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 __this_cpu_add(memcg->vmstats_local->stat[idx], val);
700 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
701 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
702 struct mem_cgroup *mi;
704 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
705 atomic_long_add(x, &mi->vmstats[idx]);
708 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
711 static struct mem_cgroup_per_node *
712 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
714 struct mem_cgroup *parent;
716 parent = parent_mem_cgroup(pn->memcg);
719 return mem_cgroup_nodeinfo(parent, nid);
723 * __mod_lruvec_state - update lruvec memory statistics
724 * @lruvec: the lruvec
725 * @idx: the stat item
726 * @val: delta to add to the counter, can be negative
728 * The lruvec is the intersection of the NUMA node and a cgroup. This
729 * function updates the all three counters that are affected by a
730 * change of state at this level: per-node, per-cgroup, per-lruvec.
732 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
735 pg_data_t *pgdat = lruvec_pgdat(lruvec);
736 struct mem_cgroup_per_node *pn;
737 struct mem_cgroup *memcg;
741 __mod_node_page_state(pgdat, idx, val);
743 if (mem_cgroup_disabled())
746 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
750 __mod_memcg_state(memcg, idx, val);
753 __this_cpu_add(pn->lruvec_stat_local->count[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;
759 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
760 atomic_long_add(x, &pi->lruvec_stat[idx]);
763 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
767 * __count_memcg_events - account VM events in a cgroup
768 * @memcg: the memory cgroup
769 * @idx: the event item
770 * @count: the number of events that occured
772 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
777 if (mem_cgroup_disabled())
780 __this_cpu_add(memcg->vmstats_local->events[idx], count);
782 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
783 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
784 struct mem_cgroup *mi;
786 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
787 atomic_long_add(x, &mi->vmevents[idx]);
790 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
793 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
795 return atomic_long_read(&memcg->vmevents[event]);
798 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
803 for_each_possible_cpu(cpu)
804 x += per_cpu(memcg->vmstats_local->events[event], cpu);
808 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
810 bool compound, int nr_pages)
813 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
814 * counted as CACHE even if it's on ANON LRU.
817 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
819 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
820 if (PageSwapBacked(page))
821 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
825 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
826 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
829 /* pagein of a big page is an event. So, ignore page size */
831 __count_memcg_events(memcg, PGPGIN, 1);
833 __count_memcg_events(memcg, PGPGOUT, 1);
834 nr_pages = -nr_pages; /* for event */
837 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
840 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
841 enum mem_cgroup_events_target target)
843 unsigned long val, next;
845 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
846 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
847 /* from time_after() in jiffies.h */
848 if ((long)(next - val) < 0) {
850 case MEM_CGROUP_TARGET_THRESH:
851 next = val + THRESHOLDS_EVENTS_TARGET;
853 case MEM_CGROUP_TARGET_SOFTLIMIT:
854 next = val + SOFTLIMIT_EVENTS_TARGET;
856 case MEM_CGROUP_TARGET_NUMAINFO:
857 next = val + NUMAINFO_EVENTS_TARGET;
862 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
869 * Check events in order.
872 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
874 /* threshold event is triggered in finer grain than soft limit */
875 if (unlikely(mem_cgroup_event_ratelimit(memcg,
876 MEM_CGROUP_TARGET_THRESH))) {
878 bool do_numainfo __maybe_unused;
880 do_softlimit = mem_cgroup_event_ratelimit(memcg,
881 MEM_CGROUP_TARGET_SOFTLIMIT);
883 do_numainfo = mem_cgroup_event_ratelimit(memcg,
884 MEM_CGROUP_TARGET_NUMAINFO);
886 mem_cgroup_threshold(memcg);
887 if (unlikely(do_softlimit))
888 mem_cgroup_update_tree(memcg, page);
890 if (unlikely(do_numainfo))
891 atomic_inc(&memcg->numainfo_events);
896 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
899 * mm_update_next_owner() may clear mm->owner to NULL
900 * if it races with swapoff, page migration, etc.
901 * So this can be called with p == NULL.
906 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
908 EXPORT_SYMBOL(mem_cgroup_from_task);
911 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
912 * @mm: mm from which memcg should be extracted. It can be NULL.
914 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
915 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
918 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
920 struct mem_cgroup *memcg;
922 if (mem_cgroup_disabled())
928 * Page cache insertions can happen withou an
929 * actual mm context, e.g. during disk probing
930 * on boot, loopback IO, acct() writes etc.
933 memcg = root_mem_cgroup;
935 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
936 if (unlikely(!memcg))
937 memcg = root_mem_cgroup;
939 } while (!css_tryget_online(&memcg->css));
943 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
946 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
947 * @page: page from which memcg should be extracted.
949 * Obtain a reference on page->memcg and returns it if successful. Otherwise
950 * root_mem_cgroup is returned.
952 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
954 struct mem_cgroup *memcg = page->mem_cgroup;
956 if (mem_cgroup_disabled())
960 if (!memcg || !css_tryget_online(&memcg->css))
961 memcg = root_mem_cgroup;
965 EXPORT_SYMBOL(get_mem_cgroup_from_page);
968 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
970 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
972 if (unlikely(current->active_memcg)) {
973 struct mem_cgroup *memcg = root_mem_cgroup;
976 if (css_tryget_online(¤t->active_memcg->css))
977 memcg = current->active_memcg;
981 return get_mem_cgroup_from_mm(current->mm);
985 * mem_cgroup_iter - iterate over memory cgroup hierarchy
986 * @root: hierarchy root
987 * @prev: previously returned memcg, NULL on first invocation
988 * @reclaim: cookie for shared reclaim walks, NULL for full walks
990 * Returns references to children of the hierarchy below @root, or
991 * @root itself, or %NULL after a full round-trip.
993 * Caller must pass the return value in @prev on subsequent
994 * invocations for reference counting, or use mem_cgroup_iter_break()
995 * to cancel a hierarchy walk before the round-trip is complete.
997 * Reclaimers can specify a node and a priority level in @reclaim to
998 * divide up the memcgs in the hierarchy among all concurrent
999 * reclaimers operating on the same node and priority.
1001 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1002 struct mem_cgroup *prev,
1003 struct mem_cgroup_reclaim_cookie *reclaim)
1005 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1006 struct cgroup_subsys_state *css = NULL;
1007 struct mem_cgroup *memcg = NULL;
1008 struct mem_cgroup *pos = NULL;
1010 if (mem_cgroup_disabled())
1014 root = root_mem_cgroup;
1016 if (prev && !reclaim)
1019 if (!root->use_hierarchy && root != root_mem_cgroup) {
1028 struct mem_cgroup_per_node *mz;
1030 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1031 iter = &mz->iter[reclaim->priority];
1033 if (prev && reclaim->generation != iter->generation)
1037 pos = READ_ONCE(iter->position);
1038 if (!pos || css_tryget(&pos->css))
1041 * css reference reached zero, so iter->position will
1042 * be cleared by ->css_released. However, we should not
1043 * rely on this happening soon, because ->css_released
1044 * is called from a work queue, and by busy-waiting we
1045 * might block it. So we clear iter->position right
1048 (void)cmpxchg(&iter->position, pos, NULL);
1056 css = css_next_descendant_pre(css, &root->css);
1059 * Reclaimers share the hierarchy walk, and a
1060 * new one might jump in right at the end of
1061 * the hierarchy - make sure they see at least
1062 * one group and restart from the beginning.
1070 * Verify the css and acquire a reference. The root
1071 * is provided by the caller, so we know it's alive
1072 * and kicking, and don't take an extra reference.
1074 memcg = mem_cgroup_from_css(css);
1076 if (css == &root->css)
1079 if (css_tryget(css))
1087 * The position could have already been updated by a competing
1088 * thread, so check that the value hasn't changed since we read
1089 * it to avoid reclaiming from the same cgroup twice.
1091 (void)cmpxchg(&iter->position, pos, memcg);
1099 reclaim->generation = iter->generation;
1105 if (prev && prev != root)
1106 css_put(&prev->css);
1112 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1113 * @root: hierarchy root
1114 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1116 void mem_cgroup_iter_break(struct mem_cgroup *root,
1117 struct mem_cgroup *prev)
1120 root = root_mem_cgroup;
1121 if (prev && prev != root)
1122 css_put(&prev->css);
1125 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1127 struct mem_cgroup *memcg = dead_memcg;
1128 struct mem_cgroup_reclaim_iter *iter;
1129 struct mem_cgroup_per_node *mz;
1133 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1134 for_each_node(nid) {
1135 mz = mem_cgroup_nodeinfo(memcg, nid);
1136 for (i = 0; i <= DEF_PRIORITY; i++) {
1137 iter = &mz->iter[i];
1138 cmpxchg(&iter->position,
1146 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1147 * @memcg: hierarchy root
1148 * @fn: function to call for each task
1149 * @arg: argument passed to @fn
1151 * This function iterates over tasks attached to @memcg or to any of its
1152 * descendants and calls @fn for each task. If @fn returns a non-zero
1153 * value, the function breaks the iteration loop and returns the value.
1154 * Otherwise, it will iterate over all tasks and return 0.
1156 * This function must not be called for the root memory cgroup.
1158 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1159 int (*fn)(struct task_struct *, void *), void *arg)
1161 struct mem_cgroup *iter;
1164 BUG_ON(memcg == root_mem_cgroup);
1166 for_each_mem_cgroup_tree(iter, memcg) {
1167 struct css_task_iter it;
1168 struct task_struct *task;
1170 css_task_iter_start(&iter->css, 0, &it);
1171 while (!ret && (task = css_task_iter_next(&it)))
1172 ret = fn(task, arg);
1173 css_task_iter_end(&it);
1175 mem_cgroup_iter_break(memcg, iter);
1183 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1185 * @pgdat: pgdat of the page
1187 * This function is only safe when following the LRU page isolation
1188 * and putback protocol: the LRU lock must be held, and the page must
1189 * either be PageLRU() or the caller must have isolated/allocated it.
1191 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1193 struct mem_cgroup_per_node *mz;
1194 struct mem_cgroup *memcg;
1195 struct lruvec *lruvec;
1197 if (mem_cgroup_disabled()) {
1198 lruvec = &pgdat->lruvec;
1202 memcg = page->mem_cgroup;
1204 * Swapcache readahead pages are added to the LRU - and
1205 * possibly migrated - before they are charged.
1208 memcg = root_mem_cgroup;
1210 mz = mem_cgroup_page_nodeinfo(memcg, page);
1211 lruvec = &mz->lruvec;
1214 * Since a node can be onlined after the mem_cgroup was created,
1215 * we have to be prepared to initialize lruvec->zone here;
1216 * and if offlined then reonlined, we need to reinitialize it.
1218 if (unlikely(lruvec->pgdat != pgdat))
1219 lruvec->pgdat = pgdat;
1224 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1225 * @lruvec: mem_cgroup per zone lru vector
1226 * @lru: index of lru list the page is sitting on
1227 * @zid: zone id of the accounted pages
1228 * @nr_pages: positive when adding or negative when removing
1230 * This function must be called under lru_lock, just before a page is added
1231 * to or just after a page is removed from an lru list (that ordering being
1232 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1234 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1235 int zid, int nr_pages)
1237 struct mem_cgroup_per_node *mz;
1238 unsigned long *lru_size;
1241 if (mem_cgroup_disabled())
1244 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1245 lru_size = &mz->lru_zone_size[zid][lru];
1248 *lru_size += nr_pages;
1251 if (WARN_ONCE(size < 0,
1252 "%s(%p, %d, %d): lru_size %ld\n",
1253 __func__, lruvec, lru, nr_pages, size)) {
1259 *lru_size += nr_pages;
1262 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1264 struct mem_cgroup *task_memcg;
1265 struct task_struct *p;
1268 p = find_lock_task_mm(task);
1270 task_memcg = get_mem_cgroup_from_mm(p->mm);
1274 * All threads may have already detached their mm's, but the oom
1275 * killer still needs to detect if they have already been oom
1276 * killed to prevent needlessly killing additional tasks.
1279 task_memcg = mem_cgroup_from_task(task);
1280 css_get(&task_memcg->css);
1283 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1284 css_put(&task_memcg->css);
1289 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1290 * @memcg: the memory cgroup
1292 * Returns the maximum amount of memory @mem can be charged with, in
1295 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1297 unsigned long margin = 0;
1298 unsigned long count;
1299 unsigned long limit;
1301 count = page_counter_read(&memcg->memory);
1302 limit = READ_ONCE(memcg->memory.max);
1304 margin = limit - count;
1306 if (do_memsw_account()) {
1307 count = page_counter_read(&memcg->memsw);
1308 limit = READ_ONCE(memcg->memsw.max);
1310 margin = min(margin, limit - count);
1319 * A routine for checking "mem" is under move_account() or not.
1321 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1322 * moving cgroups. This is for waiting at high-memory pressure
1325 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1327 struct mem_cgroup *from;
1328 struct mem_cgroup *to;
1331 * Unlike task_move routines, we access mc.to, mc.from not under
1332 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1334 spin_lock(&mc.lock);
1340 ret = mem_cgroup_is_descendant(from, memcg) ||
1341 mem_cgroup_is_descendant(to, memcg);
1343 spin_unlock(&mc.lock);
1347 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1349 if (mc.moving_task && current != mc.moving_task) {
1350 if (mem_cgroup_under_move(memcg)) {
1352 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1353 /* moving charge context might have finished. */
1356 finish_wait(&mc.waitq, &wait);
1363 static char *memory_stat_format(struct mem_cgroup *memcg)
1368 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1373 * Provide statistics on the state of the memory subsystem as
1374 * well as cumulative event counters that show past behavior.
1376 * This list is ordered following a combination of these gradients:
1377 * 1) generic big picture -> specifics and details
1378 * 2) reflecting userspace activity -> reflecting kernel heuristics
1380 * Current memory state:
1383 seq_buf_printf(&s, "anon %llu\n",
1384 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1386 seq_buf_printf(&s, "file %llu\n",
1387 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1389 seq_buf_printf(&s, "kernel_stack %llu\n",
1390 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1392 seq_buf_printf(&s, "slab %llu\n",
1393 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1394 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1396 seq_buf_printf(&s, "sock %llu\n",
1397 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1400 seq_buf_printf(&s, "shmem %llu\n",
1401 (u64)memcg_page_state(memcg, NR_SHMEM) *
1403 seq_buf_printf(&s, "file_mapped %llu\n",
1404 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1406 seq_buf_printf(&s, "file_dirty %llu\n",
1407 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1409 seq_buf_printf(&s, "file_writeback %llu\n",
1410 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1414 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1415 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1416 * arse because it requires migrating the work out of rmap to a place
1417 * where the page->mem_cgroup is set up and stable.
1419 seq_buf_printf(&s, "anon_thp %llu\n",
1420 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1423 for (i = 0; i < NR_LRU_LISTS; i++)
1424 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1425 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1428 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1429 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1431 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1432 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1435 /* Accumulated memory events */
1437 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1438 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1440 seq_buf_printf(&s, "workingset_refault %lu\n",
1441 memcg_page_state(memcg, WORKINGSET_REFAULT));
1442 seq_buf_printf(&s, "workingset_activate %lu\n",
1443 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1444 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1445 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1447 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1448 seq_buf_printf(&s, "pgscan %lu\n",
1449 memcg_events(memcg, PGSCAN_KSWAPD) +
1450 memcg_events(memcg, PGSCAN_DIRECT));
1451 seq_buf_printf(&s, "pgsteal %lu\n",
1452 memcg_events(memcg, PGSTEAL_KSWAPD) +
1453 memcg_events(memcg, PGSTEAL_DIRECT));
1454 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1455 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1456 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1457 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1459 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1460 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1461 memcg_events(memcg, THP_FAULT_ALLOC));
1462 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1463 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1464 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1466 /* The above should easily fit into one page */
1467 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1472 #define K(x) ((x) << (PAGE_SHIFT-10))
1474 * mem_cgroup_print_oom_context: Print OOM information relevant to
1475 * memory controller.
1476 * @memcg: The memory cgroup that went over limit
1477 * @p: Task that is going to be killed
1479 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1482 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1487 pr_cont(",oom_memcg=");
1488 pr_cont_cgroup_path(memcg->css.cgroup);
1490 pr_cont(",global_oom");
1492 pr_cont(",task_memcg=");
1493 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1499 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1500 * memory controller.
