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 memcg = READ_ONCE(page->mem_cgroup);
490 while (memcg && !(memcg->css.flags & CSS_ONLINE))
491 memcg = parent_mem_cgroup(memcg);
493 ino = cgroup_ino(memcg->css.cgroup);
498 static struct mem_cgroup_per_node *
499 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
501 int nid = page_to_nid(page);
503 return memcg->nodeinfo[nid];
506 static struct mem_cgroup_tree_per_node *
507 soft_limit_tree_node(int nid)
509 return soft_limit_tree.rb_tree_per_node[nid];
512 static struct mem_cgroup_tree_per_node *
513 soft_limit_tree_from_page(struct page *page)
515 int nid = page_to_nid(page);
517 return soft_limit_tree.rb_tree_per_node[nid];
520 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
521 struct mem_cgroup_tree_per_node *mctz,
522 unsigned long new_usage_in_excess)
524 struct rb_node **p = &mctz->rb_root.rb_node;
525 struct rb_node *parent = NULL;
526 struct mem_cgroup_per_node *mz_node;
527 bool rightmost = true;
532 mz->usage_in_excess = new_usage_in_excess;
533 if (!mz->usage_in_excess)
537 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
539 if (mz->usage_in_excess < mz_node->usage_in_excess) {
545 * We can't avoid mem cgroups that are over their soft
546 * limit by the same amount
548 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
553 mctz->rb_rightmost = &mz->tree_node;
555 rb_link_node(&mz->tree_node, parent, p);
556 rb_insert_color(&mz->tree_node, &mctz->rb_root);
560 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
561 struct mem_cgroup_tree_per_node *mctz)
566 if (&mz->tree_node == mctz->rb_rightmost)
567 mctz->rb_rightmost = rb_prev(&mz->tree_node);
569 rb_erase(&mz->tree_node, &mctz->rb_root);
573 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
574 struct mem_cgroup_tree_per_node *mctz)
578 spin_lock_irqsave(&mctz->lock, flags);
579 __mem_cgroup_remove_exceeded(mz, mctz);
580 spin_unlock_irqrestore(&mctz->lock, flags);
583 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
585 unsigned long nr_pages = page_counter_read(&memcg->memory);
586 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
587 unsigned long excess = 0;
589 if (nr_pages > soft_limit)
590 excess = nr_pages - soft_limit;
595 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
597 unsigned long excess;
598 struct mem_cgroup_per_node *mz;
599 struct mem_cgroup_tree_per_node *mctz;
601 mctz = soft_limit_tree_from_page(page);
605 * Necessary to update all ancestors when hierarchy is used.
606 * because their event counter is not touched.
608 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
609 mz = mem_cgroup_page_nodeinfo(memcg, page);
610 excess = soft_limit_excess(memcg);
612 * We have to update the tree if mz is on RB-tree or
613 * mem is over its softlimit.
615 if (excess || mz->on_tree) {
618 spin_lock_irqsave(&mctz->lock, flags);
619 /* if on-tree, remove it */
621 __mem_cgroup_remove_exceeded(mz, mctz);
623 * Insert again. mz->usage_in_excess will be updated.
624 * If excess is 0, no tree ops.
626 __mem_cgroup_insert_exceeded(mz, mctz, excess);
627 spin_unlock_irqrestore(&mctz->lock, flags);
632 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
634 struct mem_cgroup_tree_per_node *mctz;
635 struct mem_cgroup_per_node *mz;
639 mz = mem_cgroup_nodeinfo(memcg, nid);
640 mctz = soft_limit_tree_node(nid);
642 mem_cgroup_remove_exceeded(mz, mctz);
646 static struct mem_cgroup_per_node *
647 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
649 struct mem_cgroup_per_node *mz;
653 if (!mctz->rb_rightmost)
654 goto done; /* Nothing to reclaim from */
656 mz = rb_entry(mctz->rb_rightmost,
657 struct mem_cgroup_per_node, tree_node);
659 * Remove the node now but someone else can add it back,
660 * we will to add it back at the end of reclaim to its correct
661 * position in the tree.
663 __mem_cgroup_remove_exceeded(mz, mctz);
664 if (!soft_limit_excess(mz->memcg) ||
665 !css_tryget_online(&mz->memcg->css))
671 static struct mem_cgroup_per_node *
672 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
674 struct mem_cgroup_per_node *mz;
676 spin_lock_irq(&mctz->lock);
677 mz = __mem_cgroup_largest_soft_limit_node(mctz);
678 spin_unlock_irq(&mctz->lock);
683 * __mod_memcg_state - update cgroup memory statistics
684 * @memcg: the memory cgroup
685 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
686 * @val: delta to add to the counter, can be negative
688 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
692 if (mem_cgroup_disabled())
695 __this_cpu_add(memcg->vmstats_local->stat[idx], val);
697 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
698 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
699 struct mem_cgroup *mi;
701 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
702 atomic_long_add(x, &mi->vmstats[idx]);
705 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
708 static struct mem_cgroup_per_node *
709 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
711 struct mem_cgroup *parent;
713 parent = parent_mem_cgroup(pn->memcg);
716 return mem_cgroup_nodeinfo(parent, nid);
720 * __mod_lruvec_state - update lruvec memory statistics
721 * @lruvec: the lruvec
722 * @idx: the stat item
723 * @val: delta to add to the counter, can be negative
725 * The lruvec is the intersection of the NUMA node and a cgroup. This
726 * function updates the all three counters that are affected by a
727 * change of state at this level: per-node, per-cgroup, per-lruvec.
729 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
732 pg_data_t *pgdat = lruvec_pgdat(lruvec);
733 struct mem_cgroup_per_node *pn;
734 struct mem_cgroup *memcg;
738 __mod_node_page_state(pgdat, idx, val);
740 if (mem_cgroup_disabled())
743 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
747 __mod_memcg_state(memcg, idx, val);
750 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
752 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
753 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
754 struct mem_cgroup_per_node *pi;
756 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
757 atomic_long_add(x, &pi->lruvec_stat[idx]);
760 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
764 * __count_memcg_events - account VM events in a cgroup
765 * @memcg: the memory cgroup
766 * @idx: the event item
767 * @count: the number of events that occured
769 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
774 if (mem_cgroup_disabled())
777 __this_cpu_add(memcg->vmstats_local->events[idx], count);
779 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
780 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
781 struct mem_cgroup *mi;
783 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
784 atomic_long_add(x, &mi->vmevents[idx]);
787 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
790 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
792 return atomic_long_read(&memcg->vmevents[event]);
795 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
800 for_each_possible_cpu(cpu)
801 x += per_cpu(memcg->vmstats_local->events[event], cpu);
805 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
807 bool compound, int nr_pages)
810 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
811 * counted as CACHE even if it's on ANON LRU.
814 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
816 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
817 if (PageSwapBacked(page))
818 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
822 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
823 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
826 /* pagein of a big page is an event. So, ignore page size */
828 __count_memcg_events(memcg, PGPGIN, 1);
830 __count_memcg_events(memcg, PGPGOUT, 1);
831 nr_pages = -nr_pages; /* for event */
834 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
837 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
838 enum mem_cgroup_events_target target)
840 unsigned long val, next;
842 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
843 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
844 /* from time_after() in jiffies.h */
845 if ((long)(next - val) < 0) {
847 case MEM_CGROUP_TARGET_THRESH:
848 next = val + THRESHOLDS_EVENTS_TARGET;
850 case MEM_CGROUP_TARGET_SOFTLIMIT:
851 next = val + SOFTLIMIT_EVENTS_TARGET;
853 case MEM_CGROUP_TARGET_NUMAINFO:
854 next = val + NUMAINFO_EVENTS_TARGET;
859 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
866 * Check events in order.
869 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
871 /* threshold event is triggered in finer grain than soft limit */
872 if (unlikely(mem_cgroup_event_ratelimit(memcg,
873 MEM_CGROUP_TARGET_THRESH))) {
875 bool do_numainfo __maybe_unused;
877 do_softlimit = mem_cgroup_event_ratelimit(memcg,
878 MEM_CGROUP_TARGET_SOFTLIMIT);
880 do_numainfo = mem_cgroup_event_ratelimit(memcg,
881 MEM_CGROUP_TARGET_NUMAINFO);
883 mem_cgroup_threshold(memcg);
884 if (unlikely(do_softlimit))
885 mem_cgroup_update_tree(memcg, page);
887 if (unlikely(do_numainfo))
888 atomic_inc(&memcg->numainfo_events);
893 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
896 * mm_update_next_owner() may clear mm->owner to NULL
897 * if it races with swapoff, page migration, etc.
898 * So this can be called with p == NULL.
903 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
905 EXPORT_SYMBOL(mem_cgroup_from_task);
908 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
909 * @mm: mm from which memcg should be extracted. It can be NULL.
911 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
912 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
915 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
917 struct mem_cgroup *memcg;
919 if (mem_cgroup_disabled())
925 * Page cache insertions can happen withou an
926 * actual mm context, e.g. during disk probing
927 * on boot, loopback IO, acct() writes etc.
930 memcg = root_mem_cgroup;
932 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
933 if (unlikely(!memcg))
934 memcg = root_mem_cgroup;
936 } while (!css_tryget_online(&memcg->css));
940 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
943 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
944 * @page: page from which memcg should be extracted.
946 * Obtain a reference on page->memcg and returns it if successful. Otherwise
947 * root_mem_cgroup is returned.
949 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
951 struct mem_cgroup *memcg = page->mem_cgroup;
953 if (mem_cgroup_disabled())
957 if (!memcg || !css_tryget_online(&memcg->css))
958 memcg = root_mem_cgroup;
962 EXPORT_SYMBOL(get_mem_cgroup_from_page);
965 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
967 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
969 if (unlikely(current->active_memcg)) {
970 struct mem_cgroup *memcg = root_mem_cgroup;
973 if (css_tryget_online(¤t->active_memcg->css))
974 memcg = current->active_memcg;
978 return get_mem_cgroup_from_mm(current->mm);
982 * mem_cgroup_iter - iterate over memory cgroup hierarchy
983 * @root: hierarchy root
984 * @prev: previously returned memcg, NULL on first invocation
985 * @reclaim: cookie for shared reclaim walks, NULL for full walks
987 * Returns references to children of the hierarchy below @root, or
988 * @root itself, or %NULL after a full round-trip.
990 * Caller must pass the return value in @prev on subsequent
991 * invocations for reference counting, or use mem_cgroup_iter_break()
992 * to cancel a hierarchy walk before the round-trip is complete.
994 * Reclaimers can specify a node and a priority level in @reclaim to
995 * divide up the memcgs in the hierarchy among all concurrent
996 * reclaimers operating on the same node and priority.
998 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
999 struct mem_cgroup *prev,
1000 struct mem_cgroup_reclaim_cookie *reclaim)
1002 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1003 struct cgroup_subsys_state *css = NULL;
1004 struct mem_cgroup *memcg = NULL;
1005 struct mem_cgroup *pos = NULL;
1007 if (mem_cgroup_disabled())
1011 root = root_mem_cgroup;
1013 if (prev && !reclaim)
1016 if (!root->use_hierarchy && root != root_mem_cgroup) {
1025 struct mem_cgroup_per_node *mz;
1027 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1028 iter = &mz->iter[reclaim->priority];
1030 if (prev && reclaim->generation != iter->generation)
1034 pos = READ_ONCE(iter->position);
1035 if (!pos || css_tryget(&pos->css))
1038 * css reference reached zero, so iter->position will
1039 * be cleared by ->css_released. However, we should not
1040 * rely on this happening soon, because ->css_released
1041 * is called from a work queue, and by busy-waiting we
1042 * might block it. So we clear iter->position right
1045 (void)cmpxchg(&iter->position, pos, NULL);
1053 css = css_next_descendant_pre(css, &root->css);
1056 * Reclaimers share the hierarchy walk, and a
1057 * new one might jump in right at the end of
1058 * the hierarchy - make sure they see at least
1059 * one group and restart from the beginning.
1067 * Verify the css and acquire a reference. The root
1068 * is provided by the caller, so we know it's alive
1069 * and kicking, and don't take an extra reference.
1071 memcg = mem_cgroup_from_css(css);
1073 if (css == &root->css)
1076 if (css_tryget(css))
1084 * The position could have already been updated by a competing
1085 * thread, so check that the value hasn't changed since we read
1086 * it to avoid reclaiming from the same cgroup twice.
1088 (void)cmpxchg(&iter->position, pos, memcg);
1096 reclaim->generation = iter->generation;
1102 if (prev && prev != root)
1103 css_put(&prev->css);
1109 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1110 * @root: hierarchy root
1111 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1113 void mem_cgroup_iter_break(struct mem_cgroup *root,
1114 struct mem_cgroup *prev)
1117 root = root_mem_cgroup;
1118 if (prev && prev != root)
1119 css_put(&prev->css);
1122 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1124 struct mem_cgroup *memcg = dead_memcg;
1125 struct mem_cgroup_reclaim_iter *iter;
1126 struct mem_cgroup_per_node *mz;
1130 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1131 for_each_node(nid) {
1132 mz = mem_cgroup_nodeinfo(memcg, nid);
1133 for (i = 0; i <= DEF_PRIORITY; i++) {
1134 iter = &mz->iter[i];
1135 cmpxchg(&iter->position,
1143 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1144 * @memcg: hierarchy root
1145 * @fn: function to call for each task
1146 * @arg: argument passed to @fn
1148 * This function iterates over tasks attached to @memcg or to any of its
1149 * descendants and calls @fn for each task. If @fn returns a non-zero
1150 * value, the function breaks the iteration loop and returns the value.
1151 * Otherwise, it will iterate over all tasks and return 0.
1153 * This function must not be called for the root memory cgroup.
1155 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1156 int (*fn)(struct task_struct *, void *), void *arg)
1158 struct mem_cgroup *iter;
1161 BUG_ON(memcg == root_mem_cgroup);
1163 for_each_mem_cgroup_tree(iter, memcg) {
1164 struct css_task_iter it;
1165 struct task_struct *task;
1167 css_task_iter_start(&iter->css, 0, &it);
1168 while (!ret && (task = css_task_iter_next(&it)))
1169 ret = fn(task, arg);
1170 css_task_iter_end(&it);
1172 mem_cgroup_iter_break(memcg, iter);
1180 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1182 * @pgdat: pgdat of the page
1184 * This function is only safe when following the LRU page isolation
1185 * and putback protocol: the LRU lock must be held, and the page must
1186 * either be PageLRU() or the caller must have isolated/allocated it.
1188 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1190 struct mem_cgroup_per_node *mz;
1191 struct mem_cgroup *memcg;
1192 struct lruvec *lruvec;
1194 if (mem_cgroup_disabled()) {
1195 lruvec = &pgdat->lruvec;
1199 memcg = page->mem_cgroup;
1201 * Swapcache readahead pages are added to the LRU - and
1202 * possibly migrated - before they are charged.
1205 memcg = root_mem_cgroup;
1207 mz = mem_cgroup_page_nodeinfo(memcg, page);
1208 lruvec = &mz->lruvec;
1211 * Since a node can be onlined after the mem_cgroup was created,
1212 * we have to be prepared to initialize lruvec->zone here;
1213 * and if offlined then reonlined, we need to reinitialize it.
1215 if (unlikely(lruvec->pgdat != pgdat))
1216 lruvec->pgdat = pgdat;
1221 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1222 * @lruvec: mem_cgroup per zone lru vector
1223 * @lru: index of lru list the page is sitting on
1224 * @zid: zone id of the accounted pages
1225 * @nr_pages: positive when adding or negative when removing
1227 * This function must be called under lru_lock, just before a page is added
1228 * to or just after a page is removed from an lru list (that ordering being
1229 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1231 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1232 int zid, int nr_pages)
1234 struct mem_cgroup_per_node *mz;
1235 unsigned long *lru_size;
1238 if (mem_cgroup_disabled())
1241 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1242 lru_size = &mz->lru_zone_size[zid][lru];
1245 *lru_size += nr_pages;
1248 if (WARN_ONCE(size < 0,
1249 "%s(%p, %d, %d): lru_size %ld\n",
1250 __func__, lruvec, lru, nr_pages, size)) {
1256 *lru_size += nr_pages;
1259 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1261 struct mem_cgroup *task_memcg;
1262 struct task_struct *p;
1265 p = find_lock_task_mm(task);
1267 task_memcg = get_mem_cgroup_from_mm(p->mm);
1271 * All threads may have already detached their mm's, but the oom
1272 * killer still needs to detect if they have already been oom
1273 * killed to prevent needlessly killing additional tasks.
1276 task_memcg = mem_cgroup_from_task(task);
1277 css_get(&task_memcg->css);
1280 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1281 css_put(&task_memcg->css);
1286 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1287 * @memcg: the memory cgroup
1289 * Returns the maximum amount of memory @mem can be charged with, in
1292 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1294 unsigned long margin = 0;
1295 unsigned long count;
1296 unsigned long limit;
1298 count = page_counter_read(&memcg->memory);
1299 limit = READ_ONCE(memcg->memory.max);
1301 margin = limit - count;
1303 if (do_memsw_account()) {
1304 count = page_counter_read(&memcg->memsw);
1305 limit = READ_ONCE(memcg->memsw.max);
1307 margin = min(margin, limit - count);
1316 * A routine for checking "mem" is under move_account() or not.
1318 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1319 * moving cgroups. This is for waiting at high-memory pressure
1322 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1324 struct mem_cgroup *from;
1325 struct mem_cgroup *to;
1328 * Unlike task_move routines, we access mc.to, mc.from not under
1329 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1331 spin_lock(&mc.lock);
1337 ret = mem_cgroup_is_descendant(from, memcg) ||
1338 mem_cgroup_is_descendant(to, memcg);
1340 spin_unlock(&mc.lock);
1344 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1346 if (mc.moving_task && current != mc.moving_task) {
1347 if (mem_cgroup_under_move(memcg)) {
1349 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1350 /* moving charge context might have finished. */
1353 finish_wait(&mc.waitq, &wait);
1360 static char *memory_stat_format(struct mem_cgroup *memcg)
1365 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1370 * Provide statistics on the state of the memory subsystem as
1371 * well as cumulative event counters that show past behavior.
1373 * This list is ordered following a combination of these gradients:
1374 * 1) generic big picture -> specifics and details
1375 * 2) reflecting userspace activity -> reflecting kernel heuristics
1377 * Current memory state:
1380 seq_buf_printf(&s, "anon %llu\n",
1381 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1383 seq_buf_printf(&s, "file %llu\n",
1384 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1386 seq_buf_printf(&s, "kernel_stack %llu\n",
1387 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1389 seq_buf_printf(&s, "slab %llu\n",
1390 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1391 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1393 seq_buf_printf(&s, "sock %llu\n",
1394 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1397 seq_buf_printf(&s, "shmem %llu\n",
1398 (u64)memcg_page_state(memcg, NR_SHMEM) *
1400 seq_buf_printf(&s, "file_mapped %llu\n",
1401 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1403 seq_buf_printf(&s, "file_dirty %llu\n",
1404 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1406 seq_buf_printf(&s, "file_writeback %llu\n",
1407 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1411 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1412 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1413 * arse because it requires migrating the work out of rmap to a place
1414 * where the page->mem_cgroup is set up and stable.
1416 seq_buf_printf(&s, "anon_thp %llu\n",
1417 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1420 for (i = 0; i < NR_LRU_LISTS; i++)
1421 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1422 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1425 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1426 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1428 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1429 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1432 /* Accumulated memory events */
1434 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1435 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1437 seq_buf_printf(&s, "workingset_refault %lu\n",
1438 memcg_page_state(memcg, WORKINGSET_REFAULT));
1439 seq_buf_printf(&s, "workingset_activate %lu\n",
1440 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1441 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1442 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1444 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1445 seq_buf_printf(&s, "pgscan %lu\n",
1446 memcg_events(memcg, PGSCAN_KSWAPD) +
1447 memcg_events(memcg, PGSCAN_DIRECT));
1448 seq_buf_printf(&s, "pgsteal %lu\n",
1449 memcg_events(memcg, PGSTEAL_KSWAPD) +
1450 memcg_events(memcg, PGSTEAL_DIRECT));
1451 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1452 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1453 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1454 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1456 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1457 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1458 memcg_events(memcg, THP_FAULT_ALLOC));
1459 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1460 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1461 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1463 /* The above should easily fit into one page */
1464 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1469 #define K(x) ((x) << (PAGE_SHIFT-10))
1471 * mem_cgroup_print_oom_context: Print OOM information relevant to
1472 * memory controller.
1473 * @memcg: The memory cgroup that went over limit
1474 * @p: Task that is going to be killed
1476 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1479 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1484 pr_cont(",oom_memcg=");
1485 pr_cont_cgroup_path(memcg->css.cgroup);
1487 pr_cont(",global_oom");
1489 pr_cont(",task_memcg=");
1490 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1496 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1497 * memory controller.
1498 * @memcg: The memory cgroup that went over limit
1500 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1504 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1505 K((u64)page_counter_read(&memcg->memory)),
1506 K((u64)memcg->memory.max), memcg->memory.failcnt);
1507 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1508 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1509 K((u64)page_counter_read(&memcg->swap)),
1510 K((u64)memcg->swap.max), memcg->swap.failcnt);
1512 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1513 K((u64)page_counter_read(&memcg->memsw)),
1514 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1515 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1516 K((u64)page_counter_read(&memcg->kmem)),
1517 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1520 pr_info("Memory cgroup stats for ");
1521 pr_cont_cgroup_path(memcg->css.cgroup);
1523 buf = memory_stat_format(memcg);
1531 * Return the memory (and swap, if configured) limit for a memcg.
1533 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1537 max = memcg->memory.max;
1538 if (mem_cgroup_swappiness(memcg)) {
1539 unsigned long memsw_max;
1540 unsigned long swap_max;
1542 memsw_max = memcg->memsw.max;
1543 swap_max = memcg->swap.max;
1544 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1545 max = min(max + swap_max, memsw_max);
1550 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1553 struct oom_control oc = {
1557 .gfp_mask = gfp_mask,
1562 if (mutex_lock_killable(&oom_lock))
1565 * A few threads which were not waiting at mutex_lock_killable() can
1566 * fail to bail out. Therefore, check again after holding oom_lock.
1568 ret = should_force_charge() || out_of_memory(&oc);
1569 mutex_unlock(&oom_lock);
1573 #if MAX_NUMNODES > 1
1576 * test_mem_cgroup_node_reclaimable
1577 * @memcg: the target memcg
1578 * @nid: the node ID to be checked.
1579 * @noswap : specify true here if the user wants flle only information.
1581 * This function returns whether the specified memcg contains any
1582 * reclaimable pages on a node. Returns true if there are any reclaimable
1583 * pages in the node.
1585 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1586 int nid, bool noswap)
1588 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1590 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1591 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1593 if (noswap || !total_swap_pages)
1595 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1596 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1603 * Always updating the nodemask is not very good - even if we have an empty
1604 * list or the wrong list here, we can start from some node and traverse all
1605 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1608 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1612 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1613 * pagein/pageout changes since the last update.
1615 if (!atomic_read(&memcg->numainfo_events))
1617 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1620 /* make a nodemask where this memcg uses memory from */
1621 memcg->scan_nodes = node_states[N_MEMORY];
1623 for_each_node_mask(nid, node_states[N_MEMORY]) {
1625 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1626 node_clear(nid, memcg->scan_nodes);
1629 atomic_set(&memcg->numainfo_events, 0);
1630 atomic_set(&memcg->numainfo_updating, 0);
1634 * Selecting a node where we start reclaim from. Because what we need is just
1635 * reducing usage counter, start from anywhere is O,K. Considering
1636 * memory reclaim from current node, there are pros. and cons.
1638 * Freeing memory from current node means freeing memory from a node which
1639 * we'll use or we've used. So, it may make LRU bad. And if several threads
1640 * hit limits, it will see a contention on a node. But freeing from remote
1641 * node means more costs for memory reclaim because of memory latency.
1643 * Now, we use round-robin. Better algorithm is welcomed.
1645 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1649 mem_cgroup_may_update_nodemask(memcg);
1650 node = memcg->last_scanned_node;
1652 node = next_node_in(node, memcg->scan_nodes);
1654 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1655 * last time it really checked all the LRUs due to rate limiting.
1656 * Fallback to the current node in that case for simplicity.
1658 if (unlikely(node == MAX_NUMNODES))
1659 node = numa_node_id();
1661 memcg->last_scanned_node = node;
1665 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1671 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1674 unsigned long *total_scanned)
1676 struct mem_cgroup *victim = NULL;
1679 unsigned long excess;
1680 unsigned long nr_scanned;
1681 struct mem_cgroup_reclaim_cookie reclaim = {
1686 excess = soft_limit_excess(root_memcg);
1689 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1694 * If we have not been able to reclaim
1695 * anything, it might because there are
1696 * no reclaimable pages under this hierarchy
1701 * We want to do more targeted reclaim.
1702 * excess >> 2 is not to excessive so as to
1703 * reclaim too much, nor too less that we keep
1704 * coming back to reclaim from this cgroup
1706 if (total >= (excess >> 2) ||
1707 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1712 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1713 pgdat, &nr_scanned);
1714 *total_scanned += nr_scanned;
1715 if (!soft_limit_excess(root_memcg))
1718 mem_cgroup_iter_break(root_memcg, victim);
1722 #ifdef CONFIG_LOCKDEP
1723 static struct lockdep_map memcg_oom_lock_dep_map = {
1724 .name = "memcg_oom_lock",
1728 static DEFINE_SPINLOCK(memcg_oom_lock);
1731 * Check OOM-Killer is already running under our hierarchy.
1732 * If someone is running, return false.
1734 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1736 struct mem_cgroup *iter, *failed = NULL;
1738 spin_lock(&memcg_oom_lock);
1740 for_each_mem_cgroup_tree(iter, memcg) {
1741 if (iter->oom_lock) {
1743 * this subtree of our hierarchy is already locked
1744 * so we cannot give a lock.
1747 mem_cgroup_iter_break(memcg, iter);
1750 iter->oom_lock = true;
1755 * OK, we failed to lock the whole subtree so we have
1756 * to clean up what we set up to the failing subtree
1758 for_each_mem_cgroup_tree(iter, memcg) {
1759 if (iter == failed) {
1760 mem_cgroup_iter_break(memcg, iter);
1763 iter->oom_lock = false;
1766 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1768 spin_unlock(&memcg_oom_lock);
1773 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1775 struct mem_cgroup *iter;
1777 spin_lock(&memcg_oom_lock);
1778 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1779 for_each_mem_cgroup_tree(iter, memcg)
1780 iter->oom_lock = false;
1781 spin_unlock(&memcg_oom_lock);
1784 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1786 struct mem_cgroup *iter;
1788 spin_lock(&memcg_oom_lock);
1789 for_each_mem_cgroup_tree(iter, memcg)
1791 spin_unlock(&memcg_oom_lock);
1794 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1796 struct mem_cgroup *iter;
1799 * When a new child is created while the hierarchy is under oom,
1800 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1802 spin_lock(&memcg_oom_lock);
1803 for_each_mem_cgroup_tree(iter, memcg)
1804 if (iter->under_oom > 0)
1806 spin_unlock(&memcg_oom_lock);
1809 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1811 struct oom_wait_info {
1812 struct mem_cgroup *memcg;
1813 wait_queue_entry_t wait;
1816 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1817 unsigned mode, int sync, void *arg)
1819 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1820 struct mem_cgroup *oom_wait_memcg;
1821 struct oom_wait_info *oom_wait_info;
1823 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1824 oom_wait_memcg = oom_wait_info->memcg;
1826 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1827 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1829 return autoremove_wake_function(wait, mode, sync, arg);
1832 static void memcg_oom_recover(struct mem_cgroup *memcg)
1835 * For the following lockless ->under_oom test, the only required
1836 * guarantee is that it must see the state asserted by an OOM when
1837 * this function is called as a result of userland actions
1838 * triggered by the notification of the OOM. This is trivially
1839 * achieved by invoking mem_cgroup_mark_under_oom() before
1840 * triggering notification.
1842 if (memcg && memcg->under_oom)
1843 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1853 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1855 enum oom_status ret;
1858 if (order > PAGE_ALLOC_COSTLY_ORDER)
1861 memcg_memory_event(memcg, MEMCG_OOM);
1864 * We are in the middle of the charge context here, so we
1865 * don't want to block when potentially sitting on a callstack
1866 * that holds all kinds of filesystem and mm locks.
1868 * cgroup1 allows disabling the OOM killer and waiting for outside
1869 * handling until the charge can succeed; remember the context and put
1870 * the task to sleep at the end of the page fault when all locks are
1873 * On the other hand, in-kernel OOM killer allows for an async victim
1874 * memory reclaim (oom_reaper) and that means that we are not solely
1875 * relying on the oom victim to make a forward progress and we can
1876 * invoke the oom killer here.
1878 * Please note that mem_cgroup_out_of_memory might fail to find a
1879 * victim and then we have to bail out from the charge path.
1881 if (memcg->oom_kill_disable) {
1882 if (!current->in_user_fault)
1884 css_get(&memcg->css);
1885 current->memcg_in_oom = memcg;
1886 current->memcg_oom_gfp_mask = mask;
1887 current->memcg_oom_order = order;
1892 mem_cgroup_mark_under_oom(memcg);
1894 locked = mem_cgroup_oom_trylock(memcg);
1897 mem_cgroup_oom_notify(memcg);
1899 mem_cgroup_unmark_under_oom(memcg);
1900 if (mem_cgroup_out_of_memory(memcg, mask, order))
1906 mem_cgroup_oom_unlock(memcg);
1912 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1913 * @handle: actually kill/wait or just clean up the OOM state
1915 * This has to be called at the end of a page fault if the memcg OOM
1916 * handler was enabled.
1918 * Memcg supports userspace OOM handling where failed allocations must
1919 * sleep on a waitqueue until the userspace task resolves the
1920 * situation. Sleeping directly in the charge context with all kinds
1921 * of locks held is not a good idea, instead we remember an OOM state
1922 * in the task and mem_cgroup_oom_synchronize() has to be called at
1923 * the end of the page fault to complete the OOM handling.
1925 * Returns %true if an ongoing memcg OOM situation was detected and
1926 * completed, %false otherwise.
1928 bool mem_cgroup_oom_synchronize(bool handle)
1930 struct mem_cgroup *memcg = current->memcg_in_oom;
1931 struct oom_wait_info owait;
1934 /* OOM is global, do not handle */
1941 owait.memcg = memcg;
1942 owait.wait.flags = 0;
1943 owait.wait.func = memcg_oom_wake_function;
1944 owait.wait.private = current;
1945 INIT_LIST_HEAD(&owait.wait.entry);
1947 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1948 mem_cgroup_mark_under_oom(memcg);
1950 locked = mem_cgroup_oom_trylock(memcg);
1953 mem_cgroup_oom_notify(memcg);
1955 if (locked && !memcg->oom_kill_disable) {
1956 mem_cgroup_unmark_under_oom(memcg);
1957 finish_wait(&memcg_oom_waitq, &owait.wait);
1958 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1959 current->memcg_oom_order);
1962 mem_cgroup_unmark_under_oom(memcg);
1963 finish_wait(&memcg_oom_waitq, &owait.wait);
1967 mem_cgroup_oom_unlock(memcg);
1969 * There is no guarantee that an OOM-lock contender
1970 * sees the wakeups triggered by the OOM kill
1971 * uncharges. Wake any sleepers explicitely.
1973 memcg_oom_recover(memcg);
1976 current->memcg_in_oom = NULL;
1977 css_put(&memcg->css);
1982 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1983 * @victim: task to be killed by the OOM killer
1984 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1986 * Returns a pointer to a memory cgroup, which has to be cleaned up
1987 * by killing all belonging OOM-killable tasks.
1989 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1991 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1992 struct mem_cgroup *oom_domain)
1994 struct mem_cgroup *oom_group = NULL;
1995 struct mem_cgroup *memcg;
1997 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2001 oom_domain = root_mem_cgroup;
2005 memcg = mem_cgroup_from_task(victim);
2006 if (memcg == root_mem_cgroup)
2010 * Traverse the memory cgroup hierarchy from the victim task's
2011 * cgroup up to the OOMing cgroup (or root) to find the
2012 * highest-level memory cgroup with oom.group set.
2014 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2015 if (memcg->oom_group)
2018 if (memcg == oom_domain)
2023 css_get(&oom_group->css);
2030 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2032 pr_info("Tasks in ");
2033 pr_cont_cgroup_path(memcg->css.cgroup);
2034 pr_cont(" are going to be killed due to memory.oom.group set\n");
2038 * lock_page_memcg - lock a page->mem_cgroup binding
2041 * This function protects unlocked LRU pages from being moved to
2044 * It ensures lifetime of the returned memcg. Caller is responsible
2045 * for the lifetime of the page; __unlock_page_memcg() is available
2046 * when @page might get freed inside the locked section.
2048 struct mem_cgroup *lock_page_memcg(struct page *page)
2050 struct mem_cgroup *memcg;
2051 unsigned long flags;
2054 * The RCU lock is held throughout the transaction. The fast
2055 * path can get away without acquiring the memcg->move_lock
2056 * because page moving starts with an RCU grace period.
2058 * The RCU lock also protects the memcg from being freed when
2059 * the page state that is going to change is the only thing
2060 * preventing the page itself from being freed. E.g. writeback
2061 * doesn't hold a page reference and relies on PG_writeback to
2062 * keep off truncation, migration and so forth.
2066 if (mem_cgroup_disabled())
2069 memcg = page->mem_cgroup;
2070 if (unlikely(!memcg))
2073 if (atomic_read(&memcg->moving_account) <= 0)
2076 spin_lock_irqsave(&memcg->move_lock, flags);
2077 if (memcg != page->mem_cgroup) {
2078 spin_unlock_irqrestore(&memcg->move_lock, flags);
2083 * When charge migration first begins, we can have locked and
2084 * unlocked page stat updates happening concurrently. Track
2085 * the task who has the lock for unlock_page_memcg().
2087 memcg->move_lock_task = current;
2088 memcg->move_lock_flags = flags;
2092 EXPORT_SYMBOL(lock_page_memcg);
2095 * __unlock_page_memcg - unlock and unpin a memcg
2098 * Unlock and unpin a memcg returned by lock_page_memcg().
2100 void __unlock_page_memcg(struct mem_cgroup *memcg)
2102 if (memcg && memcg->move_lock_task == current) {
2103 unsigned long flags = memcg->move_lock_flags;
2105 memcg->move_lock_task = NULL;
2106 memcg->move_lock_flags = 0;
2108 spin_unlock_irqrestore(&memcg->move_lock, flags);
2115 * unlock_page_memcg - unlock a page->mem_cgroup binding
2118 void unlock_page_memcg(struct page *page)
2120 __unlock_page_memcg(page->mem_cgroup);
2122 EXPORT_SYMBOL(unlock_page_memcg);
2124 struct memcg_stock_pcp {
2125 struct mem_cgroup *cached; /* this never be root cgroup */
2126 unsigned int nr_pages;
2127 struct work_struct work;
2128 unsigned long flags;
2129 #define FLUSHING_CACHED_CHARGE 0
2131 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2132 static DEFINE_MUTEX(percpu_charge_mutex);
2135 * consume_stock: Try to consume stocked charge on this cpu.
2136 * @memcg: memcg to consume from.
2137 * @nr_pages: how many pages to charge.
2139 * The charges will only happen if @memcg matches the current cpu's memcg
2140 * stock, and at least @nr_pages are available in that stock. Failure to
2141 * service an allocation will refill the stock.
2143 * returns true if successful, false otherwise.
2145 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2147 struct memcg_stock_pcp *stock;
2148 unsigned long flags;
2151 if (nr_pages > MEMCG_CHARGE_BATCH)
2154 local_irq_save(flags);
2156 stock = this_cpu_ptr(&memcg_stock);
2157 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2158 stock->nr_pages -= nr_pages;
2162 local_irq_restore(flags);
2168 * Returns stocks cached in percpu and reset cached information.
2170 static void drain_stock(struct memcg_stock_pcp *stock)
2172 struct mem_cgroup *old = stock->cached;
2174 if (stock->nr_pages) {
2175 page_counter_uncharge(&old->memory, stock->nr_pages);
2176 if (do_memsw_account())
2177 page_counter_uncharge(&old->memsw, stock->nr_pages);
2178 css_put_many(&old->css, stock->nr_pages);
2179 stock->nr_pages = 0;
2181 stock->cached = NULL;
2184 static void drain_local_stock(struct work_struct *dummy)
2186 struct memcg_stock_pcp *stock;
2187 unsigned long flags;
2190 * The only protection from memory hotplug vs. drain_stock races is
2191 * that we always operate on local CPU stock here with IRQ disabled
2193 local_irq_save(flags);
2195 stock = this_cpu_ptr(&memcg_stock);
2197 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2199 local_irq_restore(flags);
2203 * Cache charges(val) to local per_cpu area.
2204 * This will be consumed by consume_stock() function, later.
2206 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2208 struct memcg_stock_pcp *stock;
2209 unsigned long flags;
2211 local_irq_save(flags);
2213 stock = this_cpu_ptr(&memcg_stock);
2214 if (stock->cached != memcg) { /* reset if necessary */
2216 stock->cached = memcg;
2218 stock->nr_pages += nr_pages;
2220 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2223 local_irq_restore(flags);
2227 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2228 * of the hierarchy under it.
2230 static void drain_all_stock(struct mem_cgroup *root_memcg)
2234 /* If someone's already draining, avoid adding running more workers. */
2235 if (!mutex_trylock(&percpu_charge_mutex))
2238 * Notify other cpus that system-wide "drain" is running
2239 * We do not care about races with the cpu hotplug because cpu down
2240 * as well as workers from this path always operate on the local
2241 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2244 for_each_online_cpu(cpu) {
2245 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2246 struct mem_cgroup *memcg;
2248 memcg = stock->cached;
2249 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2251 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2252 css_put(&memcg->css);
2255 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2257 drain_local_stock(&stock->work);
2259 schedule_work_on(cpu, &stock->work);
2261 css_put(&memcg->css);
2264 mutex_unlock(&percpu_charge_mutex);
2267 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2269 struct memcg_stock_pcp *stock;
2270 struct mem_cgroup *memcg, *mi;
2272 stock = &per_cpu(memcg_stock, cpu);
2275 for_each_mem_cgroup(memcg) {
2278 for (i = 0; i < MEMCG_NR_STAT; i++) {
2282 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2284 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2285 atomic_long_add(x, &memcg->vmstats[i]);
2287 if (i >= NR_VM_NODE_STAT_ITEMS)
2290 for_each_node(nid) {
2291 struct mem_cgroup_per_node *pn;
2293 pn = mem_cgroup_nodeinfo(memcg, nid);
2294 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2297 atomic_long_add(x, &pn->lruvec_stat[i]);
2298 } while ((pn = parent_nodeinfo(pn, nid)));
2302 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2305 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2307 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2308 atomic_long_add(x, &memcg->vmevents[i]);
2315 static void reclaim_high(struct mem_cgroup *memcg,
2316 unsigned int nr_pages,
2320 if (page_counter_read(&memcg->memory) <= memcg->high)
2322 memcg_memory_event(memcg, MEMCG_HIGH);
2323 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2324 } while ((memcg = parent_mem_cgroup(memcg)));
2327 static void high_work_func(struct work_struct *work)
2329 struct mem_cgroup *memcg;
2331 memcg = container_of(work, struct mem_cgroup, high_work);
2332 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2336 * Scheduled by try_charge() to be executed from the userland return path
2337 * and reclaims memory over the high limit.
2339 void mem_cgroup_handle_over_high(void)
2341 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2342 struct mem_cgroup *memcg;
2344 if (likely(!nr_pages))
2347 memcg = get_mem_cgroup_from_mm(current->mm);
2348 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2349 css_put(&memcg->css);
2350 current->memcg_nr_pages_over_high = 0;
2353 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2354 unsigned int nr_pages)
2356 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2357 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2358 struct mem_cgroup *mem_over_limit;
2359 struct page_counter *counter;
2360 unsigned long nr_reclaimed;
2361 bool may_swap = true;
2362 bool drained = false;
2363 enum oom_status oom_status;
2365 if (mem_cgroup_is_root(memcg))
2368 if (consume_stock(memcg, nr_pages))
2371 if (!do_memsw_account() ||
2372 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2373 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2375 if (do_memsw_account())
2376 page_counter_uncharge(&memcg->memsw, batch);
2377 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2379 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2383 if (batch > nr_pages) {
2389 * Unlike in global OOM situations, memcg is not in a physical
2390 * memory shortage. Allow dying and OOM-killed tasks to
2391 * bypass the last charges so that they can exit quickly and
2392 * free their memory.
2394 if (unlikely(should_force_charge()))
2398 * Prevent unbounded recursion when reclaim operations need to
2399 * allocate memory. This might exceed the limits temporarily,
2400 * but we prefer facilitating memory reclaim and getting back
2401 * under the limit over triggering OOM kills in these cases.
2403 if (unlikely(current->flags & PF_MEMALLOC))
2406 if (unlikely(task_in_memcg_oom(current)))
2409 if (!gfpflags_allow_blocking(gfp_mask))
2412 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2414 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2415 gfp_mask, may_swap);
2417 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2421 drain_all_stock(mem_over_limit);
2426 if (gfp_mask & __GFP_NORETRY)
2429 * Even though the limit is exceeded at this point, reclaim
2430 * may have been able to free some pages. Retry the charge
2431 * before killing the task.
2433 * Only for regular pages, though: huge pages are rather
2434 * unlikely to succeed so close to the limit, and we fall back
2435 * to regular pages anyway in case of failure.
2437 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2440 * At task move, charge accounts can be doubly counted. So, it's
2441 * better to wait until the end of task_move if something is going on.
2443 if (mem_cgroup_wait_acct_move(mem_over_limit))
2449 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2452 if (gfp_mask & __GFP_NOFAIL)
2455 if (fatal_signal_pending(current))
2459 * keep retrying as long as the memcg oom killer is able to make
2460 * a forward progress or bypass the charge if the oom killer
2461 * couldn't make any progress.
2463 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2464 get_order(nr_pages * PAGE_SIZE));
2465 switch (oom_status) {
2467 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2475 if (!(gfp_mask & __GFP_NOFAIL))
2479 * The allocation either can't fail or will lead to more memory
2480 * being freed very soon. Allow memory usage go over the limit
2481 * temporarily by force charging it.
2483 page_counter_charge(&memcg->memory, nr_pages);
2484 if (do_memsw_account())
2485 page_counter_charge(&memcg->memsw, nr_pages);
2486 css_get_many(&memcg->css, nr_pages);
2491 css_get_many(&memcg->css, batch);
2492 if (batch > nr_pages)
2493 refill_stock(memcg, batch - nr_pages);
2496 * If the hierarchy is above the normal consumption range, schedule
2497 * reclaim on returning to userland. We can perform reclaim here
2498 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2499 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2500 * not recorded as it most likely matches current's and won't
2501 * change in the meantime. As high limit is checked again before
2502 * reclaim, the cost of mismatch is negligible.
2505 if (page_counter_read(&memcg->memory) > memcg->high) {
2506 /* Don't bother a random interrupted task */
2507 if (in_interrupt()) {
2508 schedule_work(&memcg->high_work);
2511 current->memcg_nr_pages_over_high += batch;
2512 set_notify_resume(current);
2515 } while ((memcg = parent_mem_cgroup(memcg)));
2520 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2522 if (mem_cgroup_is_root(memcg))
2525 page_counter_uncharge(&memcg->memory, nr_pages);
2526 if (do_memsw_account())
2527 page_counter_uncharge(&memcg->memsw, nr_pages);
2529 css_put_many(&memcg->css, nr_pages);
2532 static void lock_page_lru(struct page *page, int *isolated)
2534 pg_data_t *pgdat = page_pgdat(page);
2536 spin_lock_irq(&pgdat->lru_lock);
2537 if (PageLRU(page)) {
2538 struct lruvec *lruvec;
2540 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2542 del_page_from_lru_list(page, lruvec, page_lru(page));
2548 static void unlock_page_lru(struct page *page, int isolated)
2550 pg_data_t *pgdat = page_pgdat(page);
2553 struct lruvec *lruvec;
2555 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2556 VM_BUG_ON_PAGE(PageLRU(page), page);
2558 add_page_to_lru_list(page, lruvec, page_lru(page));
2560 spin_unlock_irq(&pgdat->lru_lock);
2563 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2568 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2571 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2572 * may already be on some other mem_cgroup's LRU. Take care of it.
2575 lock_page_lru(page, &isolated);
2578 * Nobody should be changing or seriously looking at
2579 * page->mem_cgroup at this point:
2581 * - the page is uncharged
2583 * - the page is off-LRU
2585 * - an anonymous fault has exclusive page access, except for
2586 * a locked page table
2588 * - a page cache insertion, a swapin fault, or a migration
2589 * have the page locked
2591 page->mem_cgroup = memcg;
2594 unlock_page_lru(page, isolated);
2597 #ifdef CONFIG_MEMCG_KMEM
2598 static int memcg_alloc_cache_id(void)
2603 id = ida_simple_get(&memcg_cache_ida,
2604 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2608 if (id < memcg_nr_cache_ids)
2612 * There's no space for the new id in memcg_caches arrays,
2613 * so we have to grow them.
2615 down_write(&memcg_cache_ids_sem);
2617 size = 2 * (id + 1);
2618 if (size < MEMCG_CACHES_MIN_SIZE)
2619 size = MEMCG_CACHES_MIN_SIZE;
2620 else if (size > MEMCG_CACHES_MAX_SIZE)
2621 size = MEMCG_CACHES_MAX_SIZE;
2623 err = memcg_update_all_caches(size);
2625 err = memcg_update_all_list_lrus(size);
2627 memcg_nr_cache_ids = size;
2629 up_write(&memcg_cache_ids_sem);
2632 ida_simple_remove(&memcg_cache_ida, id);
2638 static void memcg_free_cache_id(int id)
2640 ida_simple_remove(&memcg_cache_ida, id);
2643 struct memcg_kmem_cache_create_work {
2644 struct mem_cgroup *memcg;
2645 struct kmem_cache *cachep;
2646 struct work_struct work;
2649 static void memcg_kmem_cache_create_func(struct work_struct *w)
2651 struct memcg_kmem_cache_create_work *cw =
2652 container_of(w, struct memcg_kmem_cache_create_work, work);
2653 struct mem_cgroup *memcg = cw->memcg;
2654 struct kmem_cache *cachep = cw->cachep;
2656 memcg_create_kmem_cache(memcg, cachep);
2658 css_put(&memcg->css);
2663 * Enqueue the creation of a per-memcg kmem_cache.
2665 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2666 struct kmem_cache *cachep)
2668 struct memcg_kmem_cache_create_work *cw;
2670 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2674 css_get(&memcg->css);
2677 cw->cachep = cachep;
2678 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2680 queue_work(memcg_kmem_cache_wq, &cw->work);
2683 static inline bool memcg_kmem_bypass(void)
2685 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2691 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2692 * @cachep: the original global kmem cache
2694 * Return the kmem_cache we're supposed to use for a slab allocation.
2695 * We try to use the current memcg's version of the cache.
2697 * If the cache does not exist yet, if we are the first user of it, we
2698 * create it asynchronously in a workqueue and let the current allocation
2699 * go through with the original cache.
2701 * This function takes a reference to the cache it returns to assure it
2702 * won't get destroyed while we are working with it. Once the caller is
2703 * done with it, memcg_kmem_put_cache() must be called to release the
2706 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2708 struct mem_cgroup *memcg;
2709 struct kmem_cache *memcg_cachep;
2712 VM_BUG_ON(!is_root_cache(cachep));
2714 if (memcg_kmem_bypass())
2717 memcg = get_mem_cgroup_from_current();
2718 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2722 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2723 if (likely(memcg_cachep))
2724 return memcg_cachep;
2727 * If we are in a safe context (can wait, and not in interrupt
2728 * context), we could be be predictable and return right away.
2729 * This would guarantee that the allocation being performed
2730 * already belongs in the new cache.
2732 * However, there are some clashes that can arrive from locking.
2733 * For instance, because we acquire the slab_mutex while doing
2734 * memcg_create_kmem_cache, this means no further allocation
2735 * could happen with the slab_mutex held. So it's better to
2738 memcg_schedule_kmem_cache_create(memcg, cachep);
2740 css_put(&memcg->css);
2745 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2746 * @cachep: the cache returned by memcg_kmem_get_cache
2748 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2750 if (!is_root_cache(cachep))
2751 css_put(&cachep->memcg_params.memcg->css);
2755 * __memcg_kmem_charge_memcg: charge a kmem page
2756 * @page: page to charge
2757 * @gfp: reclaim mode
2758 * @order: allocation order
2759 * @memcg: memory cgroup to charge
2761 * Returns 0 on success, an error code on failure.
2763 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2764 struct mem_cgroup *memcg)
2766 unsigned int nr_pages = 1 << order;
2767 struct page_counter *counter;
2770 ret = try_charge(memcg, gfp, nr_pages);
2774 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2775 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2776 cancel_charge(memcg, nr_pages);
2780 page->mem_cgroup = memcg;
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 __SetPageKmemcg(page);
2807 css_put(&memcg->css);
2812 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2813 * @memcg: memcg to uncharge
2814 * @nr_pages: number of pages to uncharge
2816 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2817 unsigned int nr_pages)
2819 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2820 page_counter_uncharge(&memcg->kmem, nr_pages);
2822 page_counter_uncharge(&memcg->memory, nr_pages);
2823 if (do_memsw_account())
2824 page_counter_uncharge(&memcg->memsw, nr_pages);
2827 * __memcg_kmem_uncharge: uncharge a kmem page
2828 * @page: page to uncharge
2829 * @order: allocation order
2831 void __memcg_kmem_uncharge(struct page *page, int order)
2833 struct mem_cgroup *memcg = page->mem_cgroup;
2834 unsigned int nr_pages = 1 << order;
2839 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2840 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
2841 page->mem_cgroup = NULL;
2843 /* slab pages do not have PageKmemcg flag set */
2844 if (PageKmemcg(page))
2845 __ClearPageKmemcg(page);
2847 css_put_many(&memcg->css, nr_pages);
2849 #endif /* CONFIG_MEMCG_KMEM */
2851 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2854 * Because tail pages are not marked as "used", set it. We're under
2855 * pgdat->lru_lock and migration entries setup in all page mappings.
2857 void mem_cgroup_split_huge_fixup(struct page *head)
2861 if (mem_cgroup_disabled())
2864 for (i = 1; i < HPAGE_PMD_NR; i++)
2865 head[i].mem_cgroup = head->mem_cgroup;
2867 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2869 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2871 #ifdef CONFIG_MEMCG_SWAP
2873 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2874 * @entry: swap entry to be moved
2875 * @from: mem_cgroup which the entry is moved from
2876 * @to: mem_cgroup which the entry is moved to
2878 * It succeeds only when the swap_cgroup's record for this entry is the same
2879 * as the mem_cgroup's id of @from.
2881 * Returns 0 on success, -EINVAL on failure.
2883 * The caller must have charged to @to, IOW, called page_counter_charge() about
2884 * both res and memsw, and called css_get().
2886 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2887 struct mem_cgroup *from, struct mem_cgroup *to)
2889 unsigned short old_id, new_id;
2891 old_id = mem_cgroup_id(from);
2892 new_id = mem_cgroup_id(to);
2894 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2895 mod_memcg_state(from, MEMCG_SWAP, -1);
2896 mod_memcg_state(to, MEMCG_SWAP, 1);
2902 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2903 struct mem_cgroup *from, struct mem_cgroup *to)
2909 static DEFINE_MUTEX(memcg_max_mutex);
2911 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2912 unsigned long max, bool memsw)
2914 bool enlarge = false;
2915 bool drained = false;
2917 bool limits_invariant;
2918 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2921 if (signal_pending(current)) {
2926 mutex_lock(&memcg_max_mutex);
2928 * Make sure that the new limit (memsw or memory limit) doesn't
2929 * break our basic invariant rule memory.max <= memsw.max.
2931 limits_invariant = memsw ? max >= memcg->memory.max :
2932 max <= memcg->memsw.max;
2933 if (!limits_invariant) {
2934 mutex_unlock(&memcg_max_mutex);
2938 if (max > counter->max)
2940 ret = page_counter_set_max(counter, max);
2941 mutex_unlock(&memcg_max_mutex);
2947 drain_all_stock(memcg);
2952 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2953 GFP_KERNEL, !memsw)) {
2959 if (!ret && enlarge)
2960 memcg_oom_recover(memcg);
2965 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2967 unsigned long *total_scanned)
2969 unsigned long nr_reclaimed = 0;
2970 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2971 unsigned long reclaimed;
2973 struct mem_cgroup_tree_per_node *mctz;
2974 unsigned long excess;
2975 unsigned long nr_scanned;
2980 mctz = soft_limit_tree_node(pgdat->node_id);
2983 * Do not even bother to check the largest node if the root
2984 * is empty. Do it lockless to prevent lock bouncing. Races
2985 * are acceptable as soft limit is best effort anyway.
2987 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2991 * This loop can run a while, specially if mem_cgroup's continuously
2992 * keep exceeding their soft limit and putting the system under
2999 mz = mem_cgroup_largest_soft_limit_node(mctz);
3004 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3005 gfp_mask, &nr_scanned);
3006 nr_reclaimed += reclaimed;
3007 *total_scanned += nr_scanned;
3008 spin_lock_irq(&mctz->lock);
3009 __mem_cgroup_remove_exceeded(mz, mctz);
3012 * If we failed to reclaim anything from this memory cgroup
3013 * it is time to move on to the next cgroup
3017 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3019 excess = soft_limit_excess(mz->memcg);
3021 * One school of thought says that we should not add
3022 * back the node to the tree if reclaim returns 0.
3023 * But our reclaim could return 0, simply because due
3024 * to priority we are exposing a smaller subset of
3025 * memory to reclaim from. Consider this as a longer
3028 /* If excess == 0, no tree ops */
3029 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3030 spin_unlock_irq(&mctz->lock);
3031 css_put(&mz->memcg->css);
3034 * Could not reclaim anything and there are no more
3035 * mem cgroups to try or we seem to be looping without
3036 * reclaiming anything.
3038 if (!nr_reclaimed &&
3040 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3042 } while (!nr_reclaimed);
3044 css_put(&next_mz->memcg->css);
3045 return nr_reclaimed;
3049 * Test whether @memcg has children, dead or alive. Note that this
3050 * function doesn't care whether @memcg has use_hierarchy enabled and
3051 * returns %true if there are child csses according to the cgroup
3052 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3054 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3059 ret = css_next_child(NULL, &memcg->css);
3065 * Reclaims as many pages from the given memcg as possible.
3067 * Caller is responsible for holding css reference for memcg.
3069 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3071 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3073 /* we call try-to-free pages for make this cgroup empty */
3074 lru_add_drain_all();
3076 drain_all_stock(memcg);
3078 /* try to free all pages in this cgroup */
3079 while (nr_retries && page_counter_read(&memcg->memory)) {
3082 if (signal_pending(current))
3085 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3089 /* maybe some writeback is necessary */
3090 congestion_wait(BLK_RW_ASYNC, HZ/10);
3098 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3099 char *buf, size_t nbytes,
3102 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3104 if (mem_cgroup_is_root(memcg))
3106 return mem_cgroup_force_empty(memcg) ?: nbytes;
3109 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3112 return mem_cgroup_from_css(css)->use_hierarchy;
3115 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3116 struct cftype *cft, u64 val)
3119 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3120 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3122 if (memcg->use_hierarchy == val)
3126 * If parent's use_hierarchy is set, we can't make any modifications
3127 * in the child subtrees. If it is unset, then the change can
3128 * occur, provided the current cgroup has no children.
3130 * For the root cgroup, parent_mem is NULL, we allow value to be
3131 * set if there are no children.
3133 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3134 (val == 1 || val == 0)) {
3135 if (!memcg_has_children(memcg))
3136 memcg->use_hierarchy = val;
3145 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3149 if (mem_cgroup_is_root(memcg)) {
3150 val = memcg_page_state(memcg, MEMCG_CACHE) +
3151 memcg_page_state(memcg, MEMCG_RSS);
3153 val += memcg_page_state(memcg, MEMCG_SWAP);
3156 val = page_counter_read(&memcg->memory);
3158 val = page_counter_read(&memcg->memsw);
3171 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3174 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3175 struct page_counter *counter;
3177 switch (MEMFILE_TYPE(cft->private)) {
3179 counter = &memcg->memory;
3182 counter = &memcg->memsw;
3185 counter = &memcg->kmem;
3188 counter = &memcg->tcpmem;
3194 switch (MEMFILE_ATTR(cft->private)) {
3196 if (counter == &memcg->memory)
3197 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3198 if (counter == &memcg->memsw)
3199 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3200 return (u64)page_counter_read(counter) * PAGE_SIZE;
3202 return (u64)counter->max * PAGE_SIZE;
3204 return (u64)counter->watermark * PAGE_SIZE;
3206 return counter->failcnt;
3207 case RES_SOFT_LIMIT:
3208 return (u64)memcg->soft_limit * PAGE_SIZE;
3214 #ifdef CONFIG_MEMCG_KMEM
3215 static int memcg_online_kmem(struct mem_cgroup *memcg)
3219 if (cgroup_memory_nokmem)
3222 BUG_ON(memcg->kmemcg_id >= 0);
3223 BUG_ON(memcg->kmem_state);
3225 memcg_id = memcg_alloc_cache_id();
3229 static_branch_inc(&memcg_kmem_enabled_key);
3231 * A memory cgroup is considered kmem-online as soon as it gets
3232 * kmemcg_id. Setting the id after enabling static branching will
3233 * guarantee no one starts accounting before all call sites are
3236 memcg->kmemcg_id = memcg_id;
3237 memcg->kmem_state = KMEM_ONLINE;
3238 INIT_LIST_HEAD(&memcg->kmem_caches);
3243 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3245 struct cgroup_subsys_state *css;
3246 struct mem_cgroup *parent, *child;
3249 if (memcg->kmem_state != KMEM_ONLINE)
3252 * Clear the online state before clearing memcg_caches array
3253 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3254 * guarantees that no cache will be created for this cgroup
3255 * after we are done (see memcg_create_kmem_cache()).
3257 memcg->kmem_state = KMEM_ALLOCATED;
3259 memcg_deactivate_kmem_caches(memcg);
3261 kmemcg_id = memcg->kmemcg_id;
3262 BUG_ON(kmemcg_id < 0);
3264 parent = parent_mem_cgroup(memcg);
3266 parent = root_mem_cgroup;
3269 * Change kmemcg_id of this cgroup and all its descendants to the
3270 * parent's id, and then move all entries from this cgroup's list_lrus
3271 * to ones of the parent. After we have finished, all list_lrus
3272 * corresponding to this cgroup are guaranteed to remain empty. The
3273 * ordering is imposed by list_lru_node->lock taken by
3274 * memcg_drain_all_list_lrus().
3276 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3277 css_for_each_descendant_pre(css, &memcg->css) {
3278 child = mem_cgroup_from_css(css);
3279 BUG_ON(child->kmemcg_id != kmemcg_id);
3280 child->kmemcg_id = parent->kmemcg_id;
3281 if (!memcg->use_hierarchy)
3286 memcg_drain_all_list_lrus(kmemcg_id, parent);
3288 memcg_free_cache_id(kmemcg_id);
3291 static void memcg_free_kmem(struct mem_cgroup *memcg)
3293 /* css_alloc() failed, offlining didn't happen */
3294 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3295 memcg_offline_kmem(memcg);
3297 if (memcg->kmem_state == KMEM_ALLOCATED) {
3298 memcg_destroy_kmem_caches(memcg);
3299 static_branch_dec(&memcg_kmem_enabled_key);
3300 WARN_ON(page_counter_read(&memcg->kmem));
3304 static int memcg_online_kmem(struct mem_cgroup *memcg)
3308 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3311 static void memcg_free_kmem(struct mem_cgroup *memcg)
3314 #endif /* CONFIG_MEMCG_KMEM */
3316 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3321 mutex_lock(&memcg_max_mutex);
3322 ret = page_counter_set_max(&memcg->kmem, max);
3323 mutex_unlock(&memcg_max_mutex);
3327 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3331 mutex_lock(&memcg_max_mutex);
3333 ret = page_counter_set_max(&memcg->tcpmem, max);
3337 if (!memcg->tcpmem_active) {
3339 * The active flag needs to be written after the static_key
3340 * update. This is what guarantees that the socket activation
3341 * function is the last one to run. See mem_cgroup_sk_alloc()
3342 * for details, and note that we don't mark any socket as
3343 * belonging to this memcg until that flag is up.
3345 * We need to do this, because static_keys will span multiple
3346 * sites, but we can't control their order. If we mark a socket
3347 * as accounted, but the accounting functions are not patched in
3348 * yet, we'll lose accounting.
3350 * We never race with the readers in mem_cgroup_sk_alloc(),
3351 * because when this value change, the code to process it is not
3354 static_branch_inc(&memcg_sockets_enabled_key);
3355 memcg->tcpmem_active = true;
3358 mutex_unlock(&memcg_max_mutex);
3363 * The user of this function is...
3366 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3367 char *buf, size_t nbytes, loff_t off)
3369 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3370 unsigned long nr_pages;
3373 buf = strstrip(buf);
3374 ret = page_counter_memparse(buf, "-1", &nr_pages);
3378 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3380 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3384 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3386 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3389 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3392 ret = memcg_update_kmem_max(memcg, nr_pages);
3395 ret = memcg_update_tcp_max(memcg, nr_pages);
3399 case RES_SOFT_LIMIT:
3400 memcg->soft_limit = nr_pages;
3404 return ret ?: nbytes;
3407 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3408 size_t nbytes, loff_t off)
3410 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3411 struct page_counter *counter;
3413 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3415 counter = &memcg->memory;
3418 counter = &memcg->memsw;
3421 counter = &memcg->kmem;
3424 counter = &memcg->tcpmem;
3430 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3432 page_counter_reset_watermark(counter);
3435 counter->failcnt = 0;
3444 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3447 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3451 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3452 struct cftype *cft, u64 val)
3454 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3456 if (val & ~MOVE_MASK)
3460 * No kind of locking is needed in here, because ->can_attach() will
3461 * check this value once in the beginning of the process, and then carry
3462 * on with stale data. This means that changes to this value will only
3463 * affect task migrations starting after the change.
3465 memcg->move_charge_at_immigrate = val;
3469 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3470 struct cftype *cft, u64 val)
3478 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3479 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3480 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3482 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3483 int nid, unsigned int lru_mask)
3485 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3486 unsigned long nr = 0;
3489 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3492 if (!(BIT(lru) & lru_mask))
3494 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3499 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3500 unsigned int lru_mask)
3502 unsigned long nr = 0;
3506 if (!(BIT(lru) & lru_mask))
3508 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3513 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3517 unsigned int lru_mask;
3520 static const struct numa_stat stats[] = {
3521 { "total", LRU_ALL },
3522 { "file", LRU_ALL_FILE },
3523 { "anon", LRU_ALL_ANON },
3524 { "unevictable", BIT(LRU_UNEVICTABLE) },
3526 const struct numa_stat *stat;
3529 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3531 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3532 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3533 seq_printf(m, "%s=%lu", stat->name, nr);
3534 for_each_node_state(nid, N_MEMORY) {
3535 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3537 seq_printf(m, " N%d=%lu", nid, nr);
3542 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3543 struct mem_cgroup *iter;
3546 for_each_mem_cgroup_tree(iter, memcg)
3547 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3548 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3549 for_each_node_state(nid, N_MEMORY) {
3551 for_each_mem_cgroup_tree(iter, memcg)
3552 nr += mem_cgroup_node_nr_lru_pages(
3553 iter, nid, stat->lru_mask);
3554 seq_printf(m, " N%d=%lu", nid, nr);
3561 #endif /* CONFIG_NUMA */
3563 static const unsigned int memcg1_stats[] = {
3574 static const char *const memcg1_stat_names[] = {
3585 /* Universal VM events cgroup1 shows, original sort order */
3586 static const unsigned int memcg1_events[] = {
3593 static const char *const memcg1_event_names[] = {
3600 static int memcg_stat_show(struct seq_file *m, void *v)
3602 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3603 unsigned long memory, memsw;
3604 struct mem_cgroup *mi;
3607 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3608 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3610 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3611 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3613 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3614 memcg_page_state_local(memcg, memcg1_stats[i]) *
3618 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3619 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3620 memcg_events_local(memcg, memcg1_events[i]));
3622 for (i = 0; i < NR_LRU_LISTS; i++)
3623 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3624 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3627 /* Hierarchical information */
3628 memory = memsw = PAGE_COUNTER_MAX;
3629 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3630 memory = min(memory, mi->memory.max);
3631 memsw = min(memsw, mi->memsw.max);
3633 seq_printf(m, "hierarchical_memory_limit %llu\n",
3634 (u64)memory * PAGE_SIZE);
3635 if (do_memsw_account())
3636 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3637 (u64)memsw * PAGE_SIZE);
3639 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3640 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3642 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3643 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3647 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3648 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3649 (u64)memcg_events(memcg, memcg1_events[i]));
3651 for (i = 0; i < NR_LRU_LISTS; i++)
3652 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3653 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3656 #ifdef CONFIG_DEBUG_VM
3659 struct mem_cgroup_per_node *mz;
3660 struct zone_reclaim_stat *rstat;
3661 unsigned long recent_rotated[2] = {0, 0};
3662 unsigned long recent_scanned[2] = {0, 0};
3664 for_each_online_pgdat(pgdat) {
3665 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3666 rstat = &mz->lruvec.reclaim_stat;
3668 recent_rotated[0] += rstat->recent_rotated[0];
3669 recent_rotated[1] += rstat->recent_rotated[1];
3670 recent_scanned[0] += rstat->recent_scanned[0];
3671 recent_scanned[1] += rstat->recent_scanned[1];
3673 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3674 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3675 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3676 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3683 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3686 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3688 return mem_cgroup_swappiness(memcg);
3691 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3692 struct cftype *cft, u64 val)
3694 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3700 memcg->swappiness = val;
3702 vm_swappiness = val;
3707 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3709 struct mem_cgroup_threshold_ary *t;
3710 unsigned long usage;
3715 t = rcu_dereference(memcg->thresholds.primary);
3717 t = rcu_dereference(memcg->memsw_thresholds.primary);
3722 usage = mem_cgroup_usage(memcg, swap);
3725 * current_threshold points to threshold just below or equal to usage.
3726 * If it's not true, a threshold was crossed after last
3727 * call of __mem_cgroup_threshold().
3729 i = t->current_threshold;
3732 * Iterate backward over array of thresholds starting from
3733 * current_threshold and check if a threshold is crossed.
3734 * If none of thresholds below usage is crossed, we read
3735 * only one element of the array here.
3737 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3738 eventfd_signal(t->entries[i].eventfd, 1);
3740 /* i = current_threshold + 1 */
3744 * Iterate forward over array of thresholds starting from
3745 * current_threshold+1 and check if a threshold is crossed.
3746 * If none of thresholds above usage is crossed, we read
3747 * only one element of the array here.
3749 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3750 eventfd_signal(t->entries[i].eventfd, 1);
3752 /* Update current_threshold */
3753 t->current_threshold = i - 1;
3758 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3761 __mem_cgroup_threshold(memcg, false);
3762 if (do_memsw_account())
3763 __mem_cgroup_threshold(memcg, true);
3765 memcg = parent_mem_cgroup(memcg);
3769 static int compare_thresholds(const void *a, const void *b)
3771 const struct mem_cgroup_threshold *_a = a;
3772 const struct mem_cgroup_threshold *_b = b;
3774 if (_a->threshold > _b->threshold)
3777 if (_a->threshold < _b->threshold)
3783 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3785 struct mem_cgroup_eventfd_list *ev;
3787 spin_lock(&memcg_oom_lock);
3789 list_for_each_entry(ev, &memcg->oom_notify, list)
3790 eventfd_signal(ev->eventfd, 1);
3792 spin_unlock(&memcg_oom_lock);
3796 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3798 struct mem_cgroup *iter;
3800 for_each_mem_cgroup_tree(iter, memcg)
3801 mem_cgroup_oom_notify_cb(iter);
3804 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3805 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3807 struct mem_cgroup_thresholds *thresholds;
3808 struct mem_cgroup_threshold_ary *new;
3809 unsigned long threshold;
3810 unsigned long usage;
3813 ret = page_counter_memparse(args, "-1", &threshold);
3817 mutex_lock(&memcg->thresholds_lock);
3820 thresholds = &memcg->thresholds;
3821 usage = mem_cgroup_usage(memcg, false);
3822 } else if (type == _MEMSWAP) {
3823 thresholds = &memcg->memsw_thresholds;
3824 usage = mem_cgroup_usage(memcg, true);
3828 /* Check if a threshold crossed before adding a new one */
3829 if (thresholds->primary)
3830 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3832 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3834 /* Allocate memory for new array of thresholds */
3835 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3842 /* Copy thresholds (if any) to new array */
3843 if (thresholds->primary) {
3844 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3845 sizeof(struct mem_cgroup_threshold));
3848 /* Add new threshold */
3849 new->entries[size - 1].eventfd = eventfd;
3850 new->entries[size - 1].threshold = threshold;
3852 /* Sort thresholds. Registering of new threshold isn't time-critical */
3853 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3854 compare_thresholds, NULL);
3856 /* Find current threshold */
3857 new->current_threshold = -1;
3858 for (i = 0; i < size; i++) {
3859 if (new->entries[i].threshold <= usage) {
3861 * new->current_threshold will not be used until
3862 * rcu_assign_pointer(), so it's safe to increment
3865 ++new->current_threshold;
3870 /* Free old spare buffer and save old primary buffer as spare */
3871 kfree(thresholds->spare);
3872 thresholds->spare = thresholds->primary;
3874 rcu_assign_pointer(thresholds->primary, new);
3876 /* To be sure that nobody uses thresholds */
3880 mutex_unlock(&memcg->thresholds_lock);
3885 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3886 struct eventfd_ctx *eventfd, const char *args)
3888 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3891 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3892 struct eventfd_ctx *eventfd, const char *args)
3894 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3897 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3898 struct eventfd_ctx *eventfd, enum res_type type)
3900 struct mem_cgroup_thresholds *thresholds;
3901 struct mem_cgroup_threshold_ary *new;
3902 unsigned long usage;
3905 mutex_lock(&memcg->thresholds_lock);
3908 thresholds = &memcg->thresholds;
3909 usage = mem_cgroup_usage(memcg, false);
3910 } else if (type == _MEMSWAP) {
3911 thresholds = &memcg->memsw_thresholds;
3912 usage = mem_cgroup_usage(memcg, true);
3916 if (!thresholds->primary)
3919 /* Check if a threshold crossed before removing */
3920 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3922 /* Calculate new number of threshold */
3924 for (i = 0; i < thresholds->primary->size; i++) {
3925 if (thresholds->primary->entries[i].eventfd != eventfd)
3929 new = thresholds->spare;
3931 /* Set thresholds array to NULL if we don't have thresholds */
3940 /* Copy thresholds and find current threshold */
3941 new->current_threshold = -1;
3942 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3943 if (thresholds->primary->entries[i].eventfd == eventfd)
3946 new->entries[j] = thresholds->primary->entries[i];
3947 if (new->entries[j].threshold <= usage) {
3949 * new->current_threshold will not be used
3950 * until rcu_assign_pointer(), so it's safe to increment
3953 ++new->current_threshold;
3959 /* Swap primary and spare array */
3960 thresholds->spare = thresholds->primary;
3962 rcu_assign_pointer(thresholds->primary, new);
3964 /* To be sure that nobody uses thresholds */
3967 /* If all events are unregistered, free the spare array */
3969 kfree(thresholds->spare);
3970 thresholds->spare = NULL;
3973 mutex_unlock(&memcg->thresholds_lock);
3976 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3977 struct eventfd_ctx *eventfd)
3979 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3982 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3983 struct eventfd_ctx *eventfd)
3985 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3988 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3989 struct eventfd_ctx *eventfd, const char *args)
3991 struct mem_cgroup_eventfd_list *event;
3993 event = kmalloc(sizeof(*event), GFP_KERNEL);
3997 spin_lock(&memcg_oom_lock);
3999 event->eventfd = eventfd;
4000 list_add(&event->list, &memcg->oom_notify);
4002 /* already in OOM ? */
4003 if (memcg->under_oom)
4004 eventfd_signal(eventfd, 1);
4005 spin_unlock(&memcg_oom_lock);
4010 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4011 struct eventfd_ctx *eventfd)
4013 struct mem_cgroup_eventfd_list *ev, *tmp;
4015 spin_lock(&memcg_oom_lock);
4017 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4018 if (ev->eventfd == eventfd) {
4019 list_del(&ev->list);
4024 spin_unlock(&memcg_oom_lock);
4027 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4029 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4031 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4032 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4033 seq_printf(sf, "oom_kill %lu\n",
4034 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4038 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4039 struct cftype *cft, u64 val)
4041 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4043 /* cannot set to root cgroup and only 0 and 1 are allowed */
4044 if (!css->parent || !((val == 0) || (val == 1)))
4047 memcg->oom_kill_disable = val;
4049 memcg_oom_recover(memcg);
4054 #ifdef CONFIG_CGROUP_WRITEBACK
4056 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4058 return wb_domain_init(&memcg->cgwb_domain, gfp);
4061 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4063 wb_domain_exit(&memcg->cgwb_domain);
4066 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4068 wb_domain_size_changed(&memcg->cgwb_domain);
4071 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4073 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4075 if (!memcg->css.parent)
4078 return &memcg->cgwb_domain;
4082 * idx can be of type enum memcg_stat_item or node_stat_item.
4083 * Keep in sync with memcg_exact_page().
4085 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4087 long x = atomic_long_read(&memcg->vmstats[idx]);
4090 for_each_online_cpu(cpu)
4091 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4098 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4099 * @wb: bdi_writeback in question
4100 * @pfilepages: out parameter for number of file pages
4101 * @pheadroom: out parameter for number of allocatable pages according to memcg
4102 * @pdirty: out parameter for number of dirty pages
4103 * @pwriteback: out parameter for number of pages under writeback
4105 * Determine the numbers of file, headroom, dirty, and writeback pages in
4106 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4107 * is a bit more involved.
4109 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4110 * headroom is calculated as the lowest headroom of itself and the
4111 * ancestors. Note that this doesn't consider the actual amount of
4112 * available memory in the system. The caller should further cap
4113 * *@pheadroom accordingly.
4115 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4116 unsigned long *pheadroom, unsigned long *pdirty,
4117 unsigned long *pwriteback)
4119 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4120 struct mem_cgroup *parent;
4122 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4124 /* this should eventually include NR_UNSTABLE_NFS */
4125 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4126 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4127 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4128 *pheadroom = PAGE_COUNTER_MAX;
4130 while ((parent = parent_mem_cgroup(memcg))) {
4131 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4132 unsigned long used = page_counter_read(&memcg->memory);
4134 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4139 #else /* CONFIG_CGROUP_WRITEBACK */
4141 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4146 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4150 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4154 #endif /* CONFIG_CGROUP_WRITEBACK */
4157 * DO NOT USE IN NEW FILES.
4159 * "cgroup.event_control" implementation.
4161 * This is way over-engineered. It tries to support fully configurable
4162 * events for each user. Such level of flexibility is completely
4163 * unnecessary especially in the light of the planned unified hierarchy.
4165 * Please deprecate this and replace with something simpler if at all
4170 * Unregister event and free resources.
4172 * Gets called from workqueue.
4174 static void memcg_event_remove(struct work_struct *work)
4176 struct mem_cgroup_event *event =
4177 container_of(work, struct mem_cgroup_event, remove);
4178 struct mem_cgroup *memcg = event->memcg;
4180 remove_wait_queue(event->wqh, &event->wait);
4182 event->unregister_event(memcg, event->eventfd);
4184 /* Notify userspace the event is going away. */
4185 eventfd_signal(event->eventfd, 1);
4187 eventfd_ctx_put(event->eventfd);
4189 css_put(&memcg->css);
4193 * Gets called on EPOLLHUP on eventfd when user closes it.
4195 * Called with wqh->lock held and interrupts disabled.
4197 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4198 int sync, void *key)
4200 struct mem_cgroup_event *event =
4201 container_of(wait, struct mem_cgroup_event, wait);
4202 struct mem_cgroup *memcg = event->memcg;
4203 __poll_t flags = key_to_poll(key);
4205 if (flags & EPOLLHUP) {
4207 * If the event has been detached at cgroup removal, we
4208 * can simply return knowing the other side will cleanup
4211 * We can't race against event freeing since the other
4212 * side will require wqh->lock via remove_wait_queue(),
4215 spin_lock(&memcg->event_list_lock);
4216 if (!list_empty(&event->list)) {
4217 list_del_init(&event->list);
4219 * We are in atomic context, but cgroup_event_remove()
4220 * may sleep, so we have to call it in workqueue.
4222 schedule_work(&event->remove);
4224 spin_unlock(&memcg->event_list_lock);
4230 static void memcg_event_ptable_queue_proc(struct file *file,
4231 wait_queue_head_t *wqh, poll_table *pt)
4233 struct mem_cgroup_event *event =
4234 container_of(pt, struct mem_cgroup_event, pt);
4237 add_wait_queue(wqh, &event->wait);
4241 * DO NOT USE IN NEW FILES.
4243 * Parse input and register new cgroup event handler.
4245 * Input must be in format '<event_fd> <control_fd> <args>'.
4246 * Interpretation of args is defined by control file implementation.
4248 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4249 char *buf, size_t nbytes, loff_t off)
4251 struct cgroup_subsys_state *css = of_css(of);
4252 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4253 struct mem_cgroup_event *event;
4254 struct cgroup_subsys_state *cfile_css;
4255 unsigned int efd, cfd;
4262 buf = strstrip(buf);
4264 efd = simple_strtoul(buf, &endp, 10);
4269 cfd = simple_strtoul(buf, &endp, 10);
4270 if ((*endp != ' ') && (*endp != '\0'))
4274 event = kzalloc(sizeof(*event), GFP_KERNEL);
4278 event->memcg = memcg;
4279 INIT_LIST_HEAD(&event->list);
4280 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4281 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4282 INIT_WORK(&event->remove, memcg_event_remove);
4290 event->eventfd = eventfd_ctx_fileget(efile.file);
4291 if (IS_ERR(event->eventfd)) {
4292 ret = PTR_ERR(event->eventfd);
4299 goto out_put_eventfd;
4302 /* the process need read permission on control file */
4303 /* AV: shouldn't we check that it's been opened for read instead? */
4304 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4309 * Determine the event callbacks and set them in @event. This used
4310 * to be done via struct cftype but cgroup core no longer knows
4311 * about these events. The following is crude but the whole thing
4312 * is for compatibility anyway.
4314 * DO NOT ADD NEW FILES.
4316 name = cfile.file->f_path.dentry->d_name.name;
4318 if (!strcmp(name, "memory.usage_in_bytes")) {
4319 event->register_event = mem_cgroup_usage_register_event;
4320 event->unregister_event = mem_cgroup_usage_unregister_event;
4321 } else if (!strcmp(name, "memory.oom_control")) {
4322 event->register_event = mem_cgroup_oom_register_event;
4323 event->unregister_event = mem_cgroup_oom_unregister_event;
4324 } else if (!strcmp(name, "memory.pressure_level")) {
4325 event->register_event = vmpressure_register_event;
4326 event->unregister_event = vmpressure_unregister_event;
4327 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4328 event->register_event = memsw_cgroup_usage_register_event;
4329 event->unregister_event = memsw_cgroup_usage_unregister_event;
4336 * Verify @cfile should belong to @css. Also, remaining events are
4337 * automatically removed on cgroup destruction but the removal is
4338 * asynchronous, so take an extra ref on @css.
4340 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4341 &memory_cgrp_subsys);
4343 if (IS_ERR(cfile_css))
4345 if (cfile_css != css) {
4350 ret = event->register_event(memcg, event->eventfd, buf);
4354 vfs_poll(efile.file, &event->pt);
4356 spin_lock(&memcg->event_list_lock);
4357 list_add(&event->list, &memcg->event_list);
4358 spin_unlock(&memcg->event_list_lock);
4370 eventfd_ctx_put(event->eventfd);
4379 static struct cftype mem_cgroup_legacy_files[] = {
4381 .name = "usage_in_bytes",
4382 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4383 .read_u64 = mem_cgroup_read_u64,
4386 .name = "max_usage_in_bytes",
4387 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4388 .write = mem_cgroup_reset,
4389 .read_u64 = mem_cgroup_read_u64,
4392 .name = "limit_in_bytes",
4393 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4394 .write = mem_cgroup_write,
4395 .read_u64 = mem_cgroup_read_u64,
4398 .name = "soft_limit_in_bytes",
4399 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4400 .write = mem_cgroup_write,
4401 .read_u64 = mem_cgroup_read_u64,
4405 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4406 .write = mem_cgroup_reset,
4407 .read_u64 = mem_cgroup_read_u64,
4411 .seq_show = memcg_stat_show,
4414 .name = "force_empty",
4415 .write = mem_cgroup_force_empty_write,
4418 .name = "use_hierarchy",
4419 .write_u64 = mem_cgroup_hierarchy_write,
4420 .read_u64 = mem_cgroup_hierarchy_read,
4423 .name = "cgroup.event_control", /* XXX: for compat */
4424 .write = memcg_write_event_control,
4425 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4428 .name = "swappiness",
4429 .read_u64 = mem_cgroup_swappiness_read,
4430 .write_u64 = mem_cgroup_swappiness_write,
4433 .name = "move_charge_at_immigrate",
4434 .read_u64 = mem_cgroup_move_charge_read,
4435 .write_u64 = mem_cgroup_move_charge_write,
4438 .name = "oom_control",
4439 .seq_show = mem_cgroup_oom_control_read,
4440 .write_u64 = mem_cgroup_oom_control_write,
4441 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4444 .name = "pressure_level",
4448 .name = "numa_stat",
4449 .seq_show = memcg_numa_stat_show,
4453 .name = "kmem.limit_in_bytes",
4454 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4455 .write = mem_cgroup_write,
4456 .read_u64 = mem_cgroup_read_u64,
4459 .name = "kmem.usage_in_bytes",
4460 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4461 .read_u64 = mem_cgroup_read_u64,
4464 .name = "kmem.failcnt",
4465 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4466 .write = mem_cgroup_reset,
4467 .read_u64 = mem_cgroup_read_u64,
4470 .name = "kmem.max_usage_in_bytes",
4471 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4472 .write = mem_cgroup_reset,
4473 .read_u64 = mem_cgroup_read_u64,
4475 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4477 .name = "kmem.slabinfo",
4478 .seq_start = memcg_slab_start,
4479 .seq_next = memcg_slab_next,
4480 .seq_stop = memcg_slab_stop,
4481 .seq_show = memcg_slab_show,
4485 .name = "kmem.tcp.limit_in_bytes",
4486 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4487 .write = mem_cgroup_write,
4488 .read_u64 = mem_cgroup_read_u64,
4491 .name = "kmem.tcp.usage_in_bytes",
4492 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4493 .read_u64 = mem_cgroup_read_u64,
4496 .name = "kmem.tcp.failcnt",
4497 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4498 .write = mem_cgroup_reset,
4499 .read_u64 = mem_cgroup_read_u64,
4502 .name = "kmem.tcp.max_usage_in_bytes",
4503 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4504 .write = mem_cgroup_reset,
4505 .read_u64 = mem_cgroup_read_u64,
4507 { }, /* terminate */
4511 * Private memory cgroup IDR
4513 * Swap-out records and page cache shadow entries need to store memcg
4514 * references in constrained space, so we maintain an ID space that is
4515 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4516 * memory-controlled cgroups to 64k.
4518 * However, there usually are many references to the oflline CSS after
4519 * the cgroup has been destroyed, such as page cache or reclaimable
4520 * slab objects, that don't need to hang on to the ID. We want to keep
4521 * those dead CSS from occupying IDs, or we might quickly exhaust the
4522 * relatively small ID space and prevent the creation of new cgroups
4523 * even when there are much fewer than 64k cgroups - possibly none.
4525 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4526 * be freed and recycled when it's no longer needed, which is usually
4527 * when the CSS is offlined.
4529 * The only exception to that are records of swapped out tmpfs/shmem
4530 * pages that need to be attributed to live ancestors on swapin. But
4531 * those references are manageable from userspace.
4534 static DEFINE_IDR(mem_cgroup_idr);
4536 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4538 if (memcg->id.id > 0) {
4539 idr_remove(&mem_cgroup_idr, memcg->id.id);
4544 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4546 refcount_add(n, &memcg->id.ref);
4549 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4551 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4552 mem_cgroup_id_remove(memcg);
4554 /* Memcg ID pins CSS */
4555 css_put(&memcg->css);
4559 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4561 mem_cgroup_id_get_many(memcg, 1);
4564 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4566 mem_cgroup_id_put_many(memcg, 1);
4570 * mem_cgroup_from_id - look up a memcg from a memcg id
4571 * @id: the memcg id to look up
4573 * Caller must hold rcu_read_lock().
4575 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4577 WARN_ON_ONCE(!rcu_read_lock_held());
4578 return idr_find(&mem_cgroup_idr, id);
4581 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4583 struct mem_cgroup_per_node *pn;
4586 * This routine is called against possible nodes.
4587 * But it's BUG to call kmalloc() against offline node.
4589 * TODO: this routine can waste much memory for nodes which will
4590 * never be onlined. It's better to use memory hotplug callback
4593 if (!node_state(node, N_NORMAL_MEMORY))
4595 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4599 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4600 if (!pn->lruvec_stat_local) {
4605 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4606 if (!pn->lruvec_stat_cpu) {
4607 free_percpu(pn->lruvec_stat_local);
4612 lruvec_init(&pn->lruvec);
4613 pn->usage_in_excess = 0;
4614 pn->on_tree = false;
4617 memcg->nodeinfo[node] = pn;
4621 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4623 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4628 free_percpu(pn->lruvec_stat_cpu);
4629 free_percpu(pn->lruvec_stat_local);
4633 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4638 free_mem_cgroup_per_node_info(memcg, node);
4639 free_percpu(memcg->vmstats_percpu);
4640 free_percpu(memcg->vmstats_local);
4644 static void mem_cgroup_free(struct mem_cgroup *memcg)
4646 memcg_wb_domain_exit(memcg);
4647 __mem_cgroup_free(memcg);
4650 static struct mem_cgroup *mem_cgroup_alloc(void)
4652 struct mem_cgroup *memcg;
4656 size = sizeof(struct mem_cgroup);
4657 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4659 memcg = kzalloc(size, GFP_KERNEL);
4663 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4664 1, MEM_CGROUP_ID_MAX,
4666 if (memcg->id.id < 0)
4669 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
4670 if (!memcg->vmstats_local)
4673 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4674 if (!memcg->vmstats_percpu)
4678 if (alloc_mem_cgroup_per_node_info(memcg, node))
4681 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4684 INIT_WORK(&memcg->high_work, high_work_func);
4685 memcg->last_scanned_node = MAX_NUMNODES;
4686 INIT_LIST_HEAD(&memcg->oom_notify);
4687 mutex_init(&memcg->thresholds_lock);
4688 spin_lock_init(&memcg->move_lock);
4689 vmpressure_init(&memcg->vmpressure);
4690 INIT_LIST_HEAD(&memcg->event_list);
4691 spin_lock_init(&memcg->event_list_lock);
4692 memcg->socket_pressure = jiffies;
4693 #ifdef CONFIG_MEMCG_KMEM
4694 memcg->kmemcg_id = -1;
4696 #ifdef CONFIG_CGROUP_WRITEBACK
4697 INIT_LIST_HEAD(&memcg->cgwb_list);
4699 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4702 mem_cgroup_id_remove(memcg);
4703 __mem_cgroup_free(memcg);
4707 static struct cgroup_subsys_state * __ref
4708 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4710 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4711 struct mem_cgroup *memcg;
4712 long error = -ENOMEM;
4714 memcg = mem_cgroup_alloc();
4716 return ERR_PTR(error);
4718 memcg->high = PAGE_COUNTER_MAX;
4719 memcg->soft_limit = PAGE_COUNTER_MAX;
4721 memcg->swappiness = mem_cgroup_swappiness(parent);
4722 memcg->oom_kill_disable = parent->oom_kill_disable;
4724 if (parent && parent->use_hierarchy) {
4725 memcg->use_hierarchy = true;
4726 page_counter_init(&memcg->memory, &parent->memory);
4727 page_counter_init(&memcg->swap, &parent->swap);
4728 page_counter_init(&memcg->memsw, &parent->memsw);
4729 page_counter_init(&memcg->kmem, &parent->kmem);
4730 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4732 page_counter_init(&memcg->memory, NULL);
4733 page_counter_init(&memcg->swap, NULL);
4734 page_counter_init(&memcg->memsw, NULL);
4735 page_counter_init(&memcg->kmem, NULL);
4736 page_counter_init(&memcg->tcpmem, NULL);
4738 * Deeper hierachy with use_hierarchy == false doesn't make
4739 * much sense so let cgroup subsystem know about this
4740 * unfortunate state in our controller.
4742 if (parent != root_mem_cgroup)
4743 memory_cgrp_subsys.broken_hierarchy = true;
4746 /* The following stuff does not apply to the root */
4748 root_mem_cgroup = memcg;
4752 error = memcg_online_kmem(memcg);
4756 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4757 static_branch_inc(&memcg_sockets_enabled_key);
4761 mem_cgroup_id_remove(memcg);
4762 mem_cgroup_free(memcg);
4763 return ERR_PTR(-ENOMEM);
4766 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4768 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4771 * A memcg must be visible for memcg_expand_shrinker_maps()
4772 * by the time the maps are allocated. So, we allocate maps
4773 * here, when for_each_mem_cgroup() can't skip it.
4775 if (memcg_alloc_shrinker_maps(memcg)) {
4776 mem_cgroup_id_remove(memcg);
4780 /* Online state pins memcg ID, memcg ID pins CSS */
4781 refcount_set(&memcg->id.ref, 1);
4786 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4788 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4789 struct mem_cgroup_event *event, *tmp;
4792 * Unregister events and notify userspace.
4793 * Notify userspace about cgroup removing only after rmdir of cgroup
4794 * directory to avoid race between userspace and kernelspace.
4796 spin_lock(&memcg->event_list_lock);
4797 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4798 list_del_init(&event->list);
4799 schedule_work(&event->remove);
4801 spin_unlock(&memcg->event_list_lock);
4803 page_counter_set_min(&memcg->memory, 0);
4804 page_counter_set_low(&memcg->memory, 0);
4806 memcg_offline_kmem(memcg);
4807 wb_memcg_offline(memcg);
4809 drain_all_stock(memcg);
4811 mem_cgroup_id_put(memcg);
4814 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4816 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4818 invalidate_reclaim_iterators(memcg);
4821 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4823 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4825 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4826 static_branch_dec(&memcg_sockets_enabled_key);
4828 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4829 static_branch_dec(&memcg_sockets_enabled_key);
4831 vmpressure_cleanup(&memcg->vmpressure);
4832 cancel_work_sync(&memcg->high_work);
4833 mem_cgroup_remove_from_trees(memcg);
4834 memcg_free_shrinker_maps(memcg);
4835 memcg_free_kmem(memcg);
4836 mem_cgroup_free(memcg);
4840 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4841 * @css: the target css
4843 * Reset the states of the mem_cgroup associated with @css. This is
4844 * invoked when the userland requests disabling on the default hierarchy
4845 * but the memcg is pinned through dependency. The memcg should stop
4846 * applying policies and should revert to the vanilla state as it may be
4847 * made visible again.
4849 * The current implementation only resets the essential configurations.
4850 * This needs to be expanded to cover all the visible parts.
4852 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4854 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4856 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4857 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4858 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4859 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4860 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4861 page_counter_set_min(&memcg->memory, 0);
4862 page_counter_set_low(&memcg->memory, 0);
4863 memcg->high = PAGE_COUNTER_MAX;
4864 memcg->soft_limit = PAGE_COUNTER_MAX;
4865 memcg_wb_domain_size_changed(memcg);
4869 /* Handlers for move charge at task migration. */
4870 static int mem_cgroup_do_precharge(unsigned long count)
4874 /* Try a single bulk charge without reclaim first, kswapd may wake */
4875 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4877 mc.precharge += count;
4881 /* Try charges one by one with reclaim, but do not retry */
4883 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4897 enum mc_target_type {
4904 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4905 unsigned long addr, pte_t ptent)
4907 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4909 if (!page || !page_mapped(page))
4911 if (PageAnon(page)) {
4912 if (!(mc.flags & MOVE_ANON))
4915 if (!(mc.flags & MOVE_FILE))
4918 if (!get_page_unless_zero(page))
4924 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4925 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4926 pte_t ptent, swp_entry_t *entry)
4928 struct page *page = NULL;
4929 swp_entry_t ent = pte_to_swp_entry(ptent);
4931 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4935 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4936 * a device and because they are not accessible by CPU they are store
4937 * as special swap entry in the CPU page table.
4939 if (is_device_private_entry(ent)) {
4940 page = device_private_entry_to_page(ent);
4942 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4943 * a refcount of 1 when free (unlike normal page)
4945 if (!page_ref_add_unless(page, 1, 1))
4951 * Because lookup_swap_cache() updates some statistics counter,
4952 * we call find_get_page() with swapper_space directly.
4954 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4955 if (do_memsw_account())
4956 entry->val = ent.val;
4961 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4962 pte_t ptent, swp_entry_t *entry)
4968 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4969 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4971 struct page *page = NULL;
4972 struct address_space *mapping;
4975 if (!vma->vm_file) /* anonymous vma */
4977 if (!(mc.flags & MOVE_FILE))
4980 mapping = vma->vm_file->f_mapping;
4981 pgoff = linear_page_index(vma, addr);
4983 /* page is moved even if it's not RSS of this task(page-faulted). */
4985 /* shmem/tmpfs may report page out on swap: account for that too. */
4986 if (shmem_mapping(mapping)) {
4987 page = find_get_entry(mapping, pgoff);
4988 if (xa_is_value(page)) {
4989 swp_entry_t swp = radix_to_swp_entry(page);
4990 if (do_memsw_account())
4992 page = find_get_page(swap_address_space(swp),
4996 page = find_get_page(mapping, pgoff);
4998 page = find_get_page(mapping, pgoff);
5004 * mem_cgroup_move_account - move account of the page
5006 * @compound: charge the page as compound or small page
5007 * @from: mem_cgroup which the page is moved from.
5008 * @to: mem_cgroup which the page is moved to. @from != @to.
5010 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5012 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5015 static int mem_cgroup_move_account(struct page *page,
5017 struct mem_cgroup *from,
5018 struct mem_cgroup *to)
5020 unsigned long flags;
5021 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5025 VM_BUG_ON(from == to);
5026 VM_BUG_ON_PAGE(PageLRU(page), page);
5027 VM_BUG_ON(compound && !PageTransHuge(page));
5030 * Prevent mem_cgroup_migrate() from looking at
5031 * page->mem_cgroup of its source page while we change it.
5034 if (!trylock_page(page))
5038 if (page->mem_cgroup != from)
5041 anon = PageAnon(page);
5043 spin_lock_irqsave(&from->move_lock, flags);
5045 if (!anon && page_mapped(page)) {
5046 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
5047 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
5051 * move_lock grabbed above and caller set from->moving_account, so
5052 * mod_memcg_page_state will serialize updates to PageDirty.
5053 * So mapping should be stable for dirty pages.
5055 if (!anon && PageDirty(page)) {
5056 struct address_space *mapping = page_mapping(page);
5058 if (mapping_cap_account_dirty(mapping)) {
5059 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
5060 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
5064 if (PageWriteback(page)) {
5065 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
5066 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
5070 * It is safe to change page->mem_cgroup here because the page
5071 * is referenced, charged, and isolated - we can't race with
5072 * uncharging, charging, migration, or LRU putback.
5075 /* caller should have done css_get */
5076 page->mem_cgroup = to;
5077 spin_unlock_irqrestore(&from->move_lock, flags);
5081 local_irq_disable();
5082 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5083 memcg_check_events(to, page);
5084 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5085 memcg_check_events(from, page);
5094 * get_mctgt_type - get target type of moving charge
5095 * @vma: the vma the pte to be checked belongs
5096 * @addr: the address corresponding to the pte to be checked
5097 * @ptent: the pte to be checked
5098 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5101 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5102 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5103 * move charge. if @target is not NULL, the page is stored in target->page
5104 * with extra refcnt got(Callers should handle it).
5105 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5106 * target for charge migration. if @target is not NULL, the entry is stored
5108 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
5109 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
5110 * For now we such page is charge like a regular page would be as for all
5111 * intent and purposes it is just special memory taking the place of a
5114 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5116 * Called with pte lock held.
5119 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5120 unsigned long addr, pte_t ptent, union mc_target *target)
5122 struct page *page = NULL;
5123 enum mc_target_type ret = MC_TARGET_NONE;
5124 swp_entry_t ent = { .val = 0 };
5126 if (pte_present(ptent))
5127 page = mc_handle_present_pte(vma, addr, ptent);
5128 else if (is_swap_pte(ptent))
5129 page = mc_handle_swap_pte(vma, ptent, &ent);
5130 else if (pte_none(ptent))
5131 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5133 if (!page && !ent.val)
5137 * Do only loose check w/o serialization.
5138 * mem_cgroup_move_account() checks the page is valid or
5139 * not under LRU exclusion.
5141 if (page->mem_cgroup == mc.from) {
5142 ret = MC_TARGET_PAGE;
5143 if (is_device_private_page(page) ||
5144 is_device_public_page(page))
5145 ret = MC_TARGET_DEVICE;
5147 target->page = page;
5149 if (!ret || !target)
5153 * There is a swap entry and a page doesn't exist or isn't charged.
5154 * But we cannot move a tail-page in a THP.
5156 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5157 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5158 ret = MC_TARGET_SWAP;
5165 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5167 * We don't consider PMD mapped swapping or file mapped pages because THP does
5168 * not support them for now.
5169 * Caller should make sure that pmd_trans_huge(pmd) is true.
5171 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5172 unsigned long addr, pmd_t pmd, union mc_target *target)
5174 struct page *page = NULL;
5175 enum mc_target_type ret = MC_TARGET_NONE;
5177 if (unlikely(is_swap_pmd(pmd))) {
5178 VM_BUG_ON(thp_migration_supported() &&
5179 !is_pmd_migration_entry(pmd));
5182 page = pmd_page(pmd);
5183 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5184 if (!(mc.flags & MOVE_ANON))
5186 if (page->mem_cgroup == mc.from) {
5187 ret = MC_TARGET_PAGE;
5190 target->page = page;
5196 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5197 unsigned long addr, pmd_t pmd, union mc_target *target)
5199 return MC_TARGET_NONE;
5203 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5204 unsigned long addr, unsigned long end,
5205 struct mm_walk *walk)
5207 struct vm_area_struct *vma = walk->vma;
5211 ptl = pmd_trans_huge_lock(pmd, vma);
5214 * Note their can not be MC_TARGET_DEVICE for now as we do not
5215 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5216 * MEMORY_DEVICE_PRIVATE but this might change.
5218 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5219 mc.precharge += HPAGE_PMD_NR;
5224 if (pmd_trans_unstable(pmd))
5226 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5227 for (; addr != end; pte++, addr += PAGE_SIZE)
5228 if (get_mctgt_type(vma, addr, *pte, NULL))
5229 mc.precharge++; /* increment precharge temporarily */
5230 pte_unmap_unlock(pte - 1, ptl);
5236 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5238 unsigned long precharge;
5240 struct mm_walk mem_cgroup_count_precharge_walk = {
5241 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5244 down_read(&mm->mmap_sem);
5245 walk_page_range(0, mm->highest_vm_end,
5246 &mem_cgroup_count_precharge_walk);
5247 up_read(&mm->mmap_sem);
5249 precharge = mc.precharge;
5255 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5257 unsigned long precharge = mem_cgroup_count_precharge(mm);
5259 VM_BUG_ON(mc.moving_task);
5260 mc.moving_task = current;
5261 return mem_cgroup_do_precharge(precharge);
5264 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5265 static void __mem_cgroup_clear_mc(void)
5267 struct mem_cgroup *from = mc.from;
5268 struct mem_cgroup *to = mc.to;
5270 /* we must uncharge all the leftover precharges from mc.to */
5272 cancel_charge(mc.to, mc.precharge);
5276 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5277 * we must uncharge here.
5279 if (mc.moved_charge) {
5280 cancel_charge(mc.from, mc.moved_charge);
5281 mc.moved_charge = 0;
5283 /* we must fixup refcnts and charges */
5284 if (mc.moved_swap) {
5285 /* uncharge swap account from the old cgroup */
5286 if (!mem_cgroup_is_root(mc.from))
5287 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5289 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5292 * we charged both to->memory and to->memsw, so we
5293 * should uncharge to->memory.
5295 if (!mem_cgroup_is_root(mc.to))
5296 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5298 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5299 css_put_many(&mc.to->css, mc.moved_swap);
5303 memcg_oom_recover(from);
5304 memcg_oom_recover(to);
5305 wake_up_all(&mc.waitq);
5308 static void mem_cgroup_clear_mc(void)
5310 struct mm_struct *mm = mc.mm;
5313 * we must clear moving_task before waking up waiters at the end of
5316 mc.moving_task = NULL;
5317 __mem_cgroup_clear_mc();
5318 spin_lock(&mc.lock);
5322 spin_unlock(&mc.lock);
5327 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5329 struct cgroup_subsys_state *css;
5330 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5331 struct mem_cgroup *from;
5332 struct task_struct *leader, *p;
5333 struct mm_struct *mm;
5334 unsigned long move_flags;
5337 /* charge immigration isn't supported on the default hierarchy */
5338 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5342 * Multi-process migrations only happen on the default hierarchy
5343 * where charge immigration is not used. Perform charge
5344 * immigration if @tset contains a leader and whine if there are
5348 cgroup_taskset_for_each_leader(leader, css, tset) {
5351 memcg = mem_cgroup_from_css(css);
5357 * We are now commited to this value whatever it is. Changes in this
5358 * tunable will only affect upcoming migrations, not the current one.
5359 * So we need to save it, and keep it going.
5361 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5365 from = mem_cgroup_from_task(p);
5367 VM_BUG_ON(from == memcg);
5369 mm = get_task_mm(p);
5372 /* We move charges only when we move a owner of the mm */
5373 if (mm->owner == p) {
5376 VM_BUG_ON(mc.precharge);
5377 VM_BUG_ON(mc.moved_charge);
5378 VM_BUG_ON(mc.moved_swap);
5380 spin_lock(&mc.lock);
5384 mc.flags = move_flags;
5385 spin_unlock(&mc.lock);
5386 /* We set mc.moving_task later */
5388 ret = mem_cgroup_precharge_mc(mm);
5390 mem_cgroup_clear_mc();
5397 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5400 mem_cgroup_clear_mc();
5403 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5404 unsigned long addr, unsigned long end,
5405 struct mm_walk *walk)
5408 struct vm_area_struct *vma = walk->vma;
5411 enum mc_target_type target_type;
5412 union mc_target target;
5415 ptl = pmd_trans_huge_lock(pmd, vma);
5417 if (mc.precharge < HPAGE_PMD_NR) {
5421 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5422 if (target_type == MC_TARGET_PAGE) {
5424 if (!isolate_lru_page(page)) {
5425 if (!mem_cgroup_move_account(page, true,
5427 mc.precharge -= HPAGE_PMD_NR;
5428 mc.moved_charge += HPAGE_PMD_NR;
5430 putback_lru_page(page);
5433 } else if (target_type == MC_TARGET_DEVICE) {
5435 if (!mem_cgroup_move_account(page, true,
5437 mc.precharge -= HPAGE_PMD_NR;
5438 mc.moved_charge += HPAGE_PMD_NR;
5446 if (pmd_trans_unstable(pmd))
5449 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5450 for (; addr != end; addr += PAGE_SIZE) {
5451 pte_t ptent = *(pte++);
5452 bool device = false;
5458 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5459 case MC_TARGET_DEVICE:
5462 case MC_TARGET_PAGE:
5465 * We can have a part of the split pmd here. Moving it
5466 * can be done but it would be too convoluted so simply
5467 * ignore such a partial THP and keep it in original
5468 * memcg. There should be somebody mapping the head.
5470 if (PageTransCompound(page))
5472 if (!device && isolate_lru_page(page))
5474 if (!mem_cgroup_move_account(page, false,
5477 /* we uncharge from mc.from later. */
5481 putback_lru_page(page);
5482 put: /* get_mctgt_type() gets the page */
5485 case MC_TARGET_SWAP:
5487 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5489 /* we fixup refcnts and charges later. */
5497 pte_unmap_unlock(pte - 1, ptl);
5502 * We have consumed all precharges we got in can_attach().
5503 * We try charge one by one, but don't do any additional
5504 * charges to mc.to if we have failed in charge once in attach()
5507 ret = mem_cgroup_do_precharge(1);
5515 static void mem_cgroup_move_charge(void)
5517 struct mm_walk mem_cgroup_move_charge_walk = {
5518 .pmd_entry = mem_cgroup_move_charge_pte_range,
5522 lru_add_drain_all();
5524 * Signal lock_page_memcg() to take the memcg's move_lock
5525 * while we're moving its pages to another memcg. Then wait
5526 * for already started RCU-only updates to finish.
5528 atomic_inc(&mc.from->moving_account);
5531 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5533 * Someone who are holding the mmap_sem might be waiting in
5534 * waitq. So we cancel all extra charges, wake up all waiters,
5535 * and retry. Because we cancel precharges, we might not be able
5536 * to move enough charges, but moving charge is a best-effort
5537 * feature anyway, so it wouldn't be a big problem.
5539 __mem_cgroup_clear_mc();
5544 * When we have consumed all precharges and failed in doing
5545 * additional charge, the page walk just aborts.
5547 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5549 up_read(&mc.mm->mmap_sem);
5550 atomic_dec(&mc.from->moving_account);
5553 static void mem_cgroup_move_task(void)
5556 mem_cgroup_move_charge();
5557 mem_cgroup_clear_mc();
5560 #else /* !CONFIG_MMU */
5561 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5565 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5568 static void mem_cgroup_move_task(void)
5574 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5575 * to verify whether we're attached to the default hierarchy on each mount
5578 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5581 * use_hierarchy is forced on the default hierarchy. cgroup core
5582 * guarantees that @root doesn't have any children, so turning it
5583 * on for the root memcg is enough.
5585 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5586 root_mem_cgroup->use_hierarchy = true;
5588 root_mem_cgroup->use_hierarchy = false;
5591 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5593 if (value == PAGE_COUNTER_MAX)
5594 seq_puts(m, "max\n");
5596 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5601 static u64 memory_current_read(struct cgroup_subsys_state *css,
5604 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5606 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5609 static int memory_min_show(struct seq_file *m, void *v)
5611 return seq_puts_memcg_tunable(m,
5612 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5615 static ssize_t memory_min_write(struct kernfs_open_file *of,
5616 char *buf, size_t nbytes, loff_t off)
5618 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5622 buf = strstrip(buf);
5623 err = page_counter_memparse(buf, "max", &min);
5627 page_counter_set_min(&memcg->memory, min);
5632 static int memory_low_show(struct seq_file *m, void *v)
5634 return seq_puts_memcg_tunable(m,
5635 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5638 static ssize_t memory_low_write(struct kernfs_open_file *of,
5639 char *buf, size_t nbytes, loff_t off)
5641 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5645 buf = strstrip(buf);
5646 err = page_counter_memparse(buf, "max", &low);
5650 page_counter_set_low(&memcg->memory, low);
5655 static int memory_high_show(struct seq_file *m, void *v)
5657 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5660 static ssize_t memory_high_write(struct kernfs_open_file *of,
5661 char *buf, size_t nbytes, loff_t off)
5663 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5664 unsigned long nr_pages;
5668 buf = strstrip(buf);
5669 err = page_counter_memparse(buf, "max", &high);
5675 nr_pages = page_counter_read(&memcg->memory);
5676 if (nr_pages > high)
5677 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5680 memcg_wb_domain_size_changed(memcg);
5684 static int memory_max_show(struct seq_file *m, void *v)
5686 return seq_puts_memcg_tunable(m,
5687 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5690 static ssize_t memory_max_write(struct kernfs_open_file *of,
5691 char *buf, size_t nbytes, loff_t off)
5693 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5694 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5695 bool drained = false;
5699 buf = strstrip(buf);
5700 err = page_counter_memparse(buf, "max", &max);
5704 xchg(&memcg->memory.max, max);
5707 unsigned long nr_pages = page_counter_read(&memcg->memory);
5709 if (nr_pages <= max)
5712 if (signal_pending(current)) {
5718 drain_all_stock(memcg);
5724 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5730 memcg_memory_event(memcg, MEMCG_OOM);
5731 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5735 memcg_wb_domain_size_changed(memcg);
5739 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
5741 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
5742 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
5743 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
5744 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
5745 seq_printf(m, "oom_kill %lu\n",
5746 atomic_long_read(&events[MEMCG_OOM_KILL]));
5749 static int memory_events_show(struct seq_file *m, void *v)
5751 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5753 __memory_events_show(m, memcg->memory_events);
5757 static int memory_events_local_show(struct seq_file *m, void *v)
5759 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5761 __memory_events_show(m, memcg->memory_events_local);
5765 static int memory_stat_show(struct seq_file *m, void *v)
5767 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5770 buf = memory_stat_format(memcg);
5778 static int memory_oom_group_show(struct seq_file *m, void *v)
5780 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5782 seq_printf(m, "%d\n", memcg->oom_group);
5787 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5788 char *buf, size_t nbytes, loff_t off)
5790 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5793 buf = strstrip(buf);
5797 ret = kstrtoint(buf, 0, &oom_group);
5801 if (oom_group != 0 && oom_group != 1)
5804 memcg->oom_group = oom_group;
5809 static struct cftype memory_files[] = {
5812 .flags = CFTYPE_NOT_ON_ROOT,
5813 .read_u64 = memory_current_read,
5817 .flags = CFTYPE_NOT_ON_ROOT,
5818 .seq_show = memory_min_show,
5819 .write = memory_min_write,
5823 .flags = CFTYPE_NOT_ON_ROOT,
5824 .seq_show = memory_low_show,
5825 .write = memory_low_write,
5829 .flags = CFTYPE_NOT_ON_ROOT,
5830 .seq_show = memory_high_show,
5831 .write = memory_high_write,
5835 .flags = CFTYPE_NOT_ON_ROOT,
5836 .seq_show = memory_max_show,
5837 .write = memory_max_write,
5841 .flags = CFTYPE_NOT_ON_ROOT,
5842 .file_offset = offsetof(struct mem_cgroup, events_file),
5843 .seq_show = memory_events_show,
5846 .name = "events.local",
5847 .flags = CFTYPE_NOT_ON_ROOT,
5848 .file_offset = offsetof(struct mem_cgroup, events_local_file),
5849 .seq_show = memory_events_local_show,
5853 .flags = CFTYPE_NOT_ON_ROOT,
5854 .seq_show = memory_stat_show,
5857 .name = "oom.group",
5858 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5859 .seq_show = memory_oom_group_show,
5860 .write = memory_oom_group_write,
5865 struct cgroup_subsys memory_cgrp_subsys = {
5866 .css_alloc = mem_cgroup_css_alloc,
5867 .css_online = mem_cgroup_css_online,
5868 .css_offline = mem_cgroup_css_offline,
5869 .css_released = mem_cgroup_css_released,
5870 .css_free = mem_cgroup_css_free,
5871 .css_reset = mem_cgroup_css_reset,
5872 .can_attach = mem_cgroup_can_attach,
5873 .cancel_attach = mem_cgroup_cancel_attach,
5874 .post_attach = mem_cgroup_move_task,
5875 .bind = mem_cgroup_bind,
5876 .dfl_cftypes = memory_files,
5877 .legacy_cftypes = mem_cgroup_legacy_files,
5882 * mem_cgroup_protected - check if memory consumption is in the normal range
5883 * @root: the top ancestor of the sub-tree being checked
5884 * @memcg: the memory cgroup to check
5886 * WARNING: This function is not stateless! It can only be used as part
5887 * of a top-down tree iteration, not for isolated queries.
5889 * Returns one of the following:
5890 * MEMCG_PROT_NONE: cgroup memory is not protected
5891 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5892 * an unprotected supply of reclaimable memory from other cgroups.
5893 * MEMCG_PROT_MIN: cgroup memory is protected
5895 * @root is exclusive; it is never protected when looked at directly
5897 * To provide a proper hierarchical behavior, effective memory.min/low values
5898 * are used. Below is the description of how effective memory.low is calculated.
5899 * Effective memory.min values is calculated in the same way.
5901 * Effective memory.low is always equal or less than the original memory.low.
5902 * If there is no memory.low overcommittment (which is always true for
5903 * top-level memory cgroups), these two values are equal.
5904 * Otherwise, it's a part of parent's effective memory.low,
5905 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5906 * memory.low usages, where memory.low usage is the size of actually
5910 * elow = min( memory.low, parent->elow * ------------------ ),
5911 * siblings_low_usage
5913 * | memory.current, if memory.current < memory.low
5918 * Such definition of the effective memory.low provides the expected
5919 * hierarchical behavior: parent's memory.low value is limiting
5920 * children, unprotected memory is reclaimed first and cgroups,
5921 * which are not using their guarantee do not affect actual memory
5924 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5926 * A A/memory.low = 2G, A/memory.current = 6G
5928 * BC DE B/memory.low = 3G B/memory.current = 2G
5929 * C/memory.low = 1G C/memory.current = 2G
5930 * D/memory.low = 0 D/memory.current = 2G
5931 * E/memory.low = 10G E/memory.current = 0
5933 * and the memory pressure is applied, the following memory distribution
5934 * is expected (approximately):
5936 * A/memory.current = 2G
5938 * B/memory.current = 1.3G
5939 * C/memory.current = 0.6G
5940 * D/memory.current = 0
5941 * E/memory.current = 0
5943 * These calculations require constant tracking of the actual low usages
5944 * (see propagate_protected_usage()), as well as recursive calculation of
5945 * effective memory.low values. But as we do call mem_cgroup_protected()
5946 * path for each memory cgroup top-down from the reclaim,
5947 * it's possible to optimize this part, and save calculated elow
5948 * for next usage. This part is intentionally racy, but it's ok,
5949 * as memory.low is a best-effort mechanism.
5951 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5952 struct mem_cgroup *memcg)
5954 struct mem_cgroup *parent;
5955 unsigned long emin, parent_emin;
5956 unsigned long elow, parent_elow;
5957 unsigned long usage;
5959 if (mem_cgroup_disabled())
5960 return MEMCG_PROT_NONE;
5963 root = root_mem_cgroup;
5965 return MEMCG_PROT_NONE;
5967 usage = page_counter_read(&memcg->memory);
5969 return MEMCG_PROT_NONE;
5971 emin = memcg->memory.min;
5972 elow = memcg->memory.low;
5974 parent = parent_mem_cgroup(memcg);
5975 /* No parent means a non-hierarchical mode on v1 memcg */
5977 return MEMCG_PROT_NONE;
5982 parent_emin = READ_ONCE(parent->memory.emin);
5983 emin = min(emin, parent_emin);
5984 if (emin && parent_emin) {
5985 unsigned long min_usage, siblings_min_usage;
5987 min_usage = min(usage, memcg->memory.min);
5988 siblings_min_usage = atomic_long_read(
5989 &parent->memory.children_min_usage);
5991 if (min_usage && siblings_min_usage)
5992 emin = min(emin, parent_emin * min_usage /
5993 siblings_min_usage);
5996 parent_elow = READ_ONCE(parent->memory.elow);
5997 elow = min(elow, parent_elow);
5998 if (elow && parent_elow) {
5999 unsigned long low_usage, siblings_low_usage;
6001 low_usage = min(usage, memcg->memory.low);
6002 siblings_low_usage = atomic_long_read(
6003 &parent->memory.children_low_usage);
6005 if (low_usage && siblings_low_usage)
6006 elow = min(elow, parent_elow * low_usage /
6007 siblings_low_usage);
6011 memcg->memory.emin = emin;
6012 memcg->memory.elow = elow;
6015 return MEMCG_PROT_MIN;
6016 else if (usage <= elow)
6017 return MEMCG_PROT_LOW;
6019 return MEMCG_PROT_NONE;
6023 * mem_cgroup_try_charge - try charging a page
6024 * @page: page to charge
6025 * @mm: mm context of the victim
6026 * @gfp_mask: reclaim mode
6027 * @memcgp: charged memcg return
6028 * @compound: charge the page as compound or small page
6030 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6031 * pages according to @gfp_mask if necessary.
6033 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6034 * Otherwise, an error code is returned.
6036 * After page->mapping has been set up, the caller must finalize the
6037 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6038 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6040 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6041 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6044 struct mem_cgroup *memcg = NULL;
6045 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6048 if (mem_cgroup_disabled())
6051 if (PageSwapCache(page)) {
6053 * Every swap fault against a single page tries to charge the
6054 * page, bail as early as possible. shmem_unuse() encounters
6055 * already charged pages, too. The USED bit is protected by
6056 * the page lock, which serializes swap cache removal, which
6057 * in turn serializes uncharging.
6059 VM_BUG_ON_PAGE(!PageLocked(page), page);
6060 if (compound_head(page)->mem_cgroup)
6063 if (do_swap_account) {
6064 swp_entry_t ent = { .val = page_private(page), };
6065 unsigned short id = lookup_swap_cgroup_id(ent);
6068 memcg = mem_cgroup_from_id(id);
6069 if (memcg && !css_tryget_online(&memcg->css))
6076 memcg = get_mem_cgroup_from_mm(mm);
6078 ret = try_charge(memcg, gfp_mask, nr_pages);
6080 css_put(&memcg->css);
6086 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6087 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6090 struct mem_cgroup *memcg;
6093 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6095 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6100 * mem_cgroup_commit_charge - commit a page charge
6101 * @page: page to charge
6102 * @memcg: memcg to charge the page to
6103 * @lrucare: page might be on LRU already
6104 * @compound: charge the page as compound or small page
6106 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6107 * after page->mapping has been set up. This must happen atomically
6108 * as part of the page instantiation, i.e. under the page table lock
6109 * for anonymous pages, under the page lock for page and swap cache.
6111 * In addition, the page must not be on the LRU during the commit, to
6112 * prevent racing with task migration. If it might be, use @lrucare.
6114 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6116 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6117 bool lrucare, bool compound)
6119 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6121 VM_BUG_ON_PAGE(!page->mapping, page);
6122 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6124 if (mem_cgroup_disabled())
6127 * Swap faults will attempt to charge the same page multiple
6128 * times. But reuse_swap_page() might have removed the page
6129 * from swapcache already, so we can't check PageSwapCache().
6134 commit_charge(page, memcg, lrucare);
6136 local_irq_disable();
6137 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6138 memcg_check_events(memcg, page);
6141 if (do_memsw_account() && PageSwapCache(page)) {
6142 swp_entry_t entry = { .val = page_private(page) };
6144 * The swap entry might not get freed for a long time,
6145 * let's not wait for it. The page already received a
6146 * memory+swap charge, drop the swap entry duplicate.
6148 mem_cgroup_uncharge_swap(entry, nr_pages);
6153 * mem_cgroup_cancel_charge - cancel a page charge
6154 * @page: page to charge
6155 * @memcg: memcg to charge the page to
6156 * @compound: charge the page as compound or small page
6158 * Cancel a charge transaction started by mem_cgroup_try_charge().
6160 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6163 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6165 if (mem_cgroup_disabled())
6168 * Swap faults will attempt to charge the same page multiple
6169 * times. But reuse_swap_page() might have removed the page
6170 * from swapcache already, so we can't check PageSwapCache().
6175 cancel_charge(memcg, nr_pages);
6178 struct uncharge_gather {
6179 struct mem_cgroup *memcg;
6180 unsigned long pgpgout;
6181 unsigned long nr_anon;
6182 unsigned long nr_file;
6183 unsigned long nr_kmem;
6184 unsigned long nr_huge;
6185 unsigned long nr_shmem;
6186 struct page *dummy_page;
6189 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6191 memset(ug, 0, sizeof(*ug));
6194 static void uncharge_batch(const struct uncharge_gather *ug)
6196 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6197 unsigned long flags;
6199 if (!mem_cgroup_is_root(ug->memcg)) {
6200 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6201 if (do_memsw_account())
6202 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6203 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6204 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6205 memcg_oom_recover(ug->memcg);
6208 local_irq_save(flags);
6209 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6210 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6211 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6212 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6213 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6214 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6215 memcg_check_events(ug->memcg, ug->dummy_page);
6216 local_irq_restore(flags);
6218 if (!mem_cgroup_is_root(ug->memcg))
6219 css_put_many(&ug->memcg->css, nr_pages);
6222 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6224 VM_BUG_ON_PAGE(PageLRU(page), page);
6225 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6226 !PageHWPoison(page) , page);
6228 if (!page->mem_cgroup)
6232 * Nobody should be changing or seriously looking at
6233 * page->mem_cgroup at this point, we have fully
6234 * exclusive access to the page.
6237 if (ug->memcg != page->mem_cgroup) {
6240 uncharge_gather_clear(ug);
6242 ug->memcg = page->mem_cgroup;
6245 if (!PageKmemcg(page)) {
6246 unsigned int nr_pages = 1;
6248 if (PageTransHuge(page)) {
6249 nr_pages <<= compound_order(page);
6250 ug->nr_huge += nr_pages;
6253 ug->nr_anon += nr_pages;
6255 ug->nr_file += nr_pages;
6256 if (PageSwapBacked(page))
6257 ug->nr_shmem += nr_pages;
6261 ug->nr_kmem += 1 << compound_order(page);
6262 __ClearPageKmemcg(page);
6265 ug->dummy_page = page;
6266 page->mem_cgroup = NULL;
6269 static void uncharge_list(struct list_head *page_list)
6271 struct uncharge_gather ug;
6272 struct list_head *next;
6274 uncharge_gather_clear(&ug);
6277 * Note that the list can be a single page->lru; hence the
6278 * do-while loop instead of a simple list_for_each_entry().
6280 next = page_list->next;
6284 page = list_entry(next, struct page, lru);
6285 next = page->lru.next;
6287 uncharge_page(page, &ug);
6288 } while (next != page_list);
6291 uncharge_batch(&ug);
6295 * mem_cgroup_uncharge - uncharge a page
6296 * @page: page to uncharge
6298 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6299 * mem_cgroup_commit_charge().
6301 void mem_cgroup_uncharge(struct page *page)
6303 struct uncharge_gather ug;
6305 if (mem_cgroup_disabled())
6308 /* Don't touch page->lru of any random page, pre-check: */
6309 if (!page->mem_cgroup)
6312 uncharge_gather_clear(&ug);
6313 uncharge_page(page, &ug);
6314 uncharge_batch(&ug);
6318 * mem_cgroup_uncharge_list - uncharge a list of page
6319 * @page_list: list of pages to uncharge
6321 * Uncharge a list of pages previously charged with
6322 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6324 void mem_cgroup_uncharge_list(struct list_head *page_list)
6326 if (mem_cgroup_disabled())
6329 if (!list_empty(page_list))
6330 uncharge_list(page_list);
6334 * mem_cgroup_migrate - charge a page's replacement
6335 * @oldpage: currently circulating page
6336 * @newpage: replacement page
6338 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6339 * be uncharged upon free.
6341 * Both pages must be locked, @newpage->mapping must be set up.
6343 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6345 struct mem_cgroup *memcg;
6346 unsigned int nr_pages;
6348 unsigned long flags;
6350 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6351 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6352 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6353 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6356 if (mem_cgroup_disabled())
6359 /* Page cache replacement: new page already charged? */
6360 if (newpage->mem_cgroup)
6363 /* Swapcache readahead pages can get replaced before being charged */
6364 memcg = oldpage->mem_cgroup;
6368 /* Force-charge the new page. The old one will be freed soon */
6369 compound = PageTransHuge(newpage);
6370 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6372 page_counter_charge(&memcg->memory, nr_pages);
6373 if (do_memsw_account())
6374 page_counter_charge(&memcg->memsw, nr_pages);
6375 css_get_many(&memcg->css, nr_pages);
6377 commit_charge(newpage, memcg, false);
6379 local_irq_save(flags);
6380 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6381 memcg_check_events(memcg, newpage);
6382 local_irq_restore(flags);
6385 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6386 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6388 void mem_cgroup_sk_alloc(struct sock *sk)
6390 struct mem_cgroup *memcg;
6392 if (!mem_cgroup_sockets_enabled)
6396 * Socket cloning can throw us here with sk_memcg already
6397 * filled. It won't however, necessarily happen from
6398 * process context. So the test for root memcg given
6399 * the current task's memcg won't help us in this case.
6401 * Respecting the original socket's memcg is a better
6402 * decision in this case.
6405 css_get(&sk->sk_memcg->css);
6410 memcg = mem_cgroup_from_task(current);
6411 if (memcg == root_mem_cgroup)
6413 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6415 if (css_tryget_online(&memcg->css))
6416 sk->sk_memcg = memcg;
6421 void mem_cgroup_sk_free(struct sock *sk)
6424 css_put(&sk->sk_memcg->css);
6428 * mem_cgroup_charge_skmem - charge socket memory
6429 * @memcg: memcg to charge
6430 * @nr_pages: number of pages to charge
6432 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6433 * @memcg's configured limit, %false if the charge had to be forced.
6435 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6437 gfp_t gfp_mask = GFP_KERNEL;
6439 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6440 struct page_counter *fail;
6442 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6443 memcg->tcpmem_pressure = 0;
6446 page_counter_charge(&memcg->tcpmem, nr_pages);
6447 memcg->tcpmem_pressure = 1;
6451 /* Don't block in the packet receive path */
6453 gfp_mask = GFP_NOWAIT;
6455 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6457 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6460 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6465 * mem_cgroup_uncharge_skmem - uncharge socket memory
6466 * @memcg: memcg to uncharge
6467 * @nr_pages: number of pages to uncharge
6469 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6471 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6472 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6476 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6478 refill_stock(memcg, nr_pages);
6481 static int __init cgroup_memory(char *s)
6485 while ((token = strsep(&s, ",")) != NULL) {
6488 if (!strcmp(token, "nosocket"))
6489 cgroup_memory_nosocket = true;
6490 if (!strcmp(token, "nokmem"))
6491 cgroup_memory_nokmem = true;
6495 __setup("cgroup.memory=", cgroup_memory);
6498 * subsys_initcall() for memory controller.
6500 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6501 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6502 * basically everything that doesn't depend on a specific mem_cgroup structure
6503 * should be initialized from here.
6505 static int __init mem_cgroup_init(void)
6509 #ifdef CONFIG_MEMCG_KMEM
6511 * Kmem cache creation is mostly done with the slab_mutex held,
6512 * so use a workqueue with limited concurrency to avoid stalling
6513 * all worker threads in case lots of cgroups are created and
6514 * destroyed simultaneously.
6516 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6517 BUG_ON(!memcg_kmem_cache_wq);
6520 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6521 memcg_hotplug_cpu_dead);
6523 for_each_possible_cpu(cpu)
6524 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6527 for_each_node(node) {
6528 struct mem_cgroup_tree_per_node *rtpn;
6530 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6531 node_online(node) ? node : NUMA_NO_NODE);
6533 rtpn->rb_root = RB_ROOT;
6534 rtpn->rb_rightmost = NULL;
6535 spin_lock_init(&rtpn->lock);
6536 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6541 subsys_initcall(mem_cgroup_init);
6543 #ifdef CONFIG_MEMCG_SWAP
6544 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6546 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6548 * The root cgroup cannot be destroyed, so it's refcount must
6551 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6555 memcg = parent_mem_cgroup(memcg);
6557 memcg = root_mem_cgroup;
6563 * mem_cgroup_swapout - transfer a memsw charge to swap
6564 * @page: page whose memsw charge to transfer
6565 * @entry: swap entry to move the charge to
6567 * Transfer the memsw charge of @page to @entry.
6569 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6571 struct mem_cgroup *memcg, *swap_memcg;
6572 unsigned int nr_entries;
6573 unsigned short oldid;
6575 VM_BUG_ON_PAGE(PageLRU(page), page);
6576 VM_BUG_ON_PAGE(page_count(page), page);
6578 if (!do_memsw_account())
6581 memcg = page->mem_cgroup;
6583 /* Readahead page, never charged */
6588 * In case the memcg owning these pages has been offlined and doesn't
6589 * have an ID allocated to it anymore, charge the closest online
6590 * ancestor for the swap instead and transfer the memory+swap charge.
6592 swap_memcg = mem_cgroup_id_get_online(memcg);
6593 nr_entries = hpage_nr_pages(page);
6594 /* Get references for the tail pages, too */
6596 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6597 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6599 VM_BUG_ON_PAGE(oldid, page);
6600 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6602 page->mem_cgroup = NULL;
6604 if (!mem_cgroup_is_root(memcg))
6605 page_counter_uncharge(&memcg->memory, nr_entries);
6607 if (memcg != swap_memcg) {
6608 if (!mem_cgroup_is_root(swap_memcg))
6609 page_counter_charge(&swap_memcg->memsw, nr_entries);
6610 page_counter_uncharge(&memcg->memsw, nr_entries);
6614 * Interrupts should be disabled here because the caller holds the
6615 * i_pages lock which is taken with interrupts-off. It is
6616 * important here to have the interrupts disabled because it is the
6617 * only synchronisation we have for updating the per-CPU variables.
6619 VM_BUG_ON(!irqs_disabled());
6620 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6622 memcg_check_events(memcg, page);
6624 if (!mem_cgroup_is_root(memcg))
6625 css_put_many(&memcg->css, nr_entries);
6629 * mem_cgroup_try_charge_swap - try charging swap space for a page
6630 * @page: page being added to swap
6631 * @entry: swap entry to charge
6633 * Try to charge @page's memcg for the swap space at @entry.
6635 * Returns 0 on success, -ENOMEM on failure.
6637 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6639 unsigned int nr_pages = hpage_nr_pages(page);
6640 struct page_counter *counter;
6641 struct mem_cgroup *memcg;
6642 unsigned short oldid;
6644 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6647 memcg = page->mem_cgroup;
6649 /* Readahead page, never charged */
6654 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6658 memcg = mem_cgroup_id_get_online(memcg);
6660 if (!mem_cgroup_is_root(memcg) &&
6661 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6662 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6663 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6664 mem_cgroup_id_put(memcg);
6668 /* Get references for the tail pages, too */
6670 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6671 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6672 VM_BUG_ON_PAGE(oldid, page);
6673 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6679 * mem_cgroup_uncharge_swap - uncharge swap space
6680 * @entry: swap entry to uncharge
6681 * @nr_pages: the amount of swap space to uncharge
6683 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6685 struct mem_cgroup *memcg;
6688 if (!do_swap_account)
6691 id = swap_cgroup_record(entry, 0, nr_pages);
6693 memcg = mem_cgroup_from_id(id);
6695 if (!mem_cgroup_is_root(memcg)) {
6696 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6697 page_counter_uncharge(&memcg->swap, nr_pages);
6699 page_counter_uncharge(&memcg->memsw, nr_pages);
6701 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6702 mem_cgroup_id_put_many(memcg, nr_pages);
6707 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6709 long nr_swap_pages = get_nr_swap_pages();
6711 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6712 return nr_swap_pages;
6713 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6714 nr_swap_pages = min_t(long, nr_swap_pages,
6715 READ_ONCE(memcg->swap.max) -
6716 page_counter_read(&memcg->swap));
6717 return nr_swap_pages;
6720 bool mem_cgroup_swap_full(struct page *page)
6722 struct mem_cgroup *memcg;
6724 VM_BUG_ON_PAGE(!PageLocked(page), page);
6728 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6731 memcg = page->mem_cgroup;
6735 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6736 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6742 /* for remember boot option*/
6743 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6744 static int really_do_swap_account __initdata = 1;
6746 static int really_do_swap_account __initdata;
6749 static int __init enable_swap_account(char *s)
6751 if (!strcmp(s, "1"))
6752 really_do_swap_account = 1;
6753 else if (!strcmp(s, "0"))
6754 really_do_swap_account = 0;
6757 __setup("swapaccount=", enable_swap_account);
6759 static u64 swap_current_read(struct cgroup_subsys_state *css,
6762 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6764 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6767 static int swap_max_show(struct seq_file *m, void *v)
6769 return seq_puts_memcg_tunable(m,
6770 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6773 static ssize_t swap_max_write(struct kernfs_open_file *of,
6774 char *buf, size_t nbytes, loff_t off)
6776 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6780 buf = strstrip(buf);
6781 err = page_counter_memparse(buf, "max", &max);
6785 xchg(&memcg->swap.max, max);
6790 static int swap_events_show(struct seq_file *m, void *v)
6792 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6794 seq_printf(m, "max %lu\n",
6795 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6796 seq_printf(m, "fail %lu\n",
6797 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6802 static struct cftype swap_files[] = {
6804 .name = "swap.current",
6805 .flags = CFTYPE_NOT_ON_ROOT,
6806 .read_u64 = swap_current_read,
6810 .flags = CFTYPE_NOT_ON_ROOT,
6811 .seq_show = swap_max_show,
6812 .write = swap_max_write,
6815 .name = "swap.events",
6816 .flags = CFTYPE_NOT_ON_ROOT,
6817 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6818 .seq_show = swap_events_show,
6823 static struct cftype memsw_cgroup_files[] = {
6825 .name = "memsw.usage_in_bytes",
6826 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6827 .read_u64 = mem_cgroup_read_u64,
6830 .name = "memsw.max_usage_in_bytes",
6831 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6832 .write = mem_cgroup_reset,
6833 .read_u64 = mem_cgroup_read_u64,
6836 .name = "memsw.limit_in_bytes",
6837 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6838 .write = mem_cgroup_write,
6839 .read_u64 = mem_cgroup_read_u64,
6842 .name = "memsw.failcnt",
6843 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6844 .write = mem_cgroup_reset,
6845 .read_u64 = mem_cgroup_read_u64,
6847 { }, /* terminate */
6850 static int __init mem_cgroup_swap_init(void)
6852 if (!mem_cgroup_disabled() && really_do_swap_account) {
6853 do_swap_account = 1;
6854 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6856 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6857 memsw_cgroup_files));
6861 subsys_initcall(mem_cgroup_swap_init);
6863 #endif /* CONFIG_MEMCG_SWAP */