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, CSS_TASK_ITER_PROCS, &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;
1263 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1264 * @memcg: the memory cgroup
1266 * Returns the maximum amount of memory @mem can be charged with, in
1269 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1271 unsigned long margin = 0;
1272 unsigned long count;
1273 unsigned long limit;
1275 count = page_counter_read(&memcg->memory);
1276 limit = READ_ONCE(memcg->memory.max);
1278 margin = limit - count;
1280 if (do_memsw_account()) {
1281 count = page_counter_read(&memcg->memsw);
1282 limit = READ_ONCE(memcg->memsw.max);
1284 margin = min(margin, limit - count);
1293 * A routine for checking "mem" is under move_account() or not.
1295 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1296 * moving cgroups. This is for waiting at high-memory pressure
1299 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1301 struct mem_cgroup *from;
1302 struct mem_cgroup *to;
1305 * Unlike task_move routines, we access mc.to, mc.from not under
1306 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1308 spin_lock(&mc.lock);
1314 ret = mem_cgroup_is_descendant(from, memcg) ||
1315 mem_cgroup_is_descendant(to, memcg);
1317 spin_unlock(&mc.lock);
1321 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1323 if (mc.moving_task && current != mc.moving_task) {
1324 if (mem_cgroup_under_move(memcg)) {
1326 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1327 /* moving charge context might have finished. */
1330 finish_wait(&mc.waitq, &wait);
1337 static char *memory_stat_format(struct mem_cgroup *memcg)
1342 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1347 * Provide statistics on the state of the memory subsystem as
1348 * well as cumulative event counters that show past behavior.
1350 * This list is ordered following a combination of these gradients:
1351 * 1) generic big picture -> specifics and details
1352 * 2) reflecting userspace activity -> reflecting kernel heuristics
1354 * Current memory state:
1357 seq_buf_printf(&s, "anon %llu\n",
1358 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1360 seq_buf_printf(&s, "file %llu\n",
1361 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1363 seq_buf_printf(&s, "kernel_stack %llu\n",
1364 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1366 seq_buf_printf(&s, "slab %llu\n",
1367 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1368 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1370 seq_buf_printf(&s, "sock %llu\n",
1371 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1374 seq_buf_printf(&s, "shmem %llu\n",
1375 (u64)memcg_page_state(memcg, NR_SHMEM) *
1377 seq_buf_printf(&s, "file_mapped %llu\n",
1378 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1380 seq_buf_printf(&s, "file_dirty %llu\n",
1381 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1383 seq_buf_printf(&s, "file_writeback %llu\n",
1384 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1388 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1389 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1390 * arse because it requires migrating the work out of rmap to a place
1391 * where the page->mem_cgroup is set up and stable.
1393 seq_buf_printf(&s, "anon_thp %llu\n",
1394 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1397 for (i = 0; i < NR_LRU_LISTS; i++)
1398 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1399 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1402 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1403 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1405 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1406 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1409 /* Accumulated memory events */
1411 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1412 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1414 seq_buf_printf(&s, "workingset_refault %lu\n",
1415 memcg_page_state(memcg, WORKINGSET_REFAULT));
1416 seq_buf_printf(&s, "workingset_activate %lu\n",
1417 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1418 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1419 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1421 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1422 seq_buf_printf(&s, "pgscan %lu\n",
1423 memcg_events(memcg, PGSCAN_KSWAPD) +
1424 memcg_events(memcg, PGSCAN_DIRECT));
1425 seq_buf_printf(&s, "pgsteal %lu\n",
1426 memcg_events(memcg, PGSTEAL_KSWAPD) +
1427 memcg_events(memcg, PGSTEAL_DIRECT));
1428 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1429 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1430 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1431 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1433 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1434 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1435 memcg_events(memcg, THP_FAULT_ALLOC));
1436 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1437 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1438 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1440 /* The above should easily fit into one page */
1441 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1446 #define K(x) ((x) << (PAGE_SHIFT-10))
1448 * mem_cgroup_print_oom_context: Print OOM information relevant to
1449 * memory controller.
1450 * @memcg: The memory cgroup that went over limit
1451 * @p: Task that is going to be killed
1453 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1456 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1461 pr_cont(",oom_memcg=");
1462 pr_cont_cgroup_path(memcg->css.cgroup);
1464 pr_cont(",global_oom");
1466 pr_cont(",task_memcg=");
1467 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1473 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1474 * memory controller.
1475 * @memcg: The memory cgroup that went over limit
1477 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1481 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1482 K((u64)page_counter_read(&memcg->memory)),
1483 K((u64)memcg->memory.max), memcg->memory.failcnt);
1484 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1485 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1486 K((u64)page_counter_read(&memcg->swap)),
1487 K((u64)memcg->swap.max), memcg->swap.failcnt);
1489 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1490 K((u64)page_counter_read(&memcg->memsw)),
1491 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1492 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1493 K((u64)page_counter_read(&memcg->kmem)),
1494 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1497 pr_info("Memory cgroup stats for ");
1498 pr_cont_cgroup_path(memcg->css.cgroup);
1500 buf = memory_stat_format(memcg);
1508 * Return the memory (and swap, if configured) limit for a memcg.
1510 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1514 max = memcg->memory.max;
1515 if (mem_cgroup_swappiness(memcg)) {
1516 unsigned long memsw_max;
1517 unsigned long swap_max;
1519 memsw_max = memcg->memsw.max;
1520 swap_max = memcg->swap.max;
1521 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1522 max = min(max + swap_max, memsw_max);
1527 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1530 struct oom_control oc = {
1534 .gfp_mask = gfp_mask,
1539 if (mutex_lock_killable(&oom_lock))
1542 * A few threads which were not waiting at mutex_lock_killable() can
1543 * fail to bail out. Therefore, check again after holding oom_lock.
1545 ret = should_force_charge() || out_of_memory(&oc);
1546 mutex_unlock(&oom_lock);
1550 #if MAX_NUMNODES > 1
1553 * test_mem_cgroup_node_reclaimable
1554 * @memcg: the target memcg
1555 * @nid: the node ID to be checked.
1556 * @noswap : specify true here if the user wants flle only information.
1558 * This function returns whether the specified memcg contains any
1559 * reclaimable pages on a node. Returns true if there are any reclaimable
1560 * pages in the node.
1562 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1563 int nid, bool noswap)
1565 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1567 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1568 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1570 if (noswap || !total_swap_pages)
1572 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1573 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1580 * Always updating the nodemask is not very good - even if we have an empty
1581 * list or the wrong list here, we can start from some node and traverse all
1582 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1585 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1589 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1590 * pagein/pageout changes since the last update.
1592 if (!atomic_read(&memcg->numainfo_events))
1594 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1597 /* make a nodemask where this memcg uses memory from */
1598 memcg->scan_nodes = node_states[N_MEMORY];
1600 for_each_node_mask(nid, node_states[N_MEMORY]) {
1602 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1603 node_clear(nid, memcg->scan_nodes);
1606 atomic_set(&memcg->numainfo_events, 0);
1607 atomic_set(&memcg->numainfo_updating, 0);
1611 * Selecting a node where we start reclaim from. Because what we need is just
1612 * reducing usage counter, start from anywhere is O,K. Considering
1613 * memory reclaim from current node, there are pros. and cons.
1615 * Freeing memory from current node means freeing memory from a node which
1616 * we'll use or we've used. So, it may make LRU bad. And if several threads
1617 * hit limits, it will see a contention on a node. But freeing from remote
1618 * node means more costs for memory reclaim because of memory latency.
1620 * Now, we use round-robin. Better algorithm is welcomed.
1622 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1626 mem_cgroup_may_update_nodemask(memcg);
1627 node = memcg->last_scanned_node;
1629 node = next_node_in(node, memcg->scan_nodes);
1631 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1632 * last time it really checked all the LRUs due to rate limiting.
1633 * Fallback to the current node in that case for simplicity.
1635 if (unlikely(node == MAX_NUMNODES))
1636 node = numa_node_id();
1638 memcg->last_scanned_node = node;
1642 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1648 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1651 unsigned long *total_scanned)
1653 struct mem_cgroup *victim = NULL;
1656 unsigned long excess;
1657 unsigned long nr_scanned;
1658 struct mem_cgroup_reclaim_cookie reclaim = {
1663 excess = soft_limit_excess(root_memcg);
1666 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1671 * If we have not been able to reclaim
1672 * anything, it might because there are
1673 * no reclaimable pages under this hierarchy
1678 * We want to do more targeted reclaim.
1679 * excess >> 2 is not to excessive so as to
1680 * reclaim too much, nor too less that we keep
1681 * coming back to reclaim from this cgroup
1683 if (total >= (excess >> 2) ||
1684 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1689 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1690 pgdat, &nr_scanned);
1691 *total_scanned += nr_scanned;
1692 if (!soft_limit_excess(root_memcg))
1695 mem_cgroup_iter_break(root_memcg, victim);
1699 #ifdef CONFIG_LOCKDEP
1700 static struct lockdep_map memcg_oom_lock_dep_map = {
1701 .name = "memcg_oom_lock",
1705 static DEFINE_SPINLOCK(memcg_oom_lock);
1708 * Check OOM-Killer is already running under our hierarchy.
1709 * If someone is running, return false.
1711 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1713 struct mem_cgroup *iter, *failed = NULL;
1715 spin_lock(&memcg_oom_lock);
1717 for_each_mem_cgroup_tree(iter, memcg) {
1718 if (iter->oom_lock) {
1720 * this subtree of our hierarchy is already locked
1721 * so we cannot give a lock.
1724 mem_cgroup_iter_break(memcg, iter);
1727 iter->oom_lock = true;
1732 * OK, we failed to lock the whole subtree so we have
1733 * to clean up what we set up to the failing subtree
1735 for_each_mem_cgroup_tree(iter, memcg) {
1736 if (iter == failed) {
1737 mem_cgroup_iter_break(memcg, iter);
1740 iter->oom_lock = false;
1743 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1745 spin_unlock(&memcg_oom_lock);
1750 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1752 struct mem_cgroup *iter;
1754 spin_lock(&memcg_oom_lock);
1755 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1756 for_each_mem_cgroup_tree(iter, memcg)
1757 iter->oom_lock = false;
1758 spin_unlock(&memcg_oom_lock);
1761 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1763 struct mem_cgroup *iter;
1765 spin_lock(&memcg_oom_lock);
1766 for_each_mem_cgroup_tree(iter, memcg)
1768 spin_unlock(&memcg_oom_lock);
1771 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1773 struct mem_cgroup *iter;
1776 * When a new child is created while the hierarchy is under oom,
1777 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1779 spin_lock(&memcg_oom_lock);
1780 for_each_mem_cgroup_tree(iter, memcg)
1781 if (iter->under_oom > 0)
1783 spin_unlock(&memcg_oom_lock);
1786 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1788 struct oom_wait_info {
1789 struct mem_cgroup *memcg;
1790 wait_queue_entry_t wait;
1793 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1794 unsigned mode, int sync, void *arg)
1796 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1797 struct mem_cgroup *oom_wait_memcg;
1798 struct oom_wait_info *oom_wait_info;
1800 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1801 oom_wait_memcg = oom_wait_info->memcg;
1803 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1804 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1806 return autoremove_wake_function(wait, mode, sync, arg);
1809 static void memcg_oom_recover(struct mem_cgroup *memcg)
1812 * For the following lockless ->under_oom test, the only required
1813 * guarantee is that it must see the state asserted by an OOM when
1814 * this function is called as a result of userland actions
1815 * triggered by the notification of the OOM. This is trivially
1816 * achieved by invoking mem_cgroup_mark_under_oom() before
1817 * triggering notification.
1819 if (memcg && memcg->under_oom)
1820 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1830 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1832 enum oom_status ret;
1835 if (order > PAGE_ALLOC_COSTLY_ORDER)
1838 memcg_memory_event(memcg, MEMCG_OOM);
1841 * We are in the middle of the charge context here, so we
1842 * don't want to block when potentially sitting on a callstack
1843 * that holds all kinds of filesystem and mm locks.
1845 * cgroup1 allows disabling the OOM killer and waiting for outside
1846 * handling until the charge can succeed; remember the context and put
1847 * the task to sleep at the end of the page fault when all locks are
1850 * On the other hand, in-kernel OOM killer allows for an async victim
1851 * memory reclaim (oom_reaper) and that means that we are not solely
1852 * relying on the oom victim to make a forward progress and we can
1853 * invoke the oom killer here.
1855 * Please note that mem_cgroup_out_of_memory might fail to find a
1856 * victim and then we have to bail out from the charge path.
1858 if (memcg->oom_kill_disable) {
1859 if (!current->in_user_fault)
1861 css_get(&memcg->css);
1862 current->memcg_in_oom = memcg;
1863 current->memcg_oom_gfp_mask = mask;
1864 current->memcg_oom_order = order;
1869 mem_cgroup_mark_under_oom(memcg);
1871 locked = mem_cgroup_oom_trylock(memcg);
1874 mem_cgroup_oom_notify(memcg);
1876 mem_cgroup_unmark_under_oom(memcg);
1877 if (mem_cgroup_out_of_memory(memcg, mask, order))
1883 mem_cgroup_oom_unlock(memcg);
1889 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1890 * @handle: actually kill/wait or just clean up the OOM state
1892 * This has to be called at the end of a page fault if the memcg OOM
1893 * handler was enabled.
1895 * Memcg supports userspace OOM handling where failed allocations must
1896 * sleep on a waitqueue until the userspace task resolves the
1897 * situation. Sleeping directly in the charge context with all kinds
1898 * of locks held is not a good idea, instead we remember an OOM state
1899 * in the task and mem_cgroup_oom_synchronize() has to be called at
1900 * the end of the page fault to complete the OOM handling.
1902 * Returns %true if an ongoing memcg OOM situation was detected and
1903 * completed, %false otherwise.
1905 bool mem_cgroup_oom_synchronize(bool handle)
1907 struct mem_cgroup *memcg = current->memcg_in_oom;
1908 struct oom_wait_info owait;
1911 /* OOM is global, do not handle */
1918 owait.memcg = memcg;
1919 owait.wait.flags = 0;
1920 owait.wait.func = memcg_oom_wake_function;
1921 owait.wait.private = current;
1922 INIT_LIST_HEAD(&owait.wait.entry);
1924 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1925 mem_cgroup_mark_under_oom(memcg);
1927 locked = mem_cgroup_oom_trylock(memcg);
1930 mem_cgroup_oom_notify(memcg);
1932 if (locked && !memcg->oom_kill_disable) {
1933 mem_cgroup_unmark_under_oom(memcg);
1934 finish_wait(&memcg_oom_waitq, &owait.wait);
1935 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1936 current->memcg_oom_order);
1939 mem_cgroup_unmark_under_oom(memcg);
1940 finish_wait(&memcg_oom_waitq, &owait.wait);
1944 mem_cgroup_oom_unlock(memcg);
1946 * There is no guarantee that an OOM-lock contender
1947 * sees the wakeups triggered by the OOM kill
1948 * uncharges. Wake any sleepers explicitely.
1950 memcg_oom_recover(memcg);
1953 current->memcg_in_oom = NULL;
1954 css_put(&memcg->css);
1959 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1960 * @victim: task to be killed by the OOM killer
1961 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1963 * Returns a pointer to a memory cgroup, which has to be cleaned up
1964 * by killing all belonging OOM-killable tasks.
1966 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1968 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1969 struct mem_cgroup *oom_domain)
1971 struct mem_cgroup *oom_group = NULL;
1972 struct mem_cgroup *memcg;
1974 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1978 oom_domain = root_mem_cgroup;
1982 memcg = mem_cgroup_from_task(victim);
1983 if (memcg == root_mem_cgroup)
1987 * Traverse the memory cgroup hierarchy from the victim task's
1988 * cgroup up to the OOMing cgroup (or root) to find the
1989 * highest-level memory cgroup with oom.group set.
1991 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1992 if (memcg->oom_group)
1995 if (memcg == oom_domain)
2000 css_get(&oom_group->css);
2007 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2009 pr_info("Tasks in ");
2010 pr_cont_cgroup_path(memcg->css.cgroup);
2011 pr_cont(" are going to be killed due to memory.oom.group set\n");
2015 * lock_page_memcg - lock a page->mem_cgroup binding
2018 * This function protects unlocked LRU pages from being moved to
2021 * It ensures lifetime of the returned memcg. Caller is responsible
2022 * for the lifetime of the page; __unlock_page_memcg() is available
2023 * when @page might get freed inside the locked section.
2025 struct mem_cgroup *lock_page_memcg(struct page *page)
2027 struct mem_cgroup *memcg;
2028 unsigned long flags;
2031 * The RCU lock is held throughout the transaction. The fast
2032 * path can get away without acquiring the memcg->move_lock
2033 * because page moving starts with an RCU grace period.
2035 * The RCU lock also protects the memcg from being freed when
2036 * the page state that is going to change is the only thing
2037 * preventing the page itself from being freed. E.g. writeback
2038 * doesn't hold a page reference and relies on PG_writeback to
2039 * keep off truncation, migration and so forth.
2043 if (mem_cgroup_disabled())
2046 memcg = page->mem_cgroup;
2047 if (unlikely(!memcg))
2050 if (atomic_read(&memcg->moving_account) <= 0)
2053 spin_lock_irqsave(&memcg->move_lock, flags);
2054 if (memcg != page->mem_cgroup) {
2055 spin_unlock_irqrestore(&memcg->move_lock, flags);
2060 * When charge migration first begins, we can have locked and
2061 * unlocked page stat updates happening concurrently. Track
2062 * the task who has the lock for unlock_page_memcg().
2064 memcg->move_lock_task = current;
2065 memcg->move_lock_flags = flags;
2069 EXPORT_SYMBOL(lock_page_memcg);
2072 * __unlock_page_memcg - unlock and unpin a memcg
2075 * Unlock and unpin a memcg returned by lock_page_memcg().
2077 void __unlock_page_memcg(struct mem_cgroup *memcg)
2079 if (memcg && memcg->move_lock_task == current) {
2080 unsigned long flags = memcg->move_lock_flags;
2082 memcg->move_lock_task = NULL;
2083 memcg->move_lock_flags = 0;
2085 spin_unlock_irqrestore(&memcg->move_lock, flags);
2092 * unlock_page_memcg - unlock a page->mem_cgroup binding
2095 void unlock_page_memcg(struct page *page)
2097 __unlock_page_memcg(page->mem_cgroup);
2099 EXPORT_SYMBOL(unlock_page_memcg);
2101 struct memcg_stock_pcp {
2102 struct mem_cgroup *cached; /* this never be root cgroup */
2103 unsigned int nr_pages;
2104 struct work_struct work;
2105 unsigned long flags;
2106 #define FLUSHING_CACHED_CHARGE 0
2108 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2109 static DEFINE_MUTEX(percpu_charge_mutex);
2112 * consume_stock: Try to consume stocked charge on this cpu.
2113 * @memcg: memcg to consume from.
2114 * @nr_pages: how many pages to charge.
2116 * The charges will only happen if @memcg matches the current cpu's memcg
2117 * stock, and at least @nr_pages are available in that stock. Failure to
2118 * service an allocation will refill the stock.
2120 * returns true if successful, false otherwise.
2122 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2124 struct memcg_stock_pcp *stock;
2125 unsigned long flags;
2128 if (nr_pages > MEMCG_CHARGE_BATCH)
2131 local_irq_save(flags);
2133 stock = this_cpu_ptr(&memcg_stock);
2134 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2135 stock->nr_pages -= nr_pages;
2139 local_irq_restore(flags);
2145 * Returns stocks cached in percpu and reset cached information.
2147 static void drain_stock(struct memcg_stock_pcp *stock)
2149 struct mem_cgroup *old = stock->cached;
2151 if (stock->nr_pages) {
2152 page_counter_uncharge(&old->memory, stock->nr_pages);
2153 if (do_memsw_account())
2154 page_counter_uncharge(&old->memsw, stock->nr_pages);
2155 css_put_many(&old->css, stock->nr_pages);
2156 stock->nr_pages = 0;
2158 stock->cached = NULL;
2161 static void drain_local_stock(struct work_struct *dummy)
2163 struct memcg_stock_pcp *stock;
2164 unsigned long flags;
2167 * The only protection from memory hotplug vs. drain_stock races is
2168 * that we always operate on local CPU stock here with IRQ disabled
2170 local_irq_save(flags);
2172 stock = this_cpu_ptr(&memcg_stock);
2174 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2176 local_irq_restore(flags);
2180 * Cache charges(val) to local per_cpu area.
2181 * This will be consumed by consume_stock() function, later.
2183 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2185 struct memcg_stock_pcp *stock;
2186 unsigned long flags;
2188 local_irq_save(flags);
2190 stock = this_cpu_ptr(&memcg_stock);
2191 if (stock->cached != memcg) { /* reset if necessary */
2193 stock->cached = memcg;
2195 stock->nr_pages += nr_pages;
2197 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2200 local_irq_restore(flags);
2204 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2205 * of the hierarchy under it.
2207 static void drain_all_stock(struct mem_cgroup *root_memcg)
2211 /* If someone's already draining, avoid adding running more workers. */
2212 if (!mutex_trylock(&percpu_charge_mutex))
2215 * Notify other cpus that system-wide "drain" is running
2216 * We do not care about races with the cpu hotplug because cpu down
2217 * as well as workers from this path always operate on the local
2218 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2221 for_each_online_cpu(cpu) {
2222 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2223 struct mem_cgroup *memcg;
2225 memcg = stock->cached;
2226 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2228 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2229 css_put(&memcg->css);
2232 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2234 drain_local_stock(&stock->work);
2236 schedule_work_on(cpu, &stock->work);
2238 css_put(&memcg->css);
2241 mutex_unlock(&percpu_charge_mutex);
2244 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2246 struct memcg_stock_pcp *stock;
2247 struct mem_cgroup *memcg, *mi;
2249 stock = &per_cpu(memcg_stock, cpu);
2252 for_each_mem_cgroup(memcg) {
2255 for (i = 0; i < MEMCG_NR_STAT; i++) {
2259 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2261 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2262 atomic_long_add(x, &memcg->vmstats[i]);
2264 if (i >= NR_VM_NODE_STAT_ITEMS)
2267 for_each_node(nid) {
2268 struct mem_cgroup_per_node *pn;
2270 pn = mem_cgroup_nodeinfo(memcg, nid);
2271 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2274 atomic_long_add(x, &pn->lruvec_stat[i]);
2275 } while ((pn = parent_nodeinfo(pn, nid)));
2279 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2282 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2284 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2285 atomic_long_add(x, &memcg->vmevents[i]);
2292 static void reclaim_high(struct mem_cgroup *memcg,
2293 unsigned int nr_pages,
2297 if (page_counter_read(&memcg->memory) <= memcg->high)
2299 memcg_memory_event(memcg, MEMCG_HIGH);
2300 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2301 } while ((memcg = parent_mem_cgroup(memcg)));
2304 static void high_work_func(struct work_struct *work)
2306 struct mem_cgroup *memcg;
2308 memcg = container_of(work, struct mem_cgroup, high_work);
2309 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2313 * Scheduled by try_charge() to be executed from the userland return path
2314 * and reclaims memory over the high limit.
2316 void mem_cgroup_handle_over_high(void)
2318 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2319 struct mem_cgroup *memcg;
2321 if (likely(!nr_pages))
2324 memcg = get_mem_cgroup_from_mm(current->mm);
2325 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2326 css_put(&memcg->css);
2327 current->memcg_nr_pages_over_high = 0;
2330 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2331 unsigned int nr_pages)
2333 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2334 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2335 struct mem_cgroup *mem_over_limit;
2336 struct page_counter *counter;
2337 unsigned long nr_reclaimed;
2338 bool may_swap = true;
2339 bool drained = false;
2340 enum oom_status oom_status;
2342 if (mem_cgroup_is_root(memcg))
2345 if (consume_stock(memcg, nr_pages))
2348 if (!do_memsw_account() ||
2349 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2350 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2352 if (do_memsw_account())
2353 page_counter_uncharge(&memcg->memsw, batch);
2354 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2356 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2360 if (batch > nr_pages) {
2366 * Unlike in global OOM situations, memcg is not in a physical
2367 * memory shortage. Allow dying and OOM-killed tasks to
2368 * bypass the last charges so that they can exit quickly and
2369 * free their memory.
2371 if (unlikely(should_force_charge()))
2375 * Prevent unbounded recursion when reclaim operations need to
2376 * allocate memory. This might exceed the limits temporarily,
2377 * but we prefer facilitating memory reclaim and getting back
2378 * under the limit over triggering OOM kills in these cases.
2380 if (unlikely(current->flags & PF_MEMALLOC))
2383 if (unlikely(task_in_memcg_oom(current)))
2386 if (!gfpflags_allow_blocking(gfp_mask))
2389 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2391 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2392 gfp_mask, may_swap);
2394 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2398 drain_all_stock(mem_over_limit);
2403 if (gfp_mask & __GFP_NORETRY)
2406 * Even though the limit is exceeded at this point, reclaim
2407 * may have been able to free some pages. Retry the charge
2408 * before killing the task.
2410 * Only for regular pages, though: huge pages are rather
2411 * unlikely to succeed so close to the limit, and we fall back
2412 * to regular pages anyway in case of failure.
2414 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2417 * At task move, charge accounts can be doubly counted. So, it's
2418 * better to wait until the end of task_move if something is going on.
2420 if (mem_cgroup_wait_acct_move(mem_over_limit))
2426 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2429 if (gfp_mask & __GFP_NOFAIL)
2432 if (fatal_signal_pending(current))
2436 * keep retrying as long as the memcg oom killer is able to make
2437 * a forward progress or bypass the charge if the oom killer
2438 * couldn't make any progress.
2440 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2441 get_order(nr_pages * PAGE_SIZE));
2442 switch (oom_status) {
2444 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2452 if (!(gfp_mask & __GFP_NOFAIL))
2456 * The allocation either can't fail or will lead to more memory
2457 * being freed very soon. Allow memory usage go over the limit
2458 * temporarily by force charging it.
2460 page_counter_charge(&memcg->memory, nr_pages);
2461 if (do_memsw_account())
2462 page_counter_charge(&memcg->memsw, nr_pages);
2463 css_get_many(&memcg->css, nr_pages);
2468 css_get_many(&memcg->css, batch);
2469 if (batch > nr_pages)
2470 refill_stock(memcg, batch - nr_pages);
2473 * If the hierarchy is above the normal consumption range, schedule
2474 * reclaim on returning to userland. We can perform reclaim here
2475 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2476 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2477 * not recorded as it most likely matches current's and won't
2478 * change in the meantime. As high limit is checked again before
2479 * reclaim, the cost of mismatch is negligible.
2482 if (page_counter_read(&memcg->memory) > memcg->high) {
2483 /* Don't bother a random interrupted task */
2484 if (in_interrupt()) {
2485 schedule_work(&memcg->high_work);
2488 current->memcg_nr_pages_over_high += batch;
2489 set_notify_resume(current);
2492 } while ((memcg = parent_mem_cgroup(memcg)));
2497 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2499 if (mem_cgroup_is_root(memcg))
2502 page_counter_uncharge(&memcg->memory, nr_pages);
2503 if (do_memsw_account())
2504 page_counter_uncharge(&memcg->memsw, nr_pages);
2506 css_put_many(&memcg->css, nr_pages);
2509 static void lock_page_lru(struct page *page, int *isolated)
2511 pg_data_t *pgdat = page_pgdat(page);
2513 spin_lock_irq(&pgdat->lru_lock);
2514 if (PageLRU(page)) {
2515 struct lruvec *lruvec;
2517 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2519 del_page_from_lru_list(page, lruvec, page_lru(page));
2525 static void unlock_page_lru(struct page *page, int isolated)
2527 pg_data_t *pgdat = page_pgdat(page);
2530 struct lruvec *lruvec;
2532 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2533 VM_BUG_ON_PAGE(PageLRU(page), page);
2535 add_page_to_lru_list(page, lruvec, page_lru(page));
2537 spin_unlock_irq(&pgdat->lru_lock);
2540 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2545 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2548 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2549 * may already be on some other mem_cgroup's LRU. Take care of it.
2552 lock_page_lru(page, &isolated);
2555 * Nobody should be changing or seriously looking at
2556 * page->mem_cgroup at this point:
2558 * - the page is uncharged
2560 * - the page is off-LRU
2562 * - an anonymous fault has exclusive page access, except for
2563 * a locked page table
2565 * - a page cache insertion, a swapin fault, or a migration
2566 * have the page locked
2568 page->mem_cgroup = memcg;
2571 unlock_page_lru(page, isolated);
2574 #ifdef CONFIG_MEMCG_KMEM
2575 static int memcg_alloc_cache_id(void)
2580 id = ida_simple_get(&memcg_cache_ida,
2581 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2585 if (id < memcg_nr_cache_ids)
2589 * There's no space for the new id in memcg_caches arrays,
2590 * so we have to grow them.
2592 down_write(&memcg_cache_ids_sem);
2594 size = 2 * (id + 1);
2595 if (size < MEMCG_CACHES_MIN_SIZE)
2596 size = MEMCG_CACHES_MIN_SIZE;
2597 else if (size > MEMCG_CACHES_MAX_SIZE)
2598 size = MEMCG_CACHES_MAX_SIZE;
2600 err = memcg_update_all_caches(size);
2602 err = memcg_update_all_list_lrus(size);
2604 memcg_nr_cache_ids = size;
2606 up_write(&memcg_cache_ids_sem);
2609 ida_simple_remove(&memcg_cache_ida, id);
2615 static void memcg_free_cache_id(int id)
2617 ida_simple_remove(&memcg_cache_ida, id);
2620 struct memcg_kmem_cache_create_work {
2621 struct mem_cgroup *memcg;
2622 struct kmem_cache *cachep;
2623 struct work_struct work;
2626 static void memcg_kmem_cache_create_func(struct work_struct *w)
2628 struct memcg_kmem_cache_create_work *cw =
2629 container_of(w, struct memcg_kmem_cache_create_work, work);
2630 struct mem_cgroup *memcg = cw->memcg;
2631 struct kmem_cache *cachep = cw->cachep;
2633 memcg_create_kmem_cache(memcg, cachep);
2635 css_put(&memcg->css);
2640 * Enqueue the creation of a per-memcg kmem_cache.
2642 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2643 struct kmem_cache *cachep)
2645 struct memcg_kmem_cache_create_work *cw;
2647 if (!css_tryget_online(&memcg->css))
2650 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2655 cw->cachep = cachep;
2656 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2658 queue_work(memcg_kmem_cache_wq, &cw->work);
2661 static inline bool memcg_kmem_bypass(void)
2663 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2669 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2670 * @cachep: the original global kmem cache
2672 * Return the kmem_cache we're supposed to use for a slab allocation.
2673 * We try to use the current memcg's version of the cache.
2675 * If the cache does not exist yet, if we are the first user of it, we
2676 * create it asynchronously in a workqueue and let the current allocation
2677 * go through with the original cache.
2679 * This function takes a reference to the cache it returns to assure it
2680 * won't get destroyed while we are working with it. Once the caller is
2681 * done with it, memcg_kmem_put_cache() must be called to release the
2684 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2686 struct mem_cgroup *memcg;
2687 struct kmem_cache *memcg_cachep;
2688 struct memcg_cache_array *arr;
2691 VM_BUG_ON(!is_root_cache(cachep));
2693 if (memcg_kmem_bypass())
2698 if (unlikely(current->active_memcg))
2699 memcg = current->active_memcg;
2701 memcg = mem_cgroup_from_task(current);
2703 if (!memcg || memcg == root_mem_cgroup)
2706 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2710 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2713 * Make sure we will access the up-to-date value. The code updating
2714 * memcg_caches issues a write barrier to match the data dependency
2715 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2717 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2720 * If we are in a safe context (can wait, and not in interrupt
2721 * context), we could be be predictable and return right away.
2722 * This would guarantee that the allocation being performed
2723 * already belongs in the new cache.
2725 * However, there are some clashes that can arrive from locking.
2726 * For instance, because we acquire the slab_mutex while doing
2727 * memcg_create_kmem_cache, this means no further allocation
2728 * could happen with the slab_mutex held. So it's better to
2731 * If the memcg is dying or memcg_cache is about to be released,
2732 * don't bother creating new kmem_caches. Because memcg_cachep
2733 * is ZEROed as the fist step of kmem offlining, we don't need
2734 * percpu_ref_tryget_live() here. css_tryget_online() check in
2735 * memcg_schedule_kmem_cache_create() will prevent us from
2736 * creation of a new kmem_cache.
2738 if (unlikely(!memcg_cachep))
2739 memcg_schedule_kmem_cache_create(memcg, cachep);
2740 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2741 cachep = memcg_cachep;
2748 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2749 * @cachep: the cache returned by memcg_kmem_get_cache
2751 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2753 if (!is_root_cache(cachep))
2754 percpu_ref_put(&cachep->memcg_params.refcnt);
2758 * __memcg_kmem_charge_memcg: charge a kmem page
2759 * @page: page to charge
2760 * @gfp: reclaim mode
2761 * @order: allocation order
2762 * @memcg: memory cgroup to charge
2764 * Returns 0 on success, an error code on failure.
2766 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2767 struct mem_cgroup *memcg)
2769 unsigned int nr_pages = 1 << order;
2770 struct page_counter *counter;
2773 ret = try_charge(memcg, gfp, nr_pages);
2777 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2778 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2779 cancel_charge(memcg, nr_pages);
2786 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2787 * @page: page to charge
2788 * @gfp: reclaim mode
2789 * @order: allocation order
2791 * Returns 0 on success, an error code on failure.
2793 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2795 struct mem_cgroup *memcg;
2798 if (memcg_kmem_bypass())
2801 memcg = get_mem_cgroup_from_current();
2802 if (!mem_cgroup_is_root(memcg)) {
2803 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2805 page->mem_cgroup = memcg;
2806 __SetPageKmemcg(page);
2809 css_put(&memcg->css);
2814 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2815 * @memcg: memcg to uncharge
2816 * @nr_pages: number of pages to uncharge
2818 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2819 unsigned int nr_pages)
2821 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2822 page_counter_uncharge(&memcg->kmem, nr_pages);
2824 page_counter_uncharge(&memcg->memory, nr_pages);
2825 if (do_memsw_account())
2826 page_counter_uncharge(&memcg->memsw, nr_pages);
2829 * __memcg_kmem_uncharge: uncharge a kmem page
2830 * @page: page to uncharge
2831 * @order: allocation order
2833 void __memcg_kmem_uncharge(struct page *page, int order)
2835 struct mem_cgroup *memcg = page->mem_cgroup;
2836 unsigned int nr_pages = 1 << order;
2841 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2842 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
2843 page->mem_cgroup = NULL;
2845 /* slab pages do not have PageKmemcg flag set */
2846 if (PageKmemcg(page))
2847 __ClearPageKmemcg(page);
2849 css_put_many(&memcg->css, nr_pages);
2851 #endif /* CONFIG_MEMCG_KMEM */
2853 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2856 * Because tail pages are not marked as "used", set it. We're under
2857 * pgdat->lru_lock and migration entries setup in all page mappings.
2859 void mem_cgroup_split_huge_fixup(struct page *head)
2863 if (mem_cgroup_disabled())
2866 for (i = 1; i < HPAGE_PMD_NR; i++)
2867 head[i].mem_cgroup = head->mem_cgroup;
2869 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2871 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2873 #ifdef CONFIG_MEMCG_SWAP
2875 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2876 * @entry: swap entry to be moved
2877 * @from: mem_cgroup which the entry is moved from
2878 * @to: mem_cgroup which the entry is moved to
2880 * It succeeds only when the swap_cgroup's record for this entry is the same
2881 * as the mem_cgroup's id of @from.
2883 * Returns 0 on success, -EINVAL on failure.
2885 * The caller must have charged to @to, IOW, called page_counter_charge() about
2886 * both res and memsw, and called css_get().
2888 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2889 struct mem_cgroup *from, struct mem_cgroup *to)
2891 unsigned short old_id, new_id;
2893 old_id = mem_cgroup_id(from);
2894 new_id = mem_cgroup_id(to);
2896 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2897 mod_memcg_state(from, MEMCG_SWAP, -1);
2898 mod_memcg_state(to, MEMCG_SWAP, 1);
2904 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2905 struct mem_cgroup *from, struct mem_cgroup *to)
2911 static DEFINE_MUTEX(memcg_max_mutex);
2913 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2914 unsigned long max, bool memsw)
2916 bool enlarge = false;
2917 bool drained = false;
2919 bool limits_invariant;
2920 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2923 if (signal_pending(current)) {
2928 mutex_lock(&memcg_max_mutex);
2930 * Make sure that the new limit (memsw or memory limit) doesn't
2931 * break our basic invariant rule memory.max <= memsw.max.
2933 limits_invariant = memsw ? max >= memcg->memory.max :
2934 max <= memcg->memsw.max;
2935 if (!limits_invariant) {
2936 mutex_unlock(&memcg_max_mutex);
2940 if (max > counter->max)
2942 ret = page_counter_set_max(counter, max);
2943 mutex_unlock(&memcg_max_mutex);
2949 drain_all_stock(memcg);
2954 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2955 GFP_KERNEL, !memsw)) {
2961 if (!ret && enlarge)
2962 memcg_oom_recover(memcg);
2967 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2969 unsigned long *total_scanned)
2971 unsigned long nr_reclaimed = 0;
2972 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2973 unsigned long reclaimed;
2975 struct mem_cgroup_tree_per_node *mctz;
2976 unsigned long excess;
2977 unsigned long nr_scanned;
2982 mctz = soft_limit_tree_node(pgdat->node_id);
2985 * Do not even bother to check the largest node if the root
2986 * is empty. Do it lockless to prevent lock bouncing. Races
2987 * are acceptable as soft limit is best effort anyway.
2989 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2993 * This loop can run a while, specially if mem_cgroup's continuously
2994 * keep exceeding their soft limit and putting the system under
3001 mz = mem_cgroup_largest_soft_limit_node(mctz);
3006 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3007 gfp_mask, &nr_scanned);
3008 nr_reclaimed += reclaimed;
3009 *total_scanned += nr_scanned;
3010 spin_lock_irq(&mctz->lock);
3011 __mem_cgroup_remove_exceeded(mz, mctz);
3014 * If we failed to reclaim anything from this memory cgroup
3015 * it is time to move on to the next cgroup
3019 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3021 excess = soft_limit_excess(mz->memcg);
3023 * One school of thought says that we should not add
3024 * back the node to the tree if reclaim returns 0.
3025 * But our reclaim could return 0, simply because due
3026 * to priority we are exposing a smaller subset of
3027 * memory to reclaim from. Consider this as a longer
3030 /* If excess == 0, no tree ops */
3031 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3032 spin_unlock_irq(&mctz->lock);
3033 css_put(&mz->memcg->css);
3036 * Could not reclaim anything and there are no more
3037 * mem cgroups to try or we seem to be looping without
3038 * reclaiming anything.
3040 if (!nr_reclaimed &&
3042 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3044 } while (!nr_reclaimed);
3046 css_put(&next_mz->memcg->css);
3047 return nr_reclaimed;
3051 * Test whether @memcg has children, dead or alive. Note that this
3052 * function doesn't care whether @memcg has use_hierarchy enabled and
3053 * returns %true if there are child csses according to the cgroup
3054 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3056 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3061 ret = css_next_child(NULL, &memcg->css);
3067 * Reclaims as many pages from the given memcg as possible.
3069 * Caller is responsible for holding css reference for memcg.
3071 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3073 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3075 /* we call try-to-free pages for make this cgroup empty */
3076 lru_add_drain_all();
3078 drain_all_stock(memcg);
3080 /* try to free all pages in this cgroup */
3081 while (nr_retries && page_counter_read(&memcg->memory)) {
3084 if (signal_pending(current))
3087 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3091 /* maybe some writeback is necessary */
3092 congestion_wait(BLK_RW_ASYNC, HZ/10);
3100 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3101 char *buf, size_t nbytes,
3104 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3106 if (mem_cgroup_is_root(memcg))
3108 return mem_cgroup_force_empty(memcg) ?: nbytes;
3111 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3114 return mem_cgroup_from_css(css)->use_hierarchy;
3117 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3118 struct cftype *cft, u64 val)
3121 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3122 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3124 if (memcg->use_hierarchy == val)
3128 * If parent's use_hierarchy is set, we can't make any modifications
3129 * in the child subtrees. If it is unset, then the change can
3130 * occur, provided the current cgroup has no children.
3132 * For the root cgroup, parent_mem is NULL, we allow value to be
3133 * set if there are no children.
3135 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3136 (val == 1 || val == 0)) {
3137 if (!memcg_has_children(memcg))
3138 memcg->use_hierarchy = val;
3147 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3151 if (mem_cgroup_is_root(memcg)) {
3152 val = memcg_page_state(memcg, MEMCG_CACHE) +
3153 memcg_page_state(memcg, MEMCG_RSS);
3155 val += memcg_page_state(memcg, MEMCG_SWAP);
3158 val = page_counter_read(&memcg->memory);
3160 val = page_counter_read(&memcg->memsw);
3173 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3176 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3177 struct page_counter *counter;
3179 switch (MEMFILE_TYPE(cft->private)) {
3181 counter = &memcg->memory;
3184 counter = &memcg->memsw;
3187 counter = &memcg->kmem;
3190 counter = &memcg->tcpmem;
3196 switch (MEMFILE_ATTR(cft->private)) {
3198 if (counter == &memcg->memory)
3199 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3200 if (counter == &memcg->memsw)
3201 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3202 return (u64)page_counter_read(counter) * PAGE_SIZE;
3204 return (u64)counter->max * PAGE_SIZE;
3206 return (u64)counter->watermark * PAGE_SIZE;
3208 return counter->failcnt;
3209 case RES_SOFT_LIMIT:
3210 return (u64)memcg->soft_limit * PAGE_SIZE;
3216 #ifdef CONFIG_MEMCG_KMEM
3217 static int memcg_online_kmem(struct mem_cgroup *memcg)
3221 if (cgroup_memory_nokmem)
3224 BUG_ON(memcg->kmemcg_id >= 0);
3225 BUG_ON(memcg->kmem_state);
3227 memcg_id = memcg_alloc_cache_id();
3231 static_branch_inc(&memcg_kmem_enabled_key);
3233 * A memory cgroup is considered kmem-online as soon as it gets
3234 * kmemcg_id. Setting the id after enabling static branching will
3235 * guarantee no one starts accounting before all call sites are
3238 memcg->kmemcg_id = memcg_id;
3239 memcg->kmem_state = KMEM_ONLINE;
3240 INIT_LIST_HEAD(&memcg->kmem_caches);
3245 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3247 struct cgroup_subsys_state *css;
3248 struct mem_cgroup *parent, *child;
3251 if (memcg->kmem_state != KMEM_ONLINE)
3254 * Clear the online state before clearing memcg_caches array
3255 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3256 * guarantees that no cache will be created for this cgroup
3257 * after we are done (see memcg_create_kmem_cache()).
3259 memcg->kmem_state = KMEM_ALLOCATED;
3261 parent = parent_mem_cgroup(memcg);
3263 parent = root_mem_cgroup;
3265 memcg_deactivate_kmem_caches(memcg, parent);
3267 kmemcg_id = memcg->kmemcg_id;
3268 BUG_ON(kmemcg_id < 0);
3271 * Change kmemcg_id of this cgroup and all its descendants to the
3272 * parent's id, and then move all entries from this cgroup's list_lrus
3273 * to ones of the parent. After we have finished, all list_lrus
3274 * corresponding to this cgroup are guaranteed to remain empty. The
3275 * ordering is imposed by list_lru_node->lock taken by
3276 * memcg_drain_all_list_lrus().
3278 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3279 css_for_each_descendant_pre(css, &memcg->css) {
3280 child = mem_cgroup_from_css(css);
3281 BUG_ON(child->kmemcg_id != kmemcg_id);
3282 child->kmemcg_id = parent->kmemcg_id;
3283 if (!memcg->use_hierarchy)
3288 memcg_drain_all_list_lrus(kmemcg_id, parent);
3290 memcg_free_cache_id(kmemcg_id);
3293 static void memcg_free_kmem(struct mem_cgroup *memcg)
3295 /* css_alloc() failed, offlining didn't happen */
3296 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3297 memcg_offline_kmem(memcg);
3299 if (memcg->kmem_state == KMEM_ALLOCATED) {
3300 WARN_ON(!list_empty(&memcg->kmem_caches));
3301 static_branch_dec(&memcg_kmem_enabled_key);
3305 static int memcg_online_kmem(struct mem_cgroup *memcg)
3309 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3312 static void memcg_free_kmem(struct mem_cgroup *memcg)
3315 #endif /* CONFIG_MEMCG_KMEM */
3317 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3322 mutex_lock(&memcg_max_mutex);
3323 ret = page_counter_set_max(&memcg->kmem, max);
3324 mutex_unlock(&memcg_max_mutex);
3328 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3332 mutex_lock(&memcg_max_mutex);
3334 ret = page_counter_set_max(&memcg->tcpmem, max);
3338 if (!memcg->tcpmem_active) {
3340 * The active flag needs to be written after the static_key
3341 * update. This is what guarantees that the socket activation
3342 * function is the last one to run. See mem_cgroup_sk_alloc()
3343 * for details, and note that we don't mark any socket as
3344 * belonging to this memcg until that flag is up.
3346 * We need to do this, because static_keys will span multiple
3347 * sites, but we can't control their order. If we mark a socket
3348 * as accounted, but the accounting functions are not patched in
3349 * yet, we'll lose accounting.
3351 * We never race with the readers in mem_cgroup_sk_alloc(),
3352 * because when this value change, the code to process it is not
3355 static_branch_inc(&memcg_sockets_enabled_key);
3356 memcg->tcpmem_active = true;
3359 mutex_unlock(&memcg_max_mutex);
3364 * The user of this function is...
3367 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3368 char *buf, size_t nbytes, loff_t off)
3370 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3371 unsigned long nr_pages;
3374 buf = strstrip(buf);
3375 ret = page_counter_memparse(buf, "-1", &nr_pages);
3379 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3381 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3385 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3387 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3390 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3393 ret = memcg_update_kmem_max(memcg, nr_pages);
3396 ret = memcg_update_tcp_max(memcg, nr_pages);
3400 case RES_SOFT_LIMIT:
3401 memcg->soft_limit = nr_pages;
3405 return ret ?: nbytes;
3408 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3409 size_t nbytes, loff_t off)
3411 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3412 struct page_counter *counter;
3414 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3416 counter = &memcg->memory;
3419 counter = &memcg->memsw;
3422 counter = &memcg->kmem;
3425 counter = &memcg->tcpmem;
3431 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3433 page_counter_reset_watermark(counter);
3436 counter->failcnt = 0;
3445 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3448 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3452 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3453 struct cftype *cft, u64 val)
3455 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3457 if (val & ~MOVE_MASK)
3461 * No kind of locking is needed in here, because ->can_attach() will
3462 * check this value once in the beginning of the process, and then carry
3463 * on with stale data. This means that changes to this value will only
3464 * affect task migrations starting after the change.
3466 memcg->move_charge_at_immigrate = val;
3470 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3471 struct cftype *cft, u64 val)
3479 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3480 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3481 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3483 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3484 int nid, unsigned int lru_mask)
3486 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3487 unsigned long nr = 0;
3490 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3493 if (!(BIT(lru) & lru_mask))
3495 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3500 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3501 unsigned int lru_mask)
3503 unsigned long nr = 0;
3507 if (!(BIT(lru) & lru_mask))
3509 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3514 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3518 unsigned int lru_mask;
3521 static const struct numa_stat stats[] = {
3522 { "total", LRU_ALL },
3523 { "file", LRU_ALL_FILE },
3524 { "anon", LRU_ALL_ANON },
3525 { "unevictable", BIT(LRU_UNEVICTABLE) },
3527 const struct numa_stat *stat;
3530 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3532 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3533 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3534 seq_printf(m, "%s=%lu", stat->name, nr);
3535 for_each_node_state(nid, N_MEMORY) {
3536 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3538 seq_printf(m, " N%d=%lu", nid, nr);
3543 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3544 struct mem_cgroup *iter;
3547 for_each_mem_cgroup_tree(iter, memcg)
3548 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3549 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3550 for_each_node_state(nid, N_MEMORY) {
3552 for_each_mem_cgroup_tree(iter, memcg)
3553 nr += mem_cgroup_node_nr_lru_pages(
3554 iter, nid, stat->lru_mask);
3555 seq_printf(m, " N%d=%lu", nid, nr);
3562 #endif /* CONFIG_NUMA */
3564 static const unsigned int memcg1_stats[] = {
3575 static const char *const memcg1_stat_names[] = {
3586 /* Universal VM events cgroup1 shows, original sort order */
3587 static const unsigned int memcg1_events[] = {
3594 static const char *const memcg1_event_names[] = {
3601 static int memcg_stat_show(struct seq_file *m, void *v)
3603 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3604 unsigned long memory, memsw;
3605 struct mem_cgroup *mi;
3608 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3609 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3611 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3612 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3614 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3615 memcg_page_state_local(memcg, memcg1_stats[i]) *
3619 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3620 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3621 memcg_events_local(memcg, memcg1_events[i]));
3623 for (i = 0; i < NR_LRU_LISTS; i++)
3624 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3625 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3628 /* Hierarchical information */
3629 memory = memsw = PAGE_COUNTER_MAX;
3630 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3631 memory = min(memory, mi->memory.max);
3632 memsw = min(memsw, mi->memsw.max);
3634 seq_printf(m, "hierarchical_memory_limit %llu\n",
3635 (u64)memory * PAGE_SIZE);
3636 if (do_memsw_account())
3637 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3638 (u64)memsw * PAGE_SIZE);
3640 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3641 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3643 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3644 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3648 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3649 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3650 (u64)memcg_events(memcg, memcg1_events[i]));
3652 for (i = 0; i < NR_LRU_LISTS; i++)
3653 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3654 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3657 #ifdef CONFIG_DEBUG_VM
3660 struct mem_cgroup_per_node *mz;
3661 struct zone_reclaim_stat *rstat;
3662 unsigned long recent_rotated[2] = {0, 0};
3663 unsigned long recent_scanned[2] = {0, 0};
3665 for_each_online_pgdat(pgdat) {
3666 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3667 rstat = &mz->lruvec.reclaim_stat;
3669 recent_rotated[0] += rstat->recent_rotated[0];
3670 recent_rotated[1] += rstat->recent_rotated[1];
3671 recent_scanned[0] += rstat->recent_scanned[0];
3672 recent_scanned[1] += rstat->recent_scanned[1];
3674 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3675 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3676 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3677 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3684 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3687 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3689 return mem_cgroup_swappiness(memcg);
3692 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3693 struct cftype *cft, u64 val)
3695 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3701 memcg->swappiness = val;
3703 vm_swappiness = val;
3708 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3710 struct mem_cgroup_threshold_ary *t;
3711 unsigned long usage;
3716 t = rcu_dereference(memcg->thresholds.primary);
3718 t = rcu_dereference(memcg->memsw_thresholds.primary);
3723 usage = mem_cgroup_usage(memcg, swap);
3726 * current_threshold points to threshold just below or equal to usage.
3727 * If it's not true, a threshold was crossed after last
3728 * call of __mem_cgroup_threshold().
3730 i = t->current_threshold;
3733 * Iterate backward over array of thresholds starting from
3734 * current_threshold and check if a threshold is crossed.
3735 * If none of thresholds below usage is crossed, we read
3736 * only one element of the array here.
3738 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3739 eventfd_signal(t->entries[i].eventfd, 1);
3741 /* i = current_threshold + 1 */
3745 * Iterate forward over array of thresholds starting from
3746 * current_threshold+1 and check if a threshold is crossed.
3747 * If none of thresholds above usage is crossed, we read
3748 * only one element of the array here.
3750 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3751 eventfd_signal(t->entries[i].eventfd, 1);
3753 /* Update current_threshold */
3754 t->current_threshold = i - 1;
3759 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3762 __mem_cgroup_threshold(memcg, false);
3763 if (do_memsw_account())
3764 __mem_cgroup_threshold(memcg, true);
3766 memcg = parent_mem_cgroup(memcg);
3770 static int compare_thresholds(const void *a, const void *b)
3772 const struct mem_cgroup_threshold *_a = a;
3773 const struct mem_cgroup_threshold *_b = b;
3775 if (_a->threshold > _b->threshold)
3778 if (_a->threshold < _b->threshold)
3784 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3786 struct mem_cgroup_eventfd_list *ev;
3788 spin_lock(&memcg_oom_lock);
3790 list_for_each_entry(ev, &memcg->oom_notify, list)
3791 eventfd_signal(ev->eventfd, 1);
3793 spin_unlock(&memcg_oom_lock);
3797 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3799 struct mem_cgroup *iter;
3801 for_each_mem_cgroup_tree(iter, memcg)
3802 mem_cgroup_oom_notify_cb(iter);
3805 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3806 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3808 struct mem_cgroup_thresholds *thresholds;
3809 struct mem_cgroup_threshold_ary *new;
3810 unsigned long threshold;
3811 unsigned long usage;
3814 ret = page_counter_memparse(args, "-1", &threshold);
3818 mutex_lock(&memcg->thresholds_lock);
3821 thresholds = &memcg->thresholds;
3822 usage = mem_cgroup_usage(memcg, false);
3823 } else if (type == _MEMSWAP) {
3824 thresholds = &memcg->memsw_thresholds;
3825 usage = mem_cgroup_usage(memcg, true);
3829 /* Check if a threshold crossed before adding a new one */
3830 if (thresholds->primary)
3831 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3833 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3835 /* Allocate memory for new array of thresholds */
3836 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3843 /* Copy thresholds (if any) to new array */
3844 if (thresholds->primary) {
3845 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3846 sizeof(struct mem_cgroup_threshold));
3849 /* Add new threshold */
3850 new->entries[size - 1].eventfd = eventfd;
3851 new->entries[size - 1].threshold = threshold;
3853 /* Sort thresholds. Registering of new threshold isn't time-critical */
3854 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3855 compare_thresholds, NULL);
3857 /* Find current threshold */
3858 new->current_threshold = -1;
3859 for (i = 0; i < size; i++) {
3860 if (new->entries[i].threshold <= usage) {
3862 * new->current_threshold will not be used until
3863 * rcu_assign_pointer(), so it's safe to increment
3866 ++new->current_threshold;
3871 /* Free old spare buffer and save old primary buffer as spare */
3872 kfree(thresholds->spare);
3873 thresholds->spare = thresholds->primary;
3875 rcu_assign_pointer(thresholds->primary, new);
3877 /* To be sure that nobody uses thresholds */
3881 mutex_unlock(&memcg->thresholds_lock);
3886 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3887 struct eventfd_ctx *eventfd, const char *args)
3889 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3892 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3893 struct eventfd_ctx *eventfd, const char *args)
3895 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3898 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3899 struct eventfd_ctx *eventfd, enum res_type type)
3901 struct mem_cgroup_thresholds *thresholds;
3902 struct mem_cgroup_threshold_ary *new;
3903 unsigned long usage;
3906 mutex_lock(&memcg->thresholds_lock);
3909 thresholds = &memcg->thresholds;
3910 usage = mem_cgroup_usage(memcg, false);
3911 } else if (type == _MEMSWAP) {
3912 thresholds = &memcg->memsw_thresholds;
3913 usage = mem_cgroup_usage(memcg, true);
3917 if (!thresholds->primary)
3920 /* Check if a threshold crossed before removing */
3921 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3923 /* Calculate new number of threshold */
3925 for (i = 0; i < thresholds->primary->size; i++) {
3926 if (thresholds->primary->entries[i].eventfd != eventfd)
3930 new = thresholds->spare;
3932 /* Set thresholds array to NULL if we don't have thresholds */
3941 /* Copy thresholds and find current threshold */
3942 new->current_threshold = -1;
3943 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3944 if (thresholds->primary->entries[i].eventfd == eventfd)
3947 new->entries[j] = thresholds->primary->entries[i];
3948 if (new->entries[j].threshold <= usage) {
3950 * new->current_threshold will not be used
3951 * until rcu_assign_pointer(), so it's safe to increment
3954 ++new->current_threshold;
3960 /* Swap primary and spare array */
3961 thresholds->spare = thresholds->primary;
3963 rcu_assign_pointer(thresholds->primary, new);
3965 /* To be sure that nobody uses thresholds */
3968 /* If all events are unregistered, free the spare array */
3970 kfree(thresholds->spare);
3971 thresholds->spare = NULL;
3974 mutex_unlock(&memcg->thresholds_lock);
3977 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3978 struct eventfd_ctx *eventfd)
3980 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3983 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3984 struct eventfd_ctx *eventfd)
3986 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3989 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3990 struct eventfd_ctx *eventfd, const char *args)
3992 struct mem_cgroup_eventfd_list *event;
3994 event = kmalloc(sizeof(*event), GFP_KERNEL);
3998 spin_lock(&memcg_oom_lock);
4000 event->eventfd = eventfd;
4001 list_add(&event->list, &memcg->oom_notify);
4003 /* already in OOM ? */
4004 if (memcg->under_oom)
4005 eventfd_signal(eventfd, 1);
4006 spin_unlock(&memcg_oom_lock);
4011 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4012 struct eventfd_ctx *eventfd)
4014 struct mem_cgroup_eventfd_list *ev, *tmp;
4016 spin_lock(&memcg_oom_lock);
4018 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4019 if (ev->eventfd == eventfd) {
4020 list_del(&ev->list);
4025 spin_unlock(&memcg_oom_lock);
4028 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4030 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4032 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4033 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4034 seq_printf(sf, "oom_kill %lu\n",
4035 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4039 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4040 struct cftype *cft, u64 val)
4042 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4044 /* cannot set to root cgroup and only 0 and 1 are allowed */
4045 if (!css->parent || !((val == 0) || (val == 1)))
4048 memcg->oom_kill_disable = val;
4050 memcg_oom_recover(memcg);
4055 #ifdef CONFIG_CGROUP_WRITEBACK
4057 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4059 return wb_domain_init(&memcg->cgwb_domain, gfp);
4062 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4064 wb_domain_exit(&memcg->cgwb_domain);
4067 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4069 wb_domain_size_changed(&memcg->cgwb_domain);
4072 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4074 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4076 if (!memcg->css.parent)
4079 return &memcg->cgwb_domain;
4083 * idx can be of type enum memcg_stat_item or node_stat_item.
4084 * Keep in sync with memcg_exact_page().
4086 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4088 long x = atomic_long_read(&memcg->vmstats[idx]);
4091 for_each_online_cpu(cpu)
4092 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4099 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4100 * @wb: bdi_writeback in question
4101 * @pfilepages: out parameter for number of file pages
4102 * @pheadroom: out parameter for number of allocatable pages according to memcg
4103 * @pdirty: out parameter for number of dirty pages
4104 * @pwriteback: out parameter for number of pages under writeback
4106 * Determine the numbers of file, headroom, dirty, and writeback pages in
4107 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4108 * is a bit more involved.
4110 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4111 * headroom is calculated as the lowest headroom of itself and the
4112 * ancestors. Note that this doesn't consider the actual amount of
4113 * available memory in the system. The caller should further cap
4114 * *@pheadroom accordingly.
4116 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4117 unsigned long *pheadroom, unsigned long *pdirty,
4118 unsigned long *pwriteback)
4120 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4121 struct mem_cgroup *parent;
4123 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4125 /* this should eventually include NR_UNSTABLE_NFS */
4126 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4127 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4128 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4129 *pheadroom = PAGE_COUNTER_MAX;
4131 while ((parent = parent_mem_cgroup(memcg))) {
4132 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4133 unsigned long used = page_counter_read(&memcg->memory);
4135 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4140 #else /* CONFIG_CGROUP_WRITEBACK */
4142 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4147 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4151 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4155 #endif /* CONFIG_CGROUP_WRITEBACK */
4158 * DO NOT USE IN NEW FILES.
4160 * "cgroup.event_control" implementation.
4162 * This is way over-engineered. It tries to support fully configurable
4163 * events for each user. Such level of flexibility is completely
4164 * unnecessary especially in the light of the planned unified hierarchy.
4166 * Please deprecate this and replace with something simpler if at all
4171 * Unregister event and free resources.
4173 * Gets called from workqueue.
4175 static void memcg_event_remove(struct work_struct *work)
4177 struct mem_cgroup_event *event =
4178 container_of(work, struct mem_cgroup_event, remove);
4179 struct mem_cgroup *memcg = event->memcg;
4181 remove_wait_queue(event->wqh, &event->wait);
4183 event->unregister_event(memcg, event->eventfd);
4185 /* Notify userspace the event is going away. */
4186 eventfd_signal(event->eventfd, 1);
4188 eventfd_ctx_put(event->eventfd);
4190 css_put(&memcg->css);
4194 * Gets called on EPOLLHUP on eventfd when user closes it.
4196 * Called with wqh->lock held and interrupts disabled.
4198 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4199 int sync, void *key)
4201 struct mem_cgroup_event *event =
4202 container_of(wait, struct mem_cgroup_event, wait);
4203 struct mem_cgroup *memcg = event->memcg;
4204 __poll_t flags = key_to_poll(key);
4206 if (flags & EPOLLHUP) {
4208 * If the event has been detached at cgroup removal, we
4209 * can simply return knowing the other side will cleanup
4212 * We can't race against event freeing since the other
4213 * side will require wqh->lock via remove_wait_queue(),
4216 spin_lock(&memcg->event_list_lock);
4217 if (!list_empty(&event->list)) {
4218 list_del_init(&event->list);
4220 * We are in atomic context, but cgroup_event_remove()
4221 * may sleep, so we have to call it in workqueue.
4223 schedule_work(&event->remove);
4225 spin_unlock(&memcg->event_list_lock);
4231 static void memcg_event_ptable_queue_proc(struct file *file,
4232 wait_queue_head_t *wqh, poll_table *pt)
4234 struct mem_cgroup_event *event =
4235 container_of(pt, struct mem_cgroup_event, pt);
4238 add_wait_queue(wqh, &event->wait);
4242 * DO NOT USE IN NEW FILES.
4244 * Parse input and register new cgroup event handler.
4246 * Input must be in format '<event_fd> <control_fd> <args>'.
4247 * Interpretation of args is defined by control file implementation.
4249 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4250 char *buf, size_t nbytes, loff_t off)
4252 struct cgroup_subsys_state *css = of_css(of);
4253 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4254 struct mem_cgroup_event *event;
4255 struct cgroup_subsys_state *cfile_css;
4256 unsigned int efd, cfd;
4263 buf = strstrip(buf);
4265 efd = simple_strtoul(buf, &endp, 10);
4270 cfd = simple_strtoul(buf, &endp, 10);
4271 if ((*endp != ' ') && (*endp != '\0'))
4275 event = kzalloc(sizeof(*event), GFP_KERNEL);
4279 event->memcg = memcg;
4280 INIT_LIST_HEAD(&event->list);
4281 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4282 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4283 INIT_WORK(&event->remove, memcg_event_remove);
4291 event->eventfd = eventfd_ctx_fileget(efile.file);
4292 if (IS_ERR(event->eventfd)) {
4293 ret = PTR_ERR(event->eventfd);
4300 goto out_put_eventfd;
4303 /* the process need read permission on control file */
4304 /* AV: shouldn't we check that it's been opened for read instead? */
4305 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4310 * Determine the event callbacks and set them in @event. This used
4311 * to be done via struct cftype but cgroup core no longer knows
4312 * about these events. The following is crude but the whole thing
4313 * is for compatibility anyway.
4315 * DO NOT ADD NEW FILES.
4317 name = cfile.file->f_path.dentry->d_name.name;
4319 if (!strcmp(name, "memory.usage_in_bytes")) {
4320 event->register_event = mem_cgroup_usage_register_event;
4321 event->unregister_event = mem_cgroup_usage_unregister_event;
4322 } else if (!strcmp(name, "memory.oom_control")) {
4323 event->register_event = mem_cgroup_oom_register_event;
4324 event->unregister_event = mem_cgroup_oom_unregister_event;
4325 } else if (!strcmp(name, "memory.pressure_level")) {
4326 event->register_event = vmpressure_register_event;
4327 event->unregister_event = vmpressure_unregister_event;
4328 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4329 event->register_event = memsw_cgroup_usage_register_event;
4330 event->unregister_event = memsw_cgroup_usage_unregister_event;
4337 * Verify @cfile should belong to @css. Also, remaining events are
4338 * automatically removed on cgroup destruction but the removal is
4339 * asynchronous, so take an extra ref on @css.
4341 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4342 &memory_cgrp_subsys);
4344 if (IS_ERR(cfile_css))
4346 if (cfile_css != css) {
4351 ret = event->register_event(memcg, event->eventfd, buf);
4355 vfs_poll(efile.file, &event->pt);
4357 spin_lock(&memcg->event_list_lock);
4358 list_add(&event->list, &memcg->event_list);
4359 spin_unlock(&memcg->event_list_lock);
4371 eventfd_ctx_put(event->eventfd);
4380 static struct cftype mem_cgroup_legacy_files[] = {
4382 .name = "usage_in_bytes",
4383 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4384 .read_u64 = mem_cgroup_read_u64,
4387 .name = "max_usage_in_bytes",
4388 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4389 .write = mem_cgroup_reset,
4390 .read_u64 = mem_cgroup_read_u64,
4393 .name = "limit_in_bytes",
4394 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4395 .write = mem_cgroup_write,
4396 .read_u64 = mem_cgroup_read_u64,
4399 .name = "soft_limit_in_bytes",
4400 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4401 .write = mem_cgroup_write,
4402 .read_u64 = mem_cgroup_read_u64,
4406 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4407 .write = mem_cgroup_reset,
4408 .read_u64 = mem_cgroup_read_u64,
4412 .seq_show = memcg_stat_show,
4415 .name = "force_empty",
4416 .write = mem_cgroup_force_empty_write,
4419 .name = "use_hierarchy",
4420 .write_u64 = mem_cgroup_hierarchy_write,
4421 .read_u64 = mem_cgroup_hierarchy_read,
4424 .name = "cgroup.event_control", /* XXX: for compat */
4425 .write = memcg_write_event_control,
4426 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4429 .name = "swappiness",
4430 .read_u64 = mem_cgroup_swappiness_read,
4431 .write_u64 = mem_cgroup_swappiness_write,
4434 .name = "move_charge_at_immigrate",
4435 .read_u64 = mem_cgroup_move_charge_read,
4436 .write_u64 = mem_cgroup_move_charge_write,
4439 .name = "oom_control",
4440 .seq_show = mem_cgroup_oom_control_read,
4441 .write_u64 = mem_cgroup_oom_control_write,
4442 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4445 .name = "pressure_level",
4449 .name = "numa_stat",
4450 .seq_show = memcg_numa_stat_show,
4454 .name = "kmem.limit_in_bytes",
4455 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4456 .write = mem_cgroup_write,
4457 .read_u64 = mem_cgroup_read_u64,
4460 .name = "kmem.usage_in_bytes",
4461 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4462 .read_u64 = mem_cgroup_read_u64,
4465 .name = "kmem.failcnt",
4466 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4467 .write = mem_cgroup_reset,
4468 .read_u64 = mem_cgroup_read_u64,
4471 .name = "kmem.max_usage_in_bytes",
4472 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4473 .write = mem_cgroup_reset,
4474 .read_u64 = mem_cgroup_read_u64,
4476 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4478 .name = "kmem.slabinfo",
4479 .seq_start = memcg_slab_start,
4480 .seq_next = memcg_slab_next,
4481 .seq_stop = memcg_slab_stop,
4482 .seq_show = memcg_slab_show,
4486 .name = "kmem.tcp.limit_in_bytes",
4487 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4488 .write = mem_cgroup_write,
4489 .read_u64 = mem_cgroup_read_u64,
4492 .name = "kmem.tcp.usage_in_bytes",
4493 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4494 .read_u64 = mem_cgroup_read_u64,
4497 .name = "kmem.tcp.failcnt",
4498 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4499 .write = mem_cgroup_reset,
4500 .read_u64 = mem_cgroup_read_u64,
4503 .name = "kmem.tcp.max_usage_in_bytes",
4504 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4505 .write = mem_cgroup_reset,
4506 .read_u64 = mem_cgroup_read_u64,
4508 { }, /* terminate */
4512 * Private memory cgroup IDR
4514 * Swap-out records and page cache shadow entries need to store memcg
4515 * references in constrained space, so we maintain an ID space that is
4516 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4517 * memory-controlled cgroups to 64k.
4519 * However, there usually are many references to the oflline CSS after
4520 * the cgroup has been destroyed, such as page cache or reclaimable
4521 * slab objects, that don't need to hang on to the ID. We want to keep
4522 * those dead CSS from occupying IDs, or we might quickly exhaust the
4523 * relatively small ID space and prevent the creation of new cgroups
4524 * even when there are much fewer than 64k cgroups - possibly none.
4526 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4527 * be freed and recycled when it's no longer needed, which is usually
4528 * when the CSS is offlined.
4530 * The only exception to that are records of swapped out tmpfs/shmem
4531 * pages that need to be attributed to live ancestors on swapin. But
4532 * those references are manageable from userspace.
4535 static DEFINE_IDR(mem_cgroup_idr);
4537 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4539 if (memcg->id.id > 0) {
4540 idr_remove(&mem_cgroup_idr, memcg->id.id);
4545 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4547 refcount_add(n, &memcg->id.ref);
4550 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4552 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4553 mem_cgroup_id_remove(memcg);
4555 /* Memcg ID pins CSS */
4556 css_put(&memcg->css);
4560 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4562 mem_cgroup_id_get_many(memcg, 1);
4565 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4567 mem_cgroup_id_put_many(memcg, 1);
4571 * mem_cgroup_from_id - look up a memcg from a memcg id
4572 * @id: the memcg id to look up
4574 * Caller must hold rcu_read_lock().
4576 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4578 WARN_ON_ONCE(!rcu_read_lock_held());
4579 return idr_find(&mem_cgroup_idr, id);
4582 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4584 struct mem_cgroup_per_node *pn;
4587 * This routine is called against possible nodes.
4588 * But it's BUG to call kmalloc() against offline node.
4590 * TODO: this routine can waste much memory for nodes which will
4591 * never be onlined. It's better to use memory hotplug callback
4594 if (!node_state(node, N_NORMAL_MEMORY))
4596 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4600 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4601 if (!pn->lruvec_stat_local) {
4606 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4607 if (!pn->lruvec_stat_cpu) {
4608 free_percpu(pn->lruvec_stat_local);
4613 lruvec_init(&pn->lruvec);
4614 pn->usage_in_excess = 0;
4615 pn->on_tree = false;
4618 memcg->nodeinfo[node] = pn;
4622 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4624 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4629 free_percpu(pn->lruvec_stat_cpu);
4630 free_percpu(pn->lruvec_stat_local);
4634 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4639 free_mem_cgroup_per_node_info(memcg, node);
4640 free_percpu(memcg->vmstats_percpu);
4641 free_percpu(memcg->vmstats_local);
4645 static void mem_cgroup_free(struct mem_cgroup *memcg)
4647 memcg_wb_domain_exit(memcg);
4648 __mem_cgroup_free(memcg);
4651 static struct mem_cgroup *mem_cgroup_alloc(void)
4653 struct mem_cgroup *memcg;
4657 size = sizeof(struct mem_cgroup);
4658 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4660 memcg = kzalloc(size, GFP_KERNEL);
4664 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4665 1, MEM_CGROUP_ID_MAX,
4667 if (memcg->id.id < 0)
4670 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
4671 if (!memcg->vmstats_local)
4674 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4675 if (!memcg->vmstats_percpu)
4679 if (alloc_mem_cgroup_per_node_info(memcg, node))
4682 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4685 INIT_WORK(&memcg->high_work, high_work_func);
4686 memcg->last_scanned_node = MAX_NUMNODES;
4687 INIT_LIST_HEAD(&memcg->oom_notify);
4688 mutex_init(&memcg->thresholds_lock);
4689 spin_lock_init(&memcg->move_lock);
4690 vmpressure_init(&memcg->vmpressure);
4691 INIT_LIST_HEAD(&memcg->event_list);
4692 spin_lock_init(&memcg->event_list_lock);
4693 memcg->socket_pressure = jiffies;
4694 #ifdef CONFIG_MEMCG_KMEM
4695 memcg->kmemcg_id = -1;
4697 #ifdef CONFIG_CGROUP_WRITEBACK
4698 INIT_LIST_HEAD(&memcg->cgwb_list);
4700 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4703 mem_cgroup_id_remove(memcg);
4704 __mem_cgroup_free(memcg);
4708 static struct cgroup_subsys_state * __ref
4709 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4711 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4712 struct mem_cgroup *memcg;
4713 long error = -ENOMEM;
4715 memcg = mem_cgroup_alloc();
4717 return ERR_PTR(error);
4719 memcg->high = PAGE_COUNTER_MAX;
4720 memcg->soft_limit = PAGE_COUNTER_MAX;
4722 memcg->swappiness = mem_cgroup_swappiness(parent);
4723 memcg->oom_kill_disable = parent->oom_kill_disable;
4725 if (parent && parent->use_hierarchy) {
4726 memcg->use_hierarchy = true;
4727 page_counter_init(&memcg->memory, &parent->memory);
4728 page_counter_init(&memcg->swap, &parent->swap);
4729 page_counter_init(&memcg->memsw, &parent->memsw);
4730 page_counter_init(&memcg->kmem, &parent->kmem);
4731 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4733 page_counter_init(&memcg->memory, NULL);
4734 page_counter_init(&memcg->swap, NULL);
4735 page_counter_init(&memcg->memsw, NULL);
4736 page_counter_init(&memcg->kmem, NULL);
4737 page_counter_init(&memcg->tcpmem, NULL);
4739 * Deeper hierachy with use_hierarchy == false doesn't make
4740 * much sense so let cgroup subsystem know about this
4741 * unfortunate state in our controller.
4743 if (parent != root_mem_cgroup)
4744 memory_cgrp_subsys.broken_hierarchy = true;
4747 /* The following stuff does not apply to the root */
4749 #ifdef CONFIG_MEMCG_KMEM
4750 INIT_LIST_HEAD(&memcg->kmem_caches);
4752 root_mem_cgroup = memcg;
4756 error = memcg_online_kmem(memcg);
4760 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4761 static_branch_inc(&memcg_sockets_enabled_key);
4765 mem_cgroup_id_remove(memcg);
4766 mem_cgroup_free(memcg);
4767 return ERR_PTR(-ENOMEM);
4770 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4772 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4775 * A memcg must be visible for memcg_expand_shrinker_maps()
4776 * by the time the maps are allocated. So, we allocate maps
4777 * here, when for_each_mem_cgroup() can't skip it.
4779 if (memcg_alloc_shrinker_maps(memcg)) {
4780 mem_cgroup_id_remove(memcg);
4784 /* Online state pins memcg ID, memcg ID pins CSS */
4785 refcount_set(&memcg->id.ref, 1);
4790 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4792 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4793 struct mem_cgroup_event *event, *tmp;
4796 * Unregister events and notify userspace.
4797 * Notify userspace about cgroup removing only after rmdir of cgroup
4798 * directory to avoid race between userspace and kernelspace.
4800 spin_lock(&memcg->event_list_lock);
4801 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4802 list_del_init(&event->list);
4803 schedule_work(&event->remove);
4805 spin_unlock(&memcg->event_list_lock);
4807 page_counter_set_min(&memcg->memory, 0);
4808 page_counter_set_low(&memcg->memory, 0);
4810 memcg_offline_kmem(memcg);
4811 wb_memcg_offline(memcg);
4813 drain_all_stock(memcg);
4815 mem_cgroup_id_put(memcg);
4818 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4820 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4822 invalidate_reclaim_iterators(memcg);
4825 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4827 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4829 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4830 static_branch_dec(&memcg_sockets_enabled_key);
4832 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4833 static_branch_dec(&memcg_sockets_enabled_key);
4835 vmpressure_cleanup(&memcg->vmpressure);
4836 cancel_work_sync(&memcg->high_work);
4837 mem_cgroup_remove_from_trees(memcg);
4838 memcg_free_shrinker_maps(memcg);
4839 memcg_free_kmem(memcg);
4840 mem_cgroup_free(memcg);
4844 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4845 * @css: the target css
4847 * Reset the states of the mem_cgroup associated with @css. This is
4848 * invoked when the userland requests disabling on the default hierarchy
4849 * but the memcg is pinned through dependency. The memcg should stop
4850 * applying policies and should revert to the vanilla state as it may be
4851 * made visible again.
4853 * The current implementation only resets the essential configurations.
4854 * This needs to be expanded to cover all the visible parts.
4856 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4858 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4860 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4861 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4862 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4863 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4864 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4865 page_counter_set_min(&memcg->memory, 0);
4866 page_counter_set_low(&memcg->memory, 0);
4867 memcg->high = PAGE_COUNTER_MAX;
4868 memcg->soft_limit = PAGE_COUNTER_MAX;
4869 memcg_wb_domain_size_changed(memcg);
4873 /* Handlers for move charge at task migration. */
4874 static int mem_cgroup_do_precharge(unsigned long count)
4878 /* Try a single bulk charge without reclaim first, kswapd may wake */
4879 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4881 mc.precharge += count;
4885 /* Try charges one by one with reclaim, but do not retry */
4887 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4901 enum mc_target_type {
4908 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4909 unsigned long addr, pte_t ptent)
4911 struct page *page = vm_normal_page(vma, addr, ptent);
4913 if (!page || !page_mapped(page))
4915 if (PageAnon(page)) {
4916 if (!(mc.flags & MOVE_ANON))
4919 if (!(mc.flags & MOVE_FILE))
4922 if (!get_page_unless_zero(page))
4928 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4929 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4930 pte_t ptent, swp_entry_t *entry)
4932 struct page *page = NULL;
4933 swp_entry_t ent = pte_to_swp_entry(ptent);
4935 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4939 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4940 * a device and because they are not accessible by CPU they are store
4941 * as special swap entry in the CPU page table.
4943 if (is_device_private_entry(ent)) {
4944 page = device_private_entry_to_page(ent);
4946 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4947 * a refcount of 1 when free (unlike normal page)
4949 if (!page_ref_add_unless(page, 1, 1))
4955 * Because lookup_swap_cache() updates some statistics counter,
4956 * we call find_get_page() with swapper_space directly.
4958 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4959 if (do_memsw_account())
4960 entry->val = ent.val;
4965 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4966 pte_t ptent, swp_entry_t *entry)
4972 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4973 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4975 struct page *page = NULL;
4976 struct address_space *mapping;
4979 if (!vma->vm_file) /* anonymous vma */
4981 if (!(mc.flags & MOVE_FILE))
4984 mapping = vma->vm_file->f_mapping;
4985 pgoff = linear_page_index(vma, addr);
4987 /* page is moved even if it's not RSS of this task(page-faulted). */
4989 /* shmem/tmpfs may report page out on swap: account for that too. */
4990 if (shmem_mapping(mapping)) {
4991 page = find_get_entry(mapping, pgoff);
4992 if (xa_is_value(page)) {
4993 swp_entry_t swp = radix_to_swp_entry(page);
4994 if (do_memsw_account())
4996 page = find_get_page(swap_address_space(swp),
5000 page = find_get_page(mapping, pgoff);
5002 page = find_get_page(mapping, pgoff);
5008 * mem_cgroup_move_account - move account of the page
5010 * @compound: charge the page as compound or small page
5011 * @from: mem_cgroup which the page is moved from.
5012 * @to: mem_cgroup which the page is moved to. @from != @to.
5014 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5016 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5019 static int mem_cgroup_move_account(struct page *page,
5021 struct mem_cgroup *from,
5022 struct mem_cgroup *to)
5024 unsigned long flags;
5025 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5029 VM_BUG_ON(from == to);
5030 VM_BUG_ON_PAGE(PageLRU(page), page);
5031 VM_BUG_ON(compound && !PageTransHuge(page));
5034 * Prevent mem_cgroup_migrate() from looking at
5035 * page->mem_cgroup of its source page while we change it.
5038 if (!trylock_page(page))
5042 if (page->mem_cgroup != from)
5045 anon = PageAnon(page);
5047 spin_lock_irqsave(&from->move_lock, flags);
5049 if (!anon && page_mapped(page)) {
5050 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
5051 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
5055 * move_lock grabbed above and caller set from->moving_account, so
5056 * mod_memcg_page_state will serialize updates to PageDirty.
5057 * So mapping should be stable for dirty pages.
5059 if (!anon && PageDirty(page)) {
5060 struct address_space *mapping = page_mapping(page);
5062 if (mapping_cap_account_dirty(mapping)) {
5063 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
5064 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
5068 if (PageWriteback(page)) {
5069 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
5070 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
5074 * It is safe to change page->mem_cgroup here because the page
5075 * is referenced, charged, and isolated - we can't race with
5076 * uncharging, charging, migration, or LRU putback.
5079 /* caller should have done css_get */
5080 page->mem_cgroup = to;
5081 spin_unlock_irqrestore(&from->move_lock, flags);
5085 local_irq_disable();
5086 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5087 memcg_check_events(to, page);
5088 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5089 memcg_check_events(from, page);
5098 * get_mctgt_type - get target type of moving charge
5099 * @vma: the vma the pte to be checked belongs
5100 * @addr: the address corresponding to the pte to be checked
5101 * @ptent: the pte to be checked
5102 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5105 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5106 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5107 * move charge. if @target is not NULL, the page is stored in target->page
5108 * with extra refcnt got(Callers should handle it).
5109 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5110 * target for charge migration. if @target is not NULL, the entry is stored
5112 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5113 * (so ZONE_DEVICE page and thus not on the lru).
5114 * For now we such page is charge like a regular page would be as for all
5115 * intent and purposes it is just special memory taking the place of a
5118 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5120 * Called with pte lock held.
5123 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5124 unsigned long addr, pte_t ptent, union mc_target *target)
5126 struct page *page = NULL;
5127 enum mc_target_type ret = MC_TARGET_NONE;
5128 swp_entry_t ent = { .val = 0 };
5130 if (pte_present(ptent))
5131 page = mc_handle_present_pte(vma, addr, ptent);
5132 else if (is_swap_pte(ptent))
5133 page = mc_handle_swap_pte(vma, ptent, &ent);
5134 else if (pte_none(ptent))
5135 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5137 if (!page && !ent.val)
5141 * Do only loose check w/o serialization.
5142 * mem_cgroup_move_account() checks the page is valid or
5143 * not under LRU exclusion.
5145 if (page->mem_cgroup == mc.from) {
5146 ret = MC_TARGET_PAGE;
5147 if (is_device_private_page(page))
5148 ret = MC_TARGET_DEVICE;
5150 target->page = page;
5152 if (!ret || !target)
5156 * There is a swap entry and a page doesn't exist or isn't charged.
5157 * But we cannot move a tail-page in a THP.
5159 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5160 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5161 ret = MC_TARGET_SWAP;
5168 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5170 * We don't consider PMD mapped swapping or file mapped pages because THP does
5171 * not support them for now.
5172 * Caller should make sure that pmd_trans_huge(pmd) is true.
5174 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5175 unsigned long addr, pmd_t pmd, union mc_target *target)
5177 struct page *page = NULL;
5178 enum mc_target_type ret = MC_TARGET_NONE;
5180 if (unlikely(is_swap_pmd(pmd))) {
5181 VM_BUG_ON(thp_migration_supported() &&
5182 !is_pmd_migration_entry(pmd));
5185 page = pmd_page(pmd);
5186 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5187 if (!(mc.flags & MOVE_ANON))
5189 if (page->mem_cgroup == mc.from) {
5190 ret = MC_TARGET_PAGE;
5193 target->page = page;
5199 static inline 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 return MC_TARGET_NONE;
5206 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5207 unsigned long addr, unsigned long end,
5208 struct mm_walk *walk)
5210 struct vm_area_struct *vma = walk->vma;
5214 ptl = pmd_trans_huge_lock(pmd, vma);
5217 * Note their can not be MC_TARGET_DEVICE for now as we do not
5218 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5219 * this might change.
5221 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5222 mc.precharge += HPAGE_PMD_NR;
5227 if (pmd_trans_unstable(pmd))
5229 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5230 for (; addr != end; pte++, addr += PAGE_SIZE)
5231 if (get_mctgt_type(vma, addr, *pte, NULL))
5232 mc.precharge++; /* increment precharge temporarily */
5233 pte_unmap_unlock(pte - 1, ptl);
5239 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5241 unsigned long precharge;
5243 struct mm_walk mem_cgroup_count_precharge_walk = {
5244 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5247 down_read(&mm->mmap_sem);
5248 walk_page_range(0, mm->highest_vm_end,
5249 &mem_cgroup_count_precharge_walk);
5250 up_read(&mm->mmap_sem);
5252 precharge = mc.precharge;
5258 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5260 unsigned long precharge = mem_cgroup_count_precharge(mm);
5262 VM_BUG_ON(mc.moving_task);
5263 mc.moving_task = current;
5264 return mem_cgroup_do_precharge(precharge);
5267 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5268 static void __mem_cgroup_clear_mc(void)
5270 struct mem_cgroup *from = mc.from;
5271 struct mem_cgroup *to = mc.to;
5273 /* we must uncharge all the leftover precharges from mc.to */
5275 cancel_charge(mc.to, mc.precharge);
5279 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5280 * we must uncharge here.
5282 if (mc.moved_charge) {
5283 cancel_charge(mc.from, mc.moved_charge);
5284 mc.moved_charge = 0;
5286 /* we must fixup refcnts and charges */
5287 if (mc.moved_swap) {
5288 /* uncharge swap account from the old cgroup */
5289 if (!mem_cgroup_is_root(mc.from))
5290 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5292 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5295 * we charged both to->memory and to->memsw, so we
5296 * should uncharge to->memory.
5298 if (!mem_cgroup_is_root(mc.to))
5299 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5301 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5302 css_put_many(&mc.to->css, mc.moved_swap);
5306 memcg_oom_recover(from);
5307 memcg_oom_recover(to);
5308 wake_up_all(&mc.waitq);
5311 static void mem_cgroup_clear_mc(void)
5313 struct mm_struct *mm = mc.mm;
5316 * we must clear moving_task before waking up waiters at the end of
5319 mc.moving_task = NULL;
5320 __mem_cgroup_clear_mc();
5321 spin_lock(&mc.lock);
5325 spin_unlock(&mc.lock);
5330 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5332 struct cgroup_subsys_state *css;
5333 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5334 struct mem_cgroup *from;
5335 struct task_struct *leader, *p;
5336 struct mm_struct *mm;
5337 unsigned long move_flags;
5340 /* charge immigration isn't supported on the default hierarchy */
5341 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5345 * Multi-process migrations only happen on the default hierarchy
5346 * where charge immigration is not used. Perform charge
5347 * immigration if @tset contains a leader and whine if there are
5351 cgroup_taskset_for_each_leader(leader, css, tset) {
5354 memcg = mem_cgroup_from_css(css);
5360 * We are now commited to this value whatever it is. Changes in this
5361 * tunable will only affect upcoming migrations, not the current one.
5362 * So we need to save it, and keep it going.
5364 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5368 from = mem_cgroup_from_task(p);
5370 VM_BUG_ON(from == memcg);
5372 mm = get_task_mm(p);
5375 /* We move charges only when we move a owner of the mm */
5376 if (mm->owner == p) {
5379 VM_BUG_ON(mc.precharge);
5380 VM_BUG_ON(mc.moved_charge);
5381 VM_BUG_ON(mc.moved_swap);
5383 spin_lock(&mc.lock);
5387 mc.flags = move_flags;
5388 spin_unlock(&mc.lock);
5389 /* We set mc.moving_task later */
5391 ret = mem_cgroup_precharge_mc(mm);
5393 mem_cgroup_clear_mc();
5400 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5403 mem_cgroup_clear_mc();
5406 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5407 unsigned long addr, unsigned long end,
5408 struct mm_walk *walk)
5411 struct vm_area_struct *vma = walk->vma;
5414 enum mc_target_type target_type;
5415 union mc_target target;
5418 ptl = pmd_trans_huge_lock(pmd, vma);
5420 if (mc.precharge < HPAGE_PMD_NR) {
5424 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5425 if (target_type == MC_TARGET_PAGE) {
5427 if (!isolate_lru_page(page)) {
5428 if (!mem_cgroup_move_account(page, true,
5430 mc.precharge -= HPAGE_PMD_NR;
5431 mc.moved_charge += HPAGE_PMD_NR;
5433 putback_lru_page(page);
5436 } else if (target_type == MC_TARGET_DEVICE) {
5438 if (!mem_cgroup_move_account(page, true,
5440 mc.precharge -= HPAGE_PMD_NR;
5441 mc.moved_charge += HPAGE_PMD_NR;
5449 if (pmd_trans_unstable(pmd))
5452 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5453 for (; addr != end; addr += PAGE_SIZE) {
5454 pte_t ptent = *(pte++);
5455 bool device = false;
5461 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5462 case MC_TARGET_DEVICE:
5465 case MC_TARGET_PAGE:
5468 * We can have a part of the split pmd here. Moving it
5469 * can be done but it would be too convoluted so simply
5470 * ignore such a partial THP and keep it in original
5471 * memcg. There should be somebody mapping the head.
5473 if (PageTransCompound(page))
5475 if (!device && isolate_lru_page(page))
5477 if (!mem_cgroup_move_account(page, false,
5480 /* we uncharge from mc.from later. */
5484 putback_lru_page(page);
5485 put: /* get_mctgt_type() gets the page */
5488 case MC_TARGET_SWAP:
5490 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5492 /* we fixup refcnts and charges later. */
5500 pte_unmap_unlock(pte - 1, ptl);
5505 * We have consumed all precharges we got in can_attach().
5506 * We try charge one by one, but don't do any additional
5507 * charges to mc.to if we have failed in charge once in attach()
5510 ret = mem_cgroup_do_precharge(1);
5518 static void mem_cgroup_move_charge(void)
5520 struct mm_walk mem_cgroup_move_charge_walk = {
5521 .pmd_entry = mem_cgroup_move_charge_pte_range,
5525 lru_add_drain_all();
5527 * Signal lock_page_memcg() to take the memcg's move_lock
5528 * while we're moving its pages to another memcg. Then wait
5529 * for already started RCU-only updates to finish.
5531 atomic_inc(&mc.from->moving_account);
5534 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5536 * Someone who are holding the mmap_sem might be waiting in
5537 * waitq. So we cancel all extra charges, wake up all waiters,
5538 * and retry. Because we cancel precharges, we might not be able
5539 * to move enough charges, but moving charge is a best-effort
5540 * feature anyway, so it wouldn't be a big problem.
5542 __mem_cgroup_clear_mc();
5547 * When we have consumed all precharges and failed in doing
5548 * additional charge, the page walk just aborts.
5550 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5552 up_read(&mc.mm->mmap_sem);
5553 atomic_dec(&mc.from->moving_account);
5556 static void mem_cgroup_move_task(void)
5559 mem_cgroup_move_charge();
5560 mem_cgroup_clear_mc();
5563 #else /* !CONFIG_MMU */
5564 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5568 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5571 static void mem_cgroup_move_task(void)
5577 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5578 * to verify whether we're attached to the default hierarchy on each mount
5581 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5584 * use_hierarchy is forced on the default hierarchy. cgroup core
5585 * guarantees that @root doesn't have any children, so turning it
5586 * on for the root memcg is enough.
5588 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5589 root_mem_cgroup->use_hierarchy = true;
5591 root_mem_cgroup->use_hierarchy = false;
5594 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5596 if (value == PAGE_COUNTER_MAX)
5597 seq_puts(m, "max\n");
5599 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5604 static u64 memory_current_read(struct cgroup_subsys_state *css,
5607 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5609 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5612 static int memory_min_show(struct seq_file *m, void *v)
5614 return seq_puts_memcg_tunable(m,
5615 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5618 static ssize_t memory_min_write(struct kernfs_open_file *of,
5619 char *buf, size_t nbytes, loff_t off)
5621 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5625 buf = strstrip(buf);
5626 err = page_counter_memparse(buf, "max", &min);
5630 page_counter_set_min(&memcg->memory, min);
5635 static int memory_low_show(struct seq_file *m, void *v)
5637 return seq_puts_memcg_tunable(m,
5638 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5641 static ssize_t memory_low_write(struct kernfs_open_file *of,
5642 char *buf, size_t nbytes, loff_t off)
5644 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5648 buf = strstrip(buf);
5649 err = page_counter_memparse(buf, "max", &low);
5653 page_counter_set_low(&memcg->memory, low);
5658 static int memory_high_show(struct seq_file *m, void *v)
5660 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5663 static ssize_t memory_high_write(struct kernfs_open_file *of,
5664 char *buf, size_t nbytes, loff_t off)
5666 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5667 unsigned long nr_pages;
5671 buf = strstrip(buf);
5672 err = page_counter_memparse(buf, "max", &high);
5678 nr_pages = page_counter_read(&memcg->memory);
5679 if (nr_pages > high)
5680 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5683 memcg_wb_domain_size_changed(memcg);
5687 static int memory_max_show(struct seq_file *m, void *v)
5689 return seq_puts_memcg_tunable(m,
5690 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5693 static ssize_t memory_max_write(struct kernfs_open_file *of,
5694 char *buf, size_t nbytes, loff_t off)
5696 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5697 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5698 bool drained = false;
5702 buf = strstrip(buf);
5703 err = page_counter_memparse(buf, "max", &max);
5707 xchg(&memcg->memory.max, max);
5710 unsigned long nr_pages = page_counter_read(&memcg->memory);
5712 if (nr_pages <= max)
5715 if (signal_pending(current)) {
5721 drain_all_stock(memcg);
5727 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5733 memcg_memory_event(memcg, MEMCG_OOM);
5734 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5738 memcg_wb_domain_size_changed(memcg);
5742 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
5744 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
5745 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
5746 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
5747 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
5748 seq_printf(m, "oom_kill %lu\n",
5749 atomic_long_read(&events[MEMCG_OOM_KILL]));
5752 static int memory_events_show(struct seq_file *m, void *v)
5754 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5756 __memory_events_show(m, memcg->memory_events);
5760 static int memory_events_local_show(struct seq_file *m, void *v)
5762 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5764 __memory_events_show(m, memcg->memory_events_local);
5768 static int memory_stat_show(struct seq_file *m, void *v)
5770 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5773 buf = memory_stat_format(memcg);
5781 static int memory_oom_group_show(struct seq_file *m, void *v)
5783 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5785 seq_printf(m, "%d\n", memcg->oom_group);
5790 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5791 char *buf, size_t nbytes, loff_t off)
5793 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5796 buf = strstrip(buf);
5800 ret = kstrtoint(buf, 0, &oom_group);
5804 if (oom_group != 0 && oom_group != 1)
5807 memcg->oom_group = oom_group;
5812 static struct cftype memory_files[] = {
5815 .flags = CFTYPE_NOT_ON_ROOT,
5816 .read_u64 = memory_current_read,
5820 .flags = CFTYPE_NOT_ON_ROOT,
5821 .seq_show = memory_min_show,
5822 .write = memory_min_write,
5826 .flags = CFTYPE_NOT_ON_ROOT,
5827 .seq_show = memory_low_show,
5828 .write = memory_low_write,
5832 .flags = CFTYPE_NOT_ON_ROOT,
5833 .seq_show = memory_high_show,
5834 .write = memory_high_write,
5838 .flags = CFTYPE_NOT_ON_ROOT,
5839 .seq_show = memory_max_show,
5840 .write = memory_max_write,
5844 .flags = CFTYPE_NOT_ON_ROOT,
5845 .file_offset = offsetof(struct mem_cgroup, events_file),
5846 .seq_show = memory_events_show,
5849 .name = "events.local",
5850 .flags = CFTYPE_NOT_ON_ROOT,
5851 .file_offset = offsetof(struct mem_cgroup, events_local_file),
5852 .seq_show = memory_events_local_show,
5856 .flags = CFTYPE_NOT_ON_ROOT,
5857 .seq_show = memory_stat_show,
5860 .name = "oom.group",
5861 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5862 .seq_show = memory_oom_group_show,
5863 .write = memory_oom_group_write,
5868 struct cgroup_subsys memory_cgrp_subsys = {
5869 .css_alloc = mem_cgroup_css_alloc,
5870 .css_online = mem_cgroup_css_online,
5871 .css_offline = mem_cgroup_css_offline,
5872 .css_released = mem_cgroup_css_released,
5873 .css_free = mem_cgroup_css_free,
5874 .css_reset = mem_cgroup_css_reset,
5875 .can_attach = mem_cgroup_can_attach,
5876 .cancel_attach = mem_cgroup_cancel_attach,
5877 .post_attach = mem_cgroup_move_task,
5878 .bind = mem_cgroup_bind,
5879 .dfl_cftypes = memory_files,
5880 .legacy_cftypes = mem_cgroup_legacy_files,
5885 * mem_cgroup_protected - check if memory consumption is in the normal range
5886 * @root: the top ancestor of the sub-tree being checked
5887 * @memcg: the memory cgroup to check
5889 * WARNING: This function is not stateless! It can only be used as part
5890 * of a top-down tree iteration, not for isolated queries.
5892 * Returns one of the following:
5893 * MEMCG_PROT_NONE: cgroup memory is not protected
5894 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5895 * an unprotected supply of reclaimable memory from other cgroups.
5896 * MEMCG_PROT_MIN: cgroup memory is protected
5898 * @root is exclusive; it is never protected when looked at directly
5900 * To provide a proper hierarchical behavior, effective memory.min/low values
5901 * are used. Below is the description of how effective memory.low is calculated.
5902 * Effective memory.min values is calculated in the same way.
5904 * Effective memory.low is always equal or less than the original memory.low.
5905 * If there is no memory.low overcommittment (which is always true for
5906 * top-level memory cgroups), these two values are equal.
5907 * Otherwise, it's a part of parent's effective memory.low,
5908 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5909 * memory.low usages, where memory.low usage is the size of actually
5913 * elow = min( memory.low, parent->elow * ------------------ ),
5914 * siblings_low_usage
5916 * | memory.current, if memory.current < memory.low
5921 * Such definition of the effective memory.low provides the expected
5922 * hierarchical behavior: parent's memory.low value is limiting
5923 * children, unprotected memory is reclaimed first and cgroups,
5924 * which are not using their guarantee do not affect actual memory
5927 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5929 * A A/memory.low = 2G, A/memory.current = 6G
5931 * BC DE B/memory.low = 3G B/memory.current = 2G
5932 * C/memory.low = 1G C/memory.current = 2G
5933 * D/memory.low = 0 D/memory.current = 2G
5934 * E/memory.low = 10G E/memory.current = 0
5936 * and the memory pressure is applied, the following memory distribution
5937 * is expected (approximately):
5939 * A/memory.current = 2G
5941 * B/memory.current = 1.3G
5942 * C/memory.current = 0.6G
5943 * D/memory.current = 0
5944 * E/memory.current = 0
5946 * These calculations require constant tracking of the actual low usages
5947 * (see propagate_protected_usage()), as well as recursive calculation of
5948 * effective memory.low values. But as we do call mem_cgroup_protected()
5949 * path for each memory cgroup top-down from the reclaim,
5950 * it's possible to optimize this part, and save calculated elow
5951 * for next usage. This part is intentionally racy, but it's ok,
5952 * as memory.low is a best-effort mechanism.
5954 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5955 struct mem_cgroup *memcg)
5957 struct mem_cgroup *parent;
5958 unsigned long emin, parent_emin;
5959 unsigned long elow, parent_elow;
5960 unsigned long usage;
5962 if (mem_cgroup_disabled())
5963 return MEMCG_PROT_NONE;
5966 root = root_mem_cgroup;
5968 return MEMCG_PROT_NONE;
5970 usage = page_counter_read(&memcg->memory);
5972 return MEMCG_PROT_NONE;
5974 emin = memcg->memory.min;
5975 elow = memcg->memory.low;
5977 parent = parent_mem_cgroup(memcg);
5978 /* No parent means a non-hierarchical mode on v1 memcg */
5980 return MEMCG_PROT_NONE;
5985 parent_emin = READ_ONCE(parent->memory.emin);
5986 emin = min(emin, parent_emin);
5987 if (emin && parent_emin) {
5988 unsigned long min_usage, siblings_min_usage;
5990 min_usage = min(usage, memcg->memory.min);
5991 siblings_min_usage = atomic_long_read(
5992 &parent->memory.children_min_usage);
5994 if (min_usage && siblings_min_usage)
5995 emin = min(emin, parent_emin * min_usage /
5996 siblings_min_usage);
5999 parent_elow = READ_ONCE(parent->memory.elow);
6000 elow = min(elow, parent_elow);
6001 if (elow && parent_elow) {
6002 unsigned long low_usage, siblings_low_usage;
6004 low_usage = min(usage, memcg->memory.low);
6005 siblings_low_usage = atomic_long_read(
6006 &parent->memory.children_low_usage);
6008 if (low_usage && siblings_low_usage)
6009 elow = min(elow, parent_elow * low_usage /
6010 siblings_low_usage);
6014 memcg->memory.emin = emin;
6015 memcg->memory.elow = elow;
6018 return MEMCG_PROT_MIN;
6019 else if (usage <= elow)
6020 return MEMCG_PROT_LOW;
6022 return MEMCG_PROT_NONE;
6026 * mem_cgroup_try_charge - try charging a page
6027 * @page: page to charge
6028 * @mm: mm context of the victim
6029 * @gfp_mask: reclaim mode
6030 * @memcgp: charged memcg return
6031 * @compound: charge the page as compound or small page
6033 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6034 * pages according to @gfp_mask if necessary.
6036 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6037 * Otherwise, an error code is returned.
6039 * After page->mapping has been set up, the caller must finalize the
6040 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6041 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6043 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6044 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6047 struct mem_cgroup *memcg = NULL;
6048 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6051 if (mem_cgroup_disabled())
6054 if (PageSwapCache(page)) {
6056 * Every swap fault against a single page tries to charge the
6057 * page, bail as early as possible. shmem_unuse() encounters
6058 * already charged pages, too. The USED bit is protected by
6059 * the page lock, which serializes swap cache removal, which
6060 * in turn serializes uncharging.
6062 VM_BUG_ON_PAGE(!PageLocked(page), page);
6063 if (compound_head(page)->mem_cgroup)
6066 if (do_swap_account) {
6067 swp_entry_t ent = { .val = page_private(page), };
6068 unsigned short id = lookup_swap_cgroup_id(ent);
6071 memcg = mem_cgroup_from_id(id);
6072 if (memcg && !css_tryget_online(&memcg->css))
6079 memcg = get_mem_cgroup_from_mm(mm);
6081 ret = try_charge(memcg, gfp_mask, nr_pages);
6083 css_put(&memcg->css);
6089 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6090 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6093 struct mem_cgroup *memcg;
6096 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6098 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6103 * mem_cgroup_commit_charge - commit a page charge
6104 * @page: page to charge
6105 * @memcg: memcg to charge the page to
6106 * @lrucare: page might be on LRU already
6107 * @compound: charge the page as compound or small page
6109 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6110 * after page->mapping has been set up. This must happen atomically
6111 * as part of the page instantiation, i.e. under the page table lock
6112 * for anonymous pages, under the page lock for page and swap cache.
6114 * In addition, the page must not be on the LRU during the commit, to
6115 * prevent racing with task migration. If it might be, use @lrucare.
6117 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6119 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6120 bool lrucare, bool compound)
6122 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6124 VM_BUG_ON_PAGE(!page->mapping, page);
6125 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6127 if (mem_cgroup_disabled())
6130 * Swap faults will attempt to charge the same page multiple
6131 * times. But reuse_swap_page() might have removed the page
6132 * from swapcache already, so we can't check PageSwapCache().
6137 commit_charge(page, memcg, lrucare);
6139 local_irq_disable();
6140 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6141 memcg_check_events(memcg, page);
6144 if (do_memsw_account() && PageSwapCache(page)) {
6145 swp_entry_t entry = { .val = page_private(page) };
6147 * The swap entry might not get freed for a long time,
6148 * let's not wait for it. The page already received a
6149 * memory+swap charge, drop the swap entry duplicate.
6151 mem_cgroup_uncharge_swap(entry, nr_pages);
6156 * mem_cgroup_cancel_charge - cancel a page charge
6157 * @page: page to charge
6158 * @memcg: memcg to charge the page to
6159 * @compound: charge the page as compound or small page
6161 * Cancel a charge transaction started by mem_cgroup_try_charge().
6163 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6166 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6168 if (mem_cgroup_disabled())
6171 * Swap faults will attempt to charge the same page multiple
6172 * times. But reuse_swap_page() might have removed the page
6173 * from swapcache already, so we can't check PageSwapCache().
6178 cancel_charge(memcg, nr_pages);
6181 struct uncharge_gather {
6182 struct mem_cgroup *memcg;
6183 unsigned long pgpgout;
6184 unsigned long nr_anon;
6185 unsigned long nr_file;
6186 unsigned long nr_kmem;
6187 unsigned long nr_huge;
6188 unsigned long nr_shmem;
6189 struct page *dummy_page;
6192 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6194 memset(ug, 0, sizeof(*ug));
6197 static void uncharge_batch(const struct uncharge_gather *ug)
6199 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6200 unsigned long flags;
6202 if (!mem_cgroup_is_root(ug->memcg)) {
6203 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6204 if (do_memsw_account())
6205 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6206 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6207 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6208 memcg_oom_recover(ug->memcg);
6211 local_irq_save(flags);
6212 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6213 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6214 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6215 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6216 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6217 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6218 memcg_check_events(ug->memcg, ug->dummy_page);
6219 local_irq_restore(flags);
6221 if (!mem_cgroup_is_root(ug->memcg))
6222 css_put_many(&ug->memcg->css, nr_pages);
6225 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6227 VM_BUG_ON_PAGE(PageLRU(page), page);
6228 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6229 !PageHWPoison(page) , page);
6231 if (!page->mem_cgroup)
6235 * Nobody should be changing or seriously looking at
6236 * page->mem_cgroup at this point, we have fully
6237 * exclusive access to the page.
6240 if (ug->memcg != page->mem_cgroup) {
6243 uncharge_gather_clear(ug);
6245 ug->memcg = page->mem_cgroup;
6248 if (!PageKmemcg(page)) {
6249 unsigned int nr_pages = 1;
6251 if (PageTransHuge(page)) {
6252 nr_pages <<= compound_order(page);
6253 ug->nr_huge += nr_pages;
6256 ug->nr_anon += nr_pages;
6258 ug->nr_file += nr_pages;
6259 if (PageSwapBacked(page))
6260 ug->nr_shmem += nr_pages;
6264 ug->nr_kmem += 1 << compound_order(page);
6265 __ClearPageKmemcg(page);
6268 ug->dummy_page = page;
6269 page->mem_cgroup = NULL;
6272 static void uncharge_list(struct list_head *page_list)
6274 struct uncharge_gather ug;
6275 struct list_head *next;
6277 uncharge_gather_clear(&ug);
6280 * Note that the list can be a single page->lru; hence the
6281 * do-while loop instead of a simple list_for_each_entry().
6283 next = page_list->next;
6287 page = list_entry(next, struct page, lru);
6288 next = page->lru.next;
6290 uncharge_page(page, &ug);
6291 } while (next != page_list);
6294 uncharge_batch(&ug);
6298 * mem_cgroup_uncharge - uncharge a page
6299 * @page: page to uncharge
6301 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6302 * mem_cgroup_commit_charge().
6304 void mem_cgroup_uncharge(struct page *page)
6306 struct uncharge_gather ug;
6308 if (mem_cgroup_disabled())
6311 /* Don't touch page->lru of any random page, pre-check: */
6312 if (!page->mem_cgroup)
6315 uncharge_gather_clear(&ug);
6316 uncharge_page(page, &ug);
6317 uncharge_batch(&ug);
6321 * mem_cgroup_uncharge_list - uncharge a list of page
6322 * @page_list: list of pages to uncharge
6324 * Uncharge a list of pages previously charged with
6325 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6327 void mem_cgroup_uncharge_list(struct list_head *page_list)
6329 if (mem_cgroup_disabled())
6332 if (!list_empty(page_list))
6333 uncharge_list(page_list);
6337 * mem_cgroup_migrate - charge a page's replacement
6338 * @oldpage: currently circulating page
6339 * @newpage: replacement page
6341 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6342 * be uncharged upon free.
6344 * Both pages must be locked, @newpage->mapping must be set up.
6346 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6348 struct mem_cgroup *memcg;
6349 unsigned int nr_pages;
6351 unsigned long flags;
6353 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6354 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6355 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6356 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6359 if (mem_cgroup_disabled())
6362 /* Page cache replacement: new page already charged? */
6363 if (newpage->mem_cgroup)
6366 /* Swapcache readahead pages can get replaced before being charged */
6367 memcg = oldpage->mem_cgroup;
6371 /* Force-charge the new page. The old one will be freed soon */
6372 compound = PageTransHuge(newpage);
6373 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6375 page_counter_charge(&memcg->memory, nr_pages);
6376 if (do_memsw_account())
6377 page_counter_charge(&memcg->memsw, nr_pages);
6378 css_get_many(&memcg->css, nr_pages);
6380 commit_charge(newpage, memcg, false);
6382 local_irq_save(flags);
6383 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6384 memcg_check_events(memcg, newpage);
6385 local_irq_restore(flags);
6388 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6389 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6391 void mem_cgroup_sk_alloc(struct sock *sk)
6393 struct mem_cgroup *memcg;
6395 if (!mem_cgroup_sockets_enabled)
6399 * Socket cloning can throw us here with sk_memcg already
6400 * filled. It won't however, necessarily happen from
6401 * process context. So the test for root memcg given
6402 * the current task's memcg won't help us in this case.
6404 * Respecting the original socket's memcg is a better
6405 * decision in this case.
6408 css_get(&sk->sk_memcg->css);
6413 memcg = mem_cgroup_from_task(current);
6414 if (memcg == root_mem_cgroup)
6416 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6418 if (css_tryget_online(&memcg->css))
6419 sk->sk_memcg = memcg;
6424 void mem_cgroup_sk_free(struct sock *sk)
6427 css_put(&sk->sk_memcg->css);
6431 * mem_cgroup_charge_skmem - charge socket memory
6432 * @memcg: memcg to charge
6433 * @nr_pages: number of pages to charge
6435 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6436 * @memcg's configured limit, %false if the charge had to be forced.
6438 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6440 gfp_t gfp_mask = GFP_KERNEL;
6442 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6443 struct page_counter *fail;
6445 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6446 memcg->tcpmem_pressure = 0;
6449 page_counter_charge(&memcg->tcpmem, nr_pages);
6450 memcg->tcpmem_pressure = 1;
6454 /* Don't block in the packet receive path */
6456 gfp_mask = GFP_NOWAIT;
6458 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6460 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6463 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6468 * mem_cgroup_uncharge_skmem - uncharge socket memory
6469 * @memcg: memcg to uncharge
6470 * @nr_pages: number of pages to uncharge
6472 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6474 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6475 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6479 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6481 refill_stock(memcg, nr_pages);
6484 static int __init cgroup_memory(char *s)
6488 while ((token = strsep(&s, ",")) != NULL) {
6491 if (!strcmp(token, "nosocket"))
6492 cgroup_memory_nosocket = true;
6493 if (!strcmp(token, "nokmem"))
6494 cgroup_memory_nokmem = true;
6498 __setup("cgroup.memory=", cgroup_memory);
6501 * subsys_initcall() for memory controller.
6503 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6504 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6505 * basically everything that doesn't depend on a specific mem_cgroup structure
6506 * should be initialized from here.
6508 static int __init mem_cgroup_init(void)
6512 #ifdef CONFIG_MEMCG_KMEM
6514 * Kmem cache creation is mostly done with the slab_mutex held,
6515 * so use a workqueue with limited concurrency to avoid stalling
6516 * all worker threads in case lots of cgroups are created and
6517 * destroyed simultaneously.
6519 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6520 BUG_ON(!memcg_kmem_cache_wq);
6523 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6524 memcg_hotplug_cpu_dead);
6526 for_each_possible_cpu(cpu)
6527 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6530 for_each_node(node) {
6531 struct mem_cgroup_tree_per_node *rtpn;
6533 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6534 node_online(node) ? node : NUMA_NO_NODE);
6536 rtpn->rb_root = RB_ROOT;
6537 rtpn->rb_rightmost = NULL;
6538 spin_lock_init(&rtpn->lock);
6539 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6544 subsys_initcall(mem_cgroup_init);
6546 #ifdef CONFIG_MEMCG_SWAP
6547 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6549 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6551 * The root cgroup cannot be destroyed, so it's refcount must
6554 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6558 memcg = parent_mem_cgroup(memcg);
6560 memcg = root_mem_cgroup;
6566 * mem_cgroup_swapout - transfer a memsw charge to swap
6567 * @page: page whose memsw charge to transfer
6568 * @entry: swap entry to move the charge to
6570 * Transfer the memsw charge of @page to @entry.
6572 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6574 struct mem_cgroup *memcg, *swap_memcg;
6575 unsigned int nr_entries;
6576 unsigned short oldid;
6578 VM_BUG_ON_PAGE(PageLRU(page), page);
6579 VM_BUG_ON_PAGE(page_count(page), page);
6581 if (!do_memsw_account())
6584 memcg = page->mem_cgroup;
6586 /* Readahead page, never charged */
6591 * In case the memcg owning these pages has been offlined and doesn't
6592 * have an ID allocated to it anymore, charge the closest online
6593 * ancestor for the swap instead and transfer the memory+swap charge.
6595 swap_memcg = mem_cgroup_id_get_online(memcg);
6596 nr_entries = hpage_nr_pages(page);
6597 /* Get references for the tail pages, too */
6599 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6600 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6602 VM_BUG_ON_PAGE(oldid, page);
6603 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6605 page->mem_cgroup = NULL;
6607 if (!mem_cgroup_is_root(memcg))
6608 page_counter_uncharge(&memcg->memory, nr_entries);
6610 if (memcg != swap_memcg) {
6611 if (!mem_cgroup_is_root(swap_memcg))
6612 page_counter_charge(&swap_memcg->memsw, nr_entries);
6613 page_counter_uncharge(&memcg->memsw, nr_entries);
6617 * Interrupts should be disabled here because the caller holds the
6618 * i_pages lock which is taken with interrupts-off. It is
6619 * important here to have the interrupts disabled because it is the
6620 * only synchronisation we have for updating the per-CPU variables.
6622 VM_BUG_ON(!irqs_disabled());
6623 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6625 memcg_check_events(memcg, page);
6627 if (!mem_cgroup_is_root(memcg))
6628 css_put_many(&memcg->css, nr_entries);
6632 * mem_cgroup_try_charge_swap - try charging swap space for a page
6633 * @page: page being added to swap
6634 * @entry: swap entry to charge
6636 * Try to charge @page's memcg for the swap space at @entry.
6638 * Returns 0 on success, -ENOMEM on failure.
6640 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6642 unsigned int nr_pages = hpage_nr_pages(page);
6643 struct page_counter *counter;
6644 struct mem_cgroup *memcg;
6645 unsigned short oldid;
6647 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6650 memcg = page->mem_cgroup;
6652 /* Readahead page, never charged */
6657 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6661 memcg = mem_cgroup_id_get_online(memcg);
6663 if (!mem_cgroup_is_root(memcg) &&
6664 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6665 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6666 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6667 mem_cgroup_id_put(memcg);
6671 /* Get references for the tail pages, too */
6673 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6674 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6675 VM_BUG_ON_PAGE(oldid, page);
6676 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6682 * mem_cgroup_uncharge_swap - uncharge swap space
6683 * @entry: swap entry to uncharge
6684 * @nr_pages: the amount of swap space to uncharge
6686 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6688 struct mem_cgroup *memcg;
6691 if (!do_swap_account)
6694 id = swap_cgroup_record(entry, 0, nr_pages);
6696 memcg = mem_cgroup_from_id(id);
6698 if (!mem_cgroup_is_root(memcg)) {
6699 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6700 page_counter_uncharge(&memcg->swap, nr_pages);
6702 page_counter_uncharge(&memcg->memsw, nr_pages);
6704 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6705 mem_cgroup_id_put_many(memcg, nr_pages);
6710 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6712 long nr_swap_pages = get_nr_swap_pages();
6714 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6715 return nr_swap_pages;
6716 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6717 nr_swap_pages = min_t(long, nr_swap_pages,
6718 READ_ONCE(memcg->swap.max) -
6719 page_counter_read(&memcg->swap));
6720 return nr_swap_pages;
6723 bool mem_cgroup_swap_full(struct page *page)
6725 struct mem_cgroup *memcg;
6727 VM_BUG_ON_PAGE(!PageLocked(page), page);
6731 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6734 memcg = page->mem_cgroup;
6738 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6739 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6745 /* for remember boot option*/
6746 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6747 static int really_do_swap_account __initdata = 1;
6749 static int really_do_swap_account __initdata;
6752 static int __init enable_swap_account(char *s)
6754 if (!strcmp(s, "1"))
6755 really_do_swap_account = 1;
6756 else if (!strcmp(s, "0"))
6757 really_do_swap_account = 0;
6760 __setup("swapaccount=", enable_swap_account);
6762 static u64 swap_current_read(struct cgroup_subsys_state *css,
6765 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6767 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6770 static int swap_max_show(struct seq_file *m, void *v)
6772 return seq_puts_memcg_tunable(m,
6773 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6776 static ssize_t swap_max_write(struct kernfs_open_file *of,
6777 char *buf, size_t nbytes, loff_t off)
6779 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6783 buf = strstrip(buf);
6784 err = page_counter_memparse(buf, "max", &max);
6788 xchg(&memcg->swap.max, max);
6793 static int swap_events_show(struct seq_file *m, void *v)
6795 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6797 seq_printf(m, "max %lu\n",
6798 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6799 seq_printf(m, "fail %lu\n",
6800 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6805 static struct cftype swap_files[] = {
6807 .name = "swap.current",
6808 .flags = CFTYPE_NOT_ON_ROOT,
6809 .read_u64 = swap_current_read,
6813 .flags = CFTYPE_NOT_ON_ROOT,
6814 .seq_show = swap_max_show,
6815 .write = swap_max_write,
6818 .name = "swap.events",
6819 .flags = CFTYPE_NOT_ON_ROOT,
6820 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6821 .seq_show = swap_events_show,
6826 static struct cftype memsw_cgroup_files[] = {
6828 .name = "memsw.usage_in_bytes",
6829 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6830 .read_u64 = mem_cgroup_read_u64,
6833 .name = "memsw.max_usage_in_bytes",
6834 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6835 .write = mem_cgroup_reset,
6836 .read_u64 = mem_cgroup_read_u64,
6839 .name = "memsw.limit_in_bytes",
6840 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6841 .write = mem_cgroup_write,
6842 .read_u64 = mem_cgroup_read_u64,
6845 .name = "memsw.failcnt",
6846 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6847 .write = mem_cgroup_reset,
6848 .read_u64 = mem_cgroup_read_u64,
6850 { }, /* terminate */
6853 static int __init mem_cgroup_swap_init(void)
6855 if (!mem_cgroup_disabled() && really_do_swap_account) {
6856 do_swap_account = 1;
6857 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6859 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6860 memsw_cgroup_files));
6864 subsys_initcall(mem_cgroup_swap_init);
6866 #endif /* CONFIG_MEMCG_SWAP */