1501 * @memcg: The memory cgroup that went over limit
1503 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1507 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1508 K((u64)page_counter_read(&memcg->memory)),
1509 K((u64)memcg->memory.max), memcg->memory.failcnt);
1510 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1511 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1512 K((u64)page_counter_read(&memcg->swap)),
1513 K((u64)memcg->swap.max), memcg->swap.failcnt);
1515 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1516 K((u64)page_counter_read(&memcg->memsw)),
1517 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1518 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1519 K((u64)page_counter_read(&memcg->kmem)),
1520 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1523 pr_info("Memory cgroup stats for ");
1524 pr_cont_cgroup_path(memcg->css.cgroup);
1526 buf = memory_stat_format(memcg);
1534 * Return the memory (and swap, if configured) limit for a memcg.
1536 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1540 max = memcg->memory.max;
1541 if (mem_cgroup_swappiness(memcg)) {
1542 unsigned long memsw_max;
1543 unsigned long swap_max;
1545 memsw_max = memcg->memsw.max;
1546 swap_max = memcg->swap.max;
1547 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1548 max = min(max + swap_max, memsw_max);
1553 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1556 struct oom_control oc = {
1560 .gfp_mask = gfp_mask,
1565 if (mutex_lock_killable(&oom_lock))
1568 * A few threads which were not waiting at mutex_lock_killable() can
1569 * fail to bail out. Therefore, check again after holding oom_lock.
1571 ret = should_force_charge() || out_of_memory(&oc);
1572 mutex_unlock(&oom_lock);
1576 #if MAX_NUMNODES > 1
1579 * test_mem_cgroup_node_reclaimable
1580 * @memcg: the target memcg
1581 * @nid: the node ID to be checked.
1582 * @noswap : specify true here if the user wants flle only information.
1584 * This function returns whether the specified memcg contains any
1585 * reclaimable pages on a node. Returns true if there are any reclaimable
1586 * pages in the node.
1588 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1589 int nid, bool noswap)
1591 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1593 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1594 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1596 if (noswap || !total_swap_pages)
1598 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1599 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1606 * Always updating the nodemask is not very good - even if we have an empty
1607 * list or the wrong list here, we can start from some node and traverse all
1608 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1611 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1615 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1616 * pagein/pageout changes since the last update.
1618 if (!atomic_read(&memcg->numainfo_events))
1620 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1623 /* make a nodemask where this memcg uses memory from */
1624 memcg->scan_nodes = node_states[N_MEMORY];
1626 for_each_node_mask(nid, node_states[N_MEMORY]) {
1628 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1629 node_clear(nid, memcg->scan_nodes);
1632 atomic_set(&memcg->numainfo_events, 0);
1633 atomic_set(&memcg->numainfo_updating, 0);
1637 * Selecting a node where we start reclaim from. Because what we need is just
1638 * reducing usage counter, start from anywhere is O,K. Considering
1639 * memory reclaim from current node, there are pros. and cons.
1641 * Freeing memory from current node means freeing memory from a node which
1642 * we'll use or we've used. So, it may make LRU bad. And if several threads
1643 * hit limits, it will see a contention on a node. But freeing from remote
1644 * node means more costs for memory reclaim because of memory latency.
1646 * Now, we use round-robin. Better algorithm is welcomed.
1648 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1652 mem_cgroup_may_update_nodemask(memcg);
1653 node = memcg->last_scanned_node;
1655 node = next_node_in(node, memcg->scan_nodes);
1657 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1658 * last time it really checked all the LRUs due to rate limiting.
1659 * Fallback to the current node in that case for simplicity.
1661 if (unlikely(node == MAX_NUMNODES))
1662 node = numa_node_id();
1664 memcg->last_scanned_node = node;
1668 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1674 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1677 unsigned long *total_scanned)
1679 struct mem_cgroup *victim = NULL;
1682 unsigned long excess;
1683 unsigned long nr_scanned;
1684 struct mem_cgroup_reclaim_cookie reclaim = {
1689 excess = soft_limit_excess(root_memcg);
1692 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1697 * If we have not been able to reclaim
1698 * anything, it might because there are
1699 * no reclaimable pages under this hierarchy
1704 * We want to do more targeted reclaim.
1705 * excess >> 2 is not to excessive so as to
1706 * reclaim too much, nor too less that we keep
1707 * coming back to reclaim from this cgroup
1709 if (total >= (excess >> 2) ||
1710 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1715 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1716 pgdat, &nr_scanned);
1717 *total_scanned += nr_scanned;
1718 if (!soft_limit_excess(root_memcg))
1721 mem_cgroup_iter_break(root_memcg, victim);
1725 #ifdef CONFIG_LOCKDEP
1726 static struct lockdep_map memcg_oom_lock_dep_map = {
1727 .name = "memcg_oom_lock",
1731 static DEFINE_SPINLOCK(memcg_oom_lock);
1734 * Check OOM-Killer is already running under our hierarchy.
1735 * If someone is running, return false.
1737 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1739 struct mem_cgroup *iter, *failed = NULL;
1741 spin_lock(&memcg_oom_lock);
1743 for_each_mem_cgroup_tree(iter, memcg) {
1744 if (iter->oom_lock) {
1746 * this subtree of our hierarchy is already locked
1747 * so we cannot give a lock.
1750 mem_cgroup_iter_break(memcg, iter);
1753 iter->oom_lock = true;
1758 * OK, we failed to lock the whole subtree so we have
1759 * to clean up what we set up to the failing subtree
1761 for_each_mem_cgroup_tree(iter, memcg) {
1762 if (iter == failed) {
1763 mem_cgroup_iter_break(memcg, iter);
1766 iter->oom_lock = false;
1769 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1771 spin_unlock(&memcg_oom_lock);
1776 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1778 struct mem_cgroup *iter;
1780 spin_lock(&memcg_oom_lock);
1781 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1782 for_each_mem_cgroup_tree(iter, memcg)
1783 iter->oom_lock = false;
1784 spin_unlock(&memcg_oom_lock);
1787 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1789 struct mem_cgroup *iter;
1791 spin_lock(&memcg_oom_lock);
1792 for_each_mem_cgroup_tree(iter, memcg)
1794 spin_unlock(&memcg_oom_lock);
1797 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1799 struct mem_cgroup *iter;
1802 * When a new child is created while the hierarchy is under oom,
1803 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1805 spin_lock(&memcg_oom_lock);
1806 for_each_mem_cgroup_tree(iter, memcg)
1807 if (iter->under_oom > 0)
1809 spin_unlock(&memcg_oom_lock);
1812 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1814 struct oom_wait_info {
1815 struct mem_cgroup *memcg;
1816 wait_queue_entry_t wait;
1819 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1820 unsigned mode, int sync, void *arg)
1822 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1823 struct mem_cgroup *oom_wait_memcg;
1824 struct oom_wait_info *oom_wait_info;
1826 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1827 oom_wait_memcg = oom_wait_info->memcg;
1829 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1830 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1832 return autoremove_wake_function(wait, mode, sync, arg);
1835 static void memcg_oom_recover(struct mem_cgroup *memcg)
1838 * For the following lockless ->under_oom test, the only required
1839 * guarantee is that it must see the state asserted by an OOM when
1840 * this function is called as a result of userland actions
1841 * triggered by the notification of the OOM. This is trivially
1842 * achieved by invoking mem_cgroup_mark_under_oom() before
1843 * triggering notification.
1845 if (memcg && memcg->under_oom)
1846 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1856 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1858 enum oom_status ret;
1861 if (order > PAGE_ALLOC_COSTLY_ORDER)
1864 memcg_memory_event(memcg, MEMCG_OOM);
1867 * We are in the middle of the charge context here, so we
1868 * don't want to block when potentially sitting on a callstack
1869 * that holds all kinds of filesystem and mm locks.
1871 * cgroup1 allows disabling the OOM killer and waiting for outside
1872 * handling until the charge can succeed; remember the context and put
1873 * the task to sleep at the end of the page fault when all locks are
1876 * On the other hand, in-kernel OOM killer allows for an async victim
1877 * memory reclaim (oom_reaper) and that means that we are not solely
1878 * relying on the oom victim to make a forward progress and we can
1879 * invoke the oom killer here.
1881 * Please note that mem_cgroup_out_of_memory might fail to find a
1882 * victim and then we have to bail out from the charge path.
1884 if (memcg->oom_kill_disable) {
1885 if (!current->in_user_fault)
1887 css_get(&memcg->css);
1888 current->memcg_in_oom = memcg;
1889 current->memcg_oom_gfp_mask = mask;
1890 current->memcg_oom_order = order;
1895 mem_cgroup_mark_under_oom(memcg);
1897 locked = mem_cgroup_oom_trylock(memcg);
1900 mem_cgroup_oom_notify(memcg);
1902 mem_cgroup_unmark_under_oom(memcg);
1903 if (mem_cgroup_out_of_memory(memcg, mask, order))
1909 mem_cgroup_oom_unlock(memcg);
1915 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1916 * @handle: actually kill/wait or just clean up the OOM state
1918 * This has to be called at the end of a page fault if the memcg OOM
1919 * handler was enabled.
1921 * Memcg supports userspace OOM handling where failed allocations must
1922 * sleep on a waitqueue until the userspace task resolves the
1923 * situation. Sleeping directly in the charge context with all kinds
1924 * of locks held is not a good idea, instead we remember an OOM state
1925 * in the task and mem_cgroup_oom_synchronize() has to be called at
1926 * the end of the page fault to complete the OOM handling.
1928 * Returns %true if an ongoing memcg OOM situation was detected and
1929 * completed, %false otherwise.
1931 bool mem_cgroup_oom_synchronize(bool handle)
1933 struct mem_cgroup *memcg = current->memcg_in_oom;
1934 struct oom_wait_info owait;
1937 /* OOM is global, do not handle */
1944 owait.memcg = memcg;
1945 owait.wait.flags = 0;
1946 owait.wait.func = memcg_oom_wake_function;
1947 owait.wait.private = current;
1948 INIT_LIST_HEAD(&owait.wait.entry);
1950 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1951 mem_cgroup_mark_under_oom(memcg);
1953 locked = mem_cgroup_oom_trylock(memcg);
1956 mem_cgroup_oom_notify(memcg);
1958 if (locked && !memcg->oom_kill_disable) {
1959 mem_cgroup_unmark_under_oom(memcg);
1960 finish_wait(&memcg_oom_waitq, &owait.wait);
1961 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1962 current->memcg_oom_order);
1965 mem_cgroup_unmark_under_oom(memcg);
1966 finish_wait(&memcg_oom_waitq, &owait.wait);
1970 mem_cgroup_oom_unlock(memcg);
1972 * There is no guarantee that an OOM-lock contender
1973 * sees the wakeups triggered by the OOM kill
1974 * uncharges. Wake any sleepers explicitely.
1976 memcg_oom_recover(memcg);
1979 current->memcg_in_oom = NULL;
1980 css_put(&memcg->css);
1985 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1986 * @victim: task to be killed by the OOM killer
1987 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1989 * Returns a pointer to a memory cgroup, which has to be cleaned up
1990 * by killing all belonging OOM-killable tasks.
1992 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1994 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1995 struct mem_cgroup *oom_domain)
1997 struct mem_cgroup *oom_group = NULL;
1998 struct mem_cgroup *memcg;
2000 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2004 oom_domain = root_mem_cgroup;
2008 memcg = mem_cgroup_from_task(victim);
2009 if (memcg == root_mem_cgroup)
2013 * Traverse the memory cgroup hierarchy from the victim task's
2014 * cgroup up to the OOMing cgroup (or root) to find the
2015 * highest-level memory cgroup with oom.group set.
2017 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2018 if (memcg->oom_group)
2021 if (memcg == oom_domain)
2026 css_get(&oom_group->css);
2033 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2035 pr_info("Tasks in ");
2036 pr_cont_cgroup_path(memcg->css.cgroup);
2037 pr_cont(" are going to be killed due to memory.oom.group set\n");
2041 * lock_page_memcg - lock a page->mem_cgroup binding
2044 * This function protects unlocked LRU pages from being moved to
2047 * It ensures lifetime of the returned memcg. Caller is responsible
2048 * for the lifetime of the page; __unlock_page_memcg() is available
2049 * when @page might get freed inside the locked section.
2051 struct mem_cgroup *lock_page_memcg(struct page *page)
2053 struct mem_cgroup *memcg;
2054 unsigned long flags;
2057 * The RCU lock is held throughout the transaction. The fast
2058 * path can get away without acquiring the memcg->move_lock
2059 * because page moving starts with an RCU grace period.
2061 * The RCU lock also protects the memcg from being freed when
2062 * the page state that is going to change is the only thing
2063 * preventing the page itself from being freed. E.g. writeback
2064 * doesn't hold a page reference and relies on PG_writeback to
2065 * keep off truncation, migration and so forth.
2069 if (mem_cgroup_disabled())
2072 memcg = page->mem_cgroup;
2073 if (unlikely(!memcg))
2076 if (atomic_read(&memcg->moving_account) <= 0)
2079 spin_lock_irqsave(&memcg->move_lock, flags);
2080 if (memcg != page->mem_cgroup) {
2081 spin_unlock_irqrestore(&memcg->move_lock, flags);
2086 * When charge migration first begins, we can have locked and
2087 * unlocked page stat updates happening concurrently. Track
2088 * the task who has the lock for unlock_page_memcg().
2090 memcg->move_lock_task = current;
2091 memcg->move_lock_flags = flags;
2095 EXPORT_SYMBOL(lock_page_memcg);
2098 * __unlock_page_memcg - unlock and unpin a memcg
2101 * Unlock and unpin a memcg returned by lock_page_memcg().
2103 void __unlock_page_memcg(struct mem_cgroup *memcg)
2105 if (memcg && memcg->move_lock_task == current) {
2106 unsigned long flags = memcg->move_lock_flags;
2108 memcg->move_lock_task = NULL;
2109 memcg->move_lock_flags = 0;
2111 spin_unlock_irqrestore(&memcg->move_lock, flags);
2118 * unlock_page_memcg - unlock a page->mem_cgroup binding
2121 void unlock_page_memcg(struct page *page)
2123 __unlock_page_memcg(page->mem_cgroup);
2125 EXPORT_SYMBOL(unlock_page_memcg);
2127 struct memcg_stock_pcp {
2128 struct mem_cgroup *cached; /* this never be root cgroup */
2129 unsigned int nr_pages;
2130 struct work_struct work;
2131 unsigned long flags;
2132 #define FLUSHING_CACHED_CHARGE 0
2134 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2135 static DEFINE_MUTEX(percpu_charge_mutex);
2138 * consume_stock: Try to consume stocked charge on this cpu.
2139 * @memcg: memcg to consume from.
2140 * @nr_pages: how many pages to charge.
2142 * The charges will only happen if @memcg matches the current cpu's memcg
2143 * stock, and at least @nr_pages are available in that stock. Failure to
2144 * service an allocation will refill the stock.
2146 * returns true if successful, false otherwise.
2148 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2150 struct memcg_stock_pcp *stock;
2151 unsigned long flags;
2154 if (nr_pages > MEMCG_CHARGE_BATCH)
2157 local_irq_save(flags);
2159 stock = this_cpu_ptr(&memcg_stock);
2160 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2161 stock->nr_pages -= nr_pages;
2165 local_irq_restore(flags);
2171 * Returns stocks cached in percpu and reset cached information.
2173 static void drain_stock(struct memcg_stock_pcp *stock)
2175 struct mem_cgroup *old = stock->cached;
2177 if (stock->nr_pages) {
2178 page_counter_uncharge(&old->memory, stock->nr_pages);
2179 if (do_memsw_account())
2180 page_counter_uncharge(&old->memsw, stock->nr_pages);
2181 css_put_many(&old->css, stock->nr_pages);
2182 stock->nr_pages = 0;
2184 stock->cached = NULL;
2187 static void drain_local_stock(struct work_struct *dummy)
2189 struct memcg_stock_pcp *stock;
2190 unsigned long flags;
2193 * The only protection from memory hotplug vs. drain_stock races is
2194 * that we always operate on local CPU stock here with IRQ disabled
2196 local_irq_save(flags);
2198 stock = this_cpu_ptr(&memcg_stock);
2200 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2202 local_irq_restore(flags);
2206 * Cache charges(val) to local per_cpu area.
2207 * This will be consumed by consume_stock() function, later.
2209 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2211 struct memcg_stock_pcp *stock;
2212 unsigned long flags;
2214 local_irq_save(flags);
2216 stock = this_cpu_ptr(&memcg_stock);
2217 if (stock->cached != memcg) { /* reset if necessary */
2219 stock->cached = memcg;
2221 stock->nr_pages += nr_pages;
2223 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2226 local_irq_restore(flags);
2230 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2231 * of the hierarchy under it.
2233 static void drain_all_stock(struct mem_cgroup *root_memcg)
2237 /* If someone's already draining, avoid adding running more workers. */
2238 if (!mutex_trylock(&percpu_charge_mutex))
2241 * Notify other cpus that system-wide "drain" is running
2242 * We do not care about races with the cpu hotplug because cpu down
2243 * as well as workers from this path always operate on the local
2244 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2247 for_each_online_cpu(cpu) {
2248 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2249 struct mem_cgroup *memcg;
2251 memcg = stock->cached;
2252 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2254 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2255 css_put(&memcg->css);
2258 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2260 drain_local_stock(&stock->work);
2262 schedule_work_on(cpu, &stock->work);
2264 css_put(&memcg->css);
2267 mutex_unlock(&percpu_charge_mutex);
2270 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2272 struct memcg_stock_pcp *stock;
2273 struct mem_cgroup *memcg, *mi;
2275 stock = &per_cpu(memcg_stock, cpu);
2278 for_each_mem_cgroup(memcg) {
2281 for (i = 0; i < MEMCG_NR_STAT; i++) {
2285 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2287 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2288 atomic_long_add(x, &memcg->vmstats[i]);
2290 if (i >= NR_VM_NODE_STAT_ITEMS)
2293 for_each_node(nid) {
2294 struct mem_cgroup_per_node *pn;
2296 pn = mem_cgroup_nodeinfo(memcg, nid);
2297 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2300 atomic_long_add(x, &pn->lruvec_stat[i]);
2301 } while ((pn = parent_nodeinfo(pn, nid)));
2305 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2308 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2310 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2311 atomic_long_add(x, &memcg->vmevents[i]);
2318 static void reclaim_high(struct mem_cgroup *memcg,
2319 unsigned int nr_pages,
2323 if (page_counter_read(&memcg->memory) <= memcg->high)
2325 memcg_memory_event(memcg, MEMCG_HIGH);
2326 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2327 } while ((memcg = parent_mem_cgroup(memcg)));
2330 static void high_work_func(struct work_struct *work)
2332 struct mem_cgroup *memcg;
2334 memcg = container_of(work, struct mem_cgroup, high_work);
2335 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2339 * Scheduled by try_charge() to be executed from the userland return path
2340 * and reclaims memory over the high limit.
2342 void mem_cgroup_handle_over_high(void)
2344 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2345 struct mem_cgroup *memcg;
2347 if (likely(!nr_pages))
2350 memcg = get_mem_cgroup_from_mm(current->mm);
2351 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2352 css_put(&memcg->css);
2353 current->memcg_nr_pages_over_high = 0;
2356 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2357 unsigned int nr_pages)
2359 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2360 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2361 struct mem_cgroup *mem_over_limit;
2362 struct page_counter *counter;
2363 unsigned long nr_reclaimed;
2364 bool may_swap = true;
2365 bool drained = false;
2366 enum oom_status oom_status;
2368 if (mem_cgroup_is_root(memcg))
2371 if (consume_stock(memcg, nr_pages))
2374 if (!do_memsw_account() ||
2375 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2376 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2378 if (do_memsw_account())
2379 page_counter_uncharge(&memcg->memsw, batch);
2380 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2382 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2386 if (batch > nr_pages) {
2392 * Unlike in global OOM situations, memcg is not in a physical
2393 * memory shortage. Allow dying and OOM-killed tasks to
2394 * bypass the last charges so that they can exit quickly and
2395 * free their memory.
2397 if (unlikely(should_force_charge()))
2401 * Prevent unbounded recursion when reclaim operations need to
2402 * allocate memory. This might exceed the limits temporarily,
2403 * but we prefer facilitating memory reclaim and getting back
2404 * under the limit over triggering OOM kills in these cases.
2406 if (unlikely(current->flags & PF_MEMALLOC))
2409 if (unlikely(task_in_memcg_oom(current)))
2412 if (!gfpflags_allow_blocking(gfp_mask))
2415 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2417 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2418 gfp_mask, may_swap);
2420 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2424 drain_all_stock(mem_over_limit);
2429 if (gfp_mask & __GFP_NORETRY)
2432 * Even though the limit is exceeded at this point, reclaim
2433 * may have been able to free some pages. Retry the charge
2434 * before killing the task.
2436 * Only for regular pages, though: huge pages are rather
2437 * unlikely to succeed so close to the limit, and we fall back
2438 * to regular pages anyway in case of failure.
2440 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2443 * At task move, charge accounts can be doubly counted. So, it's
2444 * better to wait until the end of task_move if something is going on.
2446 if (mem_cgroup_wait_acct_move(mem_over_limit))
2452 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2455 if (gfp_mask & __GFP_NOFAIL)
2458 if (fatal_signal_pending(current))
2462 * keep retrying as long as the memcg oom killer is able to make
2463 * a forward progress or bypass the charge if the oom killer
2464 * couldn't make any progress.
2466 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2467 get_order(nr_pages * PAGE_SIZE));
2468 switch (oom_status) {
2470 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2478 if (!(gfp_mask & __GFP_NOFAIL))
2482 * The allocation either can't fail or will lead to more memory
2483 * being freed very soon. Allow memory usage go over the limit
2484 * temporarily by force charging it.
2486 page_counter_charge(&memcg->memory, nr_pages);
2487 if (do_memsw_account())
2488 page_counter_charge(&memcg->memsw, nr_pages);
2489 css_get_many(&memcg->css, nr_pages);
2494 css_get_many(&memcg->css, batch);
2495 if (batch > nr_pages)
2496 refill_stock(memcg, batch - nr_pages);
2499 * If the hierarchy is above the normal consumption range, schedule
2500 * reclaim on returning to userland. We can perform reclaim here
2501 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2502 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2503 * not recorded as it most likely matches current's and won't
2504 * change in the meantime. As high limit is checked again before
2505 * reclaim, the cost of mismatch is negligible.
2508 if (page_counter_read(&memcg->memory) > memcg->high) {
2509 /* Don't bother a random interrupted task */
2510 if (in_interrupt()) {
2511 schedule_work(&memcg->high_work);
2514 current->memcg_nr_pages_over_high += batch;
2515 set_notify_resume(current);
2518 } while ((memcg = parent_mem_cgroup(memcg)));
2523 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2525 if (mem_cgroup_is_root(memcg))
2528 page_counter_uncharge(&memcg->memory, nr_pages);
2529 if (do_memsw_account())
2530 page_counter_uncharge(&memcg->memsw, nr_pages);
2532 css_put_many(&memcg->css, nr_pages);
2535 static void lock_page_lru(struct page *page, int *isolated)
2537 pg_data_t *pgdat = page_pgdat(page);
2539 spin_lock_irq(&pgdat->lru_lock);
2540 if (PageLRU(page)) {
2541 struct lruvec *lruvec;
2543 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2545 del_page_from_lru_list(page, lruvec, page_lru(page));
2551 static void unlock_page_lru(struct page *page, int isolated)
2553 pg_data_t *pgdat = page_pgdat(page);
2556 struct lruvec *lruvec;
2558 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2559 VM_BUG_ON_PAGE(PageLRU(page), page);
2561 add_page_to_lru_list(page, lruvec, page_lru(page));
2563 spin_unlock_irq(&pgdat->lru_lock);
2566 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2571 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2574 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2575 * may already be on some other mem_cgroup's LRU. Take care of it.
2578 lock_page_lru(page, &isolated);
2581 * Nobody should be changing or seriously looking at
2582 * page->mem_cgroup at this point:
2584 * - the page is uncharged
2586 * - the page is off-LRU
2588 * - an anonymous fault has exclusive page access, except for
2589 * a locked page table
2591 * - a page cache insertion, a swapin fault, or a migration
2592 * have the page locked
2594 page->mem_cgroup = memcg;
2597 unlock_page_lru(page, isolated);
2600 #ifdef CONFIG_MEMCG_KMEM
2601 static int memcg_alloc_cache_id(void)
2606 id = ida_simple_get(&memcg_cache_ida,
2607 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2611 if (id < memcg_nr_cache_ids)
2615 * There's no space for the new id in memcg_caches arrays,
2616 * so we have to grow them.
2618 down_write(&memcg_cache_ids_sem);
2620 size = 2 * (id + 1);
2621 if (size < MEMCG_CACHES_MIN_SIZE)
2622 size = MEMCG_CACHES_MIN_SIZE;
2623 else if (size > MEMCG_CACHES_MAX_SIZE)
2624 size = MEMCG_CACHES_MAX_SIZE;
2626 err = memcg_update_all_caches(size);
2628 err = memcg_update_all_list_lrus(size);
2630 memcg_nr_cache_ids = size;
2632 up_write(&memcg_cache_ids_sem);
2635 ida_simple_remove(&memcg_cache_ida, id);
2641 static void memcg_free_cache_id(int id)
2643 ida_simple_remove(&memcg_cache_ida, id);
2646 struct memcg_kmem_cache_create_work {
2647 struct mem_cgroup *memcg;
2648 struct kmem_cache *cachep;
2649 struct work_struct work;
2652 static void memcg_kmem_cache_create_func(struct work_struct *w)
2654 struct memcg_kmem_cache_create_work *cw =
2655 container_of(w, struct memcg_kmem_cache_create_work, work);
2656 struct mem_cgroup *memcg = cw->memcg;
2657 struct kmem_cache *cachep = cw->cachep;
2659 memcg_create_kmem_cache(memcg, cachep);
2661 css_put(&memcg->css);
2666 * Enqueue the creation of a per-memcg kmem_cache.
2668 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2669 struct kmem_cache *cachep)
2671 struct memcg_kmem_cache_create_work *cw;
2673 if (!css_tryget_online(&memcg->css))
2676 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2681 cw->cachep = cachep;
2682 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2684 queue_work(memcg_kmem_cache_wq, &cw->work);
2687 static inline bool memcg_kmem_bypass(void)
2689 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2695 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2696 * @cachep: the original global kmem cache
2698 * Return the kmem_cache we're supposed to use for a slab allocation.
2699 * We try to use the current memcg's version of the cache.
2701 * If the cache does not exist yet, if we are the first user of it, we
2702 * create it asynchronously in a workqueue and let the current allocation
2703 * go through with the original cache.
2705 * This function takes a reference to the cache it returns to assure it
2706 * won't get destroyed while we are working with it. Once the caller is
2707 * done with it, memcg_kmem_put_cache() must be called to release the
2710 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2712 struct mem_cgroup *memcg;
2713 struct kmem_cache *memcg_cachep;
2714 struct memcg_cache_array *arr;
2717 VM_BUG_ON(!is_root_cache(cachep));
2719 if (memcg_kmem_bypass())
2724 if (unlikely(current->active_memcg))
2725 memcg = current->active_memcg;
2727 memcg = mem_cgroup_from_task(current);
2729 if (!memcg || memcg == root_mem_cgroup)
2732 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2736 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2739 * Make sure we will access the up-to-date value. The code updating
2740 * memcg_caches issues a write barrier to match the data dependency
2741 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2743 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2746 * If we are in a safe context (can wait, and not in interrupt
2747 * context), we could be be predictable and return right away.
2748 * This would guarantee that the allocation being performed
2749 * already belongs in the new cache.
2751 * However, there are some clashes that can arrive from locking.
2752 * For instance, because we acquire the slab_mutex while doing
2753 * memcg_create_kmem_cache, this means no further allocation
2754 * could happen with the slab_mutex held. So it's better to
2757 * If the memcg is dying or memcg_cache is about to be released,
2758 * don't bother creating new kmem_caches. Because memcg_cachep
2759 * is ZEROed as the fist step of kmem offlining, we don't need
2760 * percpu_ref_tryget_live() here. css_tryget_online() check in
2761 * memcg_schedule_kmem_cache_create() will prevent us from
2762 * creation of a new kmem_cache.
2764 if (unlikely(!memcg_cachep))
2765 memcg_schedule_kmem_cache_create(memcg, cachep);
2766 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2767 cachep = memcg_cachep;
2774 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2775 * @cachep: the cache returned by memcg_kmem_get_cache
2777 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2779 if (!is_root_cache(cachep))
2780 percpu_ref_put(&cachep->memcg_params.refcnt);
2784 * __memcg_kmem_charge_memcg: charge a kmem page
2785 * @page: page to charge
2786 * @gfp: reclaim mode
2787 * @order: allocation order
2788 * @memcg: memory cgroup to charge
2790 * Returns 0 on success, an error code on failure.
2792 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2793 struct mem_cgroup *memcg)
2795 unsigned int nr_pages = 1 << order;
2796 struct page_counter *counter;
2799 ret = try_charge(memcg, gfp, nr_pages);
2803 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2804 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2805 cancel_charge(memcg, nr_pages);
2812 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2813 * @page: page to charge
2814 * @gfp: reclaim mode
2815 * @order: allocation order
2817 * Returns 0 on success, an error code on failure.
2819 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2821 struct mem_cgroup *memcg;
2824 if (memcg_kmem_bypass())
2827 memcg = get_mem_cgroup_from_current();
2828 if (!mem_cgroup_is_root(memcg)) {
2829 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2831 page->mem_cgroup = memcg;
2832 __SetPageKmemcg(page);
2835 css_put(&memcg->css);
2840 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2841 * @memcg: memcg to uncharge
2842 * @nr_pages: number of pages to uncharge
2844 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2845 unsigned int nr_pages)
2847 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2848 page_counter_uncharge(&memcg->kmem, nr_pages);
2850 page_counter_uncharge(&memcg->memory, nr_pages);
2851 if (do_memsw_account())
2852 page_counter_uncharge(&memcg->memsw, nr_pages);
2855 * __memcg_kmem_uncharge: uncharge a kmem page
2856 * @page: page to uncharge
2857 * @order: allocation order
2859 void __memcg_kmem_uncharge(struct page *page, int order)
2861 struct mem_cgroup *memcg = page->mem_cgroup;
2862 unsigned int nr_pages = 1 << order;
2867 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2868 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
2869 page->mem_cgroup = NULL;
2871 /* slab pages do not have PageKmemcg flag set */
2872 if (PageKmemcg(page))
2873 __ClearPageKmemcg(page);
2875 css_put_many(&memcg->css, nr_pages);
2877 #endif /* CONFIG_MEMCG_KMEM */
2879 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2882 * Because tail pages are not marked as "used", set it. We're under
2883 * pgdat->lru_lock and migration entries setup in all page mappings.
2885 void mem_cgroup_split_huge_fixup(struct page *head)
2889 if (mem_cgroup_disabled())
2892 for (i = 1; i < HPAGE_PMD_NR; i++)
2893 head[i].mem_cgroup = head->mem_cgroup;
2895 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2897 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2899 #ifdef CONFIG_MEMCG_SWAP
2901 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2902 * @entry: swap entry to be moved
2903 * @from: mem_cgroup which the entry is moved from
2904 * @to: mem_cgroup which the entry is moved to
2906 * It succeeds only when the swap_cgroup's record for this entry is the same
2907 * as the mem_cgroup's id of @from.
2909 * Returns 0 on success, -EINVAL on failure.
2911 * The caller must have charged to @to, IOW, called page_counter_charge() about
2912 * both res and memsw, and called css_get().
2914 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2915 struct mem_cgroup *from, struct mem_cgroup *to)
2917 unsigned short old_id, new_id;
2919 old_id = mem_cgroup_id(from);
2920 new_id = mem_cgroup_id(to);
2922 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2923 mod_memcg_state(from, MEMCG_SWAP, -1);
2924 mod_memcg_state(to, MEMCG_SWAP, 1);
2930 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2931 struct mem_cgroup *from, struct mem_cgroup *to)
2937 static DEFINE_MUTEX(memcg_max_mutex);
2939 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2940 unsigned long max, bool memsw)
2942 bool enlarge = false;
2943 bool drained = false;
2945 bool limits_invariant;
2946 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2949 if (signal_pending(current)) {
2954 mutex_lock(&memcg_max_mutex);
2956 * Make sure that the new limit (memsw or memory limit) doesn't
2957 * break our basic invariant rule memory.max <= memsw.max.
2959 limits_invariant = memsw ? max >= memcg->memory.max :
2960 max <= memcg->memsw.max;
2961 if (!limits_invariant) {
2962 mutex_unlock(&memcg_max_mutex);
2966 if (max > counter->max)
2968 ret = page_counter_set_max(counter, max);
2969 mutex_unlock(&memcg_max_mutex);
2975 drain_all_stock(memcg);
2980 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2981 GFP_KERNEL, !memsw)) {
2987 if (!ret && enlarge)
2988 memcg_oom_recover(memcg);
2993 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2995 unsigned long *total_scanned)
2997 unsigned long nr_reclaimed = 0;
2998 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2999 unsigned long reclaimed;
3001 struct mem_cgroup_tree_per_node *mctz;
3002 unsigned long excess;
3003 unsigned long nr_scanned;
3008 mctz = soft_limit_tree_node(pgdat->node_id);
3011 * Do not even bother to check the largest node if the root
3012 * is empty. Do it lockless to prevent lock bouncing. Races
3013 * are acceptable as soft limit is best effort anyway.
3015 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3019 * This loop can run a while, specially if mem_cgroup's continuously
3020 * keep exceeding their soft limit and putting the system under
3027 mz = mem_cgroup_largest_soft_limit_node(mctz);
3032 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3033 gfp_mask, &nr_scanned);
3034 nr_reclaimed += reclaimed;
3035 *total_scanned += nr_scanned;
3036 spin_lock_irq(&mctz->lock);
3037 __mem_cgroup_remove_exceeded(mz, mctz);
3040 * If we failed to reclaim anything from this memory cgroup
3041 * it is time to move on to the next cgroup
3045 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3047 excess = soft_limit_excess(mz->memcg);
3049 * One school of thought says that we should not add
3050 * back the node to the tree if reclaim returns 0.
3051 * But our reclaim could return 0, simply because due
3052 * to priority we are exposing a smaller subset of
3053 * memory to reclaim from. Consider this as a longer
3056 /* If excess == 0, no tree ops */
3057 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3058 spin_unlock_irq(&mctz->lock);
3059 css_put(&mz->memcg->css);
3062 * Could not reclaim anything and there are no more
3063 * mem cgroups to try or we seem to be looping without
3064 * reclaiming anything.
3066 if (!nr_reclaimed &&
3068 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3070 } while (!nr_reclaimed);
3072 css_put(&next_mz->memcg->css);
3073 return nr_reclaimed;
3077 * Test whether @memcg has children, dead or alive. Note that this
3078 * function doesn't care whether @memcg has use_hierarchy enabled and
3079 * returns %true if there are child csses according to the cgroup
3080 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3082 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3087 ret = css_next_child(NULL, &memcg->css);
3093 * Reclaims as many pages from the given memcg as possible.
3095 * Caller is responsible for holding css reference for memcg.
3097 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3099 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3101 /* we call try-to-free pages for make this cgroup empty */
3102 lru_add_drain_all();
3104 drain_all_stock(memcg);
3106 /* try to free all pages in this cgroup */
3107 while (nr_retries && page_counter_read(&memcg->memory)) {
3110 if (signal_pending(current))
3113 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3117 /* maybe some writeback is necessary */
3118 congestion_wait(BLK_RW_ASYNC, HZ/10);
3126 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3127 char *buf, size_t nbytes,
3130 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3132 if (mem_cgroup_is_root(memcg))
3134 return mem_cgroup_force_empty(memcg) ?: nbytes;
3137 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3140 return mem_cgroup_from_css(css)->use_hierarchy;
3143 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3144 struct cftype *cft, u64 val)
3147 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3148 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3150 if (memcg->use_hierarchy == val)
3154 * If parent's use_hierarchy is set, we can't make any modifications
3155 * in the child subtrees. If it is unset, then the change can
3156 * occur, provided the current cgroup has no children.
3158 * For the root cgroup, parent_mem is NULL, we allow value to be
3159 * set if there are no children.
3161 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3162 (val == 1 || val == 0)) {
3163 if (!memcg_has_children(memcg))
3164 memcg->use_hierarchy = val;
3173 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3177 if (mem_cgroup_is_root(memcg)) {
3178 val = memcg_page_state(memcg, MEMCG_CACHE) +
3179 memcg_page_state(memcg, MEMCG_RSS);
3181 val += memcg_page_state(memcg, MEMCG_SWAP);
3184 val = page_counter_read(&memcg->memory);
3186 val = page_counter_read(&memcg->memsw);
3199 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3202 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3203 struct page_counter *counter;
3205 switch (MEMFILE_TYPE(cft->private)) {
3207 counter = &memcg->memory;
3210 counter = &memcg->memsw;
3213 counter = &memcg->kmem;
3216 counter = &memcg->tcpmem;
3222 switch (MEMFILE_ATTR(cft->private)) {
3224 if (counter == &memcg->memory)
3225 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3226 if (counter == &memcg->memsw)
3227 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3228 return (u64)page_counter_read(counter) * PAGE_SIZE;
3230 return (u64)counter->max * PAGE_SIZE;
3232 return (u64)counter->watermark * PAGE_SIZE;
3234 return counter->failcnt;
3235 case RES_SOFT_LIMIT:
3236 return (u64)memcg->soft_limit * PAGE_SIZE;
3242 #ifdef CONFIG_MEMCG_KMEM
3243 static int memcg_online_kmem(struct mem_cgroup *memcg)
3247 if (cgroup_memory_nokmem)
3250 BUG_ON(memcg->kmemcg_id >= 0);
3251 BUG_ON(memcg->kmem_state);
3253 memcg_id = memcg_alloc_cache_id();
3257 static_branch_inc(&memcg_kmem_enabled_key);
3259 * A memory cgroup is considered kmem-online as soon as it gets
3260 * kmemcg_id. Setting the id after enabling static branching will
3261 * guarantee no one starts accounting before all call sites are
3264 memcg->kmemcg_id = memcg_id;
3265 memcg->kmem_state = KMEM_ONLINE;
3266 INIT_LIST_HEAD(&memcg->kmem_caches);
3271 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3273 struct cgroup_subsys_state *css;
3274 struct mem_cgroup *parent, *child;
3277 if (memcg->kmem_state != KMEM_ONLINE)
3280 * Clear the online state before clearing memcg_caches array
3281 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3282 * guarantees that no cache will be created for this cgroup
3283 * after we are done (see memcg_create_kmem_cache()).
3285 memcg->kmem_state = KMEM_ALLOCATED;
3287 memcg_deactivate_kmem_caches(memcg);
3289 kmemcg_id = memcg->kmemcg_id;
3290 BUG_ON(kmemcg_id < 0);
3292 parent = parent_mem_cgroup(memcg);
3294 parent = root_mem_cgroup;
3297 * Change kmemcg_id of this cgroup and all its descendants to the
3298 * parent's id, and then move all entries from this cgroup's list_lrus
3299 * to ones of the parent. After we have finished, all list_lrus
3300 * corresponding to this cgroup are guaranteed to remain empty. The
3301 * ordering is imposed by list_lru_node->lock taken by
3302 * memcg_drain_all_list_lrus().
3304 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3305 css_for_each_descendant_pre(css, &memcg->css) {
3306 child = mem_cgroup_from_css(css);
3307 BUG_ON(child->kmemcg_id != kmemcg_id);
3308 child->kmemcg_id = parent->kmemcg_id;
3309 if (!memcg->use_hierarchy)
3314 memcg_drain_all_list_lrus(kmemcg_id, parent);
3316 memcg_free_cache_id(kmemcg_id);
3319 static void memcg_free_kmem(struct mem_cgroup *memcg)
3321 /* css_alloc() failed, offlining didn't happen */
3322 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3323 memcg_offline_kmem(memcg);
3325 if (memcg->kmem_state == KMEM_ALLOCATED) {
3326 WARN_ON(!list_empty(&memcg->kmem_caches));
3327 static_branch_dec(&memcg_kmem_enabled_key);
3328 WARN_ON(page_counter_read(&memcg->kmem));
3332 static int memcg_online_kmem(struct mem_cgroup *memcg)
3336 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3339 static void memcg_free_kmem(struct mem_cgroup *memcg)
3342 #endif /* CONFIG_MEMCG_KMEM */
3344 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3349 mutex_lock(&memcg_max_mutex);
3350 ret = page_counter_set_max(&memcg->kmem, max);
3351 mutex_unlock(&memcg_max_mutex);
3355 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3359 mutex_lock(&memcg_max_mutex);
3361 ret = page_counter_set_max(&memcg->tcpmem, max);
3365 if (!memcg->tcpmem_active) {
3367 * The active flag needs to be written after the static_key
3368 * update. This is what guarantees that the socket activation
3369 * function is the last one to run. See mem_cgroup_sk_alloc()
3370 * for details, and note that we don't mark any socket as
3371 * belonging to this memcg until that flag is up.
3373 * We need to do this, because static_keys will span multiple
3374 * sites, but we can't control their order. If we mark a socket
3375 * as accounted, but the accounting functions are not patched in
3376 * yet, we'll lose accounting.
3378 * We never race with the readers in mem_cgroup_sk_alloc(),
3379 * because when this value change, the code to process it is not
3382 static_branch_inc(&memcg_sockets_enabled_key);
3383 memcg->tcpmem_active = true;
3386 mutex_unlock(&memcg_max_mutex);
3391 * The user of this function is...
3394 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3395 char *buf, size_t nbytes, loff_t off)
3397 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3398 unsigned long nr_pages;
3401 buf = strstrip(buf);
3402 ret = page_counter_memparse(buf, "-1", &nr_pages);
3406 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3408 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3412 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3414 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3417 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3420 ret = memcg_update_kmem_max(memcg, nr_pages);
3423 ret = memcg_update_tcp_max(memcg, nr_pages);
3427 case RES_SOFT_LIMIT:
3428 memcg->soft_limit = nr_pages;
3432 return ret ?: nbytes;
3435 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3436 size_t nbytes, loff_t off)
3438 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3439 struct page_counter *counter;
3441 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3443 counter = &memcg->memory;
3446 counter = &memcg->memsw;
3449 counter = &memcg->kmem;
3452 counter = &memcg->tcpmem;
3458 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3460 page_counter_reset_watermark(counter);
3463 counter->failcnt = 0;
3472 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3475 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3479 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3480 struct cftype *cft, u64 val)
3482 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3484 if (val & ~MOVE_MASK)
3488 * No kind of locking is needed in here, because ->can_attach() will
3489 * check this value once in the beginning of the process, and then carry
3490 * on with stale data. This means that changes to this value will only
3491 * affect task migrations starting after the change.
3493 memcg->move_charge_at_immigrate = val;
3497 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3498 struct cftype *cft, u64 val)
3506 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3507 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3508 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3510 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3511 int nid, unsigned int lru_mask)
3513 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3514 unsigned long nr = 0;
3517 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3520 if (!(BIT(lru) & lru_mask))
3522 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3527 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3528 unsigned int lru_mask)
3530 unsigned long nr = 0;
3534 if (!(BIT(lru) & lru_mask))
3536 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3541 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3545 unsigned int lru_mask;
3548 static const struct numa_stat stats[] = {
3549 { "total", LRU_ALL },
3550 { "file", LRU_ALL_FILE },
3551 { "anon", LRU_ALL_ANON },
3552 { "unevictable", BIT(LRU_UNEVICTABLE) },
3554 const struct numa_stat *stat;
3557 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3559 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3560 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3561 seq_printf(m, "%s=%lu", stat->name, nr);
3562 for_each_node_state(nid, N_MEMORY) {
3563 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3565 seq_printf(m, " N%d=%lu", nid, nr);
3570 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3571 struct mem_cgroup *iter;
3574 for_each_mem_cgroup_tree(iter, memcg)
3575 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3576 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3577 for_each_node_state(nid, N_MEMORY) {
3579 for_each_mem_cgroup_tree(iter, memcg)
3580 nr += mem_cgroup_node_nr_lru_pages(
3581 iter, nid, stat->lru_mask);
3582 seq_printf(m, " N%d=%lu", nid, nr);
3589 #endif /* CONFIG_NUMA */
3591 static const unsigned int memcg1_stats[] = {
3602 static const char *const memcg1_stat_names[] = {
3613 /* Universal VM events cgroup1 shows, original sort order */
3614 static const unsigned int memcg1_events[] = {
3621 static const char *const memcg1_event_names[] = {
3628 static int memcg_stat_show(struct seq_file *m, void *v)
3630 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3631 unsigned long memory, memsw;
3632 struct mem_cgroup *mi;
3635 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3636 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3638 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3639 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3641 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3642 memcg_page_state_local(memcg, memcg1_stats[i]) *
3646 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3647 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3648 memcg_events_local(memcg, memcg1_events[i]));
3650 for (i = 0; i < NR_LRU_LISTS; i++)
3651 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3652 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3655 /* Hierarchical information */
3656 memory = memsw = PAGE_COUNTER_MAX;
3657 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3658 memory = min(memory, mi->memory.max);
3659 memsw = min(memsw, mi->memsw.max);
3661 seq_printf(m, "hierarchical_memory_limit %llu\n",
3662 (u64)memory * PAGE_SIZE);
3663 if (do_memsw_account())
3664 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3665 (u64)memsw * PAGE_SIZE);
3667 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3668 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3670 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3671 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3675 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3676 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3677 (u64)memcg_events(memcg, memcg1_events[i]));
3679 for (i = 0; i < NR_LRU_LISTS; i++)
3680 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3681 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3684 #ifdef CONFIG_DEBUG_VM
3687 struct mem_cgroup_per_node *mz;
3688 struct zone_reclaim_stat *rstat;
3689 unsigned long recent_rotated[2] = {0, 0};
3690 unsigned long recent_scanned[2] = {0, 0};
3692 for_each_online_pgdat(pgdat) {
3693 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3694 rstat = &mz->lruvec.reclaim_stat;
3696 recent_rotated[0] += rstat->recent_rotated[0];
3697 recent_rotated[1] += rstat->recent_rotated[1];
3698 recent_scanned[0] += rstat->recent_scanned[0];
3699 recent_scanned[1] += rstat->recent_scanned[1];
3701 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3702 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3703 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3704 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3711 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3714 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3716 return mem_cgroup_swappiness(memcg);
3719 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3720 struct cftype *cft, u64 val)
3722 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3728 memcg->swappiness = val;
3730 vm_swappiness = val;
3735 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3737 struct mem_cgroup_threshold_ary *t;
3738 unsigned long usage;
3743 t = rcu_dereference(memcg->thresholds.primary);
3745 t = rcu_dereference(memcg->memsw_thresholds.primary);
3750 usage = mem_cgroup_usage(memcg, swap);
3753 * current_threshold points to threshold just below or equal to usage.
3754 * If it's not true, a threshold was crossed after last
3755 * call of __mem_cgroup_threshold().
3757 i = t->current_threshold;
3760 * Iterate backward over array of thresholds starting from
3761 * current_threshold and check if a threshold is crossed.
3762 * If none of thresholds below usage is crossed, we read
3763 * only one element of the array here.
3765 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3766 eventfd_signal(t->entries[i].eventfd, 1);
3768 /* i = current_threshold + 1 */
3772 * Iterate forward over array of thresholds starting from
3773 * current_threshold+1 and check if a threshold is crossed.
3774 * If none of thresholds above usage is crossed, we read
3775 * only one element of the array here.
3777 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3778 eventfd_signal(t->entries[i].eventfd, 1);
3780 /* Update current_threshold */
3781 t->current_threshold = i - 1;
3786 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3789 __mem_cgroup_threshold(memcg, false);
3790 if (do_memsw_account())
3791 __mem_cgroup_threshold(memcg, true);
3793 memcg = parent_mem_cgroup(memcg);
3797 static int compare_thresholds(const void *a, const void *b)
3799 const struct mem_cgroup_threshold *_a = a;
3800 const struct mem_cgroup_threshold *_b = b;
3802 if (_a->threshold > _b->threshold)
3805 if (_a->threshold < _b->threshold)
3811 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3813 struct mem_cgroup_eventfd_list *ev;
3815 spin_lock(&memcg_oom_lock);
3817 list_for_each_entry(ev, &memcg->oom_notify, list)
3818 eventfd_signal(ev->eventfd, 1);
3820 spin_unlock(&memcg_oom_lock);
3824 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3826 struct mem_cgroup *iter;
3828 for_each_mem_cgroup_tree(iter, memcg)
3829 mem_cgroup_oom_notify_cb(iter);
3832 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3833 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3835 struct mem_cgroup_thresholds *thresholds;
3836 struct mem_cgroup_threshold_ary *new;
3837 unsigned long threshold;
3838 unsigned long usage;
3841 ret = page_counter_memparse(args, "-1", &threshold);
3845 mutex_lock(&memcg->thresholds_lock);
3848 thresholds = &memcg->thresholds;
3849 usage = mem_cgroup_usage(memcg, false);
3850 } else if (type == _MEMSWAP) {
3851 thresholds = &memcg->memsw_thresholds;
3852 usage = mem_cgroup_usage(memcg, true);
3856 /* Check if a threshold crossed before adding a new one */
3857 if (thresholds->primary)
3858 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3860 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3862 /* Allocate memory for new array of thresholds */
3863 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3870 /* Copy thresholds (if any) to new array */
3871 if (thresholds->primary) {
3872 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3873 sizeof(struct mem_cgroup_threshold));
3876 /* Add new threshold */
3877 new->entries[size - 1].eventfd = eventfd;
3878 new->entries[size - 1].threshold = threshold;
3880 /* Sort thresholds. Registering of new threshold isn't time-critical */
3881 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3882 compare_thresholds, NULL);
3884 /* Find current threshold */
3885 new->current_threshold = -1;
3886 for (i = 0; i < size; i++) {
3887 if (new->entries[i].threshold <= usage) {
3889 * new->current_threshold will not be used until
3890 * rcu_assign_pointer(), so it's safe to increment
3893 ++new->current_threshold;
3898 /* Free old spare buffer and save old primary buffer as spare */
3899 kfree(thresholds->spare);
3900 thresholds->spare = thresholds->primary;
3902 rcu_assign_pointer(thresholds->primary, new);
3904 /* To be sure that nobody uses thresholds */
3908 mutex_unlock(&memcg->thresholds_lock);
3913 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3914 struct eventfd_ctx *eventfd, const char *args)
3916 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3919 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3920 struct eventfd_ctx *eventfd, const char *args)
3922 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3925 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3926 struct eventfd_ctx *eventfd, enum res_type type)
3928 struct mem_cgroup_thresholds *thresholds;
3929 struct mem_cgroup_threshold_ary *new;
3930 unsigned long usage;
3933 mutex_lock(&memcg->thresholds_lock);
3936 thresholds = &memcg->thresholds;
3937 usage = mem_cgroup_usage(memcg, false);
3938 } else if (type == _MEMSWAP) {
3939 thresholds = &memcg->memsw_thresholds;
3940 usage = mem_cgroup_usage(memcg, true);
3944 if (!thresholds->primary)
3947 /* Check if a threshold crossed before removing */
3948 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3950 /* Calculate new number of threshold */
3952 for (i = 0; i < thresholds->primary->size; i++) {
3953 if (thresholds->primary->entries[i].eventfd != eventfd)
3957 new = thresholds->spare;
3959 /* Set thresholds array to NULL if we don't have thresholds */
3968 /* Copy thresholds and find current threshold */
3969 new->current_threshold = -1;
3970 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3971 if (thresholds->primary->entries[i].eventfd == eventfd)
3974 new->entries[j] = thresholds->primary->entries[i];
3975 if (new->entries[j].threshold <= usage) {
3977 * new->current_threshold will not be used
3978 * until rcu_assign_pointer(), so it's safe to increment
3981 ++new->current_threshold;
3987 /* Swap primary and spare array */
3988 thresholds->spare = thresholds->primary;
3990 rcu_assign_pointer(thresholds->primary, new);
3992 /* To be sure that nobody uses thresholds */
3995 /* If all events are unregistered, free the spare array */
3997 kfree(thresholds->spare);
3998 thresholds->spare = NULL;
4001 mutex_unlock(&memcg->thresholds_lock);
4004 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4005 struct eventfd_ctx *eventfd)
4007 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4010 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4011 struct eventfd_ctx *eventfd)
4013 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4016 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4017 struct eventfd_ctx *eventfd, const char *args)
4019 struct mem_cgroup_eventfd_list *event;
4021 event = kmalloc(sizeof(*event), GFP_KERNEL);
4025 spin_lock(&memcg_oom_lock);
4027 event->eventfd = eventfd;
4028 list_add(&event->list, &memcg->oom_notify);
4030 /* already in OOM ? */
4031 if (memcg->under_oom)
4032 eventfd_signal(eventfd, 1);
4033 spin_unlock(&memcg_oom_lock);
4038 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4039 struct eventfd_ctx *eventfd)
4041 struct mem_cgroup_eventfd_list *ev, *tmp;
4043 spin_lock(&memcg_oom_lock);
4045 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4046 if (ev->eventfd == eventfd) {
4047 list_del(&ev->list);
4052 spin_unlock(&memcg_oom_lock);
4055 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4057 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4059 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4060 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4061 seq_printf(sf, "oom_kill %lu\n",
4062 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4066 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4067 struct cftype *cft, u64 val)
4069 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4071 /* cannot set to root cgroup and only 0 and 1 are allowed */
4072 if (!css->parent || !((val == 0) || (val == 1)))
4075 memcg->oom_kill_disable = val;
4077 memcg_oom_recover(memcg);
4082 #ifdef CONFIG_CGROUP_WRITEBACK
4084 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4086 return wb_domain_init(&memcg->cgwb_domain, gfp);
4089 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4091 wb_domain_exit(&memcg->cgwb_domain);
4094 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4096 wb_domain_size_changed(&memcg->cgwb_domain);
4099 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4101 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4103 if (!memcg->css.parent)
4106 return &memcg->cgwb_domain;
4110 * idx can be of type enum memcg_stat_item or node_stat_item.
4111 * Keep in sync with memcg_exact_page().
4113 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4115 long x = atomic_long_read(&memcg->vmstats[idx]);
4118 for_each_online_cpu(cpu)
4119 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4126 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4127 * @wb: bdi_writeback in question
4128 * @pfilepages: out parameter for number of file pages
4129 * @pheadroom: out parameter for number of allocatable pages according to memcg
4130 * @pdirty: out parameter for number of dirty pages
4131 * @pwriteback: out parameter for number of pages under writeback
4133 * Determine the numbers of file, headroom, dirty, and writeback pages in
4134 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4135 * is a bit more involved.
4137 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4138 * headroom is calculated as the lowest headroom of itself and the
4139 * ancestors. Note that this doesn't consider the actual amount of
4140 * available memory in the system. The caller should further cap
4141 * *@pheadroom accordingly.
4143 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4144 unsigned long *pheadroom, unsigned long *pdirty,
4145 unsigned long *pwriteback)
4147 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4148 struct mem_cgroup *parent;
4150 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4152 /* this should eventually include NR_UNSTABLE_NFS */
4153 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4154 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4155 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4156 *pheadroom = PAGE_COUNTER_MAX;
4158 while ((parent = parent_mem_cgroup(memcg))) {
4159 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4160 unsigned long used = page_counter_read(&memcg->memory);
4162 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4167 #else /* CONFIG_CGROUP_WRITEBACK */
4169 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4174 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4178 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4182 #endif /* CONFIG_CGROUP_WRITEBACK */
4185 * DO NOT USE IN NEW FILES.
4187 * "cgroup.event_control" implementation.
4189 * This is way over-engineered. It tries to support fully configurable
4190 * events for each user. Such level of flexibility is completely
4191 * unnecessary especially in the light of the planned unified hierarchy.
4193 * Please deprecate this and replace with something simpler if at all
4198 * Unregister event and free resources.
4200 * Gets called from workqueue.
4202 static void memcg_event_remove(struct work_struct *work)
4204 struct mem_cgroup_event *event =
4205 container_of(work, struct mem_cgroup_event, remove);
4206 struct mem_cgroup *memcg = event->memcg;
4208 remove_wait_queue(event->wqh, &event->wait);
4210 event->unregister_event(memcg, event->eventfd);
4212 /* Notify userspace the event is going away. */
4213 eventfd_signal(event->eventfd, 1);
4215 eventfd_ctx_put(event->eventfd);
4217 css_put(&memcg->css);
4221 * Gets called on EPOLLHUP on eventfd when user closes it.
4223 * Called with wqh->lock held and interrupts disabled.
4225 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4226 int sync, void *key)
4228 struct mem_cgroup_event *event =
4229 container_of(wait, struct mem_cgroup_event, wait);
4230 struct mem_cgroup *memcg = event->memcg;
4231 __poll_t flags = key_to_poll(key);
4233 if (flags & EPOLLHUP) {
4235 * If the event has been detached at cgroup removal, we
4236 * can simply return knowing the other side will cleanup
4239 * We can't race against event freeing since the other
4240 * side will require wqh->lock via remove_wait_queue(),
4243 spin_lock(&memcg->event_list_lock);
4244 if (!list_empty(&event->list)) {
4245 list_del_init(&event->list);
4247 * We are in atomic context, but cgroup_event_remove()
4248 * may sleep, so we have to call it in workqueue.
4250 schedule_work(&event->remove);
4252 spin_unlock(&memcg->event_list_lock);
4258 static void memcg_event_ptable_queue_proc(struct file *file,
4259 wait_queue_head_t *wqh, poll_table *pt)
4261 struct mem_cgroup_event *event =
4262 container_of(pt, struct mem_cgroup_event, pt);
4265 add_wait_queue(wqh, &event->wait);
4269 * DO NOT USE IN NEW FILES.
4271 * Parse input and register new cgroup event handler.
4273 * Input must be in format '<event_fd> <control_fd> <args>'.
4274 * Interpretation of args is defined by control file implementation.
4276 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4277 char *buf, size_t nbytes, loff_t off)
4279 struct cgroup_subsys_state *css = of_css(of);
4280 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4281 struct mem_cgroup_event *event;
4282 struct cgroup_subsys_state *cfile_css;
4283 unsigned int efd, cfd;
4290 buf = strstrip(buf);
4292 efd = simple_strtoul(buf, &endp, 10);
4297 cfd = simple_strtoul(buf, &endp, 10);
4298 if ((*endp != ' ') && (*endp != '\0'))
4302 event = kzalloc(sizeof(*event), GFP_KERNEL);
4306 event->memcg = memcg;
4307 INIT_LIST_HEAD(&event->list);
4308 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4309 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4310 INIT_WORK(&event->remove, memcg_event_remove);
4318 event->eventfd = eventfd_ctx_fileget(efile.file);
4319 if (IS_ERR(event->eventfd)) {
4320 ret = PTR_ERR(event->eventfd);
4327 goto out_put_eventfd;
4330 /* the process need read permission on control file */
4331 /* AV: shouldn't we check that it's been opened for read instead? */
4332 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4337 * Determine the event callbacks and set them in @event. This used
4338 * to be done via struct cftype but cgroup core no longer knows
4339 * about these events. The following is crude but the whole thing
4340 * is for compatibility anyway.
4342 * DO NOT ADD NEW FILES.
4344 name = cfile.file->f_path.dentry->d_name.name;
4346 if (!strcmp(name, "memory.usage_in_bytes")) {
4347 event->register_event = mem_cgroup_usage_register_event;
4348 event->unregister_event = mem_cgroup_usage_unregister_event;
4349 } else if (!strcmp(name, "memory.oom_control")) {
4350 event->register_event = mem_cgroup_oom_register_event;
4351 event->unregister_event = mem_cgroup_oom_unregister_event;
4352 } else if (!strcmp(name, "memory.pressure_level")) {
4353 event->register_event = vmpressure_register_event;
4354 event->unregister_event = vmpressure_unregister_event;
4355 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4356 event->register_event = memsw_cgroup_usage_register_event;
4357 event->unregister_event = memsw_cgroup_usage_unregister_event;
4364 * Verify @cfile should belong to @css. Also, remaining events are
4365 * automatically removed on cgroup destruction but the removal is
4366 * asynchronous, so take an extra ref on @css.
4368 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4369 &memory_cgrp_subsys);
4371 if (IS_ERR(cfile_css))
4373 if (cfile_css != css) {
4378 ret = event->register_event(memcg, event->eventfd, buf);
4382 vfs_poll(efile.file, &event->pt);
4384 spin_lock(&memcg->event_list_lock);
4385 list_add(&event->list, &memcg->event_list);
4386 spin_unlock(&memcg->event_list_lock);
4398 eventfd_ctx_put(event->eventfd);
4407 static struct cftype mem_cgroup_legacy_files[] = {
4409 .name = "usage_in_bytes",
4410 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4411 .read_u64 = mem_cgroup_read_u64,
4414 .name = "max_usage_in_bytes",
4415 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4416 .write = mem_cgroup_reset,
4417 .read_u64 = mem_cgroup_read_u64,
4420 .name = "limit_in_bytes",
4421 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4422 .write = mem_cgroup_write,
4423 .read_u64 = mem_cgroup_read_u64,
4426 .name = "soft_limit_in_bytes",
4427 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4428 .write = mem_cgroup_write,
4429 .read_u64 = mem_cgroup_read_u64,
4433 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4434 .write = mem_cgroup_reset,
4435 .read_u64 = mem_cgroup_read_u64,
4439 .seq_show = memcg_stat_show,
4442 .name = "force_empty",
4443 .write = mem_cgroup_force_empty_write,
4446 .name = "use_hierarchy",
4447 .write_u64 = mem_cgroup_hierarchy_write,
4448 .read_u64 = mem_cgroup_hierarchy_read,
4451 .name = "cgroup.event_control", /* XXX: for compat */
4452 .write = memcg_write_event_control,
4453 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4456 .name = "swappiness",
4457 .read_u64 = mem_cgroup_swappiness_read,
4458 .write_u64 = mem_cgroup_swappiness_write,
4461 .name = "move_charge_at_immigrate",
4462 .read_u64 = mem_cgroup_move_charge_read,
4463 .write_u64 = mem_cgroup_move_charge_write,
4466 .name = "oom_control",
4467 .seq_show = mem_cgroup_oom_control_read,
4468 .write_u64 = mem_cgroup_oom_control_write,
4469 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4472 .name = "pressure_level",
4476 .name = "numa_stat",
4477 .seq_show = memcg_numa_stat_show,
4481 .name = "kmem.limit_in_bytes",
4482 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4483 .write = mem_cgroup_write,
4484 .read_u64 = mem_cgroup_read_u64,
4487 .name = "kmem.usage_in_bytes",
4488 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4489 .read_u64 = mem_cgroup_read_u64,
4492 .name = "kmem.failcnt",
4493 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4494 .write = mem_cgroup_reset,
4495 .read_u64 = mem_cgroup_read_u64,
4498 .name = "kmem.max_usage_in_bytes",
4499 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4500 .write = mem_cgroup_reset,
4501 .read_u64 = mem_cgroup_read_u64,
4503 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4505 .name = "kmem.slabinfo",
4506 .seq_start = memcg_slab_start,
4507 .seq_next = memcg_slab_next,
4508 .seq_stop = memcg_slab_stop,
4509 .seq_show = memcg_slab_show,
4513 .name = "kmem.tcp.limit_in_bytes",
4514 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4515 .write = mem_cgroup_write,
4516 .read_u64 = mem_cgroup_read_u64,
4519 .name = "kmem.tcp.usage_in_bytes",
4520 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4521 .read_u64 = mem_cgroup_read_u64,
4524 .name = "kmem.tcp.failcnt",
4525 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4526 .write = mem_cgroup_reset,
4527 .read_u64 = mem_cgroup_read_u64,
4530 .name = "kmem.tcp.max_usage_in_bytes",
4531 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4532 .write = mem_cgroup_reset,
4533 .read_u64 = mem_cgroup_read_u64,
4535 { }, /* terminate */
4539 * Private memory cgroup IDR
4541 * Swap-out records and page cache shadow entries need to store memcg
4542 * references in constrained space, so we maintain an ID space that is
4543 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4544 * memory-controlled cgroups to 64k.
4546 * However, there usually are many references to the oflline CSS after
4547 * the cgroup has been destroyed, such as page cache or reclaimable
4548 * slab objects, that don't need to hang on to the ID. We want to keep
4549 * those dead CSS from occupying IDs, or we might quickly exhaust the
4550 * relatively small ID space and prevent the creation of new cgroups
4551 * even when there are much fewer than 64k cgroups - possibly none.
4553 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4554 * be freed and recycled when it's no longer needed, which is usually
4555 * when the CSS is offlined.
4557 * The only exception to that are records of swapped out tmpfs/shmem
4558 * pages that need to be attributed to live ancestors on swapin. But
4559 * those references are manageable from userspace.
4562 static DEFINE_IDR(mem_cgroup_idr);
4564 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4566 if (memcg->id.id > 0) {
4567 idr_remove(&mem_cgroup_idr, memcg->id.id);
4572 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4574 refcount_add(n, &memcg->id.ref);
4577 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4579 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4580 mem_cgroup_id_remove(memcg);
4582 /* Memcg ID pins CSS */
4583 css_put(&memcg->css);
4587 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4589 mem_cgroup_id_get_many(memcg, 1);
4592 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4594 mem_cgroup_id_put_many(memcg, 1);
4598 * mem_cgroup_from_id - look up a memcg from a memcg id
4599 * @id: the memcg id to look up
4601 * Caller must hold rcu_read_lock().
4603 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4605 WARN_ON_ONCE(!rcu_read_lock_held());
4606 return idr_find(&mem_cgroup_idr, id);
4609 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4611 struct mem_cgroup_per_node *pn;
4614 * This routine is called against possible nodes.
4615 * But it's BUG to call kmalloc() against offline node.
4617 * TODO: this routine can waste much memory for nodes which will
4618 * never be onlined. It's better to use memory hotplug callback
4621 if (!node_state(node, N_NORMAL_MEMORY))
4623 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4627 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4628 if (!pn->lruvec_stat_local) {
4633 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4634 if (!pn->lruvec_stat_cpu) {
4635 free_percpu(pn->lruvec_stat_local);
4640 lruvec_init(&pn->lruvec);
4641 pn->usage_in_excess = 0;
4642 pn->on_tree = false;
4645 memcg->nodeinfo[node] = pn;
4649 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4651 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4656 free_percpu(pn->lruvec_stat_cpu);
4657 free_percpu(pn->lruvec_stat_local);
4661 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4666 free_mem_cgroup_per_node_info(memcg, node);
4667 free_percpu(memcg->vmstats_percpu);
4668 free_percpu(memcg->vmstats_local);
4672 static void mem_cgroup_free(struct mem_cgroup *memcg)
4674 memcg_wb_domain_exit(memcg);
4675 __mem_cgroup_free(memcg);
4678 static struct mem_cgroup *mem_cgroup_alloc(void)
4680 struct mem_cgroup *memcg;
4684 size = sizeof(struct mem_cgroup);
4685 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4687 memcg = kzalloc(size, GFP_KERNEL);
4691 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4692 1, MEM_CGROUP_ID_MAX,
4694 if (memcg->id.id < 0)
4697 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
4698 if (!memcg->vmstats_local)
4701 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4702 if (!memcg->vmstats_percpu)
4706 if (alloc_mem_cgroup_per_node_info(memcg, node))
4709 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4712 INIT_WORK(&memcg->high_work, high_work_func);
4713 memcg->last_scanned_node = MAX_NUMNODES;
4714 INIT_LIST_HEAD(&memcg->oom_notify);
4715 mutex_init(&memcg->thresholds_lock);
4716 spin_lock_init(&memcg->move_lock);
4717 vmpressure_init(&memcg->vmpressure);
4718 INIT_LIST_HEAD(&memcg->event_list);
4719 spin_lock_init(&memcg->event_list_lock);
4720 memcg->socket_pressure = jiffies;
4721 #ifdef CONFIG_MEMCG_KMEM
4722 memcg->kmemcg_id = -1;
4724 #ifdef CONFIG_CGROUP_WRITEBACK
4725 INIT_LIST_HEAD(&memcg->cgwb_list);
4727 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4730 mem_cgroup_id_remove(memcg);
4731 __mem_cgroup_free(memcg);
4735 static struct cgroup_subsys_state * __ref
4736 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4738 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4739 struct mem_cgroup *memcg;
4740 long error = -ENOMEM;
4742 memcg = mem_cgroup_alloc();
4744 return ERR_PTR(error);
4746 memcg->high = PAGE_COUNTER_MAX;
4747 memcg->soft_limit = PAGE_COUNTER_MAX;
4749 memcg->swappiness = mem_cgroup_swappiness(parent);
4750 memcg->oom_kill_disable = parent->oom_kill_disable;
4752 if (parent && parent->use_hierarchy) {
4753 memcg->use_hierarchy = true;
4754 page_counter_init(&memcg->memory, &parent->memory);
4755 page_counter_init(&memcg->swap, &parent->swap);
4756 page_counter_init(&memcg->memsw, &parent->memsw);
4757 page_counter_init(&memcg->kmem, &parent->kmem);
4758 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4760 page_counter_init(&memcg->memory, NULL);
4761 page_counter_init(&memcg->swap, NULL);
4762 page_counter_init(&memcg->memsw, NULL);
4763 page_counter_init(&memcg->kmem, NULL);
4764 page_counter_init(&memcg->tcpmem, NULL);
4766 * Deeper hierachy with use_hierarchy == false doesn't make
4767 * much sense so let cgroup subsystem know about this
4768 * unfortunate state in our controller.
4770 if (parent != root_mem_cgroup)
4771 memory_cgrp_subsys.broken_hierarchy = true;
4774 /* The following stuff does not apply to the root */
4776 root_mem_cgroup = memcg;
4780 error = memcg_online_kmem(memcg);
4784 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4785 static_branch_inc(&memcg_sockets_enabled_key);
4789 mem_cgroup_id_remove(memcg);
4790 mem_cgroup_free(memcg);
4791 return ERR_PTR(-ENOMEM);
4794 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4796 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4799 * A memcg must be visible for memcg_expand_shrinker_maps()
4800 * by the time the maps are allocated. So, we allocate maps
4801 * here, when for_each_mem_cgroup() can't skip it.
4803 if (memcg_alloc_shrinker_maps(memcg)) {
4804 mem_cgroup_id_remove(memcg);
4808 /* Online state pins memcg ID, memcg ID pins CSS */
4809 refcount_set(&memcg->id.ref, 1);
4814 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4816 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4817 struct mem_cgroup_event *event, *tmp;
4820 * Unregister events and notify userspace.
4821 * Notify userspace about cgroup removing only after rmdir of cgroup
4822 * directory to avoid race between userspace and kernelspace.
4824 spin_lock(&memcg->event_list_lock);
4825 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4826 list_del_init(&event->list);
4827 schedule_work(&event->remove);
4829 spin_unlock(&memcg->event_list_lock);
4831 page_counter_set_min(&memcg->memory, 0);
4832 page_counter_set_low(&memcg->memory, 0);
4834 memcg_offline_kmem(memcg);
4835 wb_memcg_offline(memcg);
4837 drain_all_stock(memcg);
4839 mem_cgroup_id_put(memcg);
4842 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4844 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4846 invalidate_reclaim_iterators(memcg);
4849 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4851 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4853 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4854 static_branch_dec(&memcg_sockets_enabled_key);
4856 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4857 static_branch_dec(&memcg_sockets_enabled_key);
4859 vmpressure_cleanup(&memcg->vmpressure);
4860 cancel_work_sync(&memcg->high_work);
4861 mem_cgroup_remove_from_trees(memcg);
4862 memcg_free_shrinker_maps(memcg);
4863 memcg_free_kmem(memcg);
4864 mem_cgroup_free(memcg);
4868 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4869 * @css: the target css
4871 * Reset the states of the mem_cgroup associated with @css. This is
4872 * invoked when the userland requests disabling on the default hierarchy
4873 * but the memcg is pinned through dependency. The memcg should stop
4874 * applying policies and should revert to the vanilla state as it may be
4875 * made visible again.
4877 * The current implementation only resets the essential configurations.
4878 * This needs to be expanded to cover all the visible parts.
4880 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4882 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4884 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4885 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4886 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4887 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4888 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4889 page_counter_set_min(&memcg->memory, 0);
4890 page_counter_set_low(&memcg->memory, 0);
4891 memcg->high = PAGE_COUNTER_MAX;
4892 memcg->soft_limit = PAGE_COUNTER_MAX;
4893 memcg_wb_domain_size_changed(memcg);
4897 /* Handlers for move charge at task migration. */
4898 static int mem_cgroup_do_precharge(unsigned long count)
4902 /* Try a single bulk charge without reclaim first, kswapd may wake */
4903 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4905 mc.precharge += count;
4909 /* Try charges one by one with reclaim, but do not retry */
4911 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4925 enum mc_target_type {
4932 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4933 unsigned long addr, pte_t ptent)
4935 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4937 if (!page || !page_mapped(page))
4939 if (PageAnon(page)) {
4940 if (!(mc.flags & MOVE_ANON))
4943 if (!(mc.flags & MOVE_FILE))
4946 if (!get_page_unless_zero(page))
4952 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4953 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4954 pte_t ptent, swp_entry_t *entry)
4956 struct page *page = NULL;
4957 swp_entry_t ent = pte_to_swp_entry(ptent);
4959 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4963 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4964 * a device and because they are not accessible by CPU they are store
4965 * as special swap entry in the CPU page table.
4967 if (is_device_private_entry(ent)) {
4968 page = device_private_entry_to_page(ent);
4970 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4971 * a refcount of 1 when free (unlike normal page)
4973 if (!page_ref_add_unless(page, 1, 1))
4979 * Because lookup_swap_cache() updates some statistics counter,
4980 * we call find_get_page() with swapper_space directly.
4982 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4983 if (do_memsw_account())
4984 entry->val = ent.val;
4989 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4990 pte_t ptent, swp_entry_t *entry)
4996 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4997 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4999 struct page *page = NULL;
5000 struct address_space *mapping;
5003 if (!vma->vm_file) /* anonymous vma */
5005 if (!(mc.flags & MOVE_FILE))
5008 mapping = vma->vm_file->f_mapping;
5009 pgoff = linear_page_index(vma, addr);
5011 /* page is moved even if it's not RSS of this task(page-faulted). */
5013 /* shmem/tmpfs may report page out on swap: account for that too. */
5014 if (shmem_mapping(mapping)) {
5015 page = find_get_entry(mapping, pgoff);
5016 if (xa_is_value(page)) {
5017 swp_entry_t swp = radix_to_swp_entry(page);
5018 if (do_memsw_account())
5020 page = find_get_page(swap_address_space(swp),
5024 page = find_get_page(mapping, pgoff);
5026 page = find_get_page(mapping, pgoff);
5032 * mem_cgroup_move_account - move account of the page
5034 * @compound: charge the page as compound or small page
5035 * @from: mem_cgroup which the page is moved from.
5036 * @to: mem_cgroup which the page is moved to. @from != @to.
5038 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5040 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5043 static int mem_cgroup_move_account(struct page *page,
5045 struct mem_cgroup *from,
5046 struct mem_cgroup *to)
5048 unsigned long flags;
5049 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5053 VM_BUG_ON(from == to);
5054 VM_BUG_ON_PAGE(PageLRU(page), page);
5055 VM_BUG_ON(compound && !PageTransHuge(page));
5058 * Prevent mem_cgroup_migrate() from looking at
5059 * page->mem_cgroup of its source page while we change it.
5062 if (!trylock_page(page))
5066 if (page->mem_cgroup != from)
5069 anon = PageAnon(page);
5071 spin_lock_irqsave(&from->move_lock, flags);
5073 if (!anon && page_mapped(page)) {
5074 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
5075 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
5079 * move_lock grabbed above and caller set from->moving_account, so
5080 * mod_memcg_page_state will serialize updates to PageDirty.
5081 * So mapping should be stable for dirty pages.
5083 if (!anon && PageDirty(page)) {
5084 struct address_space *mapping = page_mapping(page);
5086 if (mapping_cap_account_dirty(mapping)) {
5087 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
5088 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
5092 if (PageWriteback(page)) {
5093 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
5094 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
5098 * It is safe to change page->mem_cgroup here because the page
5099 * is referenced, charged, and isolated - we can't race with
5100 * uncharging, charging, migration, or LRU putback.
5103 /* caller should have done css_get */
5104 page->mem_cgroup = to;
5105 spin_unlock_irqrestore(&from->move_lock, flags);
5109 local_irq_disable();
5110 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5111 memcg_check_events(to, page);
5112 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5113 memcg_check_events(from, page);
5122 * get_mctgt_type - get target type of moving charge
5123 * @vma: the vma the pte to be checked belongs
5124 * @addr: the address corresponding to the pte to be checked
5125 * @ptent: the pte to be checked
5126 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5129 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5130 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5131 * move charge. if @target is not NULL, the page is stored in target->page
5132 * with extra refcnt got(Callers should handle it).
5133 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5134 * target for charge migration. if @target is not NULL, the entry is stored
5136 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
5137 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
5138 * For now we such page is charge like a regular page would be as for all
5139 * intent and purposes it is just special memory taking the place of a
5142 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5144 * Called with pte lock held.
5147 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5148 unsigned long addr, pte_t ptent, union mc_target *target)
5150 struct page *page = NULL;
5151 enum mc_target_type ret = MC_TARGET_NONE;
5152 swp_entry_t ent = { .val = 0 };
5154 if (pte_present(ptent))
5155 page = mc_handle_present_pte(vma, addr, ptent);
5156 else if (is_swap_pte(ptent))
5157 page = mc_handle_swap_pte(vma, ptent, &ent);
5158 else if (pte_none(ptent))
5159 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5161 if (!page && !ent.val)
5165 * Do only loose check w/o serialization.
5166 * mem_cgroup_move_account() checks the page is valid or
5167 * not under LRU exclusion.
5169 if (page->mem_cgroup == mc.from) {
5170 ret = MC_TARGET_PAGE;
5171 if (is_device_private_page(page) ||
5172 is_device_public_page(page))
5173 ret = MC_TARGET_DEVICE;
5175 target->page = page;
5177 if (!ret || !target)
5181 * There is a swap entry and a page doesn't exist or isn't charged.
5182 * But we cannot move a tail-page in a THP.
5184 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5185 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5186 ret = MC_TARGET_SWAP;
5193 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5195 * We don't consider PMD mapped swapping or file mapped pages because THP does
5196 * not support them for now.
5197 * Caller should make sure that pmd_trans_huge(pmd) is true.
5199 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5200 unsigned long addr, pmd_t pmd, union mc_target *target)
5202 struct page *page = NULL;
5203 enum mc_target_type ret = MC_TARGET_NONE;
5205 if (unlikely(is_swap_pmd(pmd))) {
5206 VM_BUG_ON(thp_migration_supported() &&
5207 !is_pmd_migration_entry(pmd));
5210 page = pmd_page(pmd);
5211 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5212 if (!(mc.flags & MOVE_ANON))
5214 if (page->mem_cgroup == mc.from) {
5215 ret = MC_TARGET_PAGE;
5218 target->page = page;
5224 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5225 unsigned long addr, pmd_t pmd, union mc_target *target)
5227 return MC_TARGET_NONE;
5231 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5232 unsigned long addr, unsigned long end,
5233 struct mm_walk *walk)
5235 struct vm_area_struct *vma = walk->vma;
5239 ptl = pmd_trans_huge_lock(pmd, vma);
5242 * Note their can not be MC_TARGET_DEVICE for now as we do not
5243 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5244 * MEMORY_DEVICE_PRIVATE but this might change.
5246 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5247 mc.precharge += HPAGE_PMD_NR;
5252 if (pmd_trans_unstable(pmd))
5254 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5255 for (; addr != end; pte++, addr += PAGE_SIZE)
5256 if (get_mctgt_type(vma, addr, *pte, NULL))
5257 mc.precharge++; /* increment precharge temporarily */
5258 pte_unmap_unlock(pte - 1, ptl);
5264 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5266 unsigned long precharge;
5268 struct mm_walk mem_cgroup_count_precharge_walk = {
5269 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5272 down_read(&mm->mmap_sem);
5273 walk_page_range(0, mm->highest_vm_end,
5274 &mem_cgroup_count_precharge_walk);
5275 up_read(&mm->mmap_sem);
5277 precharge = mc.precharge;
5283 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5285 unsigned long precharge = mem_cgroup_count_precharge(mm);
5287 VM_BUG_ON(mc.moving_task);
5288 mc.moving_task = current;
5289 return mem_cgroup_do_precharge(precharge);
5292 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5293 static void __mem_cgroup_clear_mc(void)
5295 struct mem_cgroup *from = mc.from;
5296 struct mem_cgroup *to = mc.to;
5298 /* we must uncharge all the leftover precharges from mc.to */
5300 cancel_charge(mc.to, mc.precharge);
5304 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5305 * we must uncharge here.
5307 if (mc.moved_charge) {
5308 cancel_charge(mc.from, mc.moved_charge);
5309 mc.moved_charge = 0;
5311 /* we must fixup refcnts and charges */
5312 if (mc.moved_swap) {
5313 /* uncharge swap account from the old cgroup */
5314 if (!mem_cgroup_is_root(mc.from))
5315 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5317 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5320 * we charged both to->memory and to->memsw, so we
5321 * should uncharge to->memory.
5323 if (!mem_cgroup_is_root(mc.to))
5324 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5326 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5327 css_put_many(&mc.to->css, mc.moved_swap);
5331 memcg_oom_recover(from);
5332 memcg_oom_recover(to);
5333 wake_up_all(&mc.waitq);
5336 static void mem_cgroup_clear_mc(void)
5338 struct mm_struct *mm = mc.mm;
5341 * we must clear moving_task before waking up waiters at the end of
5344 mc.moving_task = NULL;
5345 __mem_cgroup_clear_mc();
5346 spin_lock(&mc.lock);
5350 spin_unlock(&mc.lock);
5355 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5357 struct cgroup_subsys_state *css;
5358 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5359 struct mem_cgroup *from;
5360 struct task_struct *leader, *p;
5361 struct mm_struct *mm;
5362 unsigned long move_flags;
5365 /* charge immigration isn't supported on the default hierarchy */
5366 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5370 * Multi-process migrations only happen on the default hierarchy
5371 * where charge immigration is not used. Perform charge
5372 * immigration if @tset contains a leader and whine if there are
5376 cgroup_taskset_for_each_leader(leader, css, tset) {
5379 memcg = mem_cgroup_from_css(css);
5385 * We are now commited to this value whatever it is. Changes in this
5386 * tunable will only affect upcoming migrations, not the current one.
5387 * So we need to save it, and keep it going.
5389 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5393 from = mem_cgroup_from_task(p);
5395 VM_BUG_ON(from == memcg);
5397 mm = get_task_mm(p);
5400 /* We move charges only when we move a owner of the mm */
5401 if (mm->owner == p) {
5404 VM_BUG_ON(mc.precharge);
5405 VM_BUG_ON(mc.moved_charge);
5406 VM_BUG_ON(mc.moved_swap);
5408 spin_lock(&mc.lock);
5412 mc.flags = move_flags;
5413 spin_unlock(&mc.lock);
5414 /* We set mc.moving_task later */
5416 ret = mem_cgroup_precharge_mc(mm);
5418 mem_cgroup_clear_mc();
5425 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5428 mem_cgroup_clear_mc();
5431 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5432 unsigned long addr, unsigned long end,
5433 struct mm_walk *walk)
5436 struct vm_area_struct *vma = walk->vma;
5439 enum mc_target_type target_type;
5440 union mc_target target;
5443 ptl = pmd_trans_huge_lock(pmd, vma);
5445 if (mc.precharge < HPAGE_PMD_NR) {
5449 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5450 if (target_type == MC_TARGET_PAGE) {
5452 if (!isolate_lru_page(page)) {
5453 if (!mem_cgroup_move_account(page, true,
5455 mc.precharge -= HPAGE_PMD_NR;
5456 mc.moved_charge += HPAGE_PMD_NR;
5458 putback_lru_page(page);
5461 } else if (target_type == MC_TARGET_DEVICE) {
5463 if (!mem_cgroup_move_account(page, true,
5465 mc.precharge -= HPAGE_PMD_NR;
5466 mc.moved_charge += HPAGE_PMD_NR;
5474 if (pmd_trans_unstable(pmd))
5477 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5478 for (; addr != end; addr += PAGE_SIZE) {
5479 pte_t ptent = *(pte++);
5480 bool device = false;
5486 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5487 case MC_TARGET_DEVICE:
5490 case MC_TARGET_PAGE:
5493 * We can have a part of the split pmd here. Moving it
5494 * can be done but it would be too convoluted so simply
5495 * ignore such a partial THP and keep it in original
5496 * memcg. There should be somebody mapping the head.
5498 if (PageTransCompound(page))
5500 if (!device && isolate_lru_page(page))
5502 if (!mem_cgroup_move_account(page, false,
5505 /* we uncharge from mc.from later. */
5509 putback_lru_page(page);
5510 put: /* get_mctgt_type() gets the page */
5513 case MC_TARGET_SWAP:
5515 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5517 /* we fixup refcnts and charges later. */
5525 pte_unmap_unlock(pte - 1, ptl);
5530 * We have consumed all precharges we got in can_attach().
5531 * We try charge one by one, but don't do any additional
5532 * charges to mc.to if we have failed in charge once in attach()
5535 ret = mem_cgroup_do_precharge(1);
5543 static void mem_cgroup_move_charge(void)
5545 struct mm_walk mem_cgroup_move_charge_walk = {
5546 .pmd_entry = mem_cgroup_move_charge_pte_range,
5550 lru_add_drain_all();
5552 * Signal lock_page_memcg() to take the memcg's move_lock
5553 * while we're moving its pages to another memcg. Then wait
5554 * for already started RCU-only updates to finish.
5556 atomic_inc(&mc.from->moving_account);
5559 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5561 * Someone who are holding the mmap_sem might be waiting in
5562 * waitq. So we cancel all extra charges, wake up all waiters,
5563 * and retry. Because we cancel precharges, we might not be able
5564 * to move enough charges, but moving charge is a best-effort
5565 * feature anyway, so it wouldn't be a big problem.
5567 __mem_cgroup_clear_mc();
5572 * When we have consumed all precharges and failed in doing
5573 * additional charge, the page walk just aborts.
5575 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5577 up_read(&mc.mm->mmap_sem);
5578 atomic_dec(&mc.from->moving_account);
5581 static void mem_cgroup_move_task(void)
5584 mem_cgroup_move_charge();
5585 mem_cgroup_clear_mc();
5588 #else /* !CONFIG_MMU */
5589 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5593 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5596 static void mem_cgroup_move_task(void)
5602 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5603 * to verify whether we're attached to the default hierarchy on each mount
5606 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5609 * use_hierarchy is forced on the default hierarchy. cgroup core
5610 * guarantees that @root doesn't have any children, so turning it
5611 * on for the root memcg is enough.
5613 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5614 root_mem_cgroup->use_hierarchy = true;
5616 root_mem_cgroup->use_hierarchy = false;
5619 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5621 if (value == PAGE_COUNTER_MAX)
5622 seq_puts(m, "max\n");
5624 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5629 static u64 memory_current_read(struct cgroup_subsys_state *css,
5632 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5634 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5637 static int memory_min_show(struct seq_file *m, void *v)
5639 return seq_puts_memcg_tunable(m,
5640 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5643 static ssize_t memory_min_write(struct kernfs_open_file *of,
5644 char *buf, size_t nbytes, loff_t off)
5646 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5650 buf = strstrip(buf);
5651 err = page_counter_memparse(buf, "max", &min);
5655 page_counter_set_min(&memcg->memory, min);
5660 static int memory_low_show(struct seq_file *m, void *v)
5662 return seq_puts_memcg_tunable(m,
5663 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5666 static ssize_t memory_low_write(struct kernfs_open_file *of,
5667 char *buf, size_t nbytes, loff_t off)
5669 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5673 buf = strstrip(buf);
5674 err = page_counter_memparse(buf, "max", &low);
5678 page_counter_set_low(&memcg->memory, low);
5683 static int memory_high_show(struct seq_file *m, void *v)
5685 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5688 static ssize_t memory_high_write(struct kernfs_open_file *of,
5689 char *buf, size_t nbytes, loff_t off)
5691 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5692 unsigned long nr_pages;
5696 buf = strstrip(buf);
5697 err = page_counter_memparse(buf, "max", &high);
5703 nr_pages = page_counter_read(&memcg->memory);
5704 if (nr_pages > high)
5705 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5708 memcg_wb_domain_size_changed(memcg);
5712 static int memory_max_show(struct seq_file *m, void *v)
5714 return seq_puts_memcg_tunable(m,
5715 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5718 static ssize_t memory_max_write(struct kernfs_open_file *of,
5719 char *buf, size_t nbytes, loff_t off)
5721 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5722 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5723 bool drained = false;
5727 buf = strstrip(buf);
5728 err = page_counter_memparse(buf, "max", &max);
5732 xchg(&memcg->memory.max, max);
5735 unsigned long nr_pages = page_counter_read(&memcg->memory);
5737 if (nr_pages <= max)
5740 if (signal_pending(current)) {
5746 drain_all_stock(memcg);
5752 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5758 memcg_memory_event(memcg, MEMCG_OOM);
5759 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5763 memcg_wb_domain_size_changed(memcg);
5767 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
5769 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
5770 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
5771 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
5772 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
5773 seq_printf(m, "oom_kill %lu\n",
5774 atomic_long_read(&events[MEMCG_OOM_KILL]));
5777 static int memory_events_show(struct seq_file *m, void *v)
5779 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5781 __memory_events_show(m, memcg->memory_events);
5785 static int memory_events_local_show(struct seq_file *m, void *v)
5787 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5789 __memory_events_show(m, memcg->memory_events_local);
5793 static int memory_stat_show(struct seq_file *m, void *v)
5795 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5798 buf = memory_stat_format(memcg);
5806 static int memory_oom_group_show(struct seq_file *m, void *v)
5808 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5810 seq_printf(m, "%d\n", memcg->oom_group);
5815 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5816 char *buf, size_t nbytes, loff_t off)
5818 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5821 buf = strstrip(buf);
5825 ret = kstrtoint(buf, 0, &oom_group);
5829 if (oom_group != 0 && oom_group != 1)
5832 memcg->oom_group = oom_group;
5837 static struct cftype memory_files[] = {
5840 .flags = CFTYPE_NOT_ON_ROOT,
5841 .read_u64 = memory_current_read,
5845 .flags = CFTYPE_NOT_ON_ROOT,
5846 .seq_show = memory_min_show,
5847 .write = memory_min_write,
5851 .flags = CFTYPE_NOT_ON_ROOT,
5852 .seq_show = memory_low_show,
5853 .write = memory_low_write,
5857 .flags = CFTYPE_NOT_ON_ROOT,
5858 .seq_show = memory_high_show,
5859 .write = memory_high_write,
5863 .flags = CFTYPE_NOT_ON_ROOT,
5864 .seq_show = memory_max_show,
5865 .write = memory_max_write,
5869 .flags = CFTYPE_NOT_ON_ROOT,
5870 .file_offset = offsetof(struct mem_cgroup, events_file),
5871 .seq_show = memory_events_show,
5874 .name = "events.local",
5875 .flags = CFTYPE_NOT_ON_ROOT,
5876 .file_offset = offsetof(struct mem_cgroup, events_local_file),
5877 .seq_show = memory_events_local_show,
5881 .flags = CFTYPE_NOT_ON_ROOT,
5882 .seq_show = memory_stat_show,
5885 .name = "oom.group",
5886 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5887 .seq_show = memory_oom_group_show,
5888 .write = memory_oom_group_write,
5893 struct cgroup_subsys memory_cgrp_subsys = {
5894 .css_alloc = mem_cgroup_css_alloc,
5895 .css_online = mem_cgroup_css_online,
5896 .css_offline = mem_cgroup_css_offline,
5897 .css_released = mem_cgroup_css_released,
5898 .css_free = mem_cgroup_css_free,
5899 .css_reset = mem_cgroup_css_reset,
5900 .can_attach = mem_cgroup_can_attach,
5901 .cancel_attach = mem_cgroup_cancel_attach,
5902 .post_attach = mem_cgroup_move_task,
5903 .bind = mem_cgroup_bind,
5904 .dfl_cftypes = memory_files,
5905 .legacy_cftypes = mem_cgroup_legacy_files,
5910 * mem_cgroup_protected - check if memory consumption is in the normal range
5911 * @root: the top ancestor of the sub-tree being checked
5912 * @memcg: the memory cgroup to check
5914 * WARNING: This function is not stateless! It can only be used as part
5915 * of a top-down tree iteration, not for isolated queries.
5917 * Returns one of the following:
5918 * MEMCG_PROT_NONE: cgroup memory is not protected
5919 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5920 * an unprotected supply of reclaimable memory from other cgroups.
5921 * MEMCG_PROT_MIN: cgroup memory is protected
5923 * @root is exclusive; it is never protected when looked at directly
5925 * To provide a proper hierarchical behavior, effective memory.min/low values
5926 * are used. Below is the description of how effective memory.low is calculated.
5927 * Effective memory.min values is calculated in the same way.
5929 * Effective memory.low is always equal or less than the original memory.low.
5930 * If there is no memory.low overcommittment (which is always true for
5931 * top-level memory cgroups), these two values are equal.
5932 * Otherwise, it's a part of parent's effective memory.low,
5933 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5934 * memory.low usages, where memory.low usage is the size of actually
5938 * elow = min( memory.low, parent->elow * ------------------ ),
5939 * siblings_low_usage
5941 * | memory.current, if memory.current < memory.low
5946 * Such definition of the effective memory.low provides the expected
5947 * hierarchical behavior: parent's memory.low value is limiting
5948 * children, unprotected memory is reclaimed first and cgroups,
5949 * which are not using their guarantee do not affect actual memory
5952 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5954 * A A/memory.low = 2G, A/memory.current = 6G
5956 * BC DE B/memory.low = 3G B/memory.current = 2G
5957 * C/memory.low = 1G C/memory.current = 2G
5958 * D/memory.low = 0 D/memory.current = 2G
5959 * E/memory.low = 10G E/memory.current = 0
5961 * and the memory pressure is applied, the following memory distribution
5962 * is expected (approximately):
5964 * A/memory.current = 2G
5966 * B/memory.current = 1.3G
5967 * C/memory.current = 0.6G
5968 * D/memory.current = 0
5969 * E/memory.current = 0
5971 * These calculations require constant tracking of the actual low usages
5972 * (see propagate_protected_usage()), as well as recursive calculation of
5973 * effective memory.low values. But as we do call mem_cgroup_protected()
5974 * path for each memory cgroup top-down from the reclaim,
5975 * it's possible to optimize this part, and save calculated elow
5976 * for next usage. This part is intentionally racy, but it's ok,
5977 * as memory.low is a best-effort mechanism.
5979 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5980 struct mem_cgroup *memcg)
5982 struct mem_cgroup *parent;
5983 unsigned long emin, parent_emin;
5984 unsigned long elow, parent_elow;
5985 unsigned long usage;
5987 if (mem_cgroup_disabled())
5988 return MEMCG_PROT_NONE;
5991 root = root_mem_cgroup;
5993 return MEMCG_PROT_NONE;
5995 usage = page_counter_read(&memcg->memory);
5997 return MEMCG_PROT_NONE;
5999 emin = memcg->memory.min;
6000 elow = memcg->memory.low;
6002 parent = parent_mem_cgroup(memcg);
6003 /* No parent means a non-hierarchical mode on v1 memcg */
6005 return MEMCG_PROT_NONE;
6010 parent_emin = READ_ONCE(parent->memory.emin);
6011 emin = min(emin, parent_emin);
6012 if (emin && parent_emin) {
6013 unsigned long min_usage, siblings_min_usage;
6015 min_usage = min(usage, memcg->memory.min);
6016 siblings_min_usage = atomic_long_read(
6017 &parent->memory.children_min_usage);
6019 if (min_usage && siblings_min_usage)
6020 emin = min(emin, parent_emin * min_usage /
6021 siblings_min_usage);
6024 parent_elow = READ_ONCE(parent->memory.elow);
6025 elow = min(elow, parent_elow);
6026 if (elow && parent_elow) {
6027 unsigned long low_usage, siblings_low_usage;
6029 low_usage = min(usage, memcg->memory.low);
6030 siblings_low_usage = atomic_long_read(
6031 &parent->memory.children_low_usage);
6033 if (low_usage && siblings_low_usage)
6034 elow = min(elow, parent_elow * low_usage /
6035 siblings_low_usage);
6039 memcg->memory.emin = emin;
6040 memcg->memory.elow = elow;
6043 return MEMCG_PROT_MIN;
6044 else if (usage <= elow)
6045 return MEMCG_PROT_LOW;
6047 return MEMCG_PROT_NONE;
6051 * mem_cgroup_try_charge - try charging a page
6052 * @page: page to charge
6053 * @mm: mm context of the victim
6054 * @gfp_mask: reclaim mode
6055 * @memcgp: charged memcg return
6056 * @compound: charge the page as compound or small page
6058 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6059 * pages according to @gfp_mask if necessary.
6061 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6062 * Otherwise, an error code is returned.
6064 * After page->mapping has been set up, the caller must finalize the
6065 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6066 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6068 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6069 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6072 struct mem_cgroup *memcg = NULL;
6073 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6076 if (mem_cgroup_disabled())
6079 if (PageSwapCache(page)) {
6081 * Every swap fault against a single page tries to charge the
6082 * page, bail as early as possible. shmem_unuse() encounters
6083 * already charged pages, too. The USED bit is protected by
6084 * the page lock, which serializes swap cache removal, which
6085 * in turn serializes uncharging.
6087 VM_BUG_ON_PAGE(!PageLocked(page), page);
6088 if (compound_head(page)->mem_cgroup)
6091 if (do_swap_account) {
6092 swp_entry_t ent = { .val = page_private(page), };
6093 unsigned short id = lookup_swap_cgroup_id(ent);
6096 memcg = mem_cgroup_from_id(id);
6097 if (memcg && !css_tryget_online(&memcg->css))
6104 memcg = get_mem_cgroup_from_mm(mm);
6106 ret = try_charge(memcg, gfp_mask, nr_pages);
6108 css_put(&memcg->css);
6114 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6115 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6118 struct mem_cgroup *memcg;
6121 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6123 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6128 * mem_cgroup_commit_charge - commit a page charge
6129 * @page: page to charge
6130 * @memcg: memcg to charge the page to
6131 * @lrucare: page might be on LRU already
6132 * @compound: charge the page as compound or small page
6134 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6135 * after page->mapping has been set up. This must happen atomically
6136 * as part of the page instantiation, i.e. under the page table lock
6137 * for anonymous pages, under the page lock for page and swap cache.
6139 * In addition, the page must not be on the LRU during the commit, to
6140 * prevent racing with task migration. If it might be, use @lrucare.
6142 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6144 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6145 bool lrucare, bool compound)
6147 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6149 VM_BUG_ON_PAGE(!page->mapping, page);
6150 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6152 if (mem_cgroup_disabled())
6155 * Swap faults will attempt to charge the same page multiple
6156 * times. But reuse_swap_page() might have removed the page
6157 * from swapcache already, so we can't check PageSwapCache().
6162 commit_charge(page, memcg, lrucare);
6164 local_irq_disable();
6165 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6166 memcg_check_events(memcg, page);
6169 if (do_memsw_account() && PageSwapCache(page)) {
6170 swp_entry_t entry = { .val = page_private(page) };
6172 * The swap entry might not get freed for a long time,
6173 * let's not wait for it. The page already received a
6174 * memory+swap charge, drop the swap entry duplicate.
6176 mem_cgroup_uncharge_swap(entry, nr_pages);
6181 * mem_cgroup_cancel_charge - cancel a page charge
6182 * @page: page to charge
6183 * @memcg: memcg to charge the page to
6184 * @compound: charge the page as compound or small page
6186 * Cancel a charge transaction started by mem_cgroup_try_charge().
6188 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6191 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6193 if (mem_cgroup_disabled())
6196 * Swap faults will attempt to charge the same page multiple
6197 * times. But reuse_swap_page() might have removed the page
6198 * from swapcache already, so we can't check PageSwapCache().
6203 cancel_charge(memcg, nr_pages);
6206 struct uncharge_gather {
6207 struct mem_cgroup *memcg;
6208 unsigned long pgpgout;
6209 unsigned long nr_anon;
6210 unsigned long nr_file;
6211 unsigned long nr_kmem;
6212 unsigned long nr_huge;
6213 unsigned long nr_shmem;
6214 struct page *dummy_page;
6217 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6219 memset(ug, 0, sizeof(*ug));
6222 static void uncharge_batch(const struct uncharge_gather *ug)
6224 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6225 unsigned long flags;
6227 if (!mem_cgroup_is_root(ug->memcg)) {
6228 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6229 if (do_memsw_account())
6230 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6231 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6232 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6233 memcg_oom_recover(ug->memcg);
6236 local_irq_save(flags);
6237 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6238 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6239 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6240 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6241 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6242 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6243 memcg_check_events(ug->memcg, ug->dummy_page);
6244 local_irq_restore(flags);
6246 if (!mem_cgroup_is_root(ug->memcg))
6247 css_put_many(&ug->memcg->css, nr_pages);
6250 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6252 VM_BUG_ON_PAGE(PageLRU(page), page);
6253 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6254 !PageHWPoison(page) , page);
6256 if (!page->mem_cgroup)
6260 * Nobody should be changing or seriously looking at
6261 * page->mem_cgroup at this point, we have fully
6262 * exclusive access to the page.
6265 if (ug->memcg != page->mem_cgroup) {
6268 uncharge_gather_clear(ug);
6270 ug->memcg = page->mem_cgroup;
6273 if (!PageKmemcg(page)) {
6274 unsigned int nr_pages = 1;
6276 if (PageTransHuge(page)) {
6277 nr_pages <<= compound_order(page);
6278 ug->nr_huge += nr_pages;
6281 ug->nr_anon += nr_pages;
6283 ug->nr_file += nr_pages;
6284 if (PageSwapBacked(page))
6285 ug->nr_shmem += nr_pages;
6289 ug->nr_kmem += 1 << compound_order(page);
6290 __ClearPageKmemcg(page);
6293 ug->dummy_page = page;
6294 page->mem_cgroup = NULL;
6297 static void uncharge_list(struct list_head *page_list)
6299 struct uncharge_gather ug;
6300 struct list_head *next;
6302 uncharge_gather_clear(&ug);
6305 * Note that the list can be a single page->lru; hence the
6306 * do-while loop instead of a simple list_for_each_entry().
6308 next = page_list->next;
6312 page = list_entry(next, struct page, lru);
6313 next = page->lru.next;
6315 uncharge_page(page, &ug);
6316 } while (next != page_list);
6319 uncharge_batch(&ug);
6323 * mem_cgroup_uncharge - uncharge a page
6324 * @page: page to uncharge
6326 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6327 * mem_cgroup_commit_charge().
6329 void mem_cgroup_uncharge(struct page *page)
6331 struct uncharge_gather ug;
6333 if (mem_cgroup_disabled())
6336 /* Don't touch page->lru of any random page, pre-check: */
6337 if (!page->mem_cgroup)
6340 uncharge_gather_clear(&ug);
6341 uncharge_page(page, &ug);
6342 uncharge_batch(&ug);
6346 * mem_cgroup_uncharge_list - uncharge a list of page
6347 * @page_list: list of pages to uncharge
6349 * Uncharge a list of pages previously charged with
6350 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6352 void mem_cgroup_uncharge_list(struct list_head *page_list)
6354 if (mem_cgroup_disabled())
6357 if (!list_empty(page_list))
6358 uncharge_list(page_list);
6362 * mem_cgroup_migrate - charge a page's replacement
6363 * @oldpage: currently circulating page
6364 * @newpage: replacement page
6366 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6367 * be uncharged upon free.
6369 * Both pages must be locked, @newpage->mapping must be set up.
6371 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6373 struct mem_cgroup *memcg;
6374 unsigned int nr_pages;
6376 unsigned long flags;
6378 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6379 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6380 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6381 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6384 if (mem_cgroup_disabled())
6387 /* Page cache replacement: new page already charged? */
6388 if (newpage->mem_cgroup)
6391 /* Swapcache readahead pages can get replaced before being charged */
6392 memcg = oldpage->mem_cgroup;
6396 /* Force-charge the new page. The old one will be freed soon */
6397 compound = PageTransHuge(newpage);
6398 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6400 page_counter_charge(&memcg->memory, nr_pages);
6401 if (do_memsw_account())
6402 page_counter_charge(&memcg->memsw, nr_pages);
6403 css_get_many(&memcg->css, nr_pages);
6405 commit_charge(newpage, memcg, false);
6407 local_irq_save(flags);
6408 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6409 memcg_check_events(memcg, newpage);
6410 local_irq_restore(flags);
6413 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6414 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6416 void mem_cgroup_sk_alloc(struct sock *sk)
6418 struct mem_cgroup *memcg;
6420 if (!mem_cgroup_sockets_enabled)
6424 * Socket cloning can throw us here with sk_memcg already
6425 * filled. It won't however, necessarily happen from
6426 * process context. So the test for root memcg given
6427 * the current task's memcg won't help us in this case.
6429 * Respecting the original socket's memcg is a better
6430 * decision in this case.
6433 css_get(&sk->sk_memcg->css);
6438 memcg = mem_cgroup_from_task(current);
6439 if (memcg == root_mem_cgroup)
6441 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6443 if (css_tryget_online(&memcg->css))
6444 sk->sk_memcg = memcg;
6449 void mem_cgroup_sk_free(struct sock *sk)
6452 css_put(&sk->sk_memcg->css);
6456 * mem_cgroup_charge_skmem - charge socket memory
6457 * @memcg: memcg to charge
6458 * @nr_pages: number of pages to charge
6460 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6461 * @memcg's configured limit, %false if the charge had to be forced.
6463 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6465 gfp_t gfp_mask = GFP_KERNEL;
6467 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6468 struct page_counter *fail;
6470 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6471 memcg->tcpmem_pressure = 0;
6474 page_counter_charge(&memcg->tcpmem, nr_pages);
6475 memcg->tcpmem_pressure = 1;
6479 /* Don't block in the packet receive path */
6481 gfp_mask = GFP_NOWAIT;
6483 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6485 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6488 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6493 * mem_cgroup_uncharge_skmem - uncharge socket memory
6494 * @memcg: memcg to uncharge
6495 * @nr_pages: number of pages to uncharge
6497 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6499 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6500 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6504 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6506 refill_stock(memcg, nr_pages);
6509 static int __init cgroup_memory(char *s)
6513 while ((token = strsep(&s, ",")) != NULL) {
6516 if (!strcmp(token, "nosocket"))
6517 cgroup_memory_nosocket = true;
6518 if (!strcmp(token, "nokmem"))
6519 cgroup_memory_nokmem = true;
6523 __setup("cgroup.memory=", cgroup_memory);
6526 * subsys_initcall() for memory controller.
6528 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6529 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6530 * basically everything that doesn't depend on a specific mem_cgroup structure
6531 * should be initialized from here.
6533 static int __init mem_cgroup_init(void)
6537 #ifdef CONFIG_MEMCG_KMEM
6539 * Kmem cache creation is mostly done with the slab_mutex held,
6540 * so use a workqueue with limited concurrency to avoid stalling
6541 * all worker threads in case lots of cgroups are created and
6542 * destroyed simultaneously.
6544 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6545 BUG_ON(!memcg_kmem_cache_wq);
6548 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6549 memcg_hotplug_cpu_dead);
6551 for_each_possible_cpu(cpu)
6552 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6555 for_each_node(node) {
6556 struct mem_cgroup_tree_per_node *rtpn;
6558 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6559 node_online(node) ? node : NUMA_NO_NODE);
6561 rtpn->rb_root = RB_ROOT;
6562 rtpn->rb_rightmost = NULL;
6563 spin_lock_init(&rtpn->lock);
6564 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6569 subsys_initcall(mem_cgroup_init);
6571 #ifdef CONFIG_MEMCG_SWAP
6572 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6574 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6576 * The root cgroup cannot be destroyed, so it's refcount must
6579 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6583 memcg = parent_mem_cgroup(memcg);
6585 memcg = root_mem_cgroup;
6591 * mem_cgroup_swapout - transfer a memsw charge to swap
6592 * @page: page whose memsw charge to transfer
6593 * @entry: swap entry to move the charge to
6595 * Transfer the memsw charge of @page to @entry.
6597 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6599 struct mem_cgroup *memcg, *swap_memcg;
6600 unsigned int nr_entries;
6601 unsigned short oldid;
6603 VM_BUG_ON_PAGE(PageLRU(page), page);
6604 VM_BUG_ON_PAGE(page_count(page), page);
6606 if (!do_memsw_account())
6609 memcg = page->mem_cgroup;
6611 /* Readahead page, never charged */
6616 * In case the memcg owning these pages has been offlined and doesn't
6617 * have an ID allocated to it anymore, charge the closest online
6618 * ancestor for the swap instead and transfer the memory+swap charge.
6620 swap_memcg = mem_cgroup_id_get_online(memcg);
6621 nr_entries = hpage_nr_pages(page);
6622 /* Get references for the tail pages, too */
6624 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6625 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6627 VM_BUG_ON_PAGE(oldid, page);
6628 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6630 page->mem_cgroup = NULL;
6632 if (!mem_cgroup_is_root(memcg))
6633 page_counter_uncharge(&memcg->memory, nr_entries);
6635 if (memcg != swap_memcg) {
6636 if (!mem_cgroup_is_root(swap_memcg))
6637 page_counter_charge(&swap_memcg->memsw, nr_entries);
6638 page_counter_uncharge(&memcg->memsw, nr_entries);
6642 * Interrupts should be disabled here because the caller holds the
6643 * i_pages lock which is taken with interrupts-off. It is
6644 * important here to have the interrupts disabled because it is the
6645 * only synchronisation we have for updating the per-CPU variables.
6647 VM_BUG_ON(!irqs_disabled());
6648 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6650 memcg_check_events(memcg, page);
6652 if (!mem_cgroup_is_root(memcg))
6653 css_put_many(&memcg->css, nr_entries);
6657 * mem_cgroup_try_charge_swap - try charging swap space for a page
6658 * @page: page being added to swap
6659 * @entry: swap entry to charge
6661 * Try to charge @page's memcg for the swap space at @entry.
6663 * Returns 0 on success, -ENOMEM on failure.
6665 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6667 unsigned int nr_pages = hpage_nr_pages(page);
6668 struct page_counter *counter;
6669 struct mem_cgroup *memcg;
6670 unsigned short oldid;
6672 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6675 memcg = page->mem_cgroup;
6677 /* Readahead page, never charged */
6682 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6686 memcg = mem_cgroup_id_get_online(memcg);
6688 if (!mem_cgroup_is_root(memcg) &&
6689 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6690 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6691 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6692 mem_cgroup_id_put(memcg);
6696 /* Get references for the tail pages, too */
6698 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6699 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6700 VM_BUG_ON_PAGE(oldid, page);
6701 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6707 * mem_cgroup_uncharge_swap - uncharge swap space
6708 * @entry: swap entry to uncharge
6709 * @nr_pages: the amount of swap space to uncharge
6711 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6713 struct mem_cgroup *memcg;
6716 if (!do_swap_account)
6719 id = swap_cgroup_record(entry, 0, nr_pages);
6721 memcg = mem_cgroup_from_id(id);
6723 if (!mem_cgroup_is_root(memcg)) {
6724 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6725 page_counter_uncharge(&memcg->swap, nr_pages);
6727 page_counter_uncharge(&memcg->memsw, nr_pages);
6729 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6730 mem_cgroup_id_put_many(memcg, nr_pages);
6735 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6737 long nr_swap_pages = get_nr_swap_pages();
6739 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6740 return nr_swap_pages;
6741 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6742 nr_swap_pages = min_t(long, nr_swap_pages,
6743 READ_ONCE(memcg->swap.max) -
6744 page_counter_read(&memcg->swap));
6745 return nr_swap_pages;
6748 bool mem_cgroup_swap_full(struct page *page)
6750 struct mem_cgroup *memcg;
6752 VM_BUG_ON_PAGE(!PageLocked(page), page);
6756 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6759 memcg = page->mem_cgroup;
6763 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6764 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6770 /* for remember boot option*/
6771 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6772 static int really_do_swap_account __initdata = 1;
6774 static int really_do_swap_account __initdata;
6777 static int __init enable_swap_account(char *s)
6779 if (!strcmp(s, "1"))
6780 really_do_swap_account = 1;
6781 else if (!strcmp(s, "0"))
6782 really_do_swap_account = 0;
6785 __setup("swapaccount=", enable_swap_account);
6787 static u64 swap_current_read(struct cgroup_subsys_state *css,
6790 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6792 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6795 static int swap_max_show(struct seq_file *m, void *v)
6797 return seq_puts_memcg_tunable(m,
6798 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6801 static ssize_t swap_max_write(struct kernfs_open_file *of,
6802 char *buf, size_t nbytes, loff_t off)
6804 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6808 buf = strstrip(buf);
6809 err = page_counter_memparse(buf, "max", &max);
6813 xchg(&memcg->swap.max, max);
6818 static int swap_events_show(struct seq_file *m, void *v)
6820 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6822 seq_printf(m, "max %lu\n",
6823 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6824 seq_printf(m, "fail %lu\n",
6825 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6830 static struct cftype swap_files[] = {
6832 .name = "swap.current",
6833 .flags = CFTYPE_NOT_ON_ROOT,
6834 .read_u64 = swap_current_read,
6838 .flags = CFTYPE_NOT_ON_ROOT,
6839 .seq_show = swap_max_show,
6840 .write = swap_max_write,
6843 .name = "swap.events",
6844 .flags = CFTYPE_NOT_ON_ROOT,
6845 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6846 .seq_show = swap_events_show,
6851 static struct cftype memsw_cgroup_files[] = {
6853 .name = "memsw.usage_in_bytes",
6854 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6855 .read_u64 = mem_cgroup_read_u64,
6858 .name = "memsw.max_usage_in_bytes",
6859 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6860 .write = mem_cgroup_reset,
6861 .read_u64 = mem_cgroup_read_u64,
6864 .name = "memsw.limit_in_bytes",
6865 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6866 .write = mem_cgroup_write,
6867 .read_u64 = mem_cgroup_read_u64,
6870 .name = "memsw.failcnt",
6871 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6872 .write = mem_cgroup_reset,
6873 .read_u64 = mem_cgroup_read_u64,
6875 { }, /* terminate */
6878 static int __init mem_cgroup_swap_init(void)
6880 if (!mem_cgroup_disabled() && really_do_swap_account) {
6881 do_swap_account = 1;
6882 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6884 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6885 memsw_cgroup_files));
6889 subsys_initcall(mem_cgroup_swap_init);
6891 #endif /* CONFIG_MEMCG_SWAP */