex);
3410 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3414 mutex_lock(&memcg_max_mutex);
3416 ret = page_counter_set_max(&memcg->tcpmem, max);
3420 if (!memcg->tcpmem_active) {
3422 * The active flag needs to be written after the static_key
3423 * update. This is what guarantees that the socket activation
3424 * function is the last one to run. See mem_cgroup_sk_alloc()
3425 * for details, and note that we don't mark any socket as
3426 * belonging to this memcg until that flag is up.
3428 * We need to do this, because static_keys will span multiple
3429 * sites, but we can't control their order. If we mark a socket
3430 * as accounted, but the accounting functions are not patched in
3431 * yet, we'll lose accounting.
3433 * We never race with the readers in mem_cgroup_sk_alloc(),
3434 * because when this value change, the code to process it is not
3437 static_branch_inc(&memcg_sockets_enabled_key);
3438 memcg->tcpmem_active = true;
3441 mutex_unlock(&memcg_max_mutex);
3446 * The user of this function is...
3449 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3450 char *buf, size_t nbytes, loff_t off)
3452 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3453 unsigned long nr_pages;
3456 buf = strstrip(buf);
3457 ret = page_counter_memparse(buf, "-1", &nr_pages);
3461 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3463 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3467 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3469 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3472 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3475 ret = memcg_update_kmem_max(memcg, nr_pages);
3478 ret = memcg_update_tcp_max(memcg, nr_pages);
3482 case RES_SOFT_LIMIT:
3483 memcg->soft_limit = nr_pages;
3487 return ret ?: nbytes;
3490 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3491 size_t nbytes, loff_t off)
3493 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3494 struct page_counter *counter;
3496 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3498 counter = &memcg->memory;
3501 counter = &memcg->memsw;
3504 counter = &memcg->kmem;
3507 counter = &memcg->tcpmem;
3513 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3515 page_counter_reset_watermark(counter);
3518 counter->failcnt = 0;
3527 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3530 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3534 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3535 struct cftype *cft, u64 val)
3537 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3539 if (val & ~MOVE_MASK)
3543 * No kind of locking is needed in here, because ->can_attach() will
3544 * check this value once in the beginning of the process, and then carry
3545 * on with stale data. This means that changes to this value will only
3546 * affect task migrations starting after the change.
3548 memcg->move_charge_at_immigrate = val;
3552 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3553 struct cftype *cft, u64 val)
3561 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3562 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3563 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3565 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3566 int nid, unsigned int lru_mask)
3568 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3569 unsigned long nr = 0;
3572 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3575 if (!(BIT(lru) & lru_mask))
3577 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3582 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3583 unsigned int lru_mask)
3585 unsigned long nr = 0;
3589 if (!(BIT(lru) & lru_mask))
3591 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3596 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3600 unsigned int lru_mask;
3603 static const struct numa_stat stats[] = {
3604 { "total", LRU_ALL },
3605 { "file", LRU_ALL_FILE },
3606 { "anon", LRU_ALL_ANON },
3607 { "unevictable", BIT(LRU_UNEVICTABLE) },
3609 const struct numa_stat *stat;
3612 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3614 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3615 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3616 seq_printf(m, "%s=%lu", stat->name, nr);
3617 for_each_node_state(nid, N_MEMORY) {
3618 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3620 seq_printf(m, " N%d=%lu", nid, nr);
3625 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3626 struct mem_cgroup *iter;
3629 for_each_mem_cgroup_tree(iter, memcg)
3630 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3631 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3632 for_each_node_state(nid, N_MEMORY) {
3634 for_each_mem_cgroup_tree(iter, memcg)
3635 nr += mem_cgroup_node_nr_lru_pages(
3636 iter, nid, stat->lru_mask);
3637 seq_printf(m, " N%d=%lu", nid, nr);
3644 #endif /* CONFIG_NUMA */
3646 static const unsigned int memcg1_stats[] = {
3657 static const char *const memcg1_stat_names[] = {
3668 /* Universal VM events cgroup1 shows, original sort order */
3669 static const unsigned int memcg1_events[] = {
3676 static const char *const memcg1_event_names[] = {
3683 static int memcg_stat_show(struct seq_file *m, void *v)
3685 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3686 unsigned long memory, memsw;
3687 struct mem_cgroup *mi;
3690 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3691 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3693 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3694 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3696 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3697 memcg_page_state_local(memcg, memcg1_stats[i]) *
3701 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3702 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3703 memcg_events_local(memcg, memcg1_events[i]));
3705 for (i = 0; i < NR_LRU_LISTS; i++)
3706 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3707 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3710 /* Hierarchical information */
3711 memory = memsw = PAGE_COUNTER_MAX;
3712 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3713 memory = min(memory, mi->memory.max);
3714 memsw = min(memsw, mi->memsw.max);
3716 seq_printf(m, "hierarchical_memory_limit %llu\n",
3717 (u64)memory * PAGE_SIZE);
3718 if (do_memsw_account())
3719 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3720 (u64)memsw * PAGE_SIZE);
3722 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3723 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3725 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3726 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3730 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3731 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3732 (u64)memcg_events(memcg, memcg1_events[i]));
3734 for (i = 0; i < NR_LRU_LISTS; i++)
3735 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3736 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3739 #ifdef CONFIG_DEBUG_VM
3742 struct mem_cgroup_per_node *mz;
3743 struct zone_reclaim_stat *rstat;
3744 unsigned long recent_rotated[2] = {0, 0};
3745 unsigned long recent_scanned[2] = {0, 0};
3747 for_each_online_pgdat(pgdat) {
3748 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3749 rstat = &mz->lruvec.reclaim_stat;
3751 recent_rotated[0] += rstat->recent_rotated[0];
3752 recent_rotated[1] += rstat->recent_rotated[1];
3753 recent_scanned[0] += rstat->recent_scanned[0];
3754 recent_scanned[1] += rstat->recent_scanned[1];
3756 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3757 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3758 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3759 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3766 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3769 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3771 return mem_cgroup_swappiness(memcg);
3774 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3775 struct cftype *cft, u64 val)
3777 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3783 memcg->swappiness = val;
3785 vm_swappiness = val;
3790 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3792 struct mem_cgroup_threshold_ary *t;
3793 unsigned long usage;
3798 t = rcu_dereference(memcg->thresholds.primary);
3800 t = rcu_dereference(memcg->memsw_thresholds.primary);
3805 usage = mem_cgroup_usage(memcg, swap);
3808 * current_threshold points to threshold just below or equal to usage.
3809 * If it's not true, a threshold was crossed after last
3810 * call of __mem_cgroup_threshold().
3812 i = t->current_threshold;
3815 * Iterate backward over array of thresholds starting from
3816 * current_threshold and check if a threshold is crossed.
3817 * If none of thresholds below usage is crossed, we read
3818 * only one element of the array here.
3820 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3821 eventfd_signal(t->entries[i].eventfd, 1);
3823 /* i = current_threshold + 1 */
3827 * Iterate forward over array of thresholds starting from
3828 * current_threshold+1 and check if a threshold is crossed.
3829 * If none of thresholds above usage is crossed, we read
3830 * only one element of the array here.
3832 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3833 eventfd_signal(t->entries[i].eventfd, 1);
3835 /* Update current_threshold */
3836 t->current_threshold = i - 1;
3841 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3844 __mem_cgroup_threshold(memcg, false);
3845 if (do_memsw_account())
3846 __mem_cgroup_threshold(memcg, true);
3848 memcg = parent_mem_cgroup(memcg);
3852 static int compare_thresholds(const void *a, const void *b)
3854 const struct mem_cgroup_threshold *_a = a;
3855 const struct mem_cgroup_threshold *_b = b;
3857 if (_a->threshold > _b->threshold)
3860 if (_a->threshold < _b->threshold)
3866 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3868 struct mem_cgroup_eventfd_list *ev;
3870 spin_lock(&memcg_oom_lock);
3872 list_for_each_entry(ev, &memcg->oom_notify, list)
3873 eventfd_signal(ev->eventfd, 1);
3875 spin_unlock(&memcg_oom_lock);
3879 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3881 struct mem_cgroup *iter;
3883 for_each_mem_cgroup_tree(iter, memcg)
3884 mem_cgroup_oom_notify_cb(iter);
3887 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3888 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3890 struct mem_cgroup_thresholds *thresholds;
3891 struct mem_cgroup_threshold_ary *new;
3892 unsigned long threshold;
3893 unsigned long usage;
3896 ret = page_counter_memparse(args, "-1", &threshold);
3900 mutex_lock(&memcg->thresholds_lock);
3903 thresholds = &memcg->thresholds;
3904 usage = mem_cgroup_usage(memcg, false);
3905 } else if (type == _MEMSWAP) {
3906 thresholds = &memcg->memsw_thresholds;
3907 usage = mem_cgroup_usage(memcg, true);
3911 /* Check if a threshold crossed before adding a new one */
3912 if (thresholds->primary)
3913 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3915 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3917 /* Allocate memory for new array of thresholds */
3918 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3925 /* Copy thresholds (if any) to new array */
3926 if (thresholds->primary) {
3927 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3928 sizeof(struct mem_cgroup_threshold));
3931 /* Add new threshold */
3932 new->entries[size - 1].eventfd = eventfd;
3933 new->entries[size - 1].threshold = threshold;
3935 /* Sort thresholds. Registering of new threshold isn't time-critical */
3936 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3937 compare_thresholds, NULL);
3939 /* Find current threshold */
3940 new->current_threshold = -1;
3941 for (i = 0; i < size; i++) {
3942 if (new->entries[i].threshold <= usage) {
3944 * new->current_threshold will not be used until
3945 * rcu_assign_pointer(), so it's safe to increment
3948 ++new->current_threshold;
3953 /* Free old spare buffer and save old primary buffer as spare */
3954 kfree(thresholds->spare);
3955 thresholds->spare = thresholds->primary;
3957 rcu_assign_pointer(thresholds->primary, new);
3959 /* To be sure that nobody uses thresholds */
3963 mutex_unlock(&memcg->thresholds_lock);
3968 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3969 struct eventfd_ctx *eventfd, const char *args)
3971 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3974 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3975 struct eventfd_ctx *eventfd, const char *args)
3977 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3980 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3981 struct eventfd_ctx *eventfd, enum res_type type)
3983 struct mem_cgroup_thresholds *thresholds;
3984 struct mem_cgroup_threshold_ary *new;
3985 unsigned long usage;
3988 mutex_lock(&memcg->thresholds_lock);
3991 thresholds = &memcg->thresholds;
3992 usage = mem_cgroup_usage(memcg, false);
3993 } else if (type == _MEMSWAP) {
3994 thresholds = &memcg->memsw_thresholds;
3995 usage = mem_cgroup_usage(memcg, true);
3999 if (!thresholds->primary)
4002 /* Check if a threshold crossed before removing */
4003 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4005 /* Calculate new number of threshold */
4007 for (i = 0; i < thresholds->primary->size; i++) {
4008 if (thresholds->primary->entries[i].eventfd != eventfd)
4012 new = thresholds->spare;
4014 /* Set thresholds array to NULL if we don't have thresholds */
4023 /* Copy thresholds and find current threshold */
4024 new->current_threshold = -1;
4025 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4026 if (thresholds->primary->entries[i].eventfd == eventfd)
4029 new->entries[j] = thresholds->primary->entries[i];
4030 if (new->entries[j].threshold <= usage) {
4032 * new->current_threshold will not be used
4033 * until rcu_assign_pointer(), so it's safe to increment
4036 ++new->current_threshold;
4042 /* Swap primary and spare array */
4043 thresholds->spare = thresholds->primary;
4045 rcu_assign_pointer(thresholds->primary, new);
4047 /* To be sure that nobody uses thresholds */
4050 /* If all events are unregistered, free the spare array */
4052 kfree(thresholds->spare);
4053 thresholds->spare = NULL;
4056 mutex_unlock(&memcg->thresholds_lock);
4059 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4060 struct eventfd_ctx *eventfd)
4062 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4065 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4066 struct eventfd_ctx *eventfd)
4068 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4071 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4072 struct eventfd_ctx *eventfd, const char *args)
4074 struct mem_cgroup_eventfd_list *event;
4076 event = kmalloc(sizeof(*event), GFP_KERNEL);
4080 spin_lock(&memcg_oom_lock);
4082 event->eventfd = eventfd;
4083 list_add(&event->list, &memcg->oom_notify);
4085 /* already in OOM ? */
4086 if (memcg->under_oom)
4087 eventfd_signal(eventfd, 1);
4088 spin_unlock(&memcg_oom_lock);
4093 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4094 struct eventfd_ctx *eventfd)
4096 struct mem_cgroup_eventfd_list *ev, *tmp;
4098 spin_lock(&memcg_oom_lock);
4100 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4101 if (ev->eventfd == eventfd) {
4102 list_del(&ev->list);
4107 spin_unlock(&memcg_oom_lock);
4110 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4112 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4114 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4115 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4116 seq_printf(sf, "oom_kill %lu\n",
4117 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4121 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4122 struct cftype *cft, u64 val)
4124 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4126 /* cannot set to root cgroup and only 0 and 1 are allowed */
4127 if (!css->parent || !((val == 0) || (val == 1)))
4130 memcg->oom_kill_disable = val;
4132 memcg_oom_recover(memcg);
4137 #ifdef CONFIG_CGROUP_WRITEBACK
4139 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4141 return wb_domain_init(&memcg->cgwb_domain, gfp);
4144 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4146 wb_domain_exit(&memcg->cgwb_domain);
4149 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4151 wb_domain_size_changed(&memcg->cgwb_domain);
4154 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4156 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4158 if (!memcg->css.parent)
4161 return &memcg->cgwb_domain;
4165 * idx can be of type enum memcg_stat_item or node_stat_item.
4166 * Keep in sync with memcg_exact_page().
4168 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4170 long x = atomic_long_read(&memcg->vmstats[idx]);
4173 for_each_online_cpu(cpu)
4174 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4181 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4182 * @wb: bdi_writeback in question
4183 * @pfilepages: out parameter for number of file pages
4184 * @pheadroom: out parameter for number of allocatable pages according to memcg
4185 * @pdirty: out parameter for number of dirty pages
4186 * @pwriteback: out parameter for number of pages under writeback
4188 * Determine the numbers of file, headroom, dirty, and writeback pages in
4189 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4190 * is a bit more involved.
4192 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4193 * headroom is calculated as the lowest headroom of itself and the
4194 * ancestors. Note that this doesn't consider the actual amount of
4195 * available memory in the system. The caller should further cap
4196 * *@pheadroom accordingly.
4198 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4199 unsigned long *pheadroom, unsigned long *pdirty,
4200 unsigned long *pwriteback)
4202 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4203 struct mem_cgroup *parent;
4205 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4207 /* this should eventually include NR_UNSTABLE_NFS */
4208 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4209 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4210 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4211 *pheadroom = PAGE_COUNTER_MAX;
4213 while ((parent = parent_mem_cgroup(memcg))) {
4214 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4215 unsigned long used = page_counter_read(&memcg->memory);
4217 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4222 #else /* CONFIG_CGROUP_WRITEBACK */
4224 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4229 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4233 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4237 #endif /* CONFIG_CGROUP_WRITEBACK */
4240 * DO NOT USE IN NEW FILES.
4242 * "cgroup.event_control" implementation.
4244 * This is way over-engineered. It tries to support fully configurable
4245 * events for each user. Such level of flexibility is completely
4246 * unnecessary especially in the light of the planned unified hierarchy.
4248 * Please deprecate this and replace with something simpler if at all
4253 * Unregister event and free resources.
4255 * Gets called from workqueue.
4257 static void memcg_event_remove(struct work_struct *work)
4259 struct mem_cgroup_event *event =
4260 container_of(work, struct mem_cgroup_event, remove);
4261 struct mem_cgroup *memcg = event->memcg;
4263 remove_wait_queue(event->wqh, &event->wait);
4265 event->unregister_event(memcg, event->eventfd);
4267 /* Notify userspace the event is going away. */
4268 eventfd_signal(event->eventfd, 1);
4270 eventfd_ctx_put(event->eventfd);
4272 css_put(&memcg->css);
4276 * Gets called on EPOLLHUP on eventfd when user closes it.
4278 * Called with wqh->lock held and interrupts disabled.
4280 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4281 int sync, void *key)
4283 struct mem_cgroup_event *event =
4284 container_of(wait, struct mem_cgroup_event, wait);
4285 struct mem_cgroup *memcg = event->memcg;
4286 __poll_t flags = key_to_poll(key);
4288 if (flags & EPOLLHUP) {
4290 * If the event has been detached at cgroup removal, we
4291 * can simply return knowing the other side will cleanup
4294 * We can't race against event freeing since the other
4295 * side will require wqh->lock via remove_wait_queue(),
4298 spin_lock(&memcg->event_list_lock);
4299 if (!list_empty(&event->list)) {
4300 list_del_init(&event->list);
4302 * We are in atomic context, but cgroup_event_remove()
4303 * may sleep, so we have to call it in workqueue.
4305 schedule_work(&event->remove);
4307 spin_unlock(&memcg->event_list_lock);
4313 static void memcg_event_ptable_queue_proc(struct file *file,
4314 wait_queue_head_t *wqh, poll_table *pt)
4316 struct mem_cgroup_event *event =
4317 container_of(pt, struct mem_cgroup_event, pt);
4320 add_wait_queue(wqh, &event->wait);
4324 * DO NOT USE IN NEW FILES.
4326 * Parse input and register new cgroup event handler.
4328 * Input must be in format '<event_fd> <control_fd> <args>'.
4329 * Interpretation of args is defined by control file implementation.
4331 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4332 char *buf, size_t nbytes, loff_t off)
4334 struct cgroup_subsys_state *css = of_css(of);
4335 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4336 struct mem_cgroup_event *event;
4337 struct cgroup_subsys_state *cfile_css;
4338 unsigned int efd, cfd;
4345 buf = strstrip(buf);
4347 efd = simple_strtoul(buf, &endp, 10);
4352 cfd = simple_strtoul(buf, &endp, 10);
4353 if ((*endp != ' ') && (*endp != '\0'))
4357 event = kzalloc(sizeof(*event), GFP_KERNEL);
4361 event->memcg = memcg;
4362 INIT_LIST_HEAD(&event->list);
4363 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4364 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4365 INIT_WORK(&event->remove, memcg_event_remove);
4373 event->eventfd = eventfd_ctx_fileget(efile.file);
4374 if (IS_ERR(event->eventfd)) {
4375 ret = PTR_ERR(event->eventfd);
4382 goto out_put_eventfd;
4385 /* the process need read permission on control file */
4386 /* AV: shouldn't we check that it's been opened for read instead? */
4387 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4392 * Determine the event callbacks and set them in @event. This used
4393 * to be done via struct cftype but cgroup core no longer knows
4394 * about these events. The following is crude but the whole thing
4395 * is for compatibility anyway.
4397 * DO NOT ADD NEW FILES.
4399 name = cfile.file->f_path.dentry->d_name.name;
4401 if (!strcmp(name, "memory.usage_in_bytes")) {
4402 event->register_event = mem_cgroup_usage_register_event;
4403 event->unregister_event = mem_cgroup_usage_unregister_event;
4404 } else if (!strcmp(name, "memory.oom_control")) {
4405 event->register_event = mem_cgroup_oom_register_event;
4406 event->unregister_event = mem_cgroup_oom_unregister_event;
4407 } else if (!strcmp(name, "memory.pressure_level")) {
4408 event->register_event = vmpressure_register_event;
4409 event->unregister_event = vmpressure_unregister_event;
4410 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4411 event->register_event = memsw_cgroup_usage_register_event;
4412 event->unregister_event = memsw_cgroup_usage_unregister_event;
4419 * Verify @cfile should belong to @css. Also, remaining events are
4420 * automatically removed on cgroup destruction but the removal is
4421 * asynchronous, so take an extra ref on @css.
4423 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4424 &memory_cgrp_subsys);
4426 if (IS_ERR(cfile_css))
4428 if (cfile_css != css) {
4433 ret = event->register_event(memcg, event->eventfd, buf);
4437 vfs_poll(efile.file, &event->pt);
4439 spin_lock(&memcg->event_list_lock);
4440 list_add(&event->list, &memcg->event_list);
4441 spin_unlock(&memcg->event_list_lock);
4453 eventfd_ctx_put(event->eventfd);
4462 static struct cftype mem_cgroup_legacy_files[] = {
4464 .name = "usage_in_bytes",
4465 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4466 .read_u64 = mem_cgroup_read_u64,
4469 .name = "max_usage_in_bytes",
4470 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4471 .write = mem_cgroup_reset,
4472 .read_u64 = mem_cgroup_read_u64,
4475 .name = "limit_in_bytes",
4476 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4477 .write = mem_cgroup_write,
4478 .read_u64 = mem_cgroup_read_u64,
4481 .name = "soft_limit_in_bytes",
4482 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4483 .write = mem_cgroup_write,
4484 .read_u64 = mem_cgroup_read_u64,
4488 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4489 .write = mem_cgroup_reset,
4490 .read_u64 = mem_cgroup_read_u64,
4494 .seq_show = memcg_stat_show,
4497 .name = "force_empty",
4498 .write = mem_cgroup_force_empty_write,
4501 .name = "use_hierarchy",
4502 .write_u64 = mem_cgroup_hierarchy_write,
4503 .read_u64 = mem_cgroup_hierarchy_read,
4506 .name = "cgroup.event_control", /* XXX: for compat */
4507 .write = memcg_write_event_control,
4508 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4511 .name = "swappiness",
4512 .read_u64 = mem_cgroup_swappiness_read,
4513 .write_u64 = mem_cgroup_swappiness_write,
4516 .name = "move_charge_at_immigrate",
4517 .read_u64 = mem_cgroup_move_charge_read,
4518 .write_u64 = mem_cgroup_move_charge_write,
4521 .name = "oom_control",
4522 .seq_show = mem_cgroup_oom_control_read,
4523 .write_u64 = mem_cgroup_oom_control_write,
4524 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4527 .name = "pressure_level",
4531 .name = "numa_stat",
4532 .seq_show = memcg_numa_stat_show,
4536 .name = "kmem.limit_in_bytes",
4537 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4538 .write = mem_cgroup_write,
4539 .read_u64 = mem_cgroup_read_u64,
4542 .name = "kmem.usage_in_bytes",
4543 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4544 .read_u64 = mem_cgroup_read_u64,
4547 .name = "kmem.failcnt",
4548 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4549 .write = mem_cgroup_reset,
4550 .read_u64 = mem_cgroup_read_u64,
4553 .name = "kmem.max_usage_in_bytes",
4554 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4555 .write = mem_cgroup_reset,
4556 .read_u64 = mem_cgroup_read_u64,
4558 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4560 .name = "kmem.slabinfo",
4561 .seq_start = memcg_slab_start,
4562 .seq_next = memcg_slab_next,
4563 .seq_stop = memcg_slab_stop,
4564 .seq_show = memcg_slab_show,
4568 .name = "kmem.tcp.limit_in_bytes",
4569 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4570 .write = mem_cgroup_write,
4571 .read_u64 = mem_cgroup_read_u64,
4574 .name = "kmem.tcp.usage_in_bytes",
4575 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4576 .read_u64 = mem_cgroup_read_u64,
4579 .name = "kmem.tcp.failcnt",
4580 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4581 .write = mem_cgroup_reset,
4582 .read_u64 = mem_cgroup_read_u64,
4585 .name = "kmem.tcp.max_usage_in_bytes",
4586 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4587 .write = mem_cgroup_reset,
4588 .read_u64 = mem_cgroup_read_u64,
4590 { }, /* terminate */
4594 * Private memory cgroup IDR
4596 * Swap-out records and page cache shadow entries need to store memcg
4597 * references in constrained space, so we maintain an ID space that is
4598 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4599 * memory-controlled cgroups to 64k.
4601 * However, there usually are many references to the oflline CSS after
4602 * the cgroup has been destroyed, such as page cache or reclaimable
4603 * slab objects, that don't need to hang on to the ID. We want to keep
4604 * those dead CSS from occupying IDs, or we might quickly exhaust the
4605 * relatively small ID space and prevent the creation of new cgroups
4606 * even when there are much fewer than 64k cgroups - possibly none.
4608 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4609 * be freed and recycled when it's no longer needed, which is usually
4610 * when the CSS is offlined.
4612 * The only exception to that are records of swapped out tmpfs/shmem
4613 * pages that need to be attributed to live ancestors on swapin. But
4614 * those references are manageable from userspace.
4617 static DEFINE_IDR(mem_cgroup_idr);
4619 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4621 if (memcg->id.id > 0) {
4622 idr_remove(&mem_cgroup_idr, memcg->id.id);
4627 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4629 refcount_add(n, &memcg->id.ref);
4632 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4634 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4635 mem_cgroup_id_remove(memcg);
4637 /* Memcg ID pins CSS */
4638 css_put(&memcg->css);
4642 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4644 mem_cgroup_id_get_many(memcg, 1);
4647 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4649 mem_cgroup_id_put_many(memcg, 1);
4653 * mem_cgroup_from_id - look up a memcg from a memcg id
4654 * @id: the memcg id to look up
4656 * Caller must hold rcu_read_lock().
4658 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4660 WARN_ON_ONCE(!rcu_read_lock_held());
4661 return idr_find(&mem_cgroup_idr, id);
4664 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4666 struct mem_cgroup_per_node *pn;
4669 * This routine is called against possible nodes.
4670 * But it's BUG to call kmalloc() against offline node.
4672 * TODO: this routine can waste much memory for nodes which will
4673 * never be onlined. It's better to use memory hotplug callback
4676 if (!node_state(node, N_NORMAL_MEMORY))
4678 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4682 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4683 if (!pn->lruvec_stat_local) {
4688 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4689 if (!pn->lruvec_stat_cpu) {
4690 free_percpu(pn->lruvec_stat_local);
4695 lruvec_init(&pn->lruvec);
4696 pn->usage_in_excess = 0;
4697 pn->on_tree = false;
4700 memcg->nodeinfo[node] = pn;
4704 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4706 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4711 free_percpu(pn->lruvec_stat_cpu);
4712 free_percpu(pn->lruvec_stat_local);
4716 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4721 * Flush percpu vmstats to guarantee the value correctness
4722 * on parent's and all ancestor levels.
4724 memcg_flush_percpu_vmstats(memcg);
4726 free_mem_cgroup_per_node_info(memcg, node);
4727 free_percpu(memcg->vmstats_percpu);
4728 free_percpu(memcg->vmstats_local);
4732 static void mem_cgroup_free(struct mem_cgroup *memcg)
4734 memcg_wb_domain_exit(memcg);
4735 __mem_cgroup_free(memcg);
4738 static struct mem_cgroup *mem_cgroup_alloc(void)
4740 struct mem_cgroup *memcg;
4744 size = sizeof(struct mem_cgroup);
4745 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4747 memcg = kzalloc(size, GFP_KERNEL);
4751 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4752 1, MEM_CGROUP_ID_MAX,
4754 if (memcg->id.id < 0)
4757 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
4758 if (!memcg->vmstats_local)
4761 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4762 if (!memcg->vmstats_percpu)
4766 if (alloc_mem_cgroup_per_node_info(memcg, node))
4769 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4772 INIT_WORK(&memcg->high_work, high_work_func);
4773 memcg->last_scanned_node = MAX_NUMNODES;
4774 INIT_LIST_HEAD(&memcg->oom_notify);
4775 mutex_init(&memcg->thresholds_lock);
4776 spin_lock_init(&memcg->move_lock);
4777 vmpressure_init(&memcg->vmpressure);
4778 INIT_LIST_HEAD(&memcg->event_list);
4779 spin_lock_init(&memcg->event_list_lock);
4780 memcg->socket_pressure = jiffies;
4781 #ifdef CONFIG_MEMCG_KMEM
4782 memcg->kmemcg_id = -1;
4784 #ifdef CONFIG_CGROUP_WRITEBACK
4785 INIT_LIST_HEAD(&memcg->cgwb_list);
4787 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4790 mem_cgroup_id_remove(memcg);
4791 __mem_cgroup_free(memcg);
4795 static struct cgroup_subsys_state * __ref
4796 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4798 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4799 struct mem_cgroup *memcg;
4800 long error = -ENOMEM;
4802 memcg = mem_cgroup_alloc();
4804 return ERR_PTR(error);
4806 memcg->high = PAGE_COUNTER_MAX;
4807 memcg->soft_limit = PAGE_COUNTER_MAX;
4809 memcg->swappiness = mem_cgroup_swappiness(parent);
4810 memcg->oom_kill_disable = parent->oom_kill_disable;
4812 if (parent && parent->use_hierarchy) {
4813 memcg->use_hierarchy = true;
4814 page_counter_init(&memcg->memory, &parent->memory);
4815 page_counter_init(&memcg->swap, &parent->swap);
4816 page_counter_init(&memcg->memsw, &parent->memsw);
4817 page_counter_init(&memcg->kmem, &parent->kmem);
4818 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4820 page_counter_init(&memcg->memory, NULL);
4821 page_counter_init(&memcg->swap, NULL);
4822 page_counter_init(&memcg->memsw, NULL);
4823 page_counter_init(&memcg->kmem, NULL);
4824 page_counter_init(&memcg->tcpmem, NULL);
4826 * Deeper hierachy with use_hierarchy == false doesn't make
4827 * much sense so let cgroup subsystem know about this
4828 * unfortunate state in our controller.
4830 if (parent != root_mem_cgroup)
4831 memory_cgrp_subsys.broken_hierarchy = true;
4834 /* The following stuff does not apply to the root */
4836 #ifdef CONFIG_MEMCG_KMEM
4837 INIT_LIST_HEAD(&memcg->kmem_caches);
4839 root_mem_cgroup = memcg;
4843 error = memcg_online_kmem(memcg);
4847 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4848 static_branch_inc(&memcg_sockets_enabled_key);
4852 mem_cgroup_id_remove(memcg);
4853 mem_cgroup_free(memcg);
4854 return ERR_PTR(-ENOMEM);
4857 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4859 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4862 * A memcg must be visible for memcg_expand_shrinker_maps()
4863 * by the time the maps are allocated. So, we allocate maps
4864 * here, when for_each_mem_cgroup() can't skip it.
4866 if (memcg_alloc_shrinker_maps(memcg)) {
4867 mem_cgroup_id_remove(memcg);
4871 /* Online state pins memcg ID, memcg ID pins CSS */
4872 refcount_set(&memcg->id.ref, 1);
4877 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4879 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4880 struct mem_cgroup_event *event, *tmp;
4883 * Unregister events and notify userspace.
4884 * Notify userspace about cgroup removing only after rmdir of cgroup
4885 * directory to avoid race between userspace and kernelspace.
4887 spin_lock(&memcg->event_list_lock);
4888 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4889 list_del_init(&event->list);
4890 schedule_work(&event->remove);
4892 spin_unlock(&memcg->event_list_lock);
4894 page_counter_set_min(&memcg->memory, 0);
4895 page_counter_set_low(&memcg->memory, 0);
4897 memcg_offline_kmem(memcg);
4898 wb_memcg_offline(memcg);
4900 drain_all_stock(memcg);
4902 mem_cgroup_id_put(memcg);
4905 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4907 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4909 invalidate_reclaim_iterators(memcg);
4912 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4914 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4916 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4917 static_branch_dec(&memcg_sockets_enabled_key);
4919 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4920 static_branch_dec(&memcg_sockets_enabled_key);
4922 vmpressure_cleanup(&memcg->vmpressure);
4923 cancel_work_sync(&memcg->high_work);
4924 mem_cgroup_remove_from_trees(memcg);
4925 memcg_free_shrinker_maps(memcg);
4926 memcg_free_kmem(memcg);
4927 mem_cgroup_free(memcg);
4931 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4932 * @css: the target css
4934 * Reset the states of the mem_cgroup associated with @css. This is
4935 * invoked when the userland requests disabling on the default hierarchy
4936 * but the memcg is pinned through dependency. The memcg should stop
4937 * applying policies and should revert to the vanilla state as it may be
4938 * made visible again.
4940 * The current implementation only resets the essential configurations.
4941 * This needs to be expanded to cover all the visible parts.
4943 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4945 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4947 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4948 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4949 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4950 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4951 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4952 page_counter_set_min(&memcg->memory, 0);
4953 page_counter_set_low(&memcg->memory, 0);
4954 memcg->high = PAGE_COUNTER_MAX;
4955 memcg->soft_limit = PAGE_COUNTER_MAX;
4956 memcg_wb_domain_size_changed(memcg);
4960 /* Handlers for move charge at task migration. */
4961 static int mem_cgroup_do_precharge(unsigned long count)
4965 /* Try a single bulk charge without reclaim first, kswapd may wake */
4966 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4968 mc.precharge += count;
4972 /* Try charges one by one with reclaim, but do not retry */
4974 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4988 enum mc_target_type {
4995 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4996 unsigned long addr, pte_t ptent)
4998 struct page *page = vm_normal_page(vma, addr, ptent);
5000 if (!page || !page_mapped(page))
5002 if (PageAnon(page)) {
5003 if (!(mc.flags & MOVE_ANON))
5006 if (!(mc.flags & MOVE_FILE))
5009 if (!get_page_unless_zero(page))
5015 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5016 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5017 pte_t ptent, swp_entry_t *entry)
5019 struct page *page = NULL;
5020 swp_entry_t ent = pte_to_swp_entry(ptent);
5022 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5026 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5027 * a device and because they are not accessible by CPU they are store
5028 * as special swap entry in the CPU page table.
5030 if (is_device_private_entry(ent)) {
5031 page = device_private_entry_to_page(ent);
5033 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5034 * a refcount of 1 when free (unlike normal page)
5036 if (!page_ref_add_unless(page, 1, 1))
5042 * Because lookup_swap_cache() updates some statistics counter,
5043 * we call find_get_page() with swapper_space directly.
5045 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5046 if (do_memsw_account())
5047 entry->val = ent.val;
5052 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5053 pte_t ptent, swp_entry_t *entry)
5059 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5060 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5062 struct page *page = NULL;
5063 struct address_space *mapping;
5066 if (!vma->vm_file) /* anonymous vma */
5068 if (!(mc.flags & MOVE_FILE))
5071 mapping = vma->vm_file->f_mapping;
5072 pgoff = linear_page_index(vma, addr);
5074 /* page is moved even if it's not RSS of this task(page-faulted). */
5076 /* shmem/tmpfs may report page out on swap: account for that too. */
5077 if (shmem_mapping(mapping)) {
5078 page = find_get_entry(mapping, pgoff);
5079 if (xa_is_value(page)) {
5080 swp_entry_t swp = radix_to_swp_entry(page);
5081 if (do_memsw_account())
5083 page = find_get_page(swap_address_space(swp),
5087 page = find_get_page(mapping, pgoff);
5089 page = find_get_page(mapping, pgoff);
5095 * mem_cgroup_move_account - move account of the page
5097 * @compound: charge the page as compound or small page
5098 * @from: mem_cgroup which the page is moved from.
5099 * @to: mem_cgroup which the page is moved to. @from != @to.
5101 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5103 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5106 static int mem_cgroup_move_account(struct page *page,
5108 struct mem_cgroup *from,
5109 struct mem_cgroup *to)
5111 unsigned long flags;
5112 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5116 VM_BUG_ON(from == to);
5117 VM_BUG_ON_PAGE(PageLRU(page), page);
5118 VM_BUG_ON(compound && !PageTransHuge(page));
5121 * Prevent mem_cgroup_migrate() from looking at
5122 * page->mem_cgroup of its source page while we change it.
5125 if (!trylock_page(page))
5129 if (page->mem_cgroup != from)
5132 anon = PageAnon(page);
5134 spin_lock_irqsave(&from->move_lock, flags);
5136 if (!anon && page_mapped(page)) {
5137 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
5138 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
5142 * move_lock grabbed above and caller set from->moving_account, so
5143 * mod_memcg_page_state will serialize updates to PageDirty.
5144 * So mapping should be stable for dirty pages.
5146 if (!anon && PageDirty(page)) {
5147 struct address_space *mapping = page_mapping(page);
5149 if (mapping_cap_account_dirty(mapping)) {
5150 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
5151 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
5155 if (PageWriteback(page)) {
5156 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
5157 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
5161 * It is safe to change page->mem_cgroup here because the page
5162 * is referenced, charged, and isolated - we can't race with
5163 * uncharging, charging, migration, or LRU putback.
5166 /* caller should have done css_get */
5167 page->mem_cgroup = to;
5168 spin_unlock_irqrestore(&from->move_lock, flags);
5172 local_irq_disable();
5173 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5174 memcg_check_events(to, page);
5175 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5176 memcg_check_events(from, page);
5185 * get_mctgt_type - get target type of moving charge
5186 * @vma: the vma the pte to be checked belongs
5187 * @addr: the address corresponding to the pte to be checked
5188 * @ptent: the pte to be checked
5189 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5192 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5193 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5194 * move charge. if @target is not NULL, the page is stored in target->page
5195 * with extra refcnt got(Callers should handle it).
5196 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5197 * target for charge migration. if @target is not NULL, the entry is stored
5199 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5200 * (so ZONE_DEVICE page and thus not on the lru).
5201 * For now we such page is charge like a regular page would be as for all
5202 * intent and purposes it is just special memory taking the place of a
5205 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5207 * Called with pte lock held.
5210 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5211 unsigned long addr, pte_t ptent, union mc_target *target)
5213 struct page *page = NULL;
5214 enum mc_target_type ret = MC_TARGET_NONE;
5215 swp_entry_t ent = { .val = 0 };
5217 if (pte_present(ptent))
5218 page = mc_handle_present_pte(vma, addr, ptent);
5219 else if (is_swap_pte(ptent))
5220 page = mc_handle_swap_pte(vma, ptent, &ent);
5221 else if (pte_none(ptent))
5222 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5224 if (!page && !ent.val)
5228 * Do only loose check w/o serialization.
5229 * mem_cgroup_move_account() checks the page is valid or
5230 * not under LRU exclusion.
5232 if (page->mem_cgroup == mc.from) {
5233 ret = MC_TARGET_PAGE;
5234 if (is_device_private_page(page))
5235 ret = MC_TARGET_DEVICE;
5237 target->page = page;
5239 if (!ret || !target)
5243 * There is a swap entry and a page doesn't exist or isn't charged.
5244 * But we cannot move a tail-page in a THP.
5246 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5247 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5248 ret = MC_TARGET_SWAP;
5255 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5257 * We don't consider PMD mapped swapping or file mapped pages because THP does
5258 * not support them for now.
5259 * Caller should make sure that pmd_trans_huge(pmd) is true.
5261 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5262 unsigned long addr, pmd_t pmd, union mc_target *target)
5264 struct page *page = NULL;
5265 enum mc_target_type ret = MC_TARGET_NONE;
5267 if (unlikely(is_swap_pmd(pmd))) {
5268 VM_BUG_ON(thp_migration_supported() &&
5269 !is_pmd_migration_entry(pmd));
5272 page = pmd_page(pmd);
5273 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5274 if (!(mc.flags & MOVE_ANON))
5276 if (page->mem_cgroup == mc.from) {
5277 ret = MC_TARGET_PAGE;
5280 target->page = page;
5286 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5287 unsigned long addr, pmd_t pmd, union mc_target *target)
5289 return MC_TARGET_NONE;
5293 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5294 unsigned long addr, unsigned long end,
5295 struct mm_walk *walk)
5297 struct vm_area_struct *vma = walk->vma;
5301 ptl = pmd_trans_huge_lock(pmd, vma);
5304 * Note their can not be MC_TARGET_DEVICE for now as we do not
5305 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5306 * this might change.
5308 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5309 mc.precharge += HPAGE_PMD_NR;
5314 if (pmd_trans_unstable(pmd))
5316 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5317 for (; addr != end; pte++, addr += PAGE_SIZE)
5318 if (get_mctgt_type(vma, addr, *pte, NULL))
5319 mc.precharge++; /* increment precharge temporarily */
5320 pte_unmap_unlock(pte - 1, ptl);
5326 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5328 unsigned long precharge;
5330 struct mm_walk mem_cgroup_count_precharge_walk = {
5331 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5334 down_read(&mm->mmap_sem);
5335 walk_page_range(0, mm->highest_vm_end,
5336 &mem_cgroup_count_precharge_walk);
5337 up_read(&mm->mmap_sem);
5339 precharge = mc.precharge;
5345 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5347 unsigned long precharge = mem_cgroup_count_precharge(mm);
5349 VM_BUG_ON(mc.moving_task);
5350 mc.moving_task = current;
5351 return mem_cgroup_do_precharge(precharge);
5354 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5355 static void __mem_cgroup_clear_mc(void)
5357 struct mem_cgroup *from = mc.from;
5358 struct mem_cgroup *to = mc.to;
5360 /* we must uncharge all the leftover precharges from mc.to */
5362 cancel_charge(mc.to, mc.precharge);
5366 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5367 * we must uncharge here.
5369 if (mc.moved_charge) {
5370 cancel_charge(mc.from, mc.moved_charge);
5371 mc.moved_charge = 0;
5373 /* we must fixup refcnts and charges */
5374 if (mc.moved_swap) {
5375 /* uncharge swap account from the old cgroup */
5376 if (!mem_cgroup_is_root(mc.from))
5377 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5379 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5382 * we charged both to->memory and to->memsw, so we
5383 * should uncharge to->memory.
5385 if (!mem_cgroup_is_root(mc.to))
5386 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5388 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5389 css_put_many(&mc.to->css, mc.moved_swap);
5393 memcg_oom_recover(from);
5394 memcg_oom_recover(to);
5395 wake_up_all(&mc.waitq);
5398 static void mem_cgroup_clear_mc(void)
5400 struct mm_struct *mm = mc.mm;
5403 * we must clear moving_task before waking up waiters at the end of
5406 mc.moving_task = NULL;
5407 __mem_cgroup_clear_mc();
5408 spin_lock(&mc.lock);
5412 spin_unlock(&mc.lock);
5417 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5419 struct cgroup_subsys_state *css;
5420 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5421 struct mem_cgroup *from;
5422 struct task_struct *leader, *p;
5423 struct mm_struct *mm;
5424 unsigned long move_flags;
5427 /* charge immigration isn't supported on the default hierarchy */
5428 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5432 * Multi-process migrations only happen on the default hierarchy
5433 * where charge immigration is not used. Perform charge
5434 * immigration if @tset contains a leader and whine if there are
5438 cgroup_taskset_for_each_leader(leader, css, tset) {
5441 memcg = mem_cgroup_from_css(css);
5447 * We are now commited to this value whatever it is. Changes in this
5448 * tunable will only affect upcoming migrations, not the current one.
5449 * So we need to save it, and keep it going.
5451 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5455 from = mem_cgroup_from_task(p);
5457 VM_BUG_ON(from == memcg);
5459 mm = get_task_mm(p);
5462 /* We move charges only when we move a owner of the mm */
5463 if (mm->owner == p) {
5466 VM_BUG_ON(mc.precharge);
5467 VM_BUG_ON(mc.moved_charge);
5468 VM_BUG_ON(mc.moved_swap);
5470 spin_lock(&mc.lock);
5474 mc.flags = move_flags;
5475 spin_unlock(&mc.lock);
5476 /* We set mc.moving_task later */
5478 ret = mem_cgroup_precharge_mc(mm);
5480 mem_cgroup_clear_mc();
5487 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5490 mem_cgroup_clear_mc();
5493 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5494 unsigned long addr, unsigned long end,
5495 struct mm_walk *walk)
5498 struct vm_area_struct *vma = walk->vma;
5501 enum mc_target_type target_type;
5502 union mc_target target;
5505 ptl = pmd_trans_huge_lock(pmd, vma);
5507 if (mc.precharge < HPAGE_PMD_NR) {
5511 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5512 if (target_type == MC_TARGET_PAGE) {
5514 if (!isolate_lru_page(page)) {
5515 if (!mem_cgroup_move_account(page, true,
5517 mc.precharge -= HPAGE_PMD_NR;
5518 mc.moved_charge += HPAGE_PMD_NR;
5520 putback_lru_page(page);
5523 } else if (target_type == MC_TARGET_DEVICE) {
5525 if (!mem_cgroup_move_account(page, true,
5527 mc.precharge -= HPAGE_PMD_NR;
5528 mc.moved_charge += HPAGE_PMD_NR;
5536 if (pmd_trans_unstable(pmd))
5539 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5540 for (; addr != end; addr += PAGE_SIZE) {
5541 pte_t ptent = *(pte++);
5542 bool device = false;
5548 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5549 case MC_TARGET_DEVICE:
5552 case MC_TARGET_PAGE:
5555 * We can have a part of the split pmd here. Moving it
5556 * can be done but it would be too convoluted so simply
5557 * ignore such a partial THP and keep it in original
5558 * memcg. There should be somebody mapping the head.
5560 if (PageTransCompound(page))
5562 if (!device && isolate_lru_page(page))
5564 if (!mem_cgroup_move_account(page, false,
5567 /* we uncharge from mc.from later. */
5571 putback_lru_page(page);
5572 put: /* get_mctgt_type() gets the page */
5575 case MC_TARGET_SWAP:
5577 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5579 /* we fixup refcnts and charges later. */
5587 pte_unmap_unlock(pte - 1, ptl);
5592 * We have consumed all precharges we got in can_attach().
5593 * We try charge one by one, but don't do any additional
5594 * charges to mc.to if we have failed in charge once in attach()
5597 ret = mem_cgroup_do_precharge(1);
5605 static void mem_cgroup_move_charge(void)
5607 struct mm_walk mem_cgroup_move_charge_walk = {
5608 .pmd_entry = mem_cgroup_move_charge_pte_range,
5612 lru_add_drain_all();
5614 * Signal lock_page_memcg() to take the memcg's move_lock
5615 * while we're moving its pages to another memcg. Then wait
5616 * for already started RCU-only updates to finish.
5618 atomic_inc(&mc.from->moving_account);
5621 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5623 * Someone who are holding the mmap_sem might be waiting in
5624 * waitq. So we cancel all extra charges, wake up all waiters,
5625 * and retry. Because we cancel precharges, we might not be able
5626 * to move enough charges, but moving charge is a best-effort
5627 * feature anyway, so it wouldn't be a big problem.
5629 __mem_cgroup_clear_mc();
5634 * When we have consumed all precharges and failed in doing
5635 * additional charge, the page walk just aborts.
5637 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5639 up_read(&mc.mm->mmap_sem);
5640 atomic_dec(&mc.from->moving_account);
5643 static void mem_cgroup_move_task(void)
5646 mem_cgroup_move_charge();
5647 mem_cgroup_clear_mc();
5650 #else /* !CONFIG_MMU */
5651 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5655 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5658 static void mem_cgroup_move_task(void)
5664 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5665 * to verify whether we're attached to the default hierarchy on each mount
5668 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5671 * use_hierarchy is forced on the default hierarchy. cgroup core
5672 * guarantees that @root doesn't have any children, so turning it
5673 * on for the root memcg is enough.
5675 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5676 root_mem_cgroup->use_hierarchy = true;
5678 root_mem_cgroup->use_hierarchy = false;
5681 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5683 if (value == PAGE_COUNTER_MAX)
5684 seq_puts(m, "max\n");
5686 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5691 static u64 memory_current_read(struct cgroup_subsys_state *css,
5694 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5696 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5699 static int memory_min_show(struct seq_file *m, void *v)
5701 return seq_puts_memcg_tunable(m,
5702 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5705 static ssize_t memory_min_write(struct kernfs_open_file *of,
5706 char *buf, size_t nbytes, loff_t off)
5708 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5712 buf = strstrip(buf);
5713 err = page_counter_memparse(buf, "max", &min);
5717 page_counter_set_min(&memcg->memory, min);
5722 static int memory_low_show(struct seq_file *m, void *v)
5724 return seq_puts_memcg_tunable(m,
5725 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5728 static ssize_t memory_low_write(struct kernfs_open_file *of,
5729 char *buf, size_t nbytes, loff_t off)
5731 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5735 buf = strstrip(buf);
5736 err = page_counter_memparse(buf, "max", &low);
5740 page_counter_set_low(&memcg->memory, low);
5745 static int memory_high_show(struct seq_file *m, void *v)
5747 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5750 static ssize_t memory_high_write(struct kernfs_open_file *of,
5751 char *buf, size_t nbytes, loff_t off)
5753 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5754 unsigned long nr_pages;
5758 buf = strstrip(buf);
5759 err = page_counter_memparse(buf, "max", &high);
5765 nr_pages = page_counter_read(&memcg->memory);
5766 if (nr_pages > high)
5767 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5770 memcg_wb_domain_size_changed(memcg);
5774 static int memory_max_show(struct seq_file *m, void *v)
5776 return seq_puts_memcg_tunable(m,
5777 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5780 static ssize_t memory_max_write(struct kernfs_open_file *of,
5781 char *buf, size_t nbytes, loff_t off)
5783 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5784 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5785 bool drained = false;
5789 buf = strstrip(buf);
5790 err = page_counter_memparse(buf, "max", &max);
5794 xchg(&memcg->memory.max, max);
5797 unsigned long nr_pages = page_counter_read(&memcg->memory);
5799 if (nr_pages <= max)
5802 if (signal_pending(current)) {
5808 drain_all_stock(memcg);
5814 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5820 memcg_memory_event(memcg, MEMCG_OOM);
5821 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5825 memcg_wb_domain_size_changed(memcg);
5829 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
5831 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
5832 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
5833 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
5834 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
5835 seq_printf(m, "oom_kill %lu\n",
5836 atomic_long_read(&events[MEMCG_OOM_KILL]));
5839 static int memory_events_show(struct seq_file *m, void *v)
5841 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5843 __memory_events_show(m, memcg->memory_events);
5847 static int memory_events_local_show(struct seq_file *m, void *v)
5849 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5851 __memory_events_show(m, memcg->memory_events_local);
5855 static int memory_stat_show(struct seq_file *m, void *v)
5857 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5860 buf = memory_stat_format(memcg);
5868 static int memory_oom_group_show(struct seq_file *m, void *v)
5870 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5872 seq_printf(m, "%d\n", memcg->oom_group);
5877 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5878 char *buf, size_t nbytes, loff_t off)
5880 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5883 buf = strstrip(buf);
5887 ret = kstrtoint(buf, 0, &oom_group);
5891 if (oom_group != 0 && oom_group != 1)
5894 memcg->oom_group = oom_group;
5899 static struct cftype memory_files[] = {
5902 .flags = CFTYPE_NOT_ON_ROOT,
5903 .read_u64 = memory_current_read,
5907 .flags = CFTYPE_NOT_ON_ROOT,
5908 .seq_show = memory_min_show,
5909 .write = memory_min_write,
5913 .flags = CFTYPE_NOT_ON_ROOT,
5914 .seq_show = memory_low_show,
5915 .write = memory_low_write,
5919 .flags = CFTYPE_NOT_ON_ROOT,
5920 .seq_show = memory_high_show,
5921 .write = memory_high_write,
5925 .flags = CFTYPE_NOT_ON_ROOT,
5926 .seq_show = memory_max_show,
5927 .write = memory_max_write,
5931 .flags = CFTYPE_NOT_ON_ROOT,
5932 .file_offset = offsetof(struct mem_cgroup, events_file),
5933 .seq_show = memory_events_show,
5936 .name = "events.local",
5937 .flags = CFTYPE_NOT_ON_ROOT,
5938 .file_offset = offsetof(struct mem_cgroup, events_local_file),
5939 .seq_show = memory_events_local_show,
5943 .flags = CFTYPE_NOT_ON_ROOT,
5944 .seq_show = memory_stat_show,
5947 .name = "oom.group",
5948 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5949 .seq_show = memory_oom_group_show,
5950 .write = memory_oom_group_write,
5955 struct cgroup_subsys memory_cgrp_subsys = {
5956 .css_alloc = mem_cgroup_css_alloc,
5957 .css_online = mem_cgroup_css_online,
5958 .css_offline = mem_cgroup_css_offline,
5959 .css_released = mem_cgroup_css_released,
5960 .css_free = mem_cgroup_css_free,
5961 .css_reset = mem_cgroup_css_reset,
5962 .can_attach = mem_cgroup_can_attach,
5963 .cancel_attach = mem_cgroup_cancel_attach,
5964 .post_attach = mem_cgroup_move_task,
5965 .bind = mem_cgroup_bind,
5966 .dfl_cftypes = memory_files,
5967 .legacy_cftypes = mem_cgroup_legacy_files,
5972 * mem_cgroup_protected - check if memory consumption is in the normal range
5973 * @root: the top ancestor of the sub-tree being checked
5974 * @memcg: the memory cgroup to check
5976 * WARNING: This function is not stateless! It can only be used as part
5977 * of a top-down tree iteration, not for isolated queries.
5979 * Returns one of the following:
5980 * MEMCG_PROT_NONE: cgroup memory is not protected
5981 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5982 * an unprotected supply of reclaimable memory from other cgroups.
5983 * MEMCG_PROT_MIN: cgroup memory is protected
5985 * @root is exclusive; it is never protected when looked at directly
5987 * To provide a proper hierarchical behavior, effective memory.min/low values
5988 * are used. Below is the description of how effective memory.low is calculated.
5989 * Effective memory.min values is calculated in the same way.
5991 * Effective memory.low is always equal or less than the original memory.low.
5992 * If there is no memory.low overcommittment (which is always true for
5993 * top-level memory cgroups), these two values are equal.
5994 * Otherwise, it's a part of parent's effective memory.low,
5995 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5996 * memory.low usages, where memory.low usage is the size of actually
6000 * elow = min( memory.low, parent->elow * ------------------ ),
6001 * siblings_low_usage
6003 * | memory.current, if memory.current < memory.low
6008 * Such definition of the effective memory.low provides the expected
6009 * hierarchical behavior: parent's memory.low value is limiting
6010 * children, unprotected memory is reclaimed first and cgroups,
6011 * which are not using their guarantee do not affect actual memory
6014 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6016 * A A/memory.low = 2G, A/memory.current = 6G
6018 * BC DE B/memory.low = 3G B/memory.current = 2G
6019 * C/memory.low = 1G C/memory.current = 2G
6020 * D/memory.low = 0 D/memory.current = 2G
6021 * E/memory.low = 10G E/memory.current = 0
6023 * and the memory pressure is applied, the following memory distribution
6024 * is expected (approximately):
6026 * A/memory.current = 2G
6028 * B/memory.current = 1.3G
6029 * C/memory.current = 0.6G
6030 * D/memory.current = 0
6031 * E/memory.current = 0
6033 * These calculations require constant tracking of the actual low usages
6034 * (see propagate_protected_usage()), as well as recursive calculation of
6035 * effective memory.low values. But as we do call mem_cgroup_protected()
6036 * path for each memory cgroup top-down from the reclaim,
6037 * it's possible to optimize this part, and save calculated elow
6038 * for next usage. This part is intentionally racy, but it's ok,
6039 * as memory.low is a best-effort mechanism.
6041 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6042 struct mem_cgroup *memcg)
6044 struct mem_cgroup *parent;
6045 unsigned long emin, parent_emin;
6046 unsigned long elow, parent_elow;
6047 unsigned long usage;
6049 if (mem_cgroup_disabled())
6050 return MEMCG_PROT_NONE;
6053 root = root_mem_cgroup;
6055 return MEMCG_PROT_NONE;
6057 usage = page_counter_read(&memcg->memory);
6059 return MEMCG_PROT_NONE;
6061 emin = memcg->memory.min;
6062 elow = memcg->memory.low;
6064 parent = parent_mem_cgroup(memcg);
6065 /* No parent means a non-hierarchical mode on v1 memcg */
6067 return MEMCG_PROT_NONE;
6072 parent_emin = READ_ONCE(parent->memory.emin);
6073 emin = min(emin, parent_emin);
6074 if (emin && parent_emin) {
6075 unsigned long min_usage, siblings_min_usage;
6077 min_usage = min(usage, memcg->memory.min);
6078 siblings_min_usage = atomic_long_read(
6079 &parent->memory.children_min_usage);
6081 if (min_usage && siblings_min_usage)
6082 emin = min(emin, parent_emin * min_usage /
6083 siblings_min_usage);
6086 parent_elow = READ_ONCE(parent->memory.elow);
6087 elow = min(elow, parent_elow);
6088 if (elow && parent_elow) {
6089 unsigned long low_usage, siblings_low_usage;
6091 low_usage = min(usage, memcg->memory.low);
6092 siblings_low_usage = atomic_long_read(
6093 &parent->memory.children_low_usage);
6095 if (low_usage && siblings_low_usage)
6096 elow = min(elow, parent_elow * low_usage /
6097 siblings_low_usage);
6101 memcg->memory.emin = emin;
6102 memcg->memory.elow = elow;
6105 return MEMCG_PROT_MIN;
6106 else if (usage <= elow)
6107 return MEMCG_PROT_LOW;
6109 return MEMCG_PROT_NONE;
6113 * mem_cgroup_try_charge - try charging a page
6114 * @page: page to charge
6115 * @mm: mm context of the victim
6116 * @gfp_mask: reclaim mode
6117 * @memcgp: charged memcg return
6118 * @compound: charge the page as compound or small page
6120 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6121 * pages according to @gfp_mask if necessary.
6123 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6124 * Otherwise, an error code is returned.
6126 * After page->mapping has been set up, the caller must finalize the
6127 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6128 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6130 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6131 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6134 struct mem_cgroup *memcg = NULL;
6135 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6138 if (mem_cgroup_disabled())
6141 if (PageSwapCache(page)) {
6143 * Every swap fault against a single page tries to charge the
6144 * page, bail as early as possible. shmem_unuse() encounters
6145 * already charged pages, too. The USED bit is protected by
6146 * the page lock, which serializes swap cache removal, which
6147 * in turn serializes uncharging.
6149 VM_BUG_ON_PAGE(!PageLocked(page), page);
6150 if (compound_head(page)->mem_cgroup)
6153 if (do_swap_account) {
6154 swp_entry_t ent = { .val = page_private(page), };
6155 unsigned short id = lookup_swap_cgroup_id(ent);
6158 memcg = mem_cgroup_from_id(id);
6159 if (memcg && !css_tryget_online(&memcg->css))
6166 memcg = get_mem_cgroup_from_mm(mm);
6168 ret = try_charge(memcg, gfp_mask, nr_pages);
6170 css_put(&memcg->css);
6176 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6177 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6180 struct mem_cgroup *memcg;
6183 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6185 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6190 * mem_cgroup_commit_charge - commit a page charge
6191 * @page: page to charge
6192 * @memcg: memcg to charge the page to
6193 * @lrucare: page might be on LRU already
6194 * @compound: charge the page as compound or small page
6196 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6197 * after page->mapping has been set up. This must happen atomically
6198 * as part of the page instantiation, i.e. under the page table lock
6199 * for anonymous pages, under the page lock for page and swap cache.
6201 * In addition, the page must not be on the LRU during the commit, to
6202 * prevent racing with task migration. If it might be, use @lrucare.
6204 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6206 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6207 bool lrucare, bool compound)
6209 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6211 VM_BUG_ON_PAGE(!page->mapping, page);
6212 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6214 if (mem_cgroup_disabled())
6217 * Swap faults will attempt to charge the same page multiple
6218 * times. But reuse_swap_page() might have removed the page
6219 * from swapcache already, so we can't check PageSwapCache().
6224 commit_charge(page, memcg, lrucare);
6226 local_irq_disable();
6227 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6228 memcg_check_events(memcg, page);
6231 if (do_memsw_account() && PageSwapCache(page)) {
6232 swp_entry_t entry = { .val = page_private(page) };
6234 * The swap entry might not get freed for a long time,
6235 * let's not wait for it. The page already received a
6236 * memory+swap charge, drop the swap entry duplicate.
6238 mem_cgroup_uncharge_swap(entry, nr_pages);
6243 * mem_cgroup_cancel_charge - cancel a page charge
6244 * @page: page to charge
6245 * @memcg: memcg to charge the page to
6246 * @compound: charge the page as compound or small page
6248 * Cancel a charge transaction started by mem_cgroup_try_charge().
6250 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6253 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6255 if (mem_cgroup_disabled())
6258 * Swap faults will attempt to charge the same page multiple
6259 * times. But reuse_swap_page() might have removed the page
6260 * from swapcache already, so we can't check PageSwapCache().
6265 cancel_charge(memcg, nr_pages);
6268 struct uncharge_gather {
6269 struct mem_cgroup *memcg;
6270 unsigned long pgpgout;
6271 unsigned long nr_anon;
6272 unsigned long nr_file;
6273 unsigned long nr_kmem;
6274 unsigned long nr_huge;
6275 unsigned long nr_shmem;
6276 struct page *dummy_page;
6279 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6281 memset(ug, 0, sizeof(*ug));
6284 static void uncharge_batch(const struct uncharge_gather *ug)
6286 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6287 unsigned long flags;
6289 if (!mem_cgroup_is_root(ug->memcg)) {
6290 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6291 if (do_memsw_account())
6292 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6293 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6294 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6295 memcg_oom_recover(ug->memcg);
6298 local_irq_save(flags);
6299 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6300 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6301 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6302 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6303 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6304 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6305 memcg_check_events(ug->memcg, ug->dummy_page);
6306 local_irq_restore(flags);
6308 if (!mem_cgroup_is_root(ug->memcg))
6309 css_put_many(&ug->memcg->css, nr_pages);
6312 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6314 VM_BUG_ON_PAGE(PageLRU(page), page);
6315 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6316 !PageHWPoison(page) , page);
6318 if (!page->mem_cgroup)
6322 * Nobody should be changing or seriously looking at
6323 * page->mem_cgroup at this point, we have fully
6324 * exclusive access to the page.
6327 if (ug->memcg != page->mem_cgroup) {
6330 uncharge_gather_clear(ug);
6332 ug->memcg = page->mem_cgroup;
6335 if (!PageKmemcg(page)) {
6336 unsigned int nr_pages = 1;
6338 if (PageTransHuge(page)) {
6339 nr_pages <<= compound_order(page);
6340 ug->nr_huge += nr_pages;
6343 ug->nr_anon += nr_pages;
6345 ug->nr_file += nr_pages;
6346 if (PageSwapBacked(page))
6347 ug->nr_shmem += nr_pages;
6351 ug->nr_kmem += 1 << compound_order(page);
6352 __ClearPageKmemcg(page);
6355 ug->dummy_page = page;
6356 page->mem_cgroup = NULL;
6359 static void uncharge_list(struct list_head *page_list)
6361 struct uncharge_gather ug;
6362 struct list_head *next;
6364 uncharge_gather_clear(&ug);
6367 * Note that the list can be a single page->lru; hence the
6368 * do-while loop instead of a simple list_for_each_entry().
6370 next = page_list->next;
6374 page = list_entry(next, struct page, lru);
6375 next = page->lru.next;
6377 uncharge_page(page, &ug);
6378 } while (next != page_list);
6381 uncharge_batch(&ug);
6385 * mem_cgroup_uncharge - uncharge a page
6386 * @page: page to uncharge
6388 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6389 * mem_cgroup_commit_charge().
6391 void mem_cgroup_uncharge(struct page *page)
6393 struct uncharge_gather ug;
6395 if (mem_cgroup_disabled())
6398 /* Don't touch page->lru of any random page, pre-check: */
6399 if (!page->mem_cgroup)
6402 uncharge_gather_clear(&ug);
6403 uncharge_page(page, &ug);
6404 uncharge_batch(&ug);
6408 * mem_cgroup_uncharge_list - uncharge a list of page
6409 * @page_list: list of pages to uncharge
6411 * Uncharge a list of pages previously charged with
6412 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6414 void mem_cgroup_uncharge_list(struct list_head *page_list)
6416 if (mem_cgroup_disabled())
6419 if (!list_empty(page_list))
6420 uncharge_list(page_list);
6424 * mem_cgroup_migrate - charge a page's replacement
6425 * @oldpage: currently circulating page
6426 * @newpage: replacement page
6428 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6429 * be uncharged upon free.
6431 * Both pages must be locked, @newpage->mapping must be set up.
6433 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6435 struct mem_cgroup *memcg;
6436 unsigned int nr_pages;
6438 unsigned long flags;
6440 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6441 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6442 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6443 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6446 if (mem_cgroup_disabled())
6449 /* Page cache replacement: new page already charged? */
6450 if (newpage->mem_cgroup)
6453 /* Swapcache readahead pages can get replaced before being charged */
6454 memcg = oldpage->mem_cgroup;
6458 /* Force-charge the new page. The old one will be freed soon */
6459 compound = PageTransHuge(newpage);
6460 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6462 page_counter_charge(&memcg->memory, nr_pages);
6463 if (do_memsw_account())
6464 page_counter_charge(&memcg->memsw, nr_pages);
6465 css_get_many(&memcg->css, nr_pages);
6467 commit_charge(newpage, memcg, false);
6469 local_irq_save(flags);
6470 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6471 memcg_check_events(memcg, newpage);
6472 local_irq_restore(flags);
6475 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6476 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6478 void mem_cgroup_sk_alloc(struct sock *sk)
6480 struct mem_cgroup *memcg;
6482 if (!mem_cgroup_sockets_enabled)
6486 * Socket cloning can throw us here with sk_memcg already
6487 * filled. It won't however, necessarily happen from
6488 * process context. So the test for root memcg given
6489 * the current task's memcg won't help us in this case.
6491 * Respecting the original socket's memcg is a better
6492 * decision in this case.
6495 css_get(&sk->sk_memcg->css);
6500 memcg = mem_cgroup_from_task(current);
6501 if (memcg == root_mem_cgroup)
6503 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6505 if (css_tryget_online(&memcg->css))
6506 sk->sk_memcg = memcg;
6511 void mem_cgroup_sk_free(struct sock *sk)
6514 css_put(&sk->sk_memcg->css);
6518 * mem_cgroup_charge_skmem - charge socket memory
6519 * @memcg: memcg to charge
6520 * @nr_pages: number of pages to charge
6522 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6523 * @memcg's configured limit, %false if the charge had to be forced.
6525 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6527 gfp_t gfp_mask = GFP_KERNEL;
6529 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6530 struct page_counter *fail;
6532 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6533 memcg->tcpmem_pressure = 0;
6536 page_counter_charge(&memcg->tcpmem, nr_pages);
6537 memcg->tcpmem_pressure = 1;
6541 /* Don't block in the packet receive path */
6543 gfp_mask = GFP_NOWAIT;
6545 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6547 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6550 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6555 * mem_cgroup_uncharge_skmem - uncharge socket memory
6556 * @memcg: memcg to uncharge
6557 * @nr_pages: number of pages to uncharge
6559 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6561 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6562 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6566 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6568 refill_stock(memcg, nr_pages);
6571 static int __init cgroup_memory(char *s)
6575 while ((token = strsep(&s, ",")) != NULL) {
6578 if (!strcmp(token, "nosocket"))
6579 cgroup_memory_nosocket = true;
6580 if (!strcmp(token, "nokmem"))
6581 cgroup_memory_nokmem = true;
6585 __setup("cgroup.memory=", cgroup_memory);
6588 * subsys_initcall() for memory controller.
6590 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6591 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6592 * basically everything that doesn't depend on a specific mem_cgroup structure
6593 * should be initialized from here.
6595 static int __init mem_cgroup_init(void)
6599 #ifdef CONFIG_MEMCG_KMEM
6601 * Kmem cache creation is mostly done with the slab_mutex held,
6602 * so use a workqueue with limited concurrency to avoid stalling
6603 * all worker threads in case lots of cgroups are created and
6604 * destroyed simultaneously.
6606 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6607 BUG_ON(!memcg_kmem_cache_wq);
6610 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6611 memcg_hotplug_cpu_dead);
6613 for_each_possible_cpu(cpu)
6614 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6617 for_each_node(node) {
6618 struct mem_cgroup_tree_per_node *rtpn;
6620 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6621 node_online(node) ? node : NUMA_NO_NODE);
6623 rtpn->rb_root = RB_ROOT;
6624 rtpn->rb_rightmost = NULL;
6625 spin_lock_init(&rtpn->lock);
6626 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6631 subsys_initcall(mem_cgroup_init);
6633 #ifdef CONFIG_MEMCG_SWAP
6634 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6636 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6638 * The root cgroup cannot be destroyed, so it's refcount must
6641 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6645 memcg = parent_mem_cgroup(memcg);
6647 memcg = root_mem_cgroup;
6653 * mem_cgroup_swapout - transfer a memsw charge to swap
6654 * @page: page whose memsw charge to transfer
6655 * @entry: swap entry to move the charge to
6657 * Transfer the memsw charge of @page to @entry.
6659 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6661 struct mem_cgroup *memcg, *swap_memcg;
6662 unsigned int nr_entries;
6663 unsigned short oldid;
6665 VM_BUG_ON_PAGE(PageLRU(page), page);
6666 VM_BUG_ON_PAGE(page_count(page), page);
6668 if (!do_memsw_account())
6671 memcg = page->mem_cgroup;
6673 /* Readahead page, never charged */
6678 * In case the memcg owning these pages has been offlined and doesn't
6679 * have an ID allocated to it anymore, charge the closest online
6680 * ancestor for the swap instead and transfer the memory+swap charge.
6682 swap_memcg = mem_cgroup_id_get_online(memcg);
6683 nr_entries = hpage_nr_pages(page);
6684 /* Get references for the tail pages, too */
6686 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6687 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6689 VM_BUG_ON_PAGE(oldid, page);
6690 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6692 page->mem_cgroup = NULL;
6694 if (!mem_cgroup_is_root(memcg))
6695 page_counter_uncharge(&memcg->memory, nr_entries);
6697 if (memcg != swap_memcg) {
6698 if (!mem_cgroup_is_root(swap_memcg))
6699 page_counter_charge(&swap_memcg->memsw, nr_entries);
6700 page_counter_uncharge(&memcg->memsw, nr_entries);
6704 * Interrupts should be disabled here because the caller holds the
6705 * i_pages lock which is taken with interrupts-off. It is
6706 * important here to have the interrupts disabled because it is the
6707 * only synchronisation we have for updating the per-CPU variables.
6709 VM_BUG_ON(!irqs_disabled());
6710 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6712 memcg_check_events(memcg, page);
6714 if (!mem_cgroup_is_root(memcg))
6715 css_put_many(&memcg->css, nr_entries);
6719 * mem_cgroup_try_charge_swap - try charging swap space for a page
6720 * @page: page being added to swap
6721 * @entry: swap entry to charge
6723 * Try to charge @page's memcg for the swap space at @entry.
6725 * Returns 0 on success, -ENOMEM on failure.
6727 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6729 unsigned int nr_pages = hpage_nr_pages(page);
6730 struct page_counter *counter;
6731 struct mem_cgroup *memcg;
6732 unsigned short oldid;
6734 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6737 memcg = page->mem_cgroup;
6739 /* Readahead page, never charged */
6744 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6748 memcg = mem_cgroup_id_get_online(memcg);
6750 if (!mem_cgroup_is_root(memcg) &&
6751 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6752 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6753 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6754 mem_cgroup_id_put(memcg);
6758 /* Get references for the tail pages, too */
6760 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6761 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6762 VM_BUG_ON_PAGE(oldid, page);
6763 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6769 * mem_cgroup_uncharge_swap - uncharge swap space
6770 * @entry: swap entry to uncharge
6771 * @nr_pages: the amount of swap space to uncharge
6773 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6775 struct mem_cgroup *memcg;
6778 if (!do_swap_account)
6781 id = swap_cgroup_record(entry, 0, nr_pages);
6783 memcg = mem_cgroup_from_id(id);
6785 if (!mem_cgroup_is_root(memcg)) {
6786 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6787 page_counter_uncharge(&memcg->swap, nr_pages);
6789 page_counter_uncharge(&memcg->memsw, nr_pages);
6791 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6792 mem_cgroup_id_put_many(memcg, nr_pages);
6797 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6799 long nr_swap_pages = get_nr_swap_pages();
6801 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6802 return nr_swap_pages;
6803 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6804 nr_swap_pages = min_t(long, nr_swap_pages,
6805 READ_ONCE(memcg->swap.max) -
6806 page_counter_read(&memcg->swap));
6807 return nr_swap_pages;
6810 bool mem_cgroup_swap_full(struct page *page)
6812 struct mem_cgroup *memcg;
6814 VM_BUG_ON_PAGE(!PageLocked(page), page);
6818 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6821 memcg = page->mem_cgroup;
6825 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6826 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6832 /* for remember boot option*/
6833 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6834 static int really_do_swap_account __initdata = 1;
6836 static int really_do_swap_account __initdata;
6839 static int __init enable_swap_account(char *s)
6841 if (!strcmp(s, "1"))
6842 really_do_swap_account = 1;
6843 else if (!strcmp(s, "0"))
6844 really_do_swap_account = 0;
6847 __setup("swapaccount=", enable_swap_account);
6849 static u64 swap_current_read(struct cgroup_subsys_state *css,
6852 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6854 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6857 static int swap_max_show(struct seq_file *m, void *v)
6859 return seq_puts_memcg_tunable(m,
6860 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6863 static ssize_t swap_max_write(struct kernfs_open_file *of,
6864 char *buf, size_t nbytes, loff_t off)
6866 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6870 buf = strstrip(buf);
6871 err = page_counter_memparse(buf, "max", &max);
6875 xchg(&memcg->swap.max, max);
6880 static int swap_events_show(struct seq_file *m, void *v)
6882 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6884 seq_printf(m, "max %lu\n",
6885 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6886 seq_printf(m, "fail %lu\n",
6887 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6892 static struct cftype swap_files[] = {
6894 .name = "swap.current",
6895 .flags = CFTYPE_NOT_ON_ROOT,
6896 .read_u64 = swap_current_read,
6900 .flags = CFTYPE_NOT_ON_ROOT,
6901 .seq_show = swap_max_show,
6902 .write = swap_max_write,
6905 .name = "swap.events",
6906 .flags = CFTYPE_NOT_ON_ROOT,
6907 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6908 .seq_show = swap_events_show,
6913 static struct cftype memsw_cgroup_files[] = {
6915 .name = "memsw.usage_in_bytes",
6916 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6917 .read_u64 = mem_cgroup_read_u64,
6920 .name = "memsw.max_usage_in_bytes",
6921 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6922 .write = mem_cgroup_reset,
6923 .read_u64 = mem_cgroup_read_u64,
6926 .name = "memsw.limit_in_bytes",
6927 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6928 .write = mem_cgroup_write,
6929 .read_u64 = mem_cgroup_read_u64,
6932 .name = "memsw.failcnt",
6933 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6934 .write = mem_cgroup_reset,
6935 .read_u64 = mem_cgroup_read_u64,
6937 { }, /* terminate */
6940 static int __init mem_cgroup_swap_init(void)
6942 if (!mem_cgroup_disabled() && really_do_swap_account) {
6943 do_swap_account = 1;
6944 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6946 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6947 memsw_cgroup_files));
6951 subsys_initcall(mem_cgroup_swap_init);
6953 #endif /* CONFIG_MEMCG_SWAP